AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
// Screen control on 你的電腦 (computer use), like Grok Bot's Computer Use helper: the bot looks at
// the screen and uses the mouse and keyboard. Same tool names and arguments as the original's
// local daemon (`computer_use_screenshot`, `computer_use_click` {method, x, y, button, count,
// modifier_keys}, `computer_use_scroll`, `computer_use_mouse_move`, `computer_use_drag`,
// `computer_use_typing` {value}, `computer_use_press_key` {key, hold_duration_ms}).
//
// Screenshot sizing (local-exec-daemon main.cjs): `zct` fits the long edge to 1024 (Gct) keeping
// both edges at least 8 (Hs); images stay under 1 MiB (mg = 1024*1024); JPEG is tried at quality
// 85, 70, 55 ($3r), taking the first that fits, else the smallest.
use base64 ::Engine ;
use serde_json ::{ json , Value } ;
use std ::sync ::Mutex ;
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
pub const NAMES : [ & str ; 12 ] = [ " computer_use_screenshot " , " computer_use_zoom " , " computer_use_show " , " computer_use_click " , " computer_use_typing " , " computer_use_press_key " , " computer_use_scroll " , " computer_use_mouse_move " , " computer_use_drag " , " computer_use_check_permissions " , " computer_use_open_app " , " computer_use_batch " ] ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Wait after an action before taking the picture that comes back with it (menus open, pages
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
/// start to change; estimated). The box waits for the screen to stop changing instead, up to
/// about this long.
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
pub const SETTLE_MS : u64 = 500 ;
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
/// Most actions one `computer_use_batch` call may carry (Anthropic's computer_batch guidance:
/// short self-contained sequences, one picture at the end).
Screen steps name their controls first, and a batch takes a whole row of keys; four held-out tasks
A quick model looking numbers up pressed 1, 2, 3, 1 for 1234, and 1234 × 5678 = didn't fit in
an 8-action batch (it split it and lost a key). The click and batch descriptions now lead with
the control's name (name "4"), and a batch takes up to 16 actions.
The benchmark gets four tasks on things none of the others use (Mousepad and its save dialog,
a Chromium bookmark, ImageMagick, a page of nested web components), run only when asked for
(--tasks N1,…): a check that the changes carry over to what they weren't tuned on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:36:50 +00:00
pub const MAX_BATCH : usize = 16 ;
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
/// A batch step's short action name → the tool it runs. Screen reads, apps and permissions stay
/// single steps on purpose: a batch is for input you already planned from the current picture.
pub fn action_tool ( a : & str ) -> Result < & 'static str , String > {
match a . trim ( ) . to_lowercase ( ) . as_str ( ) {
" click " = > Ok ( " computer_use_click " ) ,
" typing " | " type " = > Ok ( " computer_use_typing " ) ,
" press_key " | " key " = > Ok ( " computer_use_press_key " ) ,
" scroll " = > Ok ( " computer_use_scroll " ) ,
" mouse_move " | " move " = > Ok ( " computer_use_mouse_move " ) ,
" drag " = > Ok ( " computer_use_drag " ) ,
" " = > Err ( " batch 的每個動作都要有 action( click、typing、press_key、scroll、mouse_move、drag) " . into ( ) ) ,
other = > Err ( format! ( " batch 裡沒有「 {other} 」這個動作( 可用: click、typing、press_key、scroll、mouse_move、drag; 要打開 App 請單獨用 computer_use_open_app) " ) ) ,
}
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// The person took control back from the bot (the floating bar): screen actions wait.
pub static HELD_BY_PERSON : std ::sync ::atomic ::AtomicBool = std ::sync ::atomic ::AtomicBool ::new ( false ) ;
/// When the bot last moved the mouse or typed on this Mac (for the floating bar), and who.
pub static LAST_ACTION : Mutex < Option < ( std ::time ::Instant , String ) > > = Mutex ::new ( None ) ;
2026-11-05 07:48:22 +00:00
/// Spawn a short-lived helper (`open`, a player, …) and reap it in the background:
/// spawn-without-wait left one zombie per call for the app's whole life.
pub fn spawn_reaped ( c : & mut std ::process ::Command ) -> std ::io ::Result < ( ) > {
let mut child = c . spawn ( ) ? ;
std ::thread ::spawn ( move | | {
let _ = child . wait ( ) ;
} ) ;
Ok ( ( ) )
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// Longest screenshot edge sent to the model (Gct).
pub const MAX_LONG_EDGE : u32 = 1024 ;
/// Shortest edge never goes below this (Hs).
pub const MIN_EDGE : u32 = 8 ;
/// Largest screenshot sent (mg = 1024*1024).
pub const MAX_IMAGE_BYTES : usize = 1048576 ;
/// JPEG qualities tried in order ($3r).
pub const JPEG_QUALITIES : [ u8 ; 3 ] = [ 85 , 70 , 55 ] ;
/// Screenshots kept in what's sent to the model; older ones become a note (estimated).
pub const KEEP_SCREENSHOTS : usize = 3 ;
/// Most UTF-16 units per typed chunk; characters are never split (CUTypedCharacters.utf16Chunks
/// @0x1000beb00: capacity 0x14, a new chunk when the next character would pass 20).
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
pub const TYPE_CHUNK_UNITS : usize = 20 ;
/// Virtual key the typed text rides on (postChunks @0x1000beebc: CGEventCreateKeyboardEvent(_, 0x80, _)).
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
const TYPE_KEY_CODE : u16 = 128 ;
/// Key down → key up for typed chunks and pressed keys (postChunks / handlePressKey: sleep 0.035 s).
pub const KEY_HOLD_SECS : f64 = 0.035 ;
/// Between typed chunks and between keys of one press (postChunks / handlePressKey: sleep 0.02 s).
pub const KEY_GAP_SECS : f64 = 0.02 ;
/// Mouse down → up when no hold is given (CUActuator.click @0x100010120: hold > 0 ? hold : 0.03).
pub const CLICK_HOLD_SECS : f64 = 0.03 ;
/// Between the clicks of a double/triple click (CUActuator.click: sleep 0.08 s).
pub const CLICK_REPEAT_GAP_SECS : f64 = 0.08 ;
/// Interpolated dragged events per path segment (handleDrag @0x10009dee4 passes steps: 6).
pub const DRAG_STEPS : usize = 6 ;
/// Hold after pressing, before moving (handleDrag @0x10009dee4 passes holdSeconds: 0.04).
pub const DRAG_HOLD_SECS : f64 = 0.04 ;
/// The dragged-event points: each segment split into `DRAG_STEPS` equal parts, ending on its end
/// point (CURemoteEventPoster.drag @0x10008c34c: start + delta * i / steps, i = 1…steps).
pub fn drag_points ( path : & [ ( f64 , f64 ) ] ) -> Vec < ( f64 , f64 ) > {
path . windows ( 2 )
. flat_map ( | w | {
let ( ( ax , ay ) , ( bx , by ) ) = ( w [ 0 ] , w [ 1 ] ) ;
( 1 ..= DRAG_STEPS ) . map ( move | i | {
let t = i as f64 / DRAG_STEPS as f64 ;
( ax + ( bx - ax ) * t , ay + ( by - ay ) * t )
} )
} )
. collect ( )
}
/// Scroll steps when none is given, and the allowed range (handleScroll @0x10009e200: default 3,
/// max(amount, 1), min(…, 20)).
pub const SCROLL_DEFAULT_AMOUNT : i64 = 3 ;
pub const SCROLL_MAX_AMOUNT : i64 = 20 ;
/// Pixels per scroll step (handleScroll @0x10009e200: (amount + amount) << 3).
pub const SCROLL_PIXELS_PER_STEP : i32 = 16 ;
pub fn scroll_amount ( amount : Option < i64 > ) -> i32 {
amount . unwrap_or ( SCROLL_DEFAULT_AMOUNT ) . clamp ( 1 , SCROLL_MAX_AMOUNT ) as i32
}
/// (vertical, horizontal) pixel deltas: up and left positive, down and right negative.
pub fn scroll_pixels ( direction : & str , amount : i32 ) -> Option < ( i32 , i32 ) > {
let px = amount * SCROLL_PIXELS_PER_STEP ;
match direction {
" up " = > Some ( ( px , 0 ) ) ,
" down " = > Some ( ( - px , 0 ) ) ,
" left " = > Some ( ( 0 , px ) ) ,
" right " = > Some ( ( 0 , - px ) ) ,
_ = > None ,
}
}
/// Pause between mouse move / drag steps (estimated).
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
const EVENT_GAP_MS : u64 = 8 ;
/// Longest key hold allowed (estimated).
const MAX_HOLD_MS : u64 = 5000 ;
/// utf16Chunks: split text into chunks of at most `TYPE_CHUNK_UNITS` UTF-16 units without
/// splitting a character (a character longer than that gets its own chunk).
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
pub fn utf16_chunks ( text : & str ) -> Vec < Vec < u16 > > {
let mut out : Vec < Vec < u16 > > = vec! [ ] ;
let mut cur : Vec < u16 > = Vec ::with_capacity ( TYPE_CHUNK_UNITS ) ;
for c in text . chars ( ) {
let mut buf = [ 0 u16 ; 2 ] ;
let units = c . encode_utf16 ( & mut buf ) ;
if ! cur . is_empty ( ) & & cur . len ( ) + units . len ( ) > TYPE_CHUNK_UNITS {
out . push ( std ::mem ::take ( & mut cur ) ) ;
}
cur . extend_from_slice ( units ) ;
}
if ! cur . is_empty ( ) {
out . push ( cur ) ;
}
out
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// One key of a press: a key with its modifiers, or a character typed (「@」「?」 have no key of
/// their own on every keyboard; refusing them sent the bot looking for workarounds).
#[ derive(Debug, PartialEq) ]
pub enum Key {
Combo ( u16 , u64 ) ,
Char ( String ) ,
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// "cmd+a delete" → each key of one press, in order (handlePressKey presses a list of keys).
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
pub fn key_plan ( keys : & str ) -> Result < Vec < Key > , String > {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let list : Vec < & str > = keys . split_whitespace ( ) . collect ( ) ;
if list . is_empty ( ) {
return Err ( " 請給要按的鍵 " . into ( ) ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
list . into_iter ( )
. map ( | k | match parse_combo ( k ) {
// A letter, digit or symbol without a modifier is a character: typed, so an input
// method (注音) can't turn it into a syllable. Named symbol keys mean their character.
Ok ( ( c , 0 ) ) if printable_of ( c ) . is_some ( ) = > Ok ( Key ::Char ( printable_of ( c ) . unwrap_or_default ( ) ) ) ,
Ok ( ( c , f ) ) = > Ok ( Key ::Combo ( c , f ) ) ,
Err ( _ ) if k . chars ( ) . count ( ) = = 1 = > Ok ( Key ::Char ( k . to_string ( ) ) ) ,
Err ( e ) = > Err ( e ) ,
} )
. collect ( )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// The character a printing key gives without shift (US layout), or None for keys like return.
fn printable_of ( code : u16 ) -> Option < String > {
const LETTERS : [ u16 ; 26 ] = [ 0 , 11 , 8 , 2 , 14 , 3 , 5 , 4 , 34 , 38 , 40 , 37 , 46 , 45 , 31 , 35 , 12 , 15 , 1 , 17 , 32 , 9 , 13 , 7 , 16 , 6 ] ;
const DIGITS : [ u16 ; 10 ] = [ 29 , 18 , 19 , 20 , 21 , 23 , 22 , 26 , 28 , 25 ] ;
if let Some ( i ) = LETTERS . iter ( ) . position ( | c | * c = = code ) {
return Some ( ( ( b 'a' + i as u8 ) as char ) . to_string ( ) ) ;
}
if let Some ( i ) = DIGITS . iter ( ) . position ( | c | * c = = code ) {
return Some ( ( ( b '0' + i as u8 ) as char ) . to_string ( ) ) ;
}
let c = match code {
27 = > '-' ,
24 = > '=' ,
33 = > '[' ,
30 = > ']' ,
41 = > ';' ,
39 = > '\'' ,
43 = > ',' ,
47 = > '.' ,
44 = > '/' ,
42 = > '\\' ,
50 = > '`' ,
_ = > return None ,
} ;
Some ( c . to_string ( ) )
}
/// The picture the bot and the phone get of this screen: its size, always the same for the same
/// screen. Coordinates on it map to the screen without remembering any earlier screenshot (that
/// hidden state made clicks fail with "screenshot first" and land wrong for a viewer whose
/// picture was a different size).
pub fn shot_size ( ) -> ( u32 , u32 ) {
let ( px , py ) = mac ::screen_pixels ( ) ;
let ( w , h , _ ) = fit ( px . max ( 1 ) , py . max ( 1 ) ) ;
( w , h )
}
/// Screen points (the mouse's units) per picture pixel.
pub fn scale_here ( ) -> ( f64 , f64 ) {
let ( pw , ph ) = mac ::screen_points ( ) ;
let ( w , h ) = shot_size ( ) ;
( pw / w . max ( 1 ) as f64 , ph / h . max ( 1 ) as f64 )
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
/// A click, scroll or drag: a step whose whole effect should show on the screen. Only these are
/// called out when the picture didn't change (typing a few letters or a key can change too little
/// for the coarse fingerprint, and the bot was told to stop typing that worked).
pub fn is_pointer_step ( name : & str ) -> bool {
matches! ( name , " computer_use_click " | " computer_use_scroll " | " computer_use_drag " )
}
Real-use fixes: the person sees the screens the bot shows, files reach the phone, no blank-tab loops
From real runs on the cloud brain with the person's Mac (Threads patrol, a post draft, "show me my
screen"):
- computer_use_show: a plain picture of any computer's screen goes into the chat (the bot sees it
too). The bot used to describe a screenshot in words and fail to send it.
- The phone shows the bot's attachments: pictures in place, other files download on a tap (api/file,
only from the chats' file folder). It showed nothing for the bot and names only for the person.
- send_file works from the person's other computers (fetch_file over the relay, 20 MB, secrets stay).
- The same call with the same answer again and again is warned the 3rd time and stopped from the
4th (a blank tab was read 11 times, then the bot said Threads was logged out without opening it);
a blank tab says so; checking a login means opening the site.
- Screenshot lists: a nameless row around a named button isn't numbered, a nameless row takes the
words inside it, empty spacer rows aren't numbered (System Settings: 63 numbers -> 33, all named).
- Writing "in my voice" reads the person's own recent posts first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 11:37:45 +00:00
/// What the bot is told after computer_use_show (the picture itself goes into the chat).
pub const SHOWN_NOTE : & str = " 已把這台電腦現在的畫面(使用者看到的樣子,沒有編號)放進聊天,使用者看得到這張圖;你也看得到,可以用一兩句話說明畫面上有什麼 " ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
pub fn is_computer_tool ( name : & str ) -> bool {
NAMES . contains ( & name )
}
pub fn label ( name : & str ) -> & 'static str {
match name {
" computer_use_screenshot " = > " 看你的螢幕 " ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" computer_use_zoom " = > " 放大看 " ,
Real-use fixes: the person sees the screens the bot shows, files reach the phone, no blank-tab loops
From real runs on the cloud brain with the person's Mac (Threads patrol, a post draft, "show me my
screen"):
- computer_use_show: a plain picture of any computer's screen goes into the chat (the bot sees it
too). The bot used to describe a screenshot in words and fail to send it.
- The phone shows the bot's attachments: pictures in place, other files download on a tap (api/file,
only from the chats' file folder). It showed nothing for the bot and names only for the person.
- send_file works from the person's other computers (fetch_file over the relay, 20 MB, secrets stay).
- The same call with the same answer again and again is warned the 3rd time and stopped from the
4th (a blank tab was read 11 times, then the bot said Threads was logged out without opening it);
a blank tab says so; checking a login means opening the site.
- Screenshot lists: a nameless row around a named button isn't numbered, a nameless row takes the
words inside it, empty spacer rows aren't numbered (System Settings: 63 numbers -> 33, all named).
- Writing "in my voice" reads the person's own recent posts first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 11:37:45 +00:00
" computer_use_show " = > " 把畫面傳給你看 " ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
" computer_use_click " = > " 點一下 " ,
" computer_use_typing " = > " 打字 " ,
" computer_use_press_key " = > " 按鍵 " ,
" computer_use_scroll " = > " 捲動 " ,
" computer_use_mouse_move " = > " 移動滑鼠 " ,
" computer_use_drag " = > " 拖曳 " ,
" computer_use_check_permissions " = > " 檢查螢幕權限 " ,
2026-11-04 09:19:59 +00:00
" computer_use_open_app " = > " 打開 App " ,
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
" computer_use_batch " = > " 依序操作 " ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
_ = > " 操作畫面 " ,
}
}
pub fn describe ( name : & str , args : & str ) -> String {
Remove what nothing uses and keep one copy of what was written several times
Unused (found by taking away the dead-code allowances; nothing else used them, tests aside):
computer::specs, quiet_button, the never-made Poke::Spin, ten icons and their SVGs, the
theme's chrome/bubble_agent colours, ImportRow.line, voice's multipart/transcribe/speak
wrappers, and the old keychain writers vault::save/remove/index_add with their live test
(the vault store replaced them; reading old keychain entries for the import stays).
One copy each: percent-encoding (httpd::url_encode, was four), tool arguments (tools::args_of,
was three), the WebSocket handshake (httpd::accept_websocket, was three), the read pacing
(wire::Pace, was two: relay and app), docker with input and a time limit (dk_with_input, was
two). Docs: FEATURES §看畫面 (VNC), protocol (v* frames, /screen, tickets), architecture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:21:47 +00:00
let v = crate ::tools ::args_of ( args ) ;
2026-11-05 10:25:46 +00:00
// By number from the app's list (e7), else where on the screen (it said "(0, 0)" for every
// click by number).
let xy = | | match v [ " ref " ] . as_str ( ) . map ( str ::trim ) . filter ( | r | ! r . is_empty ( ) ) {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
Some ( r ) = > format! ( " # {} " , r . trim_start_matches ( [ 'e' , '#' ] ) ) ,
2026-11-05 10:25:46 +00:00
None = > format! ( " ( {} , {} ) " , v [ " x " ] . as_f64 ( ) . unwrap_or ( 0.0 ) , v [ " y " ] . as_f64 ( ) . unwrap_or ( 0.0 ) ) ,
} ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
match name {
" computer_use_screenshot " = > " 螢幕截圖 " . into ( ) ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" computer_use_zoom " = > format! ( " 放大 ( {} , {} ) 起 {} × {} " , v [ " x " ] . as_f64 ( ) . unwrap_or ( 0.0 ) , v [ " y " ] . as_f64 ( ) . unwrap_or ( 0.0 ) , v [ " width " ] . as_f64 ( ) . unwrap_or ( 0.0 ) , v [ " height " ] . as_f64 ( ) . unwrap_or ( 0.0 ) ) ,
Real-use fixes: the person sees the screens the bot shows, files reach the phone, no blank-tab loops
From real runs on the cloud brain with the person's Mac (Threads patrol, a post draft, "show me my
screen"):
- computer_use_show: a plain picture of any computer's screen goes into the chat (the bot sees it
too). The bot used to describe a screenshot in words and fail to send it.
- The phone shows the bot's attachments: pictures in place, other files download on a tap (api/file,
only from the chats' file folder). It showed nothing for the bot and names only for the person.
- send_file works from the person's other computers (fetch_file over the relay, 20 MB, secrets stay).
- The same call with the same answer again and again is warned the 3rd time and stopped from the
4th (a blank tab was read 11 times, then the bot said Threads was logged out without opening it);
a blank tab says so; checking a login means opening the site.
- Screenshot lists: a nameless row around a named button isn't numbered, a nameless row takes the
words inside it, empty spacer rows aren't numbered (System Settings: 63 numbers -> 33, all named).
- Writing "in my voice" reads the person's own recent posts first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 11:37:45 +00:00
" computer_use_show " = > " 畫面傳進聊天 " . into ( ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
" computer_use_click " = > {
let b = v [ " button " ] . as_str ( ) . unwrap_or ( " left " ) ;
let n = v [ " count " ] . as_u64 ( ) . unwrap_or ( 1 ) ;
format! ( " {} {} {} {} " , if b = = " right " { " 右鍵 " } else { " " } , if n > = 2 { " 連點 " } else { " 點 " } , if n > = 2 { format! ( " {n} 下 " ) } else { String ::new ( ) } , xy ( ) )
}
" computer_use_typing " = > {
let t = v [ " value " ] . as_str ( ) . unwrap_or ( " " ) ;
let short : String = t . chars ( ) . take ( 60 ) . collect ( ) ;
2026-11-05 10:25:46 +00:00
let at = v [ " ref " ] . as_str ( ) . filter ( | r | ! r . trim ( ) . is_empty ( ) ) . map ( | r | format! ( " 在 {r} " ) ) . unwrap_or_default ( ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
format! ( " {at} 「 {short} {} 」 {} " , if t . chars ( ) . count ( ) > 60 { " … " } else { " " } , if v [ " submit " ] . as_bool ( ) . unwrap_or ( false ) { " 再按 return " } else { " " } )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_press_key " = > v [ " key " ] . as_str ( ) . unwrap_or ( " " ) . to_string ( ) ,
" computer_use_scroll " = > format! ( " {} {} 格,在 {} " , v [ " direction " ] . as_str ( ) . unwrap_or ( " down " ) , v [ " amount " ] . as_i64 ( ) . unwrap_or ( 3 ) , xy ( ) ) ,
" computer_use_mouse_move " = > xy ( ) ,
" computer_use_drag " = > format! ( " {} 個點 " , v [ " path " ] . as_array ( ) . map ( | a | a . len ( ) ) . unwrap_or ( 0 ) ) ,
" computer_use_check_permissions " = > " 螢幕錄製、輔助使用 " . into ( ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
" computer_use_open_app " = > v [ " app " ] . as_str ( ) . filter ( | a | ! a . is_empty ( ) ) . unwrap_or ( " " ) . to_string ( ) ,
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
" computer_use_batch " = > format! ( " 依序 {} 個動作 " , v [ " actions " ] . as_array ( ) . map ( | a | a . len ( ) ) . unwrap_or ( 0 ) ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
_ = > String ::new ( ) ,
}
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// The screen tools, each with a `computer` choice. One way to see (the numbered picture), one
/// way to act (a real click or keystroke, best by number).
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
pub fn specs_for ( computers : & [ String ] , default : & str , chosen : bool ) -> Vec < Value > {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let mods = json! ( { " type " : " array " , " items " : { " type " : " string " , " enum " : [ " cmd " , " shift " , " option " , " control " ] } } ) ;
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
let rf = json! ( { " type " : " string " , " description " : " The number of a control on the latest picture (e.g. 12). " } ) ;
let named = json! ( { " type " : " string " , " description " : " Or the control's name exactly as the latest list shows it (e.g. 4, Close, 自動更新): the numbered control with that name. Easiest for keys and buttons with plain names; use ref when two have the same name. " } ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let mut v = vec! [
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_screenshot " , " description " : " See a computer's screen: a picture with every button, link, field and menu item boxed and numbered, and the same numbers in a list with their names. Act on a control by its number (ref). Every action below already returns the next picture, so call this only to start or to look again. " ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
" parameters " : { " type " : " object " , " properties " : { } } } ) ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
json! ( { " name " : " computer_use_zoom " , " description " : " Look closer at part of the latest picture, at the screen's full sharpness: small icons (the menu bar, a toolbar, a tray), small print, a tiny control. Give the area in the latest picture's pixels. You get a bigger, sharper picture of just that area; to act there, use the numbers on the normal picture, or turn a point in the zoomed picture back into the normal picture's pixels as the result explains. Use it instead of guessing, and never make your own zoom with commands. " ,
" parameters " : { " type " : " object " , " properties " : { " x " : { " type " : " number " , " description " : " Left edge, in the latest picture's pixels " } , " y " : { " type " : " number " , " description " : " Top edge " } , " width " : { " type " : " number " } , " height " : { " type " : " number " } } , " required " : [ " x " , " y " , " width " , " height " ] } } ) ,
Real-use fixes: the person sees the screens the bot shows, files reach the phone, no blank-tab loops
From real runs on the cloud brain with the person's Mac (Threads patrol, a post draft, "show me my
screen"):
- computer_use_show: a plain picture of any computer's screen goes into the chat (the bot sees it
too). The bot used to describe a screenshot in words and fail to send it.
- The phone shows the bot's attachments: pictures in place, other files download on a tap (api/file,
only from the chats' file folder). It showed nothing for the bot and names only for the person.
- send_file works from the person's other computers (fetch_file over the relay, 20 MB, secrets stay).
- The same call with the same answer again and again is warned the 3rd time and stopped from the
4th (a blank tab was read 11 times, then the bot said Threads was logged out without opening it);
a blank tab says so; checking a login means opening the site.
- Screenshot lists: a nameless row around a named button isn't numbered, a nameless row takes the
words inside it, empty spacer rows aren't numbered (System Settings: 63 numbers -> 33, all named).
- Writing "in my voice" reads the person's own recent posts first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 11:37:45 +00:00
json! ( { " name " : " computer_use_show " , " description " : " Show the person a computer's screen: a picture of it as they would see it (no numbers) goes into the chat for them, and you see it too. Use it when they ask to see a screen (截圖給我看、讓我看看畫面). One call, no file needed. " ,
" parameters " : { " type " : " object " , " properties " : { } } } ) ,
Screen steps name their controls first, and a batch takes a whole row of keys; four held-out tasks
A quick model looking numbers up pressed 1, 2, 3, 1 for 1234, and 1234 × 5678 = didn't fit in
an 8-action batch (it split it and lost a key). The click and batch descriptions now lead with
the control's name (name "4"), and a batch takes up to 16 actions.
The benchmark gets four tasks on things none of the others use (Mousepad and its save dialog,
a Chromium bookmark, ImageMagick, a page of nested web components), run only when asked for
(--tasks N1,…): a check that the changes carry over to what they weren't tuned on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:36:50 +00:00
json! ( { " name " : " computer_use_click " , " description " : " Click a control: by its name as the latest list shows it (name, e.g. 4, Close, 自動更新 — easiest, no number to look up), or by its number (ref) when two share a name. Only for something with neither (a spot in a picture or game): x, y in the picture's pixels. Returns the next picture. " ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" parameters " : { " type " : " object " , " properties " : { " ref " : rf , " name " : named , " x " : { " type " : " number " } , " y " : { " type " : " number " } , " button " : { " type " : " string " , " enum " : [ " left " , " right " , " middle " ] } , " count " : { " type " : " integer " , " description " : " 2 for a double click " } , " modifier_keys " : mods } } } ) ,
2026-11-09 02:19:18 +00:00
json! ( { " name " : " computer_use_typing " , " description " : " Type text (any language) into a field by its name or number (it is clicked first; a one-line field's old text is replaced, as if selected first — append=true adds to it instead), or where the cursor is. submit=true presses return after it (send, search, open the address). Returns the next picture. " ,
" parameters " : { " type " : " object " , " properties " : { " value " : { " type " : " string " } , " ref " : rf , " name " : named , " submit " : { " type " : " boolean " } , " append " : { " type " : " boolean " , " description " : " Keep the field's text and add to it " } } , " required " : [ " value " ] } } ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_press_key " , " description " : " Press a key or shortcut in the app in front: return, tab, escape, delete, up, slash, cmd+c, cmd+shift+t, cmd+space. Several in order: separate with spaces ( \" cmd+a delete \" ). On a Linux computer cmd means ctrl. Returns the next picture. " ,
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
" parameters " : { " type " : " object " , " properties " : { " key " : { " type " : " string " } , " hold_duration_ms " : { " type " : " integer " } } , " required " : [ " key " ] } } ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_scroll " , " description " : " Scroll a list or page: on a control by number (ref), or at x, y, or where the pointer is. Returns the next picture. " ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" parameters " : { " type " : " object " , " properties " : { " ref " : rf , " name " : named , " x " : { " type " : " number " } , " y " : { " type " : " number " } , " direction " : { " type " : " string " , " enum " : [ " up " , " down " , " left " , " right " ] } , " amount " : { " type " : " integer " , " description " : " Scroll steps of 16 pixels, 1-20, default 3 " } , " modifier_keys " : mods } , " required " : [ " direction " ] } } ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_open_app " , " description " : " Open an app, or bring it to the front when it is already open (its name as shown, e.g. Ghostty, Calculator, 計算機, Chromium, xterm). Returns the next picture. " ,
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
" parameters " : { " type " : " object " , " properties " : { " app " : { " type " : " string " } } , " required " : [ " app " ] } } ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_mouse_move " , " description " : " Move the mouse onto a control by number (ref) or to x, y (to open a hover menu). Returns the next picture. " ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" parameters " : { " type " : " object " , " properties " : { " ref " : rf , " name " : named , " x " : { " type " : " number " } , " y " : { " type " : " number " } } } } ) ,
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
json! ( { " name " : " computer_use_drag " , " description " : " Drag along a path of points in the picture's pixels (first = press, last = release). Returns the next picture. " ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
" parameters " : { " type " : " object " , " properties " : { " path " : { " type " : " array " , " items " : { " type " : " object " , " properties " : { " x " : { " type " : " number " } , " y " : { " type " : " number " } } , " required " : [ " x " , " y " ] } } , " button " : { " type " : " string " , " enum " : [ " left " , " right " ] } , " modifier_keys " : mods } , " required " : [ " path " ] } } ) ,
json! ( { " name " : " computer_use_check_permissions " , " description " : " Check whether the screen can be seen and the mouse and keyboard used on that computer. " ,
" parameters " : { " type " : " object " , " properties " : { } } } ) ,
Screen steps name their controls first, and a batch takes a whole row of keys; four held-out tasks
A quick model looking numbers up pressed 1, 2, 3, 1 for 1234, and 1234 × 5678 = didn't fit in
an 8-action batch (it split it and lost a key). The click and batch descriptions now lead with
the control's name (name "4"), and a batch takes up to 16 actions.
The benchmark gets four tasks on things none of the others use (Mousepad and its save dialog,
a Chromium bookmark, ImageMagick, a page of nested web components), run only when asked for
(--tasks N1,…): a check that the changes carry over to what they weren't tuned on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:36:50 +00:00
json! ( { " name " : " computer_use_batch " , " description " : format ! ( " Do a planned sequence of steps in one call (max {MAX_BATCH}): each item is one action with its own arguments, best naming the control, e.g. {{ \" action \" : \" click \" , \" name \" : \" 7 \" }}, {{ \" action \" : \" typing \" , \" name \" : \" Search \" , \" value \" : \" cat \" }}, {{ \" action \" : \" press_key \" , \" key \" : \" tab \" }}. Use it when the current picture already tells you every step (a row of keys to press, filling a form, a known shortcut chain): put the whole sequence in one call, they run in order, the first step that fails stops the rest, and you get one picture at the end. Several separate clicks in one step are slower and easier to get wrong than one batch. Do not use it to explore, or where step 2 depends on what step 1's screen shows — act one step at a time there. " ) ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" parameters " : { " type " : " object " , " properties " : { " actions " : { " type " : " array " , " maxItems " : MAX_BATCH , " items " : { " type " : " object " , " properties " : { " action " : { " type " : " string " , " enum " : [ " click " , " typing " , " press_key " , " scroll " , " mouse_move " , " drag " ] } , " ref " : rf . clone ( ) , " name " : named . clone ( ) , " x " : { " type " : " number " } , " y " : { " type " : " number " } , " button " : { " type " : " string " , " enum " : [ " left " , " right " , " middle " ] } , " count " : { " type " : " integer " } , " value " : { " type " : " string " } , " submit " : { " type " : " boolean " } , " key " : { " type " : " string " } , " hold_duration_ms " : { " type " : " integer " } , " direction " : { " type " : " string " , " enum " : [ " up " , " down " , " left " , " right " ] } , " amount " : { " type " : " integer " } , " path " : { " type " : " array " , " items " : { " type " : " object " , " properties " : { " x " : { " type " : " number " } , " y " : { " type " : " number " } } , " required " : [ " x " , " y " ] } } , " modifier_keys " : mods . clone ( ) } , " required " : [ " action " ] } } } , " required " : [ " actions " ] } } ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
] ;
2026-11-05 08:47:28 +00:00
// Always named: each computer goes by its own name (the instructions say which is which),
// so the bot says whose screen it means and the person can tell where it worked.
if ! computers . is_empty ( ) {
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
// The person's chosen computer may have no screen (a bot's computer without Docker): then
// the default is only the first screen there is, and the bot is told so.
let why = if chosen { " the one the person chose " } else { " the computer the person chose has no screen " } ;
let desc = format! ( " Whose screen, by the computer's name. Default: {default} ( {why} ). Use another only when the person names it. " ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
for t in v . iter_mut ( ) {
2026-11-05 08:03:34 +00:00
t [ " parameters " ] [ " properties " ] [ " computer " ] = json! ( { " type " : " string " , " enum " : computers , " description " : desc } ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
}
v
}
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
/// Longest edge of a zoomed picture (sharp enough to read a menu-bar icon, still one image).
pub const ZOOM_LONG_EDGE : u32 = 1024 ;
/// A zoom never blows up the real pixels more than this (it would only blur).
pub const ZOOM_MAX_UPSCALE : f64 = 4.0 ;
/// computer_use_zoom: an area of the latest picture (the bot's pixels, see `fit`) cut from the
/// full-sharpness screen `img` and enlarged. Returns the text for the bot and the JPEG (base64).
pub fn zoom ( img : & image ::RgbImage , v : & Value ) -> Result < ( String , String ) , String > {
let ( w , h ) = img . dimensions ( ) ;
let ( fw , fh , _ ) = fit ( w , h ) ;
let scale = w as f64 / fw . max ( 1 ) as f64 ;
let num = | k : & str | v [ k ] . as_f64 ( ) . filter ( | n | n . is_finite ( ) ) ;
let ( Some ( x ) , Some ( y ) , Some ( zw ) , Some ( zh ) ) = ( num ( " x " ) , num ( " y " ) , num ( " width " ) , num ( " height " ) ) else {
return Err ( " 放大要給 x、y、width、height( 最新那張截圖上的像素) " . into ( ) ) ;
} ;
// In the bot's picture, clamped to it; at least 8 picture pixels each way.
let x0 = x . clamp ( 0.0 , fw as f64 - 8.0 ) ;
let y0 = y . clamp ( 0.0 , fh as f64 - 8.0 ) ;
let x1 = ( x + zw ) . clamp ( x0 + 8.0 , fw as f64 ) ;
let y1 = ( y + zh ) . clamp ( y0 + 8.0 , fh as f64 ) ;
// The same area in real screen pixels.
let ( sx , sy ) = ( ( x0 * scale ) . floor ( ) as u32 , ( y0 * scale ) . floor ( ) as u32 ) ;
let sw = ( ( ( x1 - x0 ) * scale ) . ceil ( ) as u32 ) . clamp ( 1 , w - sx . min ( w - 1 ) ) ;
let sh = ( ( ( y1 - y0 ) * scale ) . ceil ( ) as u32 ) . clamp ( 1 , h - sy . min ( h - 1 ) ) ;
let crop = image ::imageops ::crop_imm ( img , sx , sy , sw , sh ) . to_image ( ) ;
let up = ( ZOOM_LONG_EDGE as f64 / sw . max ( sh ) as f64 ) . min ( ZOOM_MAX_UPSCALE ) ;
let pic = if ( up - 1.0 ) . abs ( ) > 0.05 {
let ( tw , th ) = ( ( ( sw as f64 * up ) . round ( ) as u32 ) . max ( 1 ) , ( ( sh as f64 * up ) . round ( ) as u32 ) . max ( 1 ) ) ;
image ::imageops ::resize ( & crop , tw , th , image ::imageops ::FilterType ::CatmullRom )
} else {
crop
} ;
let ( pw , ph ) = pic . dimensions ( ) ;
let jpeg = encode_jpeg ( & pic ) . ok_or ( " 放大的圖太大了,沒辦法送出 " ) ? ;
// How a point in the zoomed picture maps back to the normal picture.
let k = ( x1 - x0 ) / pw as f64 ;
let text = format! (
" 放大圖({pw}× {ph}):最新截圖上 ({:.0}, {:.0}) 到 ({:.0}, {:.0}) 的區域,用螢幕原本的清晰度。這張圖上的點 (a, b) 在一般截圖上是 ({:.0} + a× {k:.3}, {:.0} + b× {k:.3});有編號的東西還是照編號點。 " ,
x0 , y0 , x1 , y1 , x0 , y0
) ;
Ok ( ( text , base64 ::engine ::general_purpose ::STANDARD . encode ( & jpeg ) ) )
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// zct: the size to send. Long edge down to `MAX_LONG_EDGE`, never an edge under `MIN_EDGE`.
pub fn fit ( w : u32 , h : u32 ) -> ( u32 , u32 , bool ) {
if w < 1 | | h < 1 {
return ( w , h , false ) ;
}
let long = w . max ( h ) ;
if long < = MAX_LONG_EDGE {
return ( w , h , false ) ;
}
let short = w . min ( h ) ;
if short < MIN_EDGE {
return ( w , h , false ) ;
}
let mut i = MAX_LONG_EDGE as f64 / long as f64 ;
let mut o = ( ( w as f64 * i ) . round ( ) as u32 ) . max ( 1 ) ;
let mut a = ( ( h as f64 * i ) . round ( ) as u32 ) . max ( 1 ) ;
if o < MIN_EDGE | | a < MIN_EDGE {
i = MIN_EDGE as f64 / short as f64 ;
o = ( ( w as f64 * i ) . round ( ) as u32 ) . max ( MIN_EDGE ) ;
a = ( ( h as f64 * i ) . round ( ) as u32 ) . max ( MIN_EDGE ) ;
}
( o , a , o ! = w | | a ! = h )
}
/// j3r: JPEG at 85, 70, 55: the first within the budget, else the smallest if it's under the cap.
pub fn encode_jpeg ( img : & image ::RgbImage ) -> Option < Vec < u8 > > {
let mut best : Option < Vec < u8 > > = None ;
for q in JPEG_QUALITIES {
let mut out = vec! [ ] ;
let enc = image ::codecs ::jpeg ::JpegEncoder ::new_with_quality ( & mut out , q ) ;
if img . write_with_encoder ( enc ) . is_err ( ) {
continue ;
}
if out . len ( ) < = MAX_IMAGE_BYTES {
return Some ( out ) ;
}
Clippy is clean on macOS and Linux, for the app and the relay
Every warning `cargo clippy --all-targets` gave (50 on macOS, 8 more only on
Linux, 1 in the relay), fixed without changing what anything does:
- long types get a name (Peer, Members, Imported, Joining, Mention, Stream…);
- the two big enum variants are boxed (the vault's login plan, joining an
account);
- a test now waits for the copy it relaunched (it left a zombie behind);
- a hint's format had a dead `if` whose two branches were both "";
- the rest is what clippy suggests: sort_by_key, as_chunks, is_none_or,
is_multiple_of, while let, from_ref, next_back, struct init, no casts to
the same type.
Checked: cargo test (177 macOS, 181 Linux), relay tests (11), remote_e2e
28/28, and the live Linux test (X11; Wayland clicks, keys, Chinese typing,
scroll and screenshots).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-07 02:10:28 +00:00
if best . as_ref ( ) . is_none_or ( | b | out . len ( ) < b . len ( ) ) {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
best = Some ( out ) ;
}
}
best . filter ( | b | b . len ( ) < = MAX_IMAGE_BYTES )
}
/// Screenshot pixel → screen point.
pub fn to_screen ( x : f64 , y : f64 , scale : ( f64 , f64 ) ) -> ( f64 , f64 ) {
( x * scale . 0 , y * scale . 1 )
}
/// Key name → (key code, needs shift). US layout codes.
pub fn key_code ( name : & str ) -> Option < u16 > {
let n = name . trim ( ) . to_lowercase ( ) ;
let named = match n . as_str ( ) {
" return " | " enter " = > 36 ,
" tab " = > 48 ,
" space " = > 49 ,
" delete " | " backspace " = > 51 ,
" escape " | " esc " = > 53 ,
" forwarddelete " | " forward_delete " | " del " = > 117 ,
" home " = > 115 ,
" end " = > 119 ,
" pageup " | " page_up " = > 116 ,
" pagedown " | " page_down " = > 121 ,
" left " | " arrowleft " = > 123 ,
" right " | " arrowright " = > 124 ,
" down " | " arrowdown " = > 125 ,
" up " | " arrowup " = > 126 ,
" f1 " = > 122 ,
" f2 " = > 120 ,
" f3 " = > 99 ,
" f4 " = > 118 ,
" f5 " = > 96 ,
" f6 " = > 97 ,
" f7 " = > 98 ,
" f8 " = > 100 ,
" f9 " = > 101 ,
" f10 " = > 109 ,
" f11 " = > 103 ,
" f12 " = > 111 ,
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
// Symbol keys by name (a model writes "slash" as often as "/").
" slash " = > 44 ,
" backslash " = > 42 ,
" comma " = > 43 ,
" period " | " dot " = > 47 ,
" semicolon " = > 41 ,
" quote " | " apostrophe " = > 39 ,
" minus " | " hyphen " = > 27 ,
" equal " | " equals " = > 24 ,
" grave " | " backtick " | " backquote " = > 50 ,
" bracketleft " | " leftbracket " = > 33 ,
" bracketright " | " rightbracket " = > 30 ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
_ = > u16 ::MAX ,
} ;
if named ! = u16 ::MAX {
return Some ( named ) ;
}
let mut cs = n . chars ( ) ;
let ( Some ( c ) , None ) = ( cs . next ( ) , cs . next ( ) ) else { return None } ;
const LETTERS : [ u16 ; 26 ] = [ 0 , 11 , 8 , 2 , 14 , 3 , 5 , 4 , 34 , 38 , 40 , 37 , 46 , 45 , 31 , 35 , 12 , 15 , 1 , 17 , 32 , 9 , 13 , 7 , 16 , 6 ] ;
const DIGITS : [ u16 ; 10 ] = [ 29 , 18 , 19 , 20 , 21 , 23 , 22 , 26 , 28 , 25 ] ;
match c {
'a' ..= 'z' = > Some ( LETTERS [ ( c as u8 - b 'a' ) as usize ] ) ,
'0' ..= '9' = > Some ( DIGITS [ ( c as u8 - b '0' ) as usize ] ) ,
'-' = > Some ( 27 ) ,
'=' = > Some ( 24 ) ,
'[' = > Some ( 33 ) ,
']' = > Some ( 30 ) ,
';' = > Some ( 41 ) ,
'\'' = > Some ( 39 ) ,
',' = > Some ( 43 ) ,
'.' = > Some ( 47 ) ,
'/' = > Some ( 44 ) ,
'\\' = > Some ( 42 ) ,
'`' = > Some ( 50 ) ,
_ = > None ,
}
}
pub const FLAG_SHIFT : u64 = 1 < < 17 ;
pub const FLAG_CONTROL : u64 = 1 < < 18 ;
pub const FLAG_OPTION : u64 = 1 < < 19 ;
pub const FLAG_CMD : u64 = 1 < < 20 ;
pub fn modifier_flag ( m : & str ) -> Option < u64 > {
match m . trim ( ) . to_lowercase ( ) . as_str ( ) {
" cmd " | " command " | " meta " | " super " | " ⌘ " = > Some ( FLAG_CMD ) ,
" shift " = > Some ( FLAG_SHIFT ) ,
" option " | " alt " | " opt " = > Some ( FLAG_OPTION ) ,
" control " | " ctrl " = > Some ( FLAG_CONTROL ) ,
_ = > None ,
}
}
/// "cmd+shift+t" → (key code, flags).
pub fn parse_combo ( combo : & str ) -> Result < ( u16 , u64 ) , String > {
let parts : Vec < & str > = combo . split ( '+' ) . map ( str ::trim ) . filter ( | p | ! p . is_empty ( ) ) . collect ( ) ;
let Some ( ( key , mods ) ) = parts . split_last ( ) else { return Err ( " 請給要按的鍵 " . into ( ) ) } ;
let mut flags = 0 ;
for m in mods {
flags | = modifier_flag ( m ) . ok_or_else ( | | format! ( " 不認得「 {m} 」這個修飾鍵 " ) ) ? ;
}
let code = key_code ( key ) . ok_or_else ( | | format! ( " 不認得「 {key} 」這個鍵 " ) ) ? ;
Ok ( ( code , flags ) )
}
fn flags_of ( v : & Value ) -> u64 {
v [ " modifier_keys " ] . as_array ( ) . map ( | a | a . iter ( ) . filter_map ( | m | m . as_str ( ) . and_then ( modifier_flag ) ) . fold ( 0 , | f , x | f | x ) ) . unwrap_or ( 0 )
}
/// Run a computer_use tool. Returns (text, ok, screenshot as base64 JPEG).
pub fn run ( name : & str , args : & str ) -> ( String , bool , Option < String > ) {
run_by ( name , args , " " )
}
/// Same, remembering which bot acted (for the 「正在操作你的電腦」 bar).
pub fn run_by ( name : & str , args : & str , who : & str ) -> ( String , bool , Option < String > ) {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
if moves_things ( name ) {
2026-11-05 07:48:22 +00:00
* LAST_ACTION . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) = Some ( ( std ::time ::Instant ::now ( ) , who . to_string ( ) ) ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
Remove what nothing uses and keep one copy of what was written several times
Unused (found by taking away the dead-code allowances; nothing else used them, tests aside):
computer::specs, quiet_button, the never-made Poke::Spin, ten icons and their SVGs, the
theme's chrome/bubble_agent colours, ImportRow.line, voice's multipart/transcribe/speak
wrappers, and the old keychain writers vault::save/remove/index_add with their live test
(the vault store replaced them; reading old keychain entries for the import stays).
One copy each: percent-encoding (httpd::url_encode, was four), tool arguments (tools::args_of,
was three), the WebSocket handshake (httpd::accept_websocket, was three), the read pacing
(wire::Pace, was two: relay and app), docker with input and a time limit (dk_with_input, was
two). Docs: FEATURES §看畫面 (VNC), protocol (v* frames, /screen, tickets), architecture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:21:47 +00:00
match run_inner ( name , & crate ::tools ::args_of ( args ) ) {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
Ok ( ( t , img ) ) = > ( t , true , img ) ,
Err ( e ) = > ( e , false , None ) ,
}
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Steps that use the mouse or keyboard (and come back with the screen after them).
pub fn moves_things ( name : & str ) -> bool {
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
matches! ( name , " computer_use_click " | " computer_use_typing " | " computer_use_press_key " | " computer_use_scroll " | " computer_use_mouse_move " | " computer_use_drag " | " computer_use_open_app " | " computer_use_batch " )
2026-11-04 09:19:59 +00:00
}
2026-11-05 08:57:44 +00:00
/// Ask macOS for a permission (its own prompt, then 系統設定 at that pane) at most once every ten
/// minutes. Asking on every step meant a bot working through several steps — or a brain far away
/// working on this computer — covered the screen with prompts and Settings windows, and the
/// person could do nothing else.
pub fn ask_permission ( pane : & str ) {
static LAST : std ::sync ::Mutex < Vec < ( String , std ::time ::Instant ) > > = std ::sync ::Mutex ::new ( Vec ::new ( ) ) ;
{
let mut l = LAST . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) ;
if l . iter ( ) . any ( | ( p , t ) | p = = pane & & t . elapsed ( ) < std ::time ::Duration ::from_secs ( 600 ) ) {
return ;
}
l . retain ( | ( p , _ ) | p ! = pane ) ;
l . push ( ( pane . to_string ( ) , std ::time ::Instant ::now ( ) ) ) ;
}
if ! cfg! ( target_os = " macos " ) {
return ;
}
match pane {
" Privacy_Accessibility " = > {
let _ = crate ::axread ::trusted ( true ) ;
}
#[ cfg(target_os = " macos " ) ]
" Privacy_ScreenCapture " = > {
let _ = mac ::screen_capture_allowed ( ) ;
}
_ = > { }
}
Clippy is clean on macOS and Linux, for the app and the relay
Every warning `cargo clippy --all-targets` gave (50 on macOS, 8 more only on
Linux, 1 in the relay), fixed without changing what anything does:
- long types get a name (Peer, Members, Imported, Joining, Mention, Stream…);
- the two big enum variants are boxed (the vault's login plan, joining an
account);
- a test now waits for the copy it relaunched (it left a zombie behind);
- a hint's format had a dead `if` whose two branches were both "";
- the rest is what clippy suggests: sort_by_key, as_chunks, is_none_or,
is_multiple_of, while let, from_ref, next_back, struct init, no casts to
the same type.
Checked: cargo test (177 macOS, 181 Linux), relay tests (11), remote_e2e
28/28, and the live Linux test (X11; Wayland clicks, keys, Chinese typing,
scroll and screenshots).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-07 02:10:28 +00:00
let _ = spawn_reaped ( std ::process ::Command ::new ( " open " ) . arg ( format! ( " x-apple.systempreferences:com.apple.preference.security? {pane} " ) ) ) ;
2026-11-05 08:57:44 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
const NO_ACCESSIBILITY : & str = " macOS 還沒讓 AnyBot 使用「輔助使用」,所以不能讀 App, 也不能動滑鼠和鍵盤。這台電腦上已經跳出詢問並打開「系統設定 → 隱私權與安全性 → 輔助使用」:請使用者把 AnyBot 打開(只要一次),再叫我試一次。 " ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
const NO_SCREEN_RECORDING : & str = " macOS 還沒讓 AnyBot 錄製螢幕,所以看不到畫面。這台電腦上已經打開「系統設定 → 隱私權與安全性 → 螢幕與系統錄音」:請使用者把 AnyBot 打開(只要一次)。在那之前 computer_use_screenshot 還是會列出最前面 App 的文字和按鈕編號(只是沒有圖)。 " ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Reading apps and using the mouse and keyboard need 輔助使用 on macOS (other systems: always).
fn controls_allowed ( ) -> bool {
! cfg! ( target_os = " macos " ) | | crate ::axread ::trusted ( false )
}
2026-11-05 10:50:08 +00:00
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
fn need_controls ( ) -> Result < ( ) , String > {
if controls_allowed ( ) {
return Ok ( ( ) ) ;
2026-11-05 10:50:08 +00:00
}
ask_permission ( " Privacy_Accessibility " ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
Err ( NO_ACCESSIBILITY . into ( ) )
2026-11-05 10:50:08 +00:00
}
2026-11-05 08:57:44 +00:00
/// The app a shown name stands for (「計算機」 → Calculator), from the apps' display names.
#[ cfg(target_os = " macos " ) ]
fn app_named ( name : & str ) -> Option < String > {
if name . is_empty ( ) | | name . contains ( [ '"' , '\\' , '\'' ] ) {
return None ;
}
let q = format! ( " kMDItemDisplayName == \" {name} \" cd && kMDItemContentType == \" com.apple.application-bundle \" " ) ;
let out = std ::process ::Command ::new ( " mdfind " ) . arg ( q ) . output ( ) . ok ( ) ? ;
let path = String ::from_utf8_lossy ( & out . stdout ) . lines ( ) . next ( ) ? . trim ( ) . to_string ( ) ;
std ::path ::Path ::new ( & path ) . file_stem ( ) . map ( | s | s . to_string_lossy ( ) . to_string ( ) )
}
#[ cfg(not(target_os = " macos " )) ]
fn app_named ( _name : & str ) -> Option < String > {
None
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Open an app or bring it to the front. A name as the person sees it (「計算機」 on a Chinese
/// Mac) isn't one the system launches by; find the app it is and launch that.
fn open_app ( app : & str ) -> Result < String , String > {
let app = app . trim ( ) ;
if app . is_empty ( ) {
return Err ( " 請給要打開的 App 名稱 " . into ( ) ) ;
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
let real = app_named ( app ) ;
launch ( app ) . or_else ( | e | match & real {
Some ( r ) if r ! = app = > launch ( r ) ,
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
_ = > Err ( e ) ,
} ) ? ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
// Wait for its window to be in front (switching desktops takes a moment) before looking.
let start = std ::time ::Instant ::now ( ) ;
while start . elapsed ( ) < std ::time ::Duration ::from_millis ( 2500 ) {
let front = mac ::front_app ( ) . map ( | f | f . 0. to_lowercase ( ) ) . unwrap_or_default ( ) ;
if front = = app . to_lowercase ( ) | | real . as_ref ( ) . map ( | r | front = = r . to_lowercase ( ) ) . unwrap_or ( false ) {
break ;
}
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 100 ) ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
Ok ( format! ( " 已打開「 {app} 」,放在最前面。 " ) )
2026-11-04 09:19:59 +00:00
}
People watch and help on the bots' computer through VNC (noVNC), carried by the account server
The core: the bot works the computer; a person watches it like a colleague, and takes over
when it needs them. VNC is made for that (only what changed, every mouse move and key, touch on
phones), so the pictures sent one by one go, and with them their own input path.
- vnc.rs: a VNC proxy to the box's x11vnc through one docker exec (a Python bridge that ends
with either side). It signs in to the box with the box's own password (its VNC_PASSWORD) and
offers the viewer no password, so the password never leaves the brain. The person's clicks
and keys make the bot's steps wait until 8 s after they stop; text typed in the viewer's box
(any language) is pasted in the box.
- Account server: /screen (the viewer page) and /novnc.js (noVNC 1.6.0, MPL-2.0, bundled);
/relay/vnc/<host> takes the browser's WebSocket (account session, or a 10-minute ticket the
app asks for with POST /relay/screen-ticket; owners and admins) and carries it over that
AnyBot's lasting connection (vopen / vdata / vtext / vclose). Both ends of that connection
wait 2 ms instead of 40 ms to read while things move: a request and its answer took up to
160 ms more than needed.
- Phone page: 看 Bot 的電腦 and 接管 open the viewer page (also on the local network: /api/vnc on
the AnyBot's own phone server). This Mac's own screen moved to the ⋯ menu.
- Mac app: 看畫面 opens the viewer in the browser (a screen of a brain: the brain's computer by
ticket; else this computer's box, noVNC straight from it). The in-app picture viewer, the
live JPEG stream, the person's click/move/scroll/key endpoints and the input shell are gone.
- Opening a Mac app uses LaunchServices (`open -a`): agent-desktop's launch waited on the window
through accessibility and failed without 輔助使用 now and then.
Measured on this Mac (made-up account, local relay + headless brain + the real box): viewer
signed in in 0.4–0.7 s, a request and its answer through relay + brain + box 25–58 ms median;
headless Chrome shows the desktop live in /screen. remote_e2e 28/28 ×2, 149 unit tests, relay 11.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:05:44 +00:00
/// Open an app or bring it forward. macOS: LaunchServices (`open -a`), which needs no
/// permission (agent-desktop's launch waited on the window through accessibility and failed
/// without 輔助使用 now and then); elsewhere agent-desktop's.
fn launch ( app : & str ) -> Result < ( ) , String > {
#[ cfg(target_os = " macos " ) ]
{
let out = std ::process ::Command ::new ( " open " ) . args ( [ " -a " , app ] ) . output ( ) . map_err ( | e | e . to_string ( ) ) ? ;
if out . status . success ( ) {
return Ok ( ( ) ) ;
}
Err ( format! ( " 找不到「 {app} 」這個 App " ) )
}
#[ cfg(not(target_os = " macos " )) ]
crate ::desktop ::launch ( app )
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Bring the app a control belongs to in front, so a click at its spot lands in it (and not in
/// a window lying over it). Nothing to do when it is in front already.
fn bring_forward ( app : & str ) {
// macOS only: there "launch" brings a running app forward; on Linux it can start a second
// copy, and the names a window and its accessibility give differ.
if ! cfg! ( target_os = " macos " ) {
return ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let front = mac ::front_app ( ) . map ( | f | f . 0 ) . unwrap_or_default ( ) ;
if app . is_empty ( ) | | front . eq_ignore_ascii_case ( app ) {
return ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
People watch and help on the bots' computer through VNC (noVNC), carried by the account server
The core: the bot works the computer; a person watches it like a colleague, and takes over
when it needs them. VNC is made for that (only what changed, every mouse move and key, touch on
phones), so the pictures sent one by one go, and with them their own input path.
- vnc.rs: a VNC proxy to the box's x11vnc through one docker exec (a Python bridge that ends
with either side). It signs in to the box with the box's own password (its VNC_PASSWORD) and
offers the viewer no password, so the password never leaves the brain. The person's clicks
and keys make the bot's steps wait until 8 s after they stop; text typed in the viewer's box
(any language) is pasted in the box.
- Account server: /screen (the viewer page) and /novnc.js (noVNC 1.6.0, MPL-2.0, bundled);
/relay/vnc/<host> takes the browser's WebSocket (account session, or a 10-minute ticket the
app asks for with POST /relay/screen-ticket; owners and admins) and carries it over that
AnyBot's lasting connection (vopen / vdata / vtext / vclose). Both ends of that connection
wait 2 ms instead of 40 ms to read while things move: a request and its answer took up to
160 ms more than needed.
- Phone page: 看 Bot 的電腦 and 接管 open the viewer page (also on the local network: /api/vnc on
the AnyBot's own phone server). This Mac's own screen moved to the ⋯ menu.
- Mac app: 看畫面 opens the viewer in the browser (a screen of a brain: the brain's computer by
ticket; else this computer's box, noVNC straight from it). The in-app picture viewer, the
live JPEG stream, the person's click/move/scroll/key endpoints and the input shell are gone.
- Opening a Mac app uses LaunchServices (`open -a`): agent-desktop's launch waited on the window
through accessibility and failed without 輔助使用 now and then.
Measured on this Mac (made-up account, local relay + headless brain + the real box): viewer
signed in in 0.4–0.7 s, a request and its answer through relay + brain + box 25–58 ms median;
headless Chrome shows the desktop live in /screen. remote_e2e 28/28 ×2, 149 unit tests, relay 11.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:05:44 +00:00
if launch ( app ) . is_ok ( ) {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 300 ) ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// The numbered controls of this computer's latest look, and the app they belong to.
static SPOTS : Mutex < Option < ( String , crate ::marks ::Spots ) > > = Mutex ::new ( None ) ;
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
/// How long the menu bar's status icons are trusted before they are read again (they rarely move).
#[ cfg(target_os = " macos " ) ]
const MENU_BAR_FRESH_SECS : u64 = 15 ;
/// The menu bar's status icons as numbered controls, from every app's AXExtrasMenuBar.
#[ cfg(target_os = " macos " ) ]
fn menu_bar_controls ( ) -> Vec < crate ::marks ::Element > {
static CACHE : Mutex < Option < ( std ::time ::Instant , Vec < crate ::marks ::Element > ) > > = Mutex ::new ( None ) ;
static REFRESHING : std ::sync ::atomic ::AtomicBool = std ::sync ::atomic ::AtomicBool ::new ( false ) ;
let cached = CACHE . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) . clone ( ) ;
if let Some ( ( t , v ) ) = cached {
// Stale: use what there is now, and read them again in the background for the next look.
if t . elapsed ( ) . as_secs ( ) > = MENU_BAR_FRESH_SECS & & ! REFRESHING . swap ( true , std ::sync ::atomic ::Ordering ::SeqCst ) {
std ::thread ::spawn ( | | {
let fresh = read_menu_bar ( ) ;
* CACHE . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) = Some ( ( std ::time ::Instant ::now ( ) , fresh ) ) ;
REFRESHING . store ( false , std ::sync ::atomic ::Ordering ::SeqCst ) ;
} ) ;
}
return v ;
}
let els = read_menu_bar ( ) ;
* CACHE . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) = Some ( ( std ::time ::Instant ::now ( ) , els . clone ( ) ) ) ;
els
}
#[ cfg(target_os = " macos " ) ]
fn read_menu_bar ( ) -> Vec < crate ::marks ::Element > {
crate ::axread ::status_items ( & app_processes ( ) )
. into_iter ( )
. map ( | it | crate ::marks ::Element {
kind : " 選單列圖示 " . into ( ) ,
name : icon_name ( & it . app , & it . label ) ,
value : String ::new ( ) ,
x : it . rect [ 0 ] ,
y : it . rect [ 1 ] ,
w : it . rect [ 2 ] ,
h : it . rect [ 3 ] ,
active : true ,
flags : vec ! [ ] ,
window : " 選單列 " . into ( ) ,
} )
. collect ( )
}
/// How a status icon is listed: the system's own (hosted by an agent process) by their words
/// alone (「Wi‑ Fi, 已連接」, not 「MenuBarAgent: Wi‑ Fi…」), an app's as 「app: words」.
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
fn icon_name ( app : & str , label : & str ) -> String {
let system = app . ends_with ( " Agent " ) | | matches! ( app , " ControlCenter " | " Control Center " | " SystemUIServer " | " Siri AI " | " Spotlight " ) ;
match ( label . trim ( ) , system ) {
( " " , _ ) = > app . to_string ( ) ,
( l , true ) = > l . to_string ( ) ,
( l , false ) if l = = app = > l . to_string ( ) ,
( l , false ) = > format! ( " {app} : {l} " ) ,
}
}
/// This user's running apps and app helpers (anything run from an .app bundle): pid and name.
#[ cfg(target_os = " macos " ) ]
fn app_processes ( ) -> Vec < ( i32 , String ) > {
let Ok ( out ) = std ::process ::Command ::new ( " /bin/ps " ) . args ( [ " -x " , " -o " , " pid=,comm= " ] ) . output ( ) else { return vec! [ ] } ;
String ::from_utf8_lossy ( & out . stdout )
. lines ( )
. filter_map ( | l | {
let ( pid , path ) = l . trim_start ( ) . split_once ( ' ' ) ? ;
let path = path . trim ( ) ;
let i = path . find ( " .app/Contents/MacOS/ " ) ? ;
let name = path [ .. i ] . rsplit ( '/' ) . next ( ) ? . to_string ( ) ;
Some ( ( pid . parse ( ) . ok ( ) ? , name ) )
} )
. collect ( )
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// The latest look's fingerprint (to say when a step changed nothing).
static LAST_LOOK : Mutex < Option < u64 > > = Mutex ::new ( None ) ;
/// Where a step acts: a control's number on the latest picture, or x, y on the picture.
/// Returns the screen point and how to say it (「#7( 按鈕「7」) 」, 「(512, 300)」).
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
fn place_of ( v : & Value , scale : ( f64 , f64 ) ) -> Result < Option < ( f64 , f64 , String ) > , String > {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
if let Some ( r ) = v [ " ref " ] . as_str ( ) . map ( str ::trim ) . filter ( | r | ! r . is_empty ( ) ) . or_else ( | | v [ " ref " ] . as_u64 ( ) . map ( | _ | " " ) ) . map ( | r | if r . is_empty ( ) { v [ " ref " ] . to_string ( ) } else { r . to_string ( ) } ) {
let ( app , spots ) = SPOTS . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) . clone ( ) . ok_or ( " 先用 computer_use_screenshot 看一下畫面,再照圖上的編號操作。 " ) ? ;
let ( ( x , y ) , what ) = spots . get ( & r ) ? ;
bring_forward ( & app ) ;
return Ok ( Some ( ( x , y , format! ( " # {} ( {what} ) " , crate ::marks ::number_of ( & r ) . unwrap_or ( 0 ) ) ) ) ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
match ( v [ " x " ] . as_f64 ( ) , v [ " y " ] . as_f64 ( ) ) {
( Some ( x ) , Some ( y ) ) = > {
let ( sx , sy ) = to_screen ( x , y , scale ) ;
Ok ( Some ( ( sx , sy , format! ( " ( {x} , {y} ) " ) ) ) )
}
( None , None ) if v [ " x " ] . is_null ( ) & & v [ " y " ] . is_null ( ) = > Ok ( None ) ,
_ = > Err ( " x 和 y 要一起給,而且要是數字(截圖上的位置) " . into ( ) ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
/// Press each key of "cmd+a delete" in order.
fn press_keys ( keys : & str , hold : f64 ) -> Result < ( ) , String > {
for ( i , k ) in key_plan ( keys ) ? . iter ( ) . enumerate ( ) {
if i > 0 {
std ::thread ::sleep ( std ::time ::Duration ::from_secs_f64 ( KEY_GAP_SECS ) ) ;
}
match k {
Key ::Combo ( code , flags ) = > mac ::press ( * code , * flags , hold ) ,
Key ::Char ( ch ) = > mac ::type_text ( ch ) ,
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
Ok ( ( ) )
}
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
/// Run a planned sequence of actions, then one picture (in the box the box's own batch runs;
/// here, the local screen). The first action that fails stops the rest — acting on a screen
/// that isn't where you think it is makes it worse (Anthropic's computer_batch rule).
pub fn batch ( v : & Value , scale : ( f64 , f64 ) ) -> Result < String , String > {
let acts = v [ " actions " ] . as_array ( ) . ok_or ( " computer_use_batch 要給 actions: 動作陣列, 每個有 action( click、typing、press_key、scroll、mouse_move、drag) 和它的參數 " ) ? ;
if acts . is_empty ( ) {
return Err ( " actions 是空的 " . into ( ) ) ;
}
if acts . len ( ) > MAX_BATCH {
return Err ( format! ( " 一次最多 {MAX_BATCH} 個動作( {} 個太多了) " , acts . len ( ) ) ) ;
}
let mut notes = Vec ::new ( ) ;
for ( i , a ) in acts . iter ( ) . enumerate ( ) {
let name = action_tool ( a [ " action " ] . as_str ( ) . unwrap_or ( " " ) ) ? ;
match act ( name , a , scale ) {
Ok ( n ) = > notes . push ( format! ( " {} . {n} " , i + 1 ) ) ,
Err ( e ) = > {
let e = plain_words ( e ) ;
if notes . is_empty ( ) {
return Err ( format! ( " 第 {} 步就失敗: {e} " , i + 1 ) ) ;
}
notes . push ( format! ( " {} . 失敗: {e} (後面 {} 個沒有做) " , i + 1 , acts . len ( ) - i - 1 ) ) ;
break ;
}
}
}
Ok ( format! ( " 已依序做完 {} 步: \n {} " , notes . len ( ) , notes . join ( " \n " ) ) )
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Do one mouse or keyboard step at the screen's own coordinates (`scale`: screen points per
/// picture pixel). The bot's tools and a person tapping on the phone's picture both come here:
/// one way of acting on a screen.
pub fn act ( name : & str , v : & Value , scale : ( f64 , f64 ) ) -> Result < String , String > {
let button = v [ " button " ] . as_str ( ) . unwrap_or ( " left " ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
match name {
" computer_use_click " = > {
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
let ( x , y , at ) = place_of ( v , scale ) ? . ok_or ( " 請給 x、y( 截圖上的位置) 或 ref( 截圖上的編號) " ) ? ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let count = v [ " count " ] . as_u64 ( ) . unwrap_or ( 1 ) . clamp ( 1 , 3 ) as i64 ;
mac ::click ( x , y , button , count , flags_of ( v ) ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let how = format! ( " {} {} " , if button = = " right " { " 右鍵 " } else { " " } , if count > = 2 { format! ( " 連點 {count} 下 " ) } else { " 點 " . into ( ) } ) ;
Ok ( format! ( " 已 {how} {at} 。 " ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_mouse_move " = > {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let ( x , y , at ) = place_of ( v , scale ) ? . ok_or ( " 請給 x 和 y " ) ? ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
mac ::move_to ( x , y ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
Ok ( format! ( " 已把滑鼠移到 {at} 。 " ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_scroll " = > {
// CURemoteToolService.handleScroll @0x10009e200 / closure @0x10009e998.
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let at = place_of ( v , scale ) ? ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
if ! v [ " amount " ] . is_null ( ) & & v [ " amount " ] . as_i64 ( ) . is_none ( ) {
return Err ( " amount 要是整數 " . into ( ) ) ;
}
let amount = scroll_amount ( v [ " amount " ] . as_i64 ( ) ) ;
let dir = v [ " direction " ] . as_str ( ) . unwrap_or ( " " ) . trim ( ) . to_lowercase ( ) ;
let ( dy , dx ) = scroll_pixels ( & dir , amount ) . ok_or_else ( | | format! ( " 方向「 {dir} 」要是 up、down、left 或 right " ) ) ? ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
mac ::scroll ( at . as_ref ( ) . map ( | ( x , y , _ ) | ( * x , * y ) ) , dy , dx , flags_of ( v ) ) ;
Ok ( format! ( " 已捲動 {dir} × {amount} {} 。 " , at . map ( | ( _ , _ , a ) | format! ( " ,在 {a} " ) ) . unwrap_or_default ( ) ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_drag " = > {
2026-11-05 07:48:22 +00:00
// Every point must be valid: points that failed to parse used to be dropped in
// silence, so the drag ran on fewer points and reported success at the wrong place.
let raw = v [ " path " ] . as_array ( ) . ok_or ( " 拖曳要給 path: 一個點陣, 每個點有數字 x 和 y( 用截圖上的座標) " ) ? ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let pts : Vec < ( f64 , f64 ) > = raw . iter ( ) . map ( | p | match ( p [ " x " ] . as_f64 ( ) , p [ " y " ] . as_f64 ( ) ) {
( Some ( x ) , Some ( y ) ) = > Ok ( ( x , y ) ) ,
_ = > Err ( " 拖曳的每個點都要有數字 x 和 y( 用截圖上的座標) " . to_string ( ) ) ,
} ) . collect ::< Result < _ , _ > > ( ) ? ;
if pts . len ( ) < 2 {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
return Err ( " 拖曳至少要兩個點(每個點要有數字 x 和 y, 用截圖上的座標) " . into ( ) ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let path : Vec < ( f64 , f64 ) > = pts . iter ( ) . map ( | ( x , y ) | to_screen ( * x , * y , scale ) ) . collect ( ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
mac ::drag ( & path , button , flags_of ( v ) ) ;
// Like the original's "Dragged from (x, y) to (x, y) with left button."
let side = match button {
" right " = > " 右鍵 " ,
" middle " = > " 中鍵 " ,
_ = > " 左鍵 " ,
} ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let ( a , b ) = ( pts [ 0 ] , pts [ pts . len ( ) - 1 ] ) ;
Ok ( format! ( " 已從 ( {:.0} , {:.0} ) 拖曳到 ( {:.0} , {:.0} )( {side} )。 " , a . 0 , a . 1 , b . 0 , b . 1 ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_typing " = > {
let text = v [ " value " ] . as_str ( ) . unwrap_or ( " " ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let submit = v [ " submit " ] . as_bool ( ) . unwrap_or ( false ) ;
if text . is_empty ( ) & & ! submit {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
return Err ( " 請給要打的字 " . into ( ) ) ;
}
2026-11-09 02:19:18 +00:00
// Into a field by its number: click it first, like a person would; a one-line field
// gets what it holds replaced (cloudbox::replaces_field).
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let at = place_of ( & json! ( { " ref " : v [ " ref " ] } ) , scale ) ? ;
2026-11-09 02:19:18 +00:00
if let Some ( ( x , y , what ) ) = & at {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
mac ::click ( * x , * y , " left " , 1 , 0 ) ;
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 150 ) ) ;
2026-11-09 03:25:17 +00:00
if crate ::cloudbox ::replaces_field ( what , v ) {
2026-11-09 02:19:18 +00:00
press_keys ( " cmd+a " , KEY_HOLD_SECS ) ? ;
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 50 ) ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
if ! text . is_empty ( ) {
mac ::type_text ( text ) ;
}
if submit {
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 60 ) ) ;
mac ::press ( 36 , 0 , KEY_HOLD_SECS ) ;
}
Ok ( format! ( " 已打字( {} 個字) {} {} 。 " , text . chars ( ) . count ( ) , at . map ( | ( _ , _ , a ) | format! ( " 在 {a} " ) ) . unwrap_or_default ( ) , if submit { " ,再按了 return " } else { " " } ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
" computer_use_press_key " = > {
let hold = v [ " hold_duration_ms " ] . as_u64 ( ) . map ( | ms | ms . min ( MAX_HOLD_MS ) as f64 / 1000.0 ) . filter ( | s | * s > 0.0 ) . unwrap_or ( KEY_HOLD_SECS ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
press_keys ( v [ " key " ] . as_str ( ) . unwrap_or ( " " ) , hold ) ? ;
Ok ( format! ( " 已按 {} 。 " , v [ " key " ] . as_str ( ) . unwrap_or ( " " ) ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
" computer_use_open_app " = > open_app ( v [ " app " ] . as_str ( ) . unwrap_or ( " " ) ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
other = > Err ( format! ( " 沒有「 {other} 」這個工具 " ) ) ,
}
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// What the screen looks like now: the picture with the controls of the app in front boxed and
/// numbered, and the list of them. Without a picture (no 螢幕錄製), the list with the app's text.
/// After a step (`after`), it also says when nothing changed, so the bot doesn't click the same
/// empty place again and again.
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
fn look ( after : Option < String > , pointer : bool ) -> Result < ( String , Option < String > ) , String > {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
let picture = if mac ::screen_capture_preflight ( ) { Some ( capture_screen ( ) . map_err ( | e | format! ( " 看不到畫面: {e} " ) ) ? ) } else { None } ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
// The app in front: on Linux the window system's own answer when it gives one (agent-desktop
// alone goes by AT-SPI's "active" window, which some desktops never mark: no numbers, 「」).
2026-11-07 16:58:18 +00:00
// agent-desktop knows an app by its program's name (a Snap Firefox's window says
// "firefox_firefox", its program is "firefox").
#[ cfg(target_os = " linux " ) ]
let front = mac ::front_app ( ) . map ( | ( n , pid ) | crate ::desk ::program_of ( pid ) . unwrap_or ( n ) ) . unwrap_or_default ( ) ;
#[ cfg(not(target_os = " linux " )) ]
let front = String ::new ( ) ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
let ( found , why ) = if controls_allowed ( ) {
2026-11-09 05:31:25 +00:00
let read = | | crate ::desktop ::elements ( & front ) . or_else ( | e | if front . is_empty ( ) { Err ( e ) } else { crate ::desktop ::elements ( " " ) } ) ;
// Right after a click some apps have no focused window for a moment (macOS Calculator
// after 「乘」): the list came back with only the menu bar. One more look shortly after.
let first = read ( ) ;
let again = first . is_err ( ) | | first . as_ref ( ) . is_ok_and ( | f | f . 0. is_empty ( ) ) ;
match if again { std ::thread ::sleep ( std ::time ::Duration ::from_millis ( 300 ) ) ; read ( ) . or ( first ) } else { first } {
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
Ok ( f ) = > ( Some ( f ) , String ::new ( ) ) ,
Err ( e ) = > ( None , e ) ,
}
} else {
( None , String ::new ( ) )
} ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
if picture . is_none ( ) & & found . is_none ( ) {
ask_permission ( " Privacy_ScreenCapture " ) ;
return Err ( NO_SCREEN_RECORDING . into ( ) ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
2026-11-07 16:58:18 +00:00
// No controls at all, and the desktop's accessibility switch was off: turn it on (Linux).
#[ cfg(target_os = " linux " ) ]
let a11y_note = if controls_allowed ( ) & & found . as_ref ( ) . is_none_or ( | f | f . 2. is_empty ( ) ) & & crate ::desk ::accessibility_switched_on ( ) {
" \n ( 這台的無障礙介面原本是關著的, AnyBot 剛把它打開:之後打開的 App 才會有編號,已經開著的 App 要關掉重開。網頁請改用 browser_* 工具。)"
} else {
" "
} ;
#[ cfg(not(target_os = " linux " )) ]
let a11y_note = " " ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
let ( app , window , mut els ) = found . unwrap_or_default ( ) ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
// Only what is on this screen gets a number (a link below the fold was numbered at the
// picture's edge, and a click there landed on the edge).
let ( sw , sh ) = mac ::screen_points ( ) ;
els . retain ( | e | {
let ( cx , cy ) = ( e . x + e . w / 2.0 , e . y + e . h / 2.0 ) ;
sw < = 0.0 | | ( cx > = 0.0 & & cy > = 0.0 & & cx < sw & & cy < sh )
} ) ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
// A number is a promise that a click there lands in this app: a control another window
// covers (or one on another desktop) gets none.
let hidden = cover_check ( window , & mut els ) ;
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
// The menu bar's status icons (Tailscale, Wi-Fi, a VPN…) belong to no window in front, so
// they had no numbers and a bot cropped and measured screenshots to find them.
#[ cfg(target_os = " macos " ) ]
if controls_allowed ( ) {
els . extend ( menu_bar_controls ( ) ) ;
}
// Numbers kept from the last look where the same control is still there (marks::Numbering).
static NUMBERS : Mutex < Option < crate ::marks ::Numbering > > = Mutex ::new ( None ) ;
let ( numbered , lines , spots ) = {
let mut m = NUMBERS . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) ;
crate ::marks ::number_keeping ( & els , picture . is_none ( ) , m . get_or_insert_with ( Default ::default ) )
} ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
* SPOTS . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) = Some ( ( app . clone ( ) , spots ) ) ;
let mut text = after . clone ( ) . map ( | n | format! ( " {n} \n " ) ) . unwrap_or_default ( ) ;
let mut fp = crate ::marks ::fingerprint ( & els ) ;
let mut img_b64 = None ;
if let Some ( img ) = picture {
let ( fw , fh ) = shot_size ( ) ;
let mut img = if img . dimensions ( ) ! = ( fw , fh ) { image ::imageops ::resize ( & img , fw , fh , image ::imageops ::FilterType ::Triangle ) } else { img } ;
fp ^ = crate ::marks ::picture_fingerprint ( & img ) . rotate_left ( 1 ) ;
crate ::marks ::draw ( & mut img , & numbered , scale_here ( ) ) ;
let jpeg = encode_jpeg ( & img ) . ok_or ( " 截圖太大了,沒辦法送出 " ) ? ;
img_b64 = Some ( base64 ::engine ::general_purpose ::STANDARD . encode ( & jpeg ) ) ;
text . push_str ( & format! ( " 螢幕截圖( {fw} × {fh} )。 " ) ) ;
} else {
text . push_str ( " ( 這台現在拍不到畫面: macOS 還沒允許「螢幕與系統錄音」。下面是最前面 App 的內容。)" ) ;
}
let unchanged = {
let mut last = LAST_LOOK . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
let same = pointer & & after . is_some ( ) & & * last = = Some ( fp ) ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
* last = Some ( fp ) ;
same
} ;
if unchanged {
text . push_str ( " \n ⚠️ 畫面和上一步完全一樣:這一步沒有點到東西。不要再做同一步;換一個編號、換一個方法(網頁用 browser_* 工具,或直接開網址),或用 hand_over_computer 請使用者幫忙。 " ) ;
}
if hidden > 0 {
text . push_str ( & format! ( " \n (「 {app} 」有 {hidden} 個按鈕被別的視窗擋住或不在這個桌面,沒有編號:先 computer_use_open_app 把它叫到前面。) " ) ) ;
}
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
if lines . is_empty ( ) & & app . is_empty ( ) {
let why = if why . is_empty ( ) { String ::new ( ) } else { format! ( " ( {} ) " , why . lines ( ) . next ( ) . unwrap_or ( " " ) ) } ;
text . push_str ( & format! ( " \n 看不出最前面是哪個 App {why} ,所以沒有編號:只好照圖上的位置點 x、y。 " ) ) ;
} else if lines . is_empty ( ) {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
text . push_str ( & format! ( " \n 「 {app} 」沒有可以照編號點的東西(有些 App 不提供):只好照圖上的位置點 x、y。 " ) ) ;
} else {
text . push_str ( & format! ( " \n 「 {app} 」上可以操作的東西(圖上同樣的編號,用 ref 點): \n {} " , lines . join ( " \n " ) ) ) ;
}
2026-11-07 16:58:18 +00:00
text . push_str ( a11y_note ) ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
Ok ( ( text , img_b64 ) )
}
/// Take the number away from controls whose spot another app's window covers (macOS; elsewhere
/// the window system doesn't say). Returns how many.
fn cover_check ( window : Option < u32 > , els : & mut [ crate ::marks ::Element ] ) -> usize {
#[ cfg(target_os = " macos " ) ]
{
let Some ( pid ) = window . and_then ( mac ::pid_of_window ) else { return 0 } ;
let mut n = 0 ;
for e in els . iter_mut ( ) . filter ( | e | e . active ) {
let ( x , y ) = e . centre ( ) ;
if mac ::pid_on_top_at ( x , y ) . map ( | p | p ! = pid ) . unwrap_or ( false ) {
e . active = false ;
n + = 1 ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
n
}
#[ cfg(not(target_os = " macos " )) ]
{
let _ = ( window , els ) ;
0
}
}
/// After a step: wait for the screen to settle, then look.
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
fn observe ( note : String , pointer : bool ) -> ( String , Option < String > ) {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
std ::thread ::sleep ( std ::time ::Duration ::from_millis ( SETTLE_MS ) ) ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
match look ( Some ( note . clone ( ) ) , pointer ) {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
Ok ( r ) = > r ,
Err ( e ) = > ( format! ( " {note} (看不到結果: {e} ) " ) , None ) ,
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
}
fn run_inner ( name : & str , v : & Value ) -> Result < ( String , Option < String > ) , String > {
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
// A control named instead of numbered (「按 4」): its number in the latest look.
let mut named = v . clone ( ) ;
if matches! ( name , " computer_use_click " | " computer_use_typing " | " computer_use_scroll " | " computer_use_mouse_move " | " computer_use_batch " ) {
let spots = SPOTS . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) . clone ( ) . map ( | ( _ , s ) | s ) ;
crate ::marks ::names_to_numbers ( & mut named , spots . as_ref ( ) ) ? ;
}
let v = & named ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
if ! cfg! ( any ( target_os = " macos " , target_os = " linux " ) ) {
return Err ( " 操作畫面只支援 macOS 和 Linux " . into ( ) ) ;
}
#[ cfg(target_os = " linux " ) ]
if ! crate ::desk ::available ( ) {
return Err ( crate ::desk ::describe ( ) . into ( ) ) ;
}
match name {
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
" computer_use_screenshot " = > return look ( None , false ) ,
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
" computer_use_zoom " = > {
let img = capture_screen ( ) ? ;
return zoom ( & img , v ) . map ( | ( t , j ) | ( t , Some ( j ) ) ) ;
}
Real-use fixes: the person sees the screens the bot shows, files reach the phone, no blank-tab loops
From real runs on the cloud brain with the person's Mac (Threads patrol, a post draft, "show me my
screen"):
- computer_use_show: a plain picture of any computer's screen goes into the chat (the bot sees it
too). The bot used to describe a screenshot in words and fail to send it.
- The phone shows the bot's attachments: pictures in place, other files download on a tap (api/file,
only from the chats' file folder). It showed nothing for the bot and names only for the person.
- send_file works from the person's other computers (fetch_file over the relay, 20 MB, secrets stay).
- The same call with the same answer again and again is warned the 3rd time and stopped from the
4th (a blank tab was read 11 times, then the bot said Threads was logged out without opening it);
a blank tab says so; checking a login means opening the site.
- Screenshot lists: a nameless row around a named button isn't numbered, a nameless row takes the
words inside it, empty spacer rows aren't numbered (System Settings: 63 numbers -> 33, all named).
- Writing "in my voice" reads the person's own recent posts first.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 11:37:45 +00:00
" computer_use_show " = > {
let ( w , h , jpeg ) = plain_picture ( ) ? ;
return Ok ( ( format! ( " {SHOWN_NOTE} ( {w} × {h} ) " ) , Some ( base64 ::engine ::general_purpose ::STANDARD . encode ( & jpeg ) ) ) ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
" computer_use_check_permissions " = > {
#[ cfg(target_os = " linux " ) ]
return Ok ( ( crate ::desk ::describe ( ) . into ( ) , None ) ) ;
#[ allow(unreachable_code) ]
{
let see = mac ::screen_capture_preflight ( ) ;
let control = controls_allowed ( ) ;
return Ok ( ( format! ( " 螢幕錄製(看畫面): {} \n 輔助使用(滑鼠、鍵盤、讀 App) : {} {} " , if see { " 已允許 " } else { " 還沒允許 " } , if control { " 已允許 " } else { " 還沒允許 " } , if see & & control { " " } else { " \n 請使用者到「系統設定 → 隱私權與安全性」打開 AnyBot。 " } ) , None ) ) ;
}
}
_ = > { }
}
if ! moves_things ( name ) {
return Err ( format! ( " 沒有「 {name} 」這個工具 " ) ) ;
}
if HELD_BY_PERSON . load ( std ::sync ::atomic ::Ordering ::SeqCst ) {
return Err ( " 使用者正在自己操作這台電腦(已取得控制權)。請等他交還控制權後再操作;可以先截圖看看。 " . into ( ) ) ;
}
if name ! = " computer_use_open_app " {
need_controls ( ) ? ;
}
Bots do a planned sequence in one step; the box's screen settles by what it shows; no login is ever the only copy
- computer_use_batch: up to 8 planned actions run in one docker exec (steps chained with && and a
done-marker each), the first failure stops the rest, one picture comes back at the end. Typed
text rides inside the script quoted (one stdin for the chain). The prompt tells the model when
to use it (a form whose fields are all visible) and when not to (step 2 needs step 1's screen).
- The box waits for the screen to stop changing, not the clock: two matching small probes (~60 ms
each) mean settled — ~160 ms on a still screen where every step paid a fixed 500 ms, and a page
mid-load is waited for (up to ~0.9 s) instead of captured mid-change. The picture is JPEG, and
the a11y read runs in parallel with the capture (look ~307 ms; two steps with pictures 1102 ms).
(SETTLE_SH is a group command: dash needs ';' after it — found by box_screen_is_exact.)
- The home volume (every bot's logins) is never one command from loss: snapshots keep generations
(last 7 days, then one per week; the old single snapshot migrates by file date), Reset and
Restore save the running box first and refuse to wipe if saving fails, a box built on an empty
volume re-fills itself from the newest snapshot, and a sleeping holder container keeps the
volume attached so `docker volume prune` can't take it. make doctor shows volume and snapshot
state; the 電腦 card shows the box's average step times and says resets save first.
- box_batch_in_one_round_trip (ignored): three typed steps, ASCII and Chinese, land in one trip.
On the real box after upgrading v3→v5: box_update_keeps_files (files survive, snapshot 2 代),
box_numbers_are_exact, box_screen_is_exact, box_end_to_end, box_traffic_via_mac pass; 155 unit tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-06 01:47:37 +00:00
let note = match name {
" computer_use_batch " = > batch ( v , scale_here ( ) ) ? ,
_ = > act ( name , v , scale_here ( ) ) . map_err ( plain_words ) ? ,
} ;
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
Ok ( observe ( note , is_pointer_step ( name ) ) )
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
}
/// agent-desktop's error codes, as words the bot can pass on.
fn plain_words ( t : String ) -> String {
if t . contains ( " PERM_DENIED " ) {
ask_permission ( " Privacy_Accessibility " ) ;
return NO_ACCESSIBILITY . to_string ( ) ;
}
if t . contains ( " APP_NOT_FOUND " ) {
return format! ( " 找不到這個 App( 沒有開著, 或名字不對) : 先用 computer_use_open_app 打開它。( {t} ) " ) ;
}
t
}
/// The phone or another viewer acting on this screen as the person: their tap is on the picture
/// they were sent (`shot_size()`), so it maps exactly like the bot's coordinates. Not refused
/// while the person holds the screen (it is the person), and no picture back (the viewer
/// refreshes its own).
pub fn person_act ( name : & str , v : & Value ) -> Result < ( ) , String > {
need_controls ( ) ? ;
act ( name , v , scale_here ( ) ) . map ( | _ | ( ) )
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// The app a screen action will land in: the window under the point for clicks, else the
/// frontmost app. Used for the per-app permission (「AnyBot 可以看到 {app},並在其中點擊和輸入」).
pub fn target_app ( name : & str , args : & str ) -> Option < String > {
Remove what nothing uses and keep one copy of what was written several times
Unused (found by taking away the dead-code allowances; nothing else used them, tests aside):
computer::specs, quiet_button, the never-made Poke::Spin, ten icons and their SVGs, the
theme's chrome/bubble_agent colours, ImportRow.line, voice's multipart/transcribe/speak
wrappers, and the old keychain writers vault::save/remove/index_add with their live test
(the vault store replaced them; reading old keychain entries for the import stays).
One copy each: percent-encoding (httpd::url_encode, was four), tool arguments (tools::args_of,
was three), the WebSocket handshake (httpd::accept_websocket, was three), the read pacing
(wire::Pace, was two: relay and app), docker with input and a time limit (dk_with_input, was
two). Docs: FEATURES §看畫面 (VNC), protocol (v* frames, /screen, tickets), architecture.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 15:21:47 +00:00
let v = crate ::tools ::args_of ( args ) ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
if let Some ( app ) = v [ " app " ] . as_str ( ) . filter ( | a | ! a . trim ( ) . is_empty ( ) & & name = = " computer_use_open_app " ) {
2026-11-04 09:19:59 +00:00
return Some ( app . trim ( ) . to_string ( ) ) ;
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
if ! v [ " ref " ] . is_null ( ) {
return SPOTS . lock ( ) . unwrap_or_else ( | e | e . into_inner ( ) ) . as_ref ( ) . map ( | ( app , _ ) | app . clone ( ) ) ;
2026-11-04 09:19:59 +00:00
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let s = scale_here ( ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let point = match name {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
" computer_use_click " | " computer_use_mouse_move " | " computer_use_scroll " if ! v [ " x " ] . is_null ( ) = > Some ( to_screen ( v [ " x " ] . as_f64 ( ) ? , v [ " y " ] . as_f64 ( ) ? , s ) ) ,
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
" computer_use_drag " = > v [ " path " ] . as_array ( ) . and_then ( | a | a . first ( ) ) . and_then ( | p | Some ( to_screen ( p [ " x " ] . as_f64 ( ) ? , p [ " y " ] . as_f64 ( ) ? , s ) ) ) ,
_ = > None ,
} ;
match point {
Some ( ( x , y ) ) = > mac ::window_at ( x , y ) . map ( | w | w . 0 ) ,
None = > mac ::front_app ( ) . map ( | w | w . 0 ) ,
}
}
/// Screenshot pixels per screen point (for the demonstration recorder).
pub fn points_to_shot ( ) -> f64 {
let ( pw , _ ) = mac ::screen_points ( ) ;
let ( px , py ) = mac ::screen_pixels ( ) ;
let ( fw , _ , _ ) = fit ( px . max ( 1 ) , py . max ( 1 ) ) ;
if pw > 0.0 { fw as f64 / pw } else { 1.0 }
}
/// The person's own clicks and keys can be recorded (教它一項任務): macOS, Linux on X11.
/// Wayland doesn't let apps watch other apps' input.
pub fn demo_supported ( ) -> bool {
#[ cfg(target_os = " linux " ) ]
{
crate ::desk ::kind ( ) = = crate ::desk ::Kind ::X11
}
#[ cfg(not(target_os = " linux " )) ]
{
cfg! ( target_os = " macos " )
}
}
/// This system's screen can be seen and used (macOS; Linux on X11 or Wayland).
pub fn screen_supported ( ) -> bool {
#[ cfg(target_os = " linux " ) ]
{
crate ::desk ::available ( )
}
#[ cfg(not(target_os = " linux " )) ]
{
cfg! ( target_os = " macos " )
}
}
pub fn screen_capture_ok ( ) -> bool {
mac ::screen_capture_preflight ( )
}
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
pub fn frontmost_app ( ) -> Option < String > {
mac ::front_app ( ) . map ( | w | w . 0 )
}
#[ cfg_attr(not(target_os = " macos " ), allow(dead_code)) ]
pub fn app_at_point ( x : f64 , y : f64 ) -> Option < String > {
mac ::window_at ( x , y ) . map ( | w | w . 0 )
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// The screen as the person sees it (no numbers), at the bot's picture size: for the phone.
pub fn plain_picture ( ) -> Result < ( u32 , u32 , Vec < u8 > ) , String > {
2026-11-05 08:57:44 +00:00
if ! mac ::screen_capture_preflight ( ) {
ask_permission ( " Privacy_ScreenCapture " ) ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
return Err ( NO_SCREEN_RECORDING . into ( ) ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
let img = capture_screen ( ) ? ;
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
let ( fw , fh ) = shot_size ( ) ;
let img = if img . dimensions ( ) ! = ( fw , fh ) { image ::imageops ::resize ( & img , fw , fh , image ::imageops ::FilterType ::Triangle ) } else { img } ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
Ok ( ( fw , fh , encode_jpeg ( & img ) . ok_or ( " 截圖太大了,沒辦法送出 " ) ? ) )
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
/// The whole screen as RGB (backing pixels).
pub fn capture_screen ( ) -> Result < image ::RgbImage , String > {
#[ cfg(target_os = " linux " ) ]
return crate ::desk ::capture ( ) ;
#[ allow(unreachable_code) ]
{
2026-11-05 07:48:22 +00:00
// A random tail: two bots screenshotting the same millisecond must not delete each
// other's file (pid+time alone collides).
let file = std ::env ::temp_dir ( ) . join ( format! ( " anybot-screen- {} - {} - {} .png " , std ::process ::id ( ) , crate ::model ::now_ms ( ) , crate ::httpd ::random_hex ( 4 ) ) ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let st = std ::process ::Command ::new ( " screencapture " ) . args ( [ " -x " , " -m " , " -t " , " png " ] ) . arg ( & file ) . status ( ) . map_err ( | e | e . to_string ( ) ) ? ;
let bytes = std ::fs ::read ( & file ) ;
let _ = std ::fs ::remove_file ( & file ) ;
if ! st . success ( ) {
return Err ( " 截圖失敗 " . into ( ) ) ;
}
Ok ( image ::load_from_memory ( & bytes . map_err ( | e | format! ( " 截圖失敗: {e} " ) ) ? ) . map_err ( | e | format! ( " 截圖讀不出來: {e} " ) ) ? . to_rgb8 ( ) )
}
}
#[ cfg(target_os = " linux " ) ]
use crate ::desk as mac ;
#[ cfg(target_os = " macos " ) ]
mod mac {
use super ::* ;
use std ::ffi ::c_void ;
use std ::time ::Duration ;
#[ repr(C) ]
#[ derive(Clone, Copy) ]
struct CGPoint {
x : f64 ,
y : f64 ,
}
#[ repr(C) ]
#[ derive(Clone, Copy) ]
struct CGSize {
width : f64 ,
height : f64 ,
}
#[ repr(C) ]
#[ derive(Clone, Copy) ]
struct CGRect {
origin : CGPoint ,
size : CGSize ,
}
type Ev = * mut c_void ;
type Src = * mut c_void ;
#[ link(name = " CoreGraphics " , kind = " framework " ) ]
extern " C " {
fn CGMainDisplayID ( ) -> u32 ;
fn CGDisplayBounds ( d : u32 ) -> CGRect ;
fn CGPreflightScreenCaptureAccess ( ) -> bool ;
fn CGRequestScreenCaptureAccess ( ) -> bool ;
fn CGEventSourceCreate ( state : i32 ) -> Src ;
fn CGEventSourceSetLocalEventsSuppressionInterval ( src : Src , secs : f64 ) ;
fn CGEventCreateMouseEvent ( src : Src , kind : u32 , at : CGPoint , button : u32 ) -> Ev ;
fn CGEventCreateKeyboardEvent ( src : Src , key : u16 , down : bool ) -> Ev ;
fn CGEventKeyboardSetUnicodeString ( e : Ev , len : usize , s : * const u16 ) ;
fn CGEventSetFlags ( e : Ev , flags : u64 ) ;
fn CGEventSetIntegerValueField ( e : Ev , field : u32 , v : i64 ) ;
fn CGEventPost ( tap : u32 , e : Ev ) ;
fn CGEventCreateScrollWheelEvent ( src : Src , units : u32 , count : u32 , wheel1 : i32 , .. . ) -> Ev ;
fn CGWindowListCopyWindowInfo ( option : u32 , relative_to : u32 ) -> * const c_void ;
fn CGRectMakeWithDictionaryRepresentation ( dict : * const c_void , rect : * mut CGRect ) -> bool ;
static kCGWindowOwnerName : * const c_void ;
static kCGWindowOwnerPID : * const c_void ;
static kCGWindowBounds : * const c_void ;
static kCGWindowLayer : * const c_void ;
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
static kCGWindowNumber : * const c_void ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
#[ link(name = " CoreFoundation " , kind = " framework " ) ]
extern " C " {
fn CFArrayGetCount ( a : * const c_void ) -> isize ;
fn CFArrayGetValueAtIndex ( a : * const c_void , i : isize ) -> * const c_void ;
fn CFDictionaryGetValue ( d : * const c_void , k : * const c_void ) -> * const c_void ;
fn CFNumberGetValue ( n : * const c_void , kind : isize , out : * mut c_void ) -> bool ;
fn CFStringGetCString ( s : * const c_void , buf : * mut std ::ffi ::c_char , len : isize , enc : u32 ) -> bool ;
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// On-screen windows front to back: (owner app, pid, bounds, layer, window number).
fn all_windows ( ) -> Vec < ( String , i32 , CGRect , i32 , u32 ) > {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let mut out = vec! [ ] ;
unsafe {
// kCGWindowListOptionOnScreenOnly | kCGWindowListExcludeDesktopElements
let list = CGWindowListCopyWindowInfo ( 1 | 16 , 0 ) ;
if list . is_null ( ) {
return out ;
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
let num = | d : * const c_void , k : * const c_void | -> i64 {
let mut v : i64 = - 1 ;
let p = CFDictionaryGetValue ( d , k ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
if ! p . is_null ( ) {
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
CFNumberGetValue ( p , 4 , & mut v as * mut i64 as * mut c_void ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
v
} ;
for i in 0 .. CFArrayGetCount ( list ) {
let d = CFArrayGetValueAtIndex ( list , i ) ;
let ( layer , pid , number ) = ( num ( d , kCGWindowLayer ) as i32 , num ( d , kCGWindowOwnerPID ) as i32 , num ( d , kCGWindowNumber ) as u32 ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let mut name = String ::new ( ) ;
let n = CFDictionaryGetValue ( d , kCGWindowOwnerName ) ;
if ! n . is_null ( ) {
let mut buf = [ 0 u8 ; 256 ] ;
if CFStringGetCString ( n , buf . as_mut_ptr ( ) as * mut std ::ffi ::c_char , buf . len ( ) as isize , 0x0800_0100 ) {
let end = buf . iter ( ) . position ( | b | * b = = 0 ) . unwrap_or ( buf . len ( ) ) ;
name = String ::from_utf8_lossy ( & buf [ .. end ] ) . to_string ( ) ;
}
}
let mut r = CGRect { origin : CGPoint { x : 0.0 , y : 0.0 } , size : CGSize { width : 0.0 , height : 0.0 } } ;
let b = CFDictionaryGetValue ( d , kCGWindowBounds ) ;
if ! b . is_null ( ) {
CGRectMakeWithDictionaryRepresentation ( b , & mut r ) ;
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
out . push ( ( name , pid , r , layer , number ) ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
CFRelease ( list ) ;
}
out
}
Bots click by number: every control on the screen is boxed and numbered, on every computer
A bot asked to open a Threads profile clicked an empty strip ten times: it guessed coordinates
from a picture, and the web tools' description told it they didn't work on its own computer.
- marks.rs: the picture a bot gets has every button, link, field and menu item boxed and
numbered (a tiny built-in digit font, pure Rust), with the same numbers in a list. click /
typing / scroll / mouse_move take the number (ref) and land on the control's centre.
- Where the controls come from: macOS AX and Linux AT-SPI through agent-desktop on a person's
computer (desktop::elements); AT-SPI in the bots' own computer (box/a11y, after XBot's
box-a11y), web pages included. On a Mac a control whose spot another app's window covers gets
no number (topmost window at its centre must be the app's; the Dock's invisible full-screen
window doesn't count). open_app waits until the app is really in front.
- No going round in circles: a step that changed nothing says so; the same step (clicks within
16 px) a third time in a turn comes with a warning, from the fourth time it isn't run.
- The box (version 4): one TigerVNC server instead of Xvfb + x11vnc, the AT-SPI bus and registry
at boot, one browser launcher (box/browser, after XBot's box-chrome: same profile and CDP,
renderer accessibility, a launch lock, stale profile locks cleared). One docker exec per look
(list + picture, ~0.4 s). A brain updates an older box by itself when it has been idle a
minute (files and logins stay); a new box is never made from an older image.
- Gone: computer_use_read_app (the list is in every picture), computer_use_mouse_button, the
per-call picture for the phone (it gets a plain one), the stale browser_read note; the prompt
says web work goes through browser_* by number, and never to read browser cookie or history
files.
Tests: marks (numbering, drawing, fingerprints), box_numbers_are_exact on a box from the new
image (a Threads-like page: the icon links get numbers, Profile clicked by number, Chinese typed
into the numbered field, a click on nothing called out), box_screen_is_exact, VNC e2e 11/11 on
TigerVNC, Calculator 7×8=56 by numbers through the signed app on this Mac, remote_e2e 28/28,
153 unit tests; Linux build + 157 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 16:47:01 +00:00
/// On-screen windows front to back: (owner app, pid, bounds), normal layer only.
fn windows ( ) -> Vec < ( String , i32 , CGRect ) > {
all_windows ( ) . into_iter ( ) . filter ( | w | w . 3 = = 0 ) . map ( | ( n , p , r , _ , _ ) | ( n , p , r ) ) . collect ( )
}
/// The app (pid) that owns a window, by its number.
pub fn pid_of_window ( number : u32 ) -> Option < i32 > {
all_windows ( ) . into_iter ( ) . find ( | w | w . 4 = = number ) . map ( | w | w . 1 )
}
/// Whose window is on top at a screen point: app windows, floating and modal panels, pop-up
/// menus (window levels 0, 3, 8, 101). Not the menu bar, overlays, or the Dock's invisible
/// full-screen window (level 20), which lies over everything.
pub fn pid_on_top_at ( x : f64 , y : f64 ) -> Option < i32 > {
all_windows ( )
. into_iter ( )
. filter ( | w | matches! ( w . 3 , 0 | 3 | 8 | 101 ) & & w . 2. size . width > 1.0 )
. find ( | ( _ , _ , r , _ , _ ) | x > = r . origin . x & & y > = r . origin . y & & x < r . origin . x + r . size . width & & y < r . origin . y + r . size . height )
. map ( | w | w . 1 )
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
/// The app whose window is under a screen point.
pub fn window_at ( x : f64 , y : f64 ) -> Option < ( String , i32 ) > {
windows ( ) . into_iter ( ) . find ( | ( _ , _ , r ) | x > = r . origin . x & & y > = r . origin . y & & x < r . origin . x + r . size . width & & y < r . origin . y + r . size . height ) . map ( | ( n , p , _ ) | ( n , p ) )
}
/// The frontmost app (its frontmost normal window).
pub fn front_app ( ) -> Option < ( String , i32 ) > {
windows ( ) . into_iter ( ) . find ( | ( n , _ , r ) | ! n . is_empty ( ) & & r . size . width > 1.0 ) . map ( | ( n , p , _ ) | ( n , p ) )
}
pub fn screen_capture_preflight ( ) -> bool {
unsafe { CGPreflightScreenCaptureAccess ( ) }
}
pub fn screen_pixels ( ) -> ( u32 , u32 ) {
Whole control flow checked against real chats: what the bot sees is what is there, every step lands, nothing done twice
Web
- web_fetch kept pages with a <header> only up to it (taken for <head>): articles came back as their menus.
- browser_open waits for the page to load and returns it like browser_read (one step fewer per page);
actions wait for the page to hold still instead of fixed 1-2.5 s sleeps; one way to reach the tab.
The bot's computer (box v6)
- Numbers only on what can be seen: clipped to the window and the page's visible area, none under
another app's window (a link scrolled under the toolbar was numbered on Back).
- press_key with a typed character no longer swallows the keys after it (xdotool type --args 1).
- The person's input through noVNC/TigerVNC is noticed again (fence, continuous updates, QEMU keys).
- "Nothing changed" only after clicks, scrolls and drags; long text is pasted; one screen step at a
time; one docker inspect per step when the box is known ready; poppler-utils for long PDFs.
- One typing path and fewer round trips (stdin variant, unused picture path, extra batch execs gone).
The person's computers
- Linux: the window system's front app when it has one, and the reason when there is none (was 「」);
hidden and off-screen controls get no number; wording no longer names a tool that doesn't exist.
The loop
- Without Docker, the bot's own computer runs no commands where there is no sandbox (Linux, Windows):
it was a plain shell on the brain itself, keys in reach, asked nothing.
- The bot's own computer is re-checked each step; unreachable computers are failures, not safety
blocks; no direct-mode note on every result; an empty reply is asked again; scrolling and keys
aren't counted as going in circles; steps aren't moved to a computer that can't do them.
- One finish() for every step's result, keep_grant(), dead appctl list_computers/SERVING and the
impossible server branches removed; duplicated prompt instructions merged.
Tests: 162 (Mac) / 166 (Linux) unit; real box drills 6/6 on box v6; a brain without Chrome browses
on its box; remote E2E 28/28.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha2RVfSYmrhqNnqx9Z4
2026-11-06 09:44:59 +00:00
// Points times the backing scale (CGDisplayPixelsWide reports points on Retina, while
// screencapture gives backing pixels).
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let ( pw , ph ) = screen_points ( ) ;
let scale = backing_scale ( ) ;
( ( pw * scale ) as u32 , ( ph * scale ) as u32 )
}
fn backing_scale ( ) -> f64 {
static S : std ::sync ::OnceLock < f64 > = std ::sync ::OnceLock ::new ( ) ;
* S . get_or_init ( | | {
let file = std ::env ::temp_dir ( ) . join ( format! ( " anybot-scale- {} .png " , std ::process ::id ( ) ) ) ;
let ok = std ::process ::Command ::new ( " screencapture " ) . args ( [ " -x " , " -m " , " -t " , " png " , " -R0,0,10,10 " ] ) . arg ( & file ) . status ( ) . map ( | s | s . success ( ) ) . unwrap_or ( false ) ;
let w = if ok { image ::image_dimensions ( & file ) . map ( | d | d . 0 ) . unwrap_or ( 10 ) } else { 10 } ;
let _ = std ::fs ::remove_file ( & file ) ;
( w as f64 / 10.0 ) . max ( 1.0 )
} )
}
#[ link(name = " CoreFoundation " , kind = " framework " ) ]
extern " C " {
fn CFRelease ( x : * const c_void ) ;
}
const HID_TAP : u32 = 0 ;
/// kCGEventSourceStateHIDSystemState, as CUActuator uses (CGEventSourceCreate(1)).
const HID_STATE : i32 = 1 ;
const CLICK_STATE : u32 = 1 ;
const MOVED : u32 = 5 ;
/// kCGScrollEventUnitPixel, as CGEventCreateScrollWheelEvent2(_, 0, 2, …) in the original.
const SCROLL_PIXEL : u32 = 0 ;
/// An event source like the original's: HID system state, local-event suppression 0 s, so
/// the person's own mouse and keys aren't frozen after each synthetic event.
struct Source ( Src ) ;
impl Source {
fn new ( ) -> Self {
let s = unsafe { CGEventSourceCreate ( HID_STATE ) } ;
if ! s . is_null ( ) {
unsafe { CGEventSourceSetLocalEventsSuppressionInterval ( s , 0.0 ) } ;
}
Source ( s )
}
}
impl Drop for Source {
fn drop ( & mut self ) {
if ! self . 0. is_null ( ) {
unsafe { CFRelease ( self . 0 as * const c_void ) }
}
}
}
fn post ( e : Ev , flags : u64 ) {
if e . is_null ( ) {
return ;
}
unsafe {
CGEventSetFlags ( e , flags ) ;
CGEventPost ( HID_TAP , e ) ;
CFRelease ( e ) ;
}
}
fn gap ( ) {
std ::thread ::sleep ( Duration ::from_millis ( EVENT_GAP_MS ) ) ;
}
fn sleep_secs ( s : f64 ) {
std ::thread ::sleep ( Duration ::from_secs_f64 ( s ) ) ;
}
/// (down, up, dragged) event kinds and button number.
fn kinds ( button : & str ) -> ( u32 , u32 , u32 , u32 ) {
match button {
" right " = > ( 3 , 4 , 7 , 1 ) ,
" middle " = > ( 25 , 26 , 27 , 2 ) ,
_ = > ( 1 , 2 , 6 , 0 ) ,
}
}
pub fn screen_capture_allowed ( ) -> bool {
unsafe { CGPreflightScreenCaptureAccess ( ) | | CGRequestScreenCaptureAccess ( ) }
}
pub fn screen_points ( ) -> ( f64 , f64 ) {
let r = unsafe { CGDisplayBounds ( CGMainDisplayID ( ) ) } ;
( r . size . width , r . size . height )
}
pub fn move_to ( x : f64 , y : f64 ) {
let src = Source ::new ( ) ;
post ( unsafe { CGEventCreateMouseEvent ( src . 0 , MOVED , CGPoint { x , y } , 0 ) } , 0 ) ;
}
/// CUActuator.click: press, hold 0.03 s, release; 0.08 s between the clicks of a multi-click.
pub fn click ( x : f64 , y : f64 , button : & str , count : i64 , flags : u64 ) {
let ( down , up , _ , b ) = kinds ( button ) ;
let at = CGPoint { x , y } ;
move_to ( x , y ) ;
gap ( ) ;
let src = Source ::new ( ) ;
for n in 1 ..= count {
if n > 1 {
sleep_secs ( CLICK_REPEAT_GAP_SECS ) ;
}
for ( i , kind ) in [ down , up ] . into_iter ( ) . enumerate ( ) {
if i = = 1 {
sleep_secs ( CLICK_HOLD_SECS ) ;
}
let e = unsafe { CGEventCreateMouseEvent ( src . 0 , kind , at , b ) } ;
if ! e . is_null ( ) {
unsafe { CGEventSetIntegerValueField ( e , CLICK_STATE , n ) } ;
}
post ( e , flags ) ;
}
}
}
/// CURemoteEventPoster.drag(path:button:steps:holdSeconds:): press at the first point (click
/// state 1), hold, then walk each segment in `DRAG_STEPS` interpolated dragged events, then
/// release at the last point.
pub fn drag ( path : & [ ( f64 , f64 ) ] , button : & str , flags : u64 ) {
let ( down , up , dragged , b ) = kinds ( button ) ;
let ( x0 , y0 ) = path [ 0 ] ;
move_to ( x0 , y0 ) ;
gap ( ) ;
let src = Source ::new ( ) ;
let event = | kind : u32 , x : f64 , y : f64 | {
let e = unsafe { CGEventCreateMouseEvent ( src . 0 , kind , CGPoint { x , y } , b ) } ;
if ! e . is_null ( ) {
unsafe { CGEventSetIntegerValueField ( e , CLICK_STATE , 1 ) } ;
}
post ( e , flags ) ;
} ;
event ( down , x0 , y0 ) ;
sleep_secs ( DRAG_HOLD_SECS ) ;
for ( x , y ) in drag_points ( path ) {
gap ( ) ;
event ( dragged , x , y ) ;
}
let ( xn , yn ) = path [ path . len ( ) - 1 ] ;
event ( up , xn , yn ) ;
}
/// handleScroll closure @0x10009e998: a mouse-moved event at the spot (the pointer's own
/// position when no x/y), then one pixel-unit scroll event with 2 wheels at that spot.
pub fn scroll ( at : Option < ( f64 , f64 ) > , dy : i32 , dx : i32 , flags : u64 ) {
if let Some ( ( x , y ) ) = at {
move_to ( x , y ) ;
gap ( ) ;
}
let src = Source ::new ( ) ;
post ( unsafe { CGEventCreateScrollWheelEvent ( src . 0 , SCROLL_PIXEL , 2 , dy , dx ) } , flags ) ;
}
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
/// Type text so that any app takes it, whatever the keyboard layout or input method: paste
/// it (⌘V) and put the clipboard back. Text sent as key events (the original's
/// CUTypedCharacters, key 128 carrying the characters) went into the input method instead
/// when one was on — with 注音 every letter vanished while return still arrived, so a
/// command typed into a terminal ran empty. Falls back to those key events without a
/// clipboard.
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
pub fn type_text ( text : & str ) {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
if ! paste ( text ) {
type_as_keys ( text ) ;
}
}
/// Wait for the app to read the clipboard before it is put back (Electron apps read it a
/// moment after ⌘V; estimated).
const PASTE_READ_MS : u64 = 350 ;
fn paste ( text : & str ) -> bool {
let Ok ( mut cb ) = arboard ::Clipboard ::new ( ) else { return false } ;
// What the person had there comes back afterwards (text or a picture).
let saved_text = cb . get_text ( ) . ok ( ) ;
let saved_image = if saved_text . is_none ( ) { cb . get_image ( ) . ok ( ) } else { None } ;
if cb . set_text ( text . to_string ( ) ) . is_err ( ) {
return false ;
}
std ::thread ::sleep ( Duration ::from_millis ( 40 ) ) ;
press ( 9 , FLAG_CMD , KEY_HOLD_SECS ) ;
std ::thread ::sleep ( Duration ::from_millis ( PASTE_READ_MS ) ) ;
match ( saved_text , saved_image ) {
( Some ( t ) , _ ) = > {
let _ = cb . set_text ( t ) ;
}
( None , Some ( i ) ) = > {
let _ = cb . set_image ( i ) ;
}
_ = > {
let _ = cb . clear ( ) ;
}
}
true
}
/// CUTypedCharacters.postChunks: each chunk on key 128, down, 0.035 s, up; 0.02 s between chunks.
fn type_as_keys ( text : & str ) {
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
let src = Source ::new ( ) ;
let chunks = utf16_chunks ( text ) ;
for ( i , chunk ) in chunks . iter ( ) . enumerate ( ) {
if i > 0 {
sleep_secs ( KEY_GAP_SECS ) ;
}
for down in [ true , false ] {
if ! down {
sleep_secs ( KEY_HOLD_SECS ) ;
}
let e = unsafe { CGEventCreateKeyboardEvent ( src . 0 , TYPE_KEY_CODE , down ) } ;
if ! e . is_null ( ) {
unsafe { CGEventKeyboardSetUnicodeString ( e , chunk . len ( ) , chunk . as_ptr ( ) ) } ;
}
post ( e , 0 ) ;
}
}
}
/// handlePressKey: key down, hold (0.035 s unless told), key up.
pub fn press ( code : u16 , flags : u64 , hold_secs : f64 ) {
let src = Source ::new ( ) ;
post ( unsafe { CGEventCreateKeyboardEvent ( src . 0 , code , true ) } , flags ) ;
sleep_secs ( hold_secs ) ;
post ( unsafe { CGEventCreateKeyboardEvent ( src . 0 , code , false ) } , flags ) ;
}
}
#[ cfg(not(any(target_os = " macos " , target_os = " linux " ))) ]
mod mac {
pub fn screen_capture_allowed ( ) -> bool {
false
}
pub fn screen_points ( ) -> ( f64 , f64 ) {
( 0.0 , 0.0 )
}
pub fn move_to ( _ : f64 , _ : f64 ) { }
pub fn click ( _ : f64 , _ : f64 , _ : & str , _ : i64 , _ : u64 ) { }
pub fn drag ( _ : & [ ( f64 , f64 ) ] , _ : & str , _ : u64 ) { }
pub fn scroll ( _ : Option < ( f64 , f64 ) > , _ : i32 , _ : i32 , _ : u64 ) { }
pub fn type_text ( _ : & str ) { }
pub fn press ( _ : u16 , _ : u64 , _ : f64 ) { }
pub fn window_at ( _ : f64 , _ : f64 ) -> Option < ( String , i32 ) > {
None
}
pub fn front_app ( ) -> Option < ( String , i32 ) > {
None
}
pub fn screen_capture_preflight ( ) -> bool {
false
}
pub fn screen_pixels ( ) -> ( u32 , u32 ) {
( 0 , 0 )
}
}
#[ cfg(test) ]
mod tests {
use super ::* ;
Bots work like someone who knows the computer: they know each machine, reach what they see, plan, and hand the planned steps to a quick model
Measured on the bot's own box (8 tasks x 2, grok-4.7): 14/16 done, 120 steps, 1114 s before;
16/16, 102 steps, 574 s after (model time 1015 -> 471 s). A replayed skill: 49 s -> 15 s.
Knowing the machine
- facts.rs: 電腦現況 per computer (system, desktop, systemd, sudo, network, package manager,
tools, apps, a fixed TZ), read once by a small read-only probe, kept 30 min, in the
instructions at the start of a turn.
- The box's sudo is the bot's: only what follows sudo is checked against the always-denied
list (it was refused outright, while the instructions said it could sudo), and commands with
sudo run as the box's root (setuid does nothing under gVisor).
- Box image v8 drops ENV TZ=UTC (it overrode /etc/localtime). ANYBOT_BOX_IMAGE /
ANYBOT_BOX_VOLUME give a throwaway box its own image and volume, never adopting the real one.
- Mac shell sandbox: App Store app CLIs (they carry their own app sandbox) and ps/top run
outside it; plain `osascript -e` too, asked the first time even in direct mode, refusing do
shell script, terminals, keystrokes, file reads, ObjC and the secret folders. A command the
sandbox broke says so, instead of looking like a broken program.
Reaching what it sees
- The Mac menu bar's status icons are numbered (each app's AXExtrasMenuBar, named from inside
for the system's own), computer_use_zoom looks closer at full sharpness.
- Numbers stay with their controls from look to look (they were reassigned every look, so a
step's second click landed elsewhere); the box's front window is listed first; a control can
be named instead of numbered (name "4").
- Web page numbering walks open shadow DOM (Chromium settings, web components).
- A box on a Docker host elsewhere retries the next port when the usual one is taken (a retry
without it came up with no network).
Working like a person
- Instructions: pick the route, know what done looks like, one route, check once, say what's
stuck instead of faking it with a substitute, ask before cutting the connection.
- StepPlan: the bot's model takes the first step, the step after trouble (failed, refused,
repeated, nothing changed on screen, the quick model switching routes), one in four, and the
final check; a quick model (Bot.step_model, auto: grok-4.20-0309-non-reasoning on xAI) does
the planned steps in between, falling back to the bot's model if it can't be used.
- Skills keep the steps that worked (with what each acted on, not numbers); replay_skill does
them again with no model in between, finding each control again, handing back on a mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha5F5gJt93t3yEx3XcS
2026-11-09 01:01:03 +00:00
#[ test ]
fn menu_bar_icons_are_named_plainly ( ) {
assert_eq! ( icon_name ( " MenuBarAgent " , " Wi‑ Fi, 已連接, 3格 " ) , " Wi‑ Fi, 已連接, 3格 " ) ;
assert_eq! ( icon_name ( " Tailscale " , " " ) , " Tailscale " ) ;
assert_eq! ( icon_name ( " Tailscale " , " Tailscale " ) , " Tailscale " ) ;
assert_eq! ( icon_name ( " LINE " , " 好友名單 " ) , " LINE: 好友名單 " ) ;
assert_eq! ( crate ::axread ::menu_extra_name ( " com.apple.menuextra.wifi " ) , " Wi-Fi " ) ;
assert_eq! ( crate ::axread ::menu_extra_name ( " com.apple.menuextra.clock " ) , " 時鐘 " ) ;
assert_eq! ( crate ::axread ::menu_extra_name ( " org.example.thing " ) , " org.example.thing " ) ;
}
/// The menu bar's icons get numbers on a Mac (needs 輔助使用 for the test runner):
/// `cargo test menu_bar_icons_get_numbers -- --ignored --nocapture`.
#[ test ]
#[ ignore ]
#[ cfg(target_os = " macos " ) ]
fn menu_bar_icons_get_numbers ( ) {
let t = std ::time ::Instant ::now ( ) ;
let els = menu_bar_controls ( ) ;
println! ( " {} icons in {} ms " , els . len ( ) , t . elapsed ( ) . as_millis ( ) ) ;
for e in & els {
println! ( " {} 「 {} 」 ( {:.0} , {:.0} {:.0} x {:.0} ) " , e . kind , e . name , e . x , e . y , e . w , e . h ) ;
}
assert! ( ! els . is_empty ( ) & & els . iter ( ) . all ( | e | e . active & & e . y < 60.0 & & ! e . name . is_empty ( ) ) ) ;
// The second look within the freshness window is the cached list: instant.
let t = std ::time ::Instant ::now ( ) ;
assert_eq! ( menu_bar_controls ( ) . len ( ) , els . len ( ) ) ;
assert! ( t . elapsed ( ) . as_millis ( ) < 20 ) ;
}
/// A zoom cuts the asked area from the sharp screen, enlarges it, and says how to map back.
#[ test ]
fn a_zoom_is_the_asked_area_sharp_and_bigger ( ) {
// A 2048× 1152 screen (the bot sees it at 1024× 576): a red square at screen (1800..1840, 10..50).
let mut img = image ::RgbImage ::from_pixel ( 2048 , 1152 , image ::Rgb ( [ 255 , 255 , 255 ] ) ) ;
for x in 1800 .. 1840 {
for y in 10 .. 50 {
img . put_pixel ( x , y , image ::Rgb ( [ 255 , 0 , 0 ] ) ) ;
}
}
// In the bot's picture that square is at (900..920, 5..25): zoom the menu-bar strip around it.
let ( text , b64 ) = zoom ( & img , & json! ( { " x " : 880 , " y " : 0 , " width " : 64 , " height " : 32 } ) ) . unwrap ( ) ;
let pic = image ::load_from_memory ( & base64 ::engine ::general_purpose ::STANDARD . decode ( b64 ) . unwrap ( ) ) . unwrap ( ) . to_rgb8 ( ) ;
let ( w , h ) = pic . dimensions ( ) ;
// 64× 32 picture pixels = 128× 64 screen pixels, enlarged 4× (the cap) to 512× 256.
assert_eq! ( ( w , h ) , ( 512 , 256 ) , " {text} " ) ;
// The square's middle (screen 1820, 30) → crop (60, 30) → zoomed (240, 120): red.
let p = pic . get_pixel ( 240 , 120 ) ;
assert! ( p [ 0 ] > 200 & & p [ 1 ] < 80 , " {p:?} " ) ;
assert! ( text . contains ( " (880, 0) 到 (944, 32) " ) & & text . contains ( " 0.125 " ) , " {text} " ) ;
// Out of the picture or missing numbers: clamped, or said plainly.
assert! ( zoom ( & img , & json! ( { " x " : 5000 , " y " : 5000 , " width " : 10 , " height " : 10 } ) ) . is_ok ( ) ) ;
assert! ( zoom ( & img , & json! ( { " x " : 1 } ) ) . unwrap_err ( ) . contains ( " width " ) ) ;
}
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
#[ test ]
fn screenshots_fit_like_the_original ( ) {
assert_eq! ( fit ( 2940 , 1912 ) , ( 1024 , 666 , true ) ) ;
assert_eq! ( fit ( 1024 , 640 ) , ( 1024 , 640 , false ) ) ;
assert_eq! ( fit ( 800 , 600 ) , ( 800 , 600 , false ) ) ;
assert_eq! ( fit ( 4000 , 7 ) , ( 4000 , 7 , false ) ) ;
// A very wide strip keeps its short edge at 8.
assert_eq! ( fit ( 4000 , 10 ) , ( 3200 , 8 , true ) ) ;
}
#[ test ]
fn jpeg_stays_under_the_cap ( ) {
let img = image ::RgbImage ::from_fn ( 1024 , 640 , | x , y | image ::Rgb ( [ ( x % 256 ) as u8 , ( y % 256 ) as u8 , ( ( x ^ y ) % 256 ) as u8 ] ) ) ;
let j = encode_jpeg ( & img ) . unwrap ( ) ;
assert! ( j . len ( ) < = MAX_IMAGE_BYTES & & j . starts_with ( & [ 0xFF , 0xD8 ] ) ) ;
}
#[ test ]
fn shortcuts_are_parsed ( ) {
assert_eq! ( parse_combo ( " cmd+c " ) . unwrap ( ) , ( 8 , FLAG_CMD ) ) ;
assert_eq! ( parse_combo ( " cmd+shift+t " ) . unwrap ( ) , ( 17 , FLAG_CMD | FLAG_SHIFT ) ) ;
assert_eq! ( parse_combo ( " Return " ) . unwrap ( ) , ( 36 , 0 ) ) ;
assert_eq! ( parse_combo ( " ctrl+option+space " ) . unwrap ( ) , ( 49 , FLAG_CONTROL | FLAG_OPTION ) ) ;
assert! ( parse_combo ( " hyper+x " ) . is_err ( ) ) ;
assert! ( parse_combo ( " " ) . is_err ( ) ) ;
}
#[ test ]
fn typed_text_is_chunked_like_the_original ( ) {
let c = utf16_chunks ( & " a " . repeat ( 45 ) ) ;
assert_eq! ( c . iter ( ) . map ( Vec ::len ) . collect ::< Vec < _ > > ( ) , vec! [ 20 , 20 , 5 ] ) ;
// An emoji (2 UTF-16 units) is never split across chunks.
let c = utf16_chunks ( & format! ( " {} 😀b " , " a " . repeat ( 19 ) ) ) ;
assert_eq! ( c . iter ( ) . map ( Vec ::len ) . collect ::< Vec < _ > > ( ) , vec! [ 19 , 3 ] ) ;
assert_eq! ( utf16_chunks ( " 你好 " ) . len ( ) , 1 ) ;
assert! ( utf16_chunks ( " " ) . is_empty ( ) ) ;
}
#[ test ]
fn scrolls_like_the_original ( ) {
assert_eq! ( scroll_amount ( None ) , 3 ) ;
assert_eq! ( scroll_amount ( Some ( 0 ) ) , 1 ) ;
assert_eq! ( scroll_amount ( Some ( 99 ) ) , 20 ) ;
assert_eq! ( scroll_pixels ( " down " , 3 ) , Some ( ( - 48 , 0 ) ) ) ;
assert_eq! ( scroll_pixels ( " left " , 2 ) , Some ( ( 0 , 32 ) ) ) ;
assert_eq! ( scroll_pixels ( " sideways " , 1 ) , None ) ;
}
#[ test ]
fn drags_walk_six_steps_per_segment ( ) {
let pts = drag_points ( & [ ( 0.0 , 0.0 ) , ( 60.0 , 0.0 ) , ( 60.0 , 30.0 ) ] ) ;
assert_eq! ( pts . len ( ) , 2 * DRAG_STEPS ) ;
assert_eq! ( pts [ 0 ] , ( 10.0 , 0.0 ) ) ;
assert_eq! ( pts [ DRAG_STEPS - 1 ] , ( 60.0 , 0.0 ) ) ;
assert_eq! ( * pts . last ( ) . unwrap ( ) , ( 60.0 , 30.0 ) ) ;
}
#[ test ]
fn a_press_can_be_several_keys ( ) {
One way to act on a screen: a real click or keystroke, wherever the bot looked
- Acting by an app's number (read_app's e7) is now a real click at that control's spot, the
same path as screenshot coordinates. agent-desktop only reads; its own actions (a second set
with its own key names and failures: no typing in terminals, "/" and "`" failing off the main
thread) are gone, with computer_use_select, background clicks and press_key's app.
- Every action comes back with the screen after it (or, with no picture allowed, the front
app's numbered controls), so the bot no longer spends a step on a screenshot to check.
- Coordinates map without remembering a screenshot: the picture's size depends only on the
screen. The phone's taps go straight to the screen as the person's, in the pixels of the
picture it was sent; they used to borrow the bot's last screenshot scale and landed 1.25x off
on the bots' computer, or failed with "screenshot first".
- Mac typing pastes (⌘V, clipboard put back): with 注音 on, text sent as key events vanished
while return arrived, so commands typed into a terminal ran empty. Bare printing keys
("slash", "r") are typed the same way; keys like "slash" and "grave" have names.
- The bots' Docker computer: one docker exec per step (action, settle, picture), never
`mousemove --sync` (it waits forever when the pointer is already there: a second click on
the same spot hung two minutes), open_app (chromium = the browser browser_* drive, xterm).
Tests: box_screen_is_exact (real Docker: bot and phone coordinates exact, same spot twice in
under 2 s, ASCII + Chinese + named key typed), Calculator by numbers and Terminal typing under
注音 through the signed app, remote_e2e 28/28, Linux build + 151 tests + x11_live.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-05 13:51:19 +00:00
assert_eq! ( key_plan ( " cmd+a delete " ) . unwrap ( ) , vec! [ Key ::Combo ( 0 , FLAG_CMD ) , Key ::Combo ( 51 , 0 ) ] ) ;
assert! ( key_plan ( " " ) . is_err ( ) ) ;
// Keys by the names models use, and characters without a key typed instead.
assert_eq! ( key_plan ( " cmd+grave " ) . unwrap ( ) , vec! [ Key ::Combo ( 50 , FLAG_CMD ) ] ) ;
// A printing key alone is its character (typed, so 注音 can't swallow it).
assert_eq! ( key_plan ( " slash " ) . unwrap ( ) , vec! [ Key ::Char ( " / " . into ( ) ) ] ) ;
assert_eq! ( key_plan ( " / @ r return " ) . unwrap ( ) , vec! [ Key ::Char ( " / " . into ( ) ) , Key ::Char ( " @ " . into ( ) ) , Key ::Char ( " r " . into ( ) ) , Key ::Combo ( 36 , 0 ) ] ) ;
assert! ( key_plan ( " hyper+x " ) . is_err ( ) ) ;
}
#[ test ]
fn pictures_and_clicks_share_one_size ( ) {
// The picture's size depends only on the screen, so a tap or a click maps the same
// whoever saw which screenshot last (a 2940× 1912 Retina screen of 1470× 956 points).
let ( w , h , _ ) = fit ( 2940 , 1912 ) ;
let s = ( 1470.0 / w as f64 , 956.0 / h as f64 ) ;
let ( x , y ) = to_screen ( 512.0 , 333.0 , s ) ;
assert! ( ( x - 735.0 ) . abs ( ) < 0.01 & & ( y - 478.0 ) . abs ( ) < 1.0 ) ;
AnyBot: first commit of the working desktop app
Rust/egui rebuild of Grok Bot as of 2026-11-04: chat with streaming
providers (xAI, OpenAI, Claude, Gemini, OpenRouter, OpenCode Go, Ollama,
custom), tools and approvals, the bot's Docker computer, computer use on
macOS / Linux X11 / Wayland, AnyBot's browser with a full DevTools-based
toolset, the encrypted vault, pinned grid and sidebar sections, Markdown
with Mermaid diagrams, phone access through the relay, and the Linux
build and test scripts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Uirha3oEtWeHh58zA7ErHa
2026-11-04 08:51:51 +00:00
}
#[ test ]
fn screenshot_pixels_map_to_screen_points ( ) {
// A 1470-point-wide screen sent as a 1024-wide picture.
let s = ( 1470.0 / 1024.0 , 956.0 / 666.0 ) ;
let ( x , y ) = to_screen ( 512.0 , 333.0 , s ) ;
assert! ( ( x - 735.0 ) . abs ( ) < 0.01 & & ( y - 478.0 ) . abs ( ) < 0.01 ) ;
}
}
#[ cfg(test) ]
mod window_tests {
/// `cargo test --release -- --ignored finds_apps_on_screen --nocapture`
#[ test ]
#[ ignore ]
fn finds_apps_on_screen ( ) {
let front = super ::frontmost_app ( ) ;
println! ( " front: {front:?} " ) ;
println! ( " at (200,200): {:?} " , super ::app_at_point ( 200.0 , 200.0 ) ) ;
println! ( " points→shot: {} " , super ::points_to_shot ( ) ) ;
assert! ( front . is_some ( ) ) ;
}
}