diff --git a/.env.example b/.env.example
index afe55fe..7a39024 100644
--- a/.env.example
+++ b/.env.example
@@ -33,6 +33,8 @@ LAZYBOY_MEMORY_BYTE_BUDGET=6000
 LAZYBOY_EVENT_RETENTION_DAYS=30
 LAZYBOY_CHECKPOINT_RETENTION_DAYS=7
 LAZYBOY_RUN_RETENTION_DAYS=90
+# Per-run trace shown when hovering the thinking avatar (round, tool results, errors).
+LAZYBOY_RUN_ACTIVITY_RETENTION_DAYS=7
 LAZYBOY_RECORDING_RETENTION_DAYS=30
 LAZYBOY_MEMORY_HISTORY_RETENTION_DAYS=90
 LAZYBOY_DB_WARN_MB=1024
diff --git a/Cargo.lock b/Cargo.lock
index 505e86f..b945f05 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1889,7 +1889,6 @@ name = "lazyboy-api"
 version = "0.1.0"
 dependencies = [
  "aes-gcm",
- "async-trait",
  "axum",
  "base64 0.22.1",
  "cap-std",
@@ -1900,9 +1899,7 @@ dependencies = [
  "fastembed",
  "futures-util",
  "hex",
- "hmac",
  "http",
- "http-body-util",
  "lazyboy-contracts",
  "lazyboy-control",
  "lazyboy-harness",
@@ -1915,7 +1912,6 @@ dependencies = [
  "serde_json",
  "sha2",
  "sqlx",
- "thiserror",
  "tokio",
  "tokio-tungstenite 0.26.2",
  "tower-http",
@@ -1958,9 +1954,7 @@ dependencies = [
  "base64 0.22.1",
  "lazyboy-contracts",
  "lazyboy-control",
- "serde",
  "serde_json",
- "thiserror",
  "tokio",
  "tracing",
  "tracing-subscriber",
@@ -1979,7 +1973,6 @@ dependencies = [
  "rig-core",
  "rustls",
  "rustls-native-certs",
- "serde",
  "serde_json",
  "thiserror",
  "tokio",
@@ -1992,39 +1985,30 @@ version = "0.1.0"
 dependencies = [
  "async-trait",
  "base64 0.22.1",
- "chrono",
- "hex",
  "lazyboy-contracts",
  "lazyboy-control",
  "reqwest 0.12.28",
- "serde",
  "serde_json",
- "sha2",
- "tokio",
 ]
 
 [[package]]
 name = "lazyboy-supervisor"
 version = "0.1.0"
 dependencies = [
- "async-trait",
  "axum",
  "base64 0.22.1",
  "bollard",
  "futures-util",
  "hex",
  "hmac",
- "lazyboy-contracts",
  "lazyboy-control",
  "reqwest 0.12.28",
  "serde",
  "serde_json",
  "sha2",
- "thiserror",
  "tokio",
  "tracing",
  "tracing-subscriber",
- "uuid",
 ]
 
 [[package]]
diff --git a/Cargo.toml b/Cargo.toml
index c60dcd8..ffb8e7e 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -12,10 +12,28 @@ members = [
 
 [workspace.package]
 edition = "2024"
+rust-version = "1.98"
 version = "0.1.0"
 license = "Apache-2.0"
 publish = false
 
+# --- Lint policy -----------------------------------------------------------
+# A single place decides how the whole workspace is linted; every crate opts in
+# with `[lints] workspace = true`. `make lint` runs clippy with -D warnings, so a
+# new warning has to be fixed (or relaxed at the call site with a reason) before
+# it reaches main. Thresholds live in clippy.toml, which cargo-clippy only reads
+# from the directory it was started in - keep running it at the workspace root.
+[workspace.lints.rust]
+# Safety-relevant code (the vault, the docker socket) has to stay explicit.
+unsafe_code = "warn"
+unused_must_use = "deny"
+unused_crate_dependencies = "warn"
+
+[workspace.lints.clippy]
+# Debug leftovers and placeholder implementations must not reach a branch.
+dbg_macro = "warn"
+todo = "warn"
+
 [profile.release]
 lto = true
 codegen-units = 1
diff --git a/Makefile b/Makefile
index 175dd58..3a27ceb 100644
--- a/Makefile
+++ b/Makefile
@@ -11,9 +11,9 @@ DATA_DIR       ?= ./data
 
 .PHONY: help env env-force \
         up logs ps health down purge \
-        computer postgres postgres-down \
+        computer postgres postgres-down pg-collation \
         build build-api build-supervisor build-controld \
-        fmt clippy test clean \
+        fmt fmt-check clippy lint audit test clean \
         web \
         dev dev-supervisor dev-api
 
@@ -33,6 +33,7 @@ help: ## Show this help
 	@echo "    make computer       Build the heavy Debian desktop image (lazyboy/computer:local)"
 	@echo "    make postgres       Start only postgres (127.0.0.1:5434) and wait for ready"
 	@echo "    make postgres-down  Stop postgres"
+	@echo "    make pg-collation   Repair a Postgres collation version mismatch (see docs)"
 	@echo ""
 	@echo "  Local dev (postgres in Docker, Rust services on the host):"
 	@echo "    make dev            Prep .env + postgres + computer image, then print run steps"
@@ -43,8 +44,11 @@ help: ## Show this help
 	@echo "    make build          cargo build --release (whole workspace)"
 	@echo "    make build-api      cargo build --release -p lazyboy-api"
 	@echo "    make fmt            cargo fmt --all"
-	@echo "    make clippy         cargo clippy (deny warnings)"
-	@echo "    make test           cargo test --workspace"
+	@echo "    make fmt-check      Report files rustfmt would change (legacy drift exists)"
+	@echo "    make clippy         cargo clippy (deny warnings, reads clippy.toml)"
+	@echo "    make lint           The Rust gate: clippy with -D warnings"
+	@echo "    make audit          cargo deny: RustSec advisories, licenses, sources"
+	@echo "    make test           cargo test --workspace (DB tests need: make postgres)"
 	@echo "    make web            Build the frontend in $(WEB_DIR) (needs node/npm)"
 	@echo "    make clean          cargo clean"
 	@echo ""
@@ -97,6 +101,14 @@ postgres: ## Start only postgres and wait until it is ready
 postgres-down: ## Stop postgres
 	$(COMPOSE) down postgres
 
+# A pgvector image rebuilt on another glibc leaves every database recording the old
+# collation version; Postgres then refuses CREATE DATABASE and `cargo test` hangs on
+# PoolTimedOut. Stop the api first, then reindex + refresh each database in place.
+pg-collation: ## Repair a Postgres collation version mismatch after an image update
+	$(COMPOSE) exec -T postgres psql -X -v ON_ERROR_STOP=1 -U lazyboy -d template1 -c "REINDEX DATABASE template1;" -c "ALTER DATABASE template1 REFRESH COLLATION VERSION;"
+	$(COMPOSE) exec -T postgres psql -X -v ON_ERROR_STOP=1 -U lazyboy -d postgres -c "REINDEX DATABASE postgres;" -c "ALTER DATABASE postgres REFRESH COLLATION VERSION;"
+	$(COMPOSE) exec -T postgres psql -X -v ON_ERROR_STOP=1 -U lazyboy -d lazyboy -c "REINDEX DATABASE lazyboy;" -c "ALTER DATABASE lazyboy REFRESH COLLATION VERSION;"
+
 # --- Rust / web ------------------------------------------------------------
 
 build: ## Release build of the whole workspace
@@ -114,8 +126,23 @@ build-controld: ## Release build of the controld binary
 fmt: ## Format all Rust code
 	cargo fmt --all
 
-clippy: ## Run clippy, denying warnings
-	cargo clippy --all-targets -- -D warnings
+fmt-check: ## Check formatting without touching files
+	cargo fmt --all --check
+
+# Lint policy lives in [workspace.lints] in the root Cargo.toml; the thresholds
+# (e.g. too-many-arguments-threshold) live in clippy.toml, which cargo-clippy
+# only reads from the directory it is started in - keep running this at the root.
+clippy: ## Run clippy over the workspace, denying warnings
+	cargo clippy --workspace --all-targets -- -D warnings
+
+lint: clippy ## The Rust quality gate used by CI
+
+# cargo-deny is a separate CLI: cargo install --locked cargo-deny
+audit: ## Supply-chain check (RustSec advisories, licenses, dependency sources)
+	@command -v cargo-deny >/dev/null 2>&1 || { \
+	  echo "cargo-deny is not installed: cargo install --locked cargo-deny"; exit 1; }
+	@echo "(advisories need the RustSec advisory DB, cloned on first run)"
+	cargo deny check
 
 test: ## Run the test suite
 	cargo test --workspace
diff --git a/apps/web/src/App.tsx b/apps/web/src/App.tsx
index 93596e4..76d799a 100644
--- a/apps/web/src/App.tsx
+++ b/apps/web/src/App.tsx
@@ -19,6 +19,7 @@ import { Avatar, AvatarLookProvider, AvatarStack, BLOBATAR_BACKGROUNDS, BLOBATAR
 import { dateLocale, getLocale, listJoin, setLocale, t, useLocale, type MessageKey } from "./i18n";
 import type { AvatarShape, Bot, ComputerMode, ComputerStatus, FileSkill, McpCatalogEntry, McpServer, McpTransport, MemoryItem, Message, MessageFile, ModelProviderId, Playbook, PlaybookInput, PlaybookStep, Room, RoomMember, Session, TaughtSkill, VoiceSettings, WorkspaceSettings } from "./types";
 import { ChatMarkdown, CopyMessageButton } from "./markdown";
+import { RunProbe, errorActions, errorTitle } from "./run-monitor";
 import { ScheduleEditor, ScheduleList, cronFromPreset, defaultCronPreset, presetFromCron, scheduleWhen, type CronPreset, type ScheduleItem } from "./schedule";
 import { CallOverlay, PhoneIcon } from "./call";
 import { VoiceSettingsDialog } from "./voice-settings";
@@ -68,6 +69,7 @@ function localizeStep(step?:string|null){
 function localizeError(message:string){
   if(message==="示範進行中：先按「完成示範」或「取消」，再送訊息。")return t("teachInProgress");
   if(message==="AI 回應逾時（150 秒）")return t("aiTimeout");
+  if(message==="run is not retryable")return t("errorRetryFailed");
   return message;
 }
 function hudLabel(computer:ComputerStatus,connecting:boolean,handingOff:boolean){
@@ -88,8 +90,8 @@ function readAsBase64(file:File){return new Promise<string>((resolve,reject)=>{c
 function formatBytes(size:number){if(size<1024)return `${size} B`;if(size<1024*1024)return `${Math.round(size/102.4)/10} KB`;return `${Math.round(size/104857.6)/10} MB`}
 function fileExt(name:string){const dot=name.lastIndexOf(".");const ext=dot>=0?name.slice(dot+1).replace(/[^a-z0-9]/gi,""):"";return (ext||"FILE").slice(0,4).toUpperCase()}
 function messageFiles(blocks:unknown):MessageFile[]{if(!Array.isArray(blocks))return [];return blocks.flatMap(block=>{if(!block||typeof block!=="object")return [];const value=block as {kind?:string;name?:string;mimeType?:string;size?:number};if(value.kind!=="file"&&value.kind!=="image")return [];return [{kind:value.kind,name:value.name||"file",mimeType:value.mimeType,size:value.size}]})}
-type MessageChip={kind:string;site?:string;why?:string;name?:string;human?:string;cron?:string;reason?:string;turns?:number;limit?:number};
-function chipBlocks(blocks:unknown){if(!Array.isArray(blocks))return [] as MessageChip[];return blocks.flatMap(block=>{if(!block||typeof block!=="object")return [];const value=block as MessageChip;if(value.kind==="login"||value.kind==="schedule"||value.kind==="scheduleRun"||value.kind==="resume")return [value];return []})}
+type MessageChip={kind:string;site?:string;why?:string;name?:string;human?:string;cron?:string;reason?:string;turns?:number;limit?:number;code?:string;retryable?:boolean;runId?:string;turn?:number;step?:string|null};
+function chipBlocks(blocks:unknown){if(!Array.isArray(blocks))return [] as MessageChip[];return blocks.flatMap(block=>{if(!block||typeof block!=="object")return [];const value=block as MessageChip;if(value.kind==="login"||value.kind==="schedule"||value.kind==="scheduleRun"||value.kind==="resume"||value.kind==="error")return [value];return []})}
 function resumeTitle(reason?:string){return reason==="budget_exhausted"?t("resumeBudget"):t("resumeMidTask")}
 function isAutoAttachCaption(body:string,files:MessageFile[]){const text=body.trim();if(!files.length)return false;if(!text)return true;return files.some(file=>text===file.name||text===`附件 ${file.name}`||text===`Attached ${file.name}`||text===t("attachedFile",{name:file.name}))}
 function isAutoScheduleCaption(body:string,chips:{kind?:string}[]){if(!chips.some(chip=>chip.kind==="schedule"||chip.kind==="scheduleRun"))return false;const text=body.trim();return /^\[排程(試跑)?\]/.test(text)||/^\[Schedule( test)?\]/i.test(text)}
@@ -115,7 +117,7 @@ export function App(){
   const [createOpen,setCreateOpen]=useState(false); const [createMenuOpen,setCreateMenuOpen]=useState(false); const [groupOpen,setGroupOpen]=useState(false); const [deleteOpen,setDeleteOpen]=useState(false); const [computerOpen,setComputerOpen]=useState(false);
   const paneStart=readPaneStore();
   const [rightCollapsed,setRightCollapsed]=useState(paneStart.collapsed); const [rightPart,setRightPart]=useState<RightPart>(paneStart.part);
-  const [sessionMenuOpen,setSessionMenuOpen]=useState(false); const [clearOpen,setClearOpen]=useState(false); const [sessionToDelete,setSessionToDelete]=useState<string|null>(null); const [remembered,setRemembered]=useState<Record<string,boolean>>({}); const [resumedChips,setResumedChips]=useState<Record<string,boolean>>({});
+  const [sessionMenuOpen,setSessionMenuOpen]=useState(false); const [clearOpen,setClearOpen]=useState(false); const [sessionToDelete,setSessionToDelete]=useState<string|null>(null); const [remembered,setRemembered]=useState<Record<string,boolean>>({}); const [resumedChips,setResumedChips]=useState<Record<string,boolean>>({}); const [retriedRuns,setRetriedRuns]=useState<Record<string,boolean>>({});
   const [mobileNav,setMobileNav]=useState(false); const [error,setError]=useState<string|null>(null); const [busy,setBusy]=useState(false);
   useEffect(()=>{
     if(!mobileNav)return;
@@ -270,6 +272,7 @@ export function App(){
   // normal message: the backend folds it into the paused run instead of
   // starting a new task.
   async function continueRun(messageId:string){if(!activeSessionId||sendingRef.current||busy)return;sendingRef.current=true;const text=t("resumeSent");try{await action(async()=>{await api(`/api/sessions/${activeSessionId}/messages`,{method:"POST",body:JSON.stringify({text,clientNonce:clientNonce()})});setResumedChips(current=>({...current,[messageId]:true}));await loadSessions()})}finally{sendingRef.current=false}}
+  async function retryRun(runId?:string){if(!runId||busy)return;await action(()=>api(`/api/runs/${runId}/retry`,{method:"POST",body:"{}"}));setRetriedRuns(current=>({...current,[runId]:true}))}
   async function clearSession(){if(!activeSessionId)return;setClearOpen(false);setSessionMenuOpen(false);await action(async()=>{await api(`/api/sessions/${activeSessionId}/messages`,{method:"DELETE"});setMessages([]);await loadSessions()})}
   async function deleteSession(id:string){if(!sessionsPath||!sessionStoreKey)return;setSessionMenuOpen(false);setBusy(true);setError(null);try{await api(`/api/sessions/${id}`,{method:"DELETE"});const next=await api<Session[]>(sessionsPath);setSessions(next);const pick=id===activeSessionId||!next.some(session=>session.id===activeSessionId)?next[0]?.id||null:activeSessionId;setActiveSessionId(pick);if(pick)writeSessionStore(sessionStoreKey,pick)}catch(e){setError(e instanceof Error?localizeError(e.message):t("operationFailed"))}finally{setBusy(false)}}
   async function rememberMessage(message:Message){const botId=message.speakerBotId||active?.id||activeRoom?.members[0]?.id;if(!botId||!message.body.trim())return;try{await api(`/api/bots/${botId}/memories`,{method:"POST",body:JSON.stringify({content:message.body,sessionId:activeSessionId})});setRemembered(current=>({...current,[message.id]:true}))}catch(e){setError(e instanceof Error?localizeError(e.message):t("rememberFailed"))}}
@@ -366,11 +369,11 @@ export function App(){
     <main className="chat-panel">
       <header className="topbar"><button type="button" className="icon-button mobile-menu" aria-label={t("chooseBot")} aria-expanded={mobileNav} aria-controls="conversation-sidebar" onClick={()=>setMobileNav(v=>!v)}><UseAnimations animation={menu} size={18} strokeColor="var(--ink)"/></button>{activeRoom?<><AvatarStack members={activeRoom.members} size={32} online thinkingIds={busyMembers.map(member=>member.id)}/><strong>{activeRoom.name}</strong><SessionMenu sessions={sessions} activeSessionId={activeSessionId} open={sessionMenuOpen} setOpen={setSessionMenuOpen} busy={busy} onSelect={selectSession} onCreate={createSession} onDelete={id=>{setSessionMenuOpen(false);setSessionToDelete(id)}} onClear={()=>{setSessionMenuOpen(false);setClearOpen(true)}}/><span className="grow"/>{topTools}</>:active?<><Avatar lookId={active.id} name={active.name} color={active.avatarColor} shape={active.avatarShape} active online/><strong>{active.name}</strong><SessionMenu sessions={sessions} activeSessionId={activeSessionId} open={sessionMenuOpen} setOpen={setSessionMenuOpen} busy={busy} onSelect={selectSession} onCreate={createSession} onDelete={id=>{setSessionMenuOpen(false);setSessionToDelete(id)}} onClear={()=>{setSessionMenuOpen(false);setClearOpen(true)}}/><span className="grow"/>{topTools}</>:<><strong>{t("chooseBot")}</strong><span className="grow"/>{topTools}</>}</header>
       {callOpen&&voiceSettings?.enabled&&active&&activeSessionId&&!activeRoomId?<CallOverlay bot={active} sessionId={activeSessionId} takeover={computer.takeoverRequested||computer.controlHolder==="user"} onHangUp={()=>setCallOpen(false)} onTakeOver={()=>{setRightPart("computer");setRightCollapsed(false);if(active)void action(()=>api(`/api/computer/${active.id}/takeover`,{method:"POST",body:"{}"}))}}/>:null}
-      <div className="messages">{(activeRoom||active)&&messages.length===0?<div className="welcome">{activeRoom?<AvatarStack members={activeRoom.members} size={56} online/>:<Avatar lookId={active!.id} name={active!.name} color={active!.avatarColor} shape={active!.avatarShape} active online size={64}/>}<h1>{activeRoom?t("startRoomDiscussion",{name:activeRoom.name}):t("startBotWork",{name:active!.name})}</h1><p>{activeRoom?t("roomWillReply",{names:listJoin(activeRoom.members.map(member=>member.name))}):active!.description||t("botWelcome")}</p></div>:messages.map(message=>{const spoken=message.role!=="user"&&Boolean(activeRoom);const speakerName=message.speakerName||(spoken?paneBot?.name:undefined);const speakerShape=(message.speakerShape||paneBot?.avatarShape||"blob") as AvatarShape;const files=messageFiles(message.blocks);const chips=chipBlocks(message.blocks);const hideBody=isAutoAttachCaption(message.body,files)||isAutoScheduleCaption(message.body,chips);return <div key={message.id} className={`message ${message.role} ${spoken?"spoken":""} ${files.length?"with-files":""}`}>{spoken&&<span className="msg-avatar"><Avatar lookId={message.speakerBotId||paneBot?.id||undefined} name={speakerName||"agent"} color={message.speakerColor||undefined} shape={speakerShape} size={22}/></span>}{spoken&&<b className="speaker" style={{color:message.speakerColor||undefined}}>{speakerName}</b>}{files.length>0&&<div className="msg-attachments">{files.map(file=><FileCard key={file.name} file={file}/>)}</div>}{chips.map((chip,index)=>chip.kind==="resume"?<div className="login-chip resume-chip" key={`${message.id}-resume-${index}`}><div className="login-label">{resumeTitle(chip.reason)}</div>{(chip.limit||0)>0?<div className="login-why">{t("resumeProgress",{turns:chip.turns||0,limit:chip.limit||0})}</div>:null}{resumedChips[message.id]||(message.seq??0)<lastAssistantSeq?null:<div className="resume-actions"><button type="button" className="primary" disabled={busy} onClick={()=>void continueRun(message.id)}>{t("resumeContinue")}</button><button type="button" className="outline" disabled={busy} onClick={()=>setResumedChips(current=>({...current,[message.id]:true}))}>{t("resumeStop")}</button></div>}</div>:chip.kind==="login"?<div className="login-chip" key={`${message.id}-login-${index}`}><div className="login-label">{t("loginNeedsYou")}</div><div className="login-site">{chip.site||message.body}</div>{chip.why?<div className="login-why">{t("loginWhy",{why:chip.why})}</div>:null}<button type="button" className="primary" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button></div>:chip.kind==="schedule"?<div className="sched-chip" key={`${message.id}-sched-${index}`}><div className="sched-label">{t("scheduleChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>:<div className="sched-chip" key={`${message.id}-run-${index}`}><div className="sched-label">{t("scheduleRunChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>)}{!hideBody&&(message.role==="assistant"?<div className="message-stack"><div className="message-body md"><ChatMarkdown>{message.body}</ChatMarkdown></div>{message.body.trim()?<CopyMessageButton text={message.body}/>:null}</div>:<span className="message-body">{message.body}</span>)}{!hideBody&&message.body.trim()&&<button type="button" className={`remember-msg ${remembered[message.id]?"saved":""}`} title={remembered[message.id]?t("remembered"):t("remember")} disabled={!!remembered[message.id]} onClick={()=>void rememberMessage(message)}><UseAnimations animation={bookmark} size={14} strokeColor="var(--muted)"/></button>}</div>})}{pausedForUser&&paneBot&&<div className="pause-banner" role="status"><span>{computer.controlHolder==="user"?t("pausedUserControl",{name:paneBot.name}):t("pausedNeedsUser",{name:paneBot.name})}{(computer.queuedRuns||0)>0&&<> {t("queuedMessages",{count:computer.queuedRuns||0})}</>}</span>{computer.controlHolder==="user"?<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/release`,{method:"POST",body:"{}"}))}>{t("releaseAndContinue")}</button>:<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/takeover`,{method:"POST",body:"{}"}))}>{t("takeOverNow")}</button>}</div>}{teaching&&active&&<div className="teach-banner recording" role="status"><span><i className="record-dot live"/>{t("teachingLive",{goal:teaching.goal})}<small>{t("teachingHint")}{teaching.eventCount>0&&<> · {t("teachingCaptured",{count:teaching.eventCount})}</>}</small></span><button type="button" className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button type="button" className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>}{drafting&&<div className="teach-banner" role="status"><UseAnimations animation={loading} size={18} wrapperStyle={{display:"inline-block",verticalAlign:"middle"}}/><span>{t("distilling",{goal:drafting.goal})}</span></div>}{skillDraft&&active&&<SkillDraftCard key={skillDraft.id} skill={skillDraft} busy={busy} onSave={(name,playbook)=>void saveSkill(skillDraft,name,playbook)} onTest={(name,playbook)=>void testSkill(skillDraft,name,playbook)} onDiscard={()=>void discardSkill(skillDraft)} onEdit={()=>setEditingSkillId(skillDraft.id)} onExport={(name,playbook)=>downloadSkill(name,skillDraft.goal,playbook)}/>}<div ref={messageEndRef} aria-hidden="true"/></div>
+      <div className="messages">{(activeRoom||active)&&messages.length===0?<div className="welcome">{activeRoom?<AvatarStack members={activeRoom.members} size={56} online/>:<Avatar lookId={active!.id} name={active!.name} color={active!.avatarColor} shape={active!.avatarShape} active online size={64}/>}<h1>{activeRoom?t("startRoomDiscussion",{name:activeRoom.name}):t("startBotWork",{name:active!.name})}</h1><p>{activeRoom?t("roomWillReply",{names:listJoin(activeRoom.members.map(member=>member.name))}):active!.description||t("botWelcome")}</p></div>:messages.map(message=>{const spoken=message.role!=="user"&&Boolean(activeRoom);const speakerName=message.speakerName||(spoken?paneBot?.name:undefined);const speakerShape=(message.speakerShape||paneBot?.avatarShape||"blob") as AvatarShape;const files=messageFiles(message.blocks);const chips=chipBlocks(message.blocks);const hideBody=isAutoAttachCaption(message.body,files)||isAutoScheduleCaption(message.body,chips);return <div key={message.id} className={`message ${message.role} ${spoken?"spoken":""} ${files.length?"with-files":""}`}>{spoken&&<span className="msg-avatar"><Avatar lookId={message.speakerBotId||paneBot?.id||undefined} name={speakerName||"agent"} color={message.speakerColor||undefined} shape={speakerShape} size={22}/></span>}{spoken&&<b className="speaker" style={{color:message.speakerColor||undefined}}>{speakerName}</b>}{files.length>0&&<div className="msg-attachments">{files.map(file=><FileCard key={file.name} file={file}/>)}</div>}{chips.map((chip,index)=>chip.kind==="error"?<div className="login-chip error-chip" key={`${message.id}-error-${index}`}><div className="login-label">{errorTitle(chip.code)}</div><div className="error-actions">{errorActions(chip.code).map(next=>next==="retry"?<button type="button" className="primary" key="retry" title={t("errorRetryHint")} disabled={busy||!chip.runId||Boolean(retriedRuns[chip.runId])} onClick={()=>void retryRun(chip.runId)}>{retriedRuns[chip.runId||""]?t("errorRetried"):t("errorRetry")}</button>:next==="screen"?<button type="button" key="screen" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button>:<button type="button" key="settings" onClick={()=>openAccount("model")}>{t("errorOpenSettings")}</button>)}<RunProbe runId={chip.runId||null} align="end" label={t("errorDetails")}><span className="error-probe" title={t("errorDetails")}><Info/></span></RunProbe></div></div>:chip.kind==="resume"?<div className="login-chip resume-chip" key={`${message.id}-resume-${index}`}><div className="login-label">{resumeTitle(chip.reason)}</div>{(chip.limit||0)>0?<div className="login-why">{t("resumeProgress",{turns:chip.turns||0,limit:chip.limit||0})}</div>:null}{resumedChips[message.id]||(message.seq??0)<lastAssistantSeq?null:<div className="resume-actions"><button type="button" className="primary" disabled={busy} onClick={()=>void continueRun(message.id)}>{t("resumeContinue")}</button><button type="button" className="outline" disabled={busy} onClick={()=>setResumedChips(current=>({...current,[message.id]:true}))}>{t("resumeStop")}</button></div>}</div>:chip.kind==="login"?<div className="login-chip" key={`${message.id}-login-${index}`}><div className="login-label">{t("loginNeedsYou")}</div><div className="login-site">{chip.site||message.body}</div>{chip.why?<div className="login-why">{t("loginWhy",{why:chip.why})}</div>:null}<button type="button" className="primary" onClick={()=>void openLoginScreen()}>{t("loginOpenScreen")}</button></div>:chip.kind==="schedule"?<div className="sched-chip" key={`${message.id}-sched-${index}`}><div className="sched-label">{t("scheduleChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>:<div className="sched-chip" key={`${message.id}-run-${index}`}><div className="sched-label">{t("scheduleRunChip")}</div><strong>{chip.name}</strong><small>{scheduleWhen(chip.cron,chip.human)}</small></div>)}{!hideBody&&(message.role==="assistant"?<div className="message-stack"><div className="message-body md"><ChatMarkdown>{message.body}</ChatMarkdown></div>{message.body.trim()?<CopyMessageButton text={message.body}/>:null}</div>:<span className="message-body">{message.body}</span>)}{!hideBody&&message.body.trim()&&<button type="button" className={`remember-msg ${remembered[message.id]?"saved":""}`} title={remembered[message.id]?t("remembered"):t("remember")} disabled={!!remembered[message.id]} onClick={()=>void rememberMessage(message)}><UseAnimations animation={bookmark} size={14} strokeColor="var(--muted)"/></button>}</div>})}{pausedForUser&&paneBot&&<div className="pause-banner" role="status"><span>{computer.controlHolder==="user"?t("pausedUserControl",{name:paneBot.name}):t("pausedNeedsUser",{name:paneBot.name})}{(computer.queuedRuns||0)>0&&<> {t("queuedMessages",{count:computer.queuedRuns||0})}</>}</span>{computer.controlHolder==="user"?<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/release`,{method:"POST",body:"{}"}))}>{t("releaseAndContinue")}</button>:<button type="button" className="primary" disabled={busy} onClick={()=>void action(()=>api(`/api/computer/${paneBot.id}/takeover`,{method:"POST",body:"{}"}))}>{t("takeOverNow")}</button>}</div>}{teaching&&active&&<div className="teach-banner recording" role="status"><span><i className="record-dot live"/>{t("teachingLive",{goal:teaching.goal})}<small>{t("teachingHint")}{teaching.eventCount>0&&<> · {t("teachingCaptured",{count:teaching.eventCount})}</>}</small></span><button type="button" className="primary" disabled={busy} onClick={()=>void stopTeaching()}>{t("finishDemo")}</button><button type="button" className="outline" disabled={busy} onClick={()=>void cancelTeaching()}>{t("cancel")}</button></div>}{drafting&&<div className="teach-banner" role="status"><UseAnimations animation={loading} size={18} wrapperStyle={{display:"inline-block",verticalAlign:"middle"}}/><span>{t("distilling",{goal:drafting.goal})}</span></div>}{skillDraft&&active&&<SkillDraftCard key={skillDraft.id} skill={skillDraft} busy={busy} onSave={(name,playbook)=>void saveSkill(skillDraft,name,playbook)} onTest={(name,playbook)=>void testSkill(skillDraft,name,playbook)} onDiscard={()=>void discardSkill(skillDraft)} onEdit={()=>setEditingSkillId(skillDraft.id)} onExport={(name,playbook)=>downloadSkill(name,skillDraft.goal,playbook)}/>}<div ref={messageEndRef} aria-hidden="true"/></div>
       {error&&<div className="error-banner"><span>{error}</span><button onClick={()=>setError(null)}><X/></button></div>}
       {otherSessionBusy&&<div className="queue-hint">{t("anotherConversationQueued")}</div>}
       <div className={`composer-dock ${statusMembers.length?"has-status":""}`}>
-      {statusMembers.map(member=>{const step=computer.busyStep&&member.id===computer.botId?computer.busyStep:null;const transition=isTransitionStep(step);const label=t("working",{name:member.name});return <div className="thinking-row" key={member.id}><Avatar lookId={member.id} name={member.name} color={member.avatarColor} shape={member.avatarShape} thinking online/><span className="working-copy"><span className="working-label">{label}</span>{!transition&&step?<span className="working-step">{localizeStep(step)}</span>:null}</span></div>})}
+      {statusMembers.map(member=>{const step=computer.busyStep&&member.id===computer.botId?computer.busyStep:null;const transition=isTransitionStep(step);const label=t("working",{name:member.name});return <div className="thinking-row" key={member.id}><RunProbe runId={member.id===computer.botId?computer.busyRunId:null}><Avatar lookId={member.id} name={member.name} color={member.avatarColor} shape={member.avatarShape} thinking online/></RunProbe><span className="working-copy"><span className="working-label">{label}</span>{!transition&&step?<span className="working-step">{localizeStep(step)}</span>:null}</span></div>})}
       <form className={`composer ${pendingFiles.length?"has-files":""}`} onSubmit={send} onDragOver={event=>{event.preventDefault()}} onDrop={event=>{event.preventDefault();if(event.dataTransfer.files.length)addPendingFiles(event.dataTransfer.files)}}><div className="plus-menu-wrap" onClick={event=>event.stopPropagation()}><button type="button" className={`composer-plus ${plusOpen?"open":""}`} disabled={!activeSessionId} title={t("moreActions")} aria-label={t("moreActions")} aria-haspopup="menu" aria-expanded={plusOpen} onClick={()=>setPlusOpen(v=>!v)}><Plus/></button>{plusOpen&&<div className="plus-menu" role="menu"><button type="button" role="menuitem" disabled={!activeSessionId||Boolean(teaching)} title={t("attachFileHint")} onClick={()=>{setPlusOpen(false);attachRef.current?.click()}}><Paperclip/>{t("attachFile")}</button><button type="button" role="menuitem" disabled={!active||Boolean(teaching)||Boolean(drafting)} title={active?t("teachTaskHint"):t("teachNeedsBot")} onClick={()=>{setPlusOpen(false);setTeachOpen(true)}}><i className="record-dot"/>{t("teachTask")}</button><button type="button" role="menuitem" disabled={!active} title={t("importSkillHint")} onClick={()=>{setPlusOpen(false);importRef.current?.click()}}><Upload/>{t("importSkill")}</button>{savedSkills.length>0&&<><hr/><div className="plus-menu-skills"><small className="plus-menu-label">{t("taughtSkills")}{savedSkills.length>5?` · ${savedSkills.length}`:""}</small>{savedSkills.length>=6&&<input className="plus-menu-search" value={skillQuery} onChange={e=>setSkillQuery(e.target.value)} placeholder={t("searchSkills")} aria-label={t("searchSkills")} onClick={e=>e.stopPropagation()}/>}<div className="plus-menu-skill-list">{listedSkills.map(skill=><div className="plus-menu-skill" key={skill.id}><button type="button" role="menuitem" title={t("runSkillNamed",{name:skill.name})+(skill.playbook.whenToUse?`\n${skill.playbook.whenToUse}`:"")} onClick={()=>runSkill(skill)}><Sparkle/>{skill.name}</button><button type="button" className="skill-edit" title={t("exportSkillHint")} aria-label={t("exportSkill")} onClick={()=>downloadSkill(skill.name,skill.goal,skill.playbook)}><Download/></button><button type="button" className="skill-edit" title={t("editSkill")} aria-label={t("editSkill")} onClick={()=>{setPlusOpen(false);setEditingSkillId(skill.id)}}><Pencil/></button></div>)}{listedSkills.length===0&&<small className="plus-menu-empty">{t("noMatchingSkills")}</small>}</div></div></>}</div>}</div><input ref={importRef} className="skill-import-input" type="file" accept="application/json,.json" tabIndex={-1} aria-hidden="true" onChange={event=>{const file=event.target.files?.[0];event.currentTarget.value="";if(file)void importSkillFile(file)}}/><input ref={attachRef} className="skill-import-input attach-input" type="file" multiple accept={ATTACH_ACCEPT} tabIndex={-1} aria-hidden="true" onChange={event=>{const files=[...event.target.files||[]];event.currentTarget.value="";if(files.length)addPendingFiles(files)}}/>{pendingFiles.length>0&&<div className="composer-files">{pendingFiles.map(item=><FileCard key={item.id} file={{name:item.file.name,size:item.file.size}} preview={item.preview} onRemove={()=>removePendingFile(item.id)}/>)}</div>}{slashSuggestions.length>0&&<div className="slash-suggestions" role="listbox" aria-label={t("slashList")}>{slashSuggestions.map((skill,index)=><button type="button" role="option" aria-selected={index===slashIndex} key={skill.name} onClick={()=>{setDraft(`/${skill.name} `);requestAnimationFrame(()=>document.querySelector<HTMLTextAreaElement>(".composer textarea")?.focus())}}><strong>/{skill.name}</strong><small>{skill.kind}</small><span>{skill.description}</span></button>)}</div>}<textarea rows={1} value={draft} onChange={e=>setDraft(e.target.value)} onKeyDown={composerKeyDown} onPaste={event=>{const files=event.clipboardData?.files;if(files&&files.length){event.preventDefault();addPendingFiles(files)}}} placeholder={teaching?t("teachingComposerHint"):activeSessionId&&chatName?t("messageTo",{name:chatName}):t("chooseConversationFirst")} disabled={!activeSessionId||Boolean(teaching)}/>{sessionBusy?<button type="button" className="send stop-send" title={t("stopConversation")} onClick={()=>void action(stopChat)}><Square/></button>:<button className="send" disabled={!activeSessionId||(!draft.trim()&&pendingFiles.length===0)||busy}><UseAnimations animation={arrowUp} size={20} strokeColor="var(--on-send)"/></button>}</form>
       </div>
     </main>
diff --git a/apps/web/src/locales/en.ts b/apps/web/src/locales/en.ts
index 71782c7..87b3556 100644
--- a/apps/web/src/locales/en.ts
+++ b/apps/web/src/locales/en.ts
@@ -459,5 +459,53 @@ export const en: { [K in keyof typeof zhTW]: string } = {
   resumeProgress: "{turns}/{limit} turns",
   resumeContinue: "Keep going",
   resumeStop: "Stop here",
-  resumeSent: "Keep going and finish it.",
+  resumeSent: "Keep going and finish it.",  monitorTitle: "Live log",
+  monitorQueued: "Queued",
+  monitorTurnOf: "turn {turn}/{limit}",
+  monitorTurn: "turn {turn}",
+  monitorElapsed: "up {time}",
+  monitorNow: "Now: {step}",
+  monitorEmpty: "Nothing yet — the trail fills in as I work.",
+  monitorLoadFailed: "Couldn’t load the log",
+  monitorCopy: "Copy log",
+  monitorCopied: "Copied",
+  monitorPinHint: "Folds away when the mouse leaves",
+  monitorPinned: "Pinned — press Esc to close",
+  monitorClose: "Close",
+  monitorKindModel: "Model",
+  monitorKindTool: "Action",
+  monitorKindRetry: "Retry",
+  monitorKindNotice: "Note",
+  monitorKindRun: "Status",
+  monitorRunStarted: "Started: {task}",
+  monitorRunCompleted: "Done in {turns} turns",
+  monitorRunFailed: "Failed",
+  monitorRunWaiting: "Waiting for you: {reason}",
+  monitorRunPaused: "Paused: {reason}",
+  monitorRunRetry: "Re-queued",
+  monitorModelTurn: "turn {turn}, thought for {time}",
+  monitorRetryAttempt: "attempt {attempt} failed, retrying",
+  monitorGaveUp: "gave up",
+  monitorStatusOk: "ok",
+  monitorStatusError: "failed",
+  monitorStatusTimedOut: "timed out",
+  monitorStatusPaused: "needs you",
+  errorTitleInterrupted: "Interrupted",
+  errorTitleToolTimeout: "Action timed out",
+  errorTitleModelKey: "Model API key rejected",
+  errorTitleModelQuota: "Rate limited or out of quota",
+  errorTitleModelUnknown: "Model not found",
+  errorTitleModelTimeout: "Model took too long",
+  errorTitleNetwork: "Can’t reach the model",
+  errorTitleComputerGone: "Computer is gone",
+  errorTitleLeaseLost: "Work picked up elsewhere",
+  errorTitleUnknown: "Something broke",
+  errorChip: "Task failed",
+  errorRetry: "Retry",
+  errorRetryHint: "Continue where it stopped",
+  errorOpenSettings: "Model settings",
+  errorDetails: "Show the log",
+  errorRetried: "Back in the queue",
+  errorRetryFailed: "Can’t re-queue right now — another run may still be busy.",
+
 };
diff --git a/apps/web/src/locales/zh-TW.ts b/apps/web/src/locales/zh-TW.ts
index baa9a77..0a2828f 100644
--- a/apps/web/src/locales/zh-TW.ts
+++ b/apps/web/src/locales/zh-TW.ts
@@ -154,5 +154,53 @@ export const zhTW = {
   resumeProgress: "{turns}/{limit} 輪",
   resumeContinue: "繼續",
   resumeStop: "就到這裡",
-  resumeSent: "繼續，把它做完。",
+  resumeSent: "繼續，把它做完。",  monitorTitle: "即時記錄",
+  monitorQueued: "排隊中",
+  monitorTurnOf: "第 {turn}/{limit} 輪",
+  monitorTurn: "第 {turn} 輪",
+  monitorElapsed: "已運行 {time}",
+  monitorNow: "現在：{step}",
+  monitorEmpty: "還沒有記錄。開始動手後會一條條出現。",
+  monitorLoadFailed: "記錄載入失敗",
+  monitorCopy: "複製記錄",
+  monitorCopied: "已複製",
+  monitorPinHint: "滑鼠移開會自動收起來",
+  monitorPinned: "已釘住（按 Esc 關閉）",
+  monitorClose: "關閉",
+  monitorKindModel: "模型",
+  monitorKindTool: "動作",
+  monitorKindRetry: "重試",
+  monitorKindNotice: "提醒",
+  monitorKindRun: "狀態",
+  monitorRunStarted: "開始工作：{task}",
+  monitorRunCompleted: "完成了（{turns} 輪）",
+  monitorRunFailed: "失敗了",
+  monitorRunWaiting: "停下來等你：{reason}",
+  monitorRunPaused: "暫停：{reason}",
+  monitorRunRetry: "重新排進佇列",
+  monitorModelTurn: "第 {turn} 輪，想了 {time}",
+  monitorRetryAttempt: "第 {attempt} 次嘗試失敗，正在重試",
+  monitorGaveUp: "放棄重試",
+  monitorStatusOk: "成功",
+  monitorStatusError: "失敗",
+  monitorStatusTimedOut: "逾時",
+  monitorStatusPaused: "要你接手",
+  errorTitleInterrupted: "被打斷",
+  errorTitleToolTimeout: "動作逾時",
+  errorTitleModelKey: "模型金鑰不被接受",
+  errorTitleModelQuota: "限流或額度不足",
+  errorTitleModelUnknown: "模型名稱找不到",
+  errorTitleModelTimeout: "模型太久沒回話",
+  errorTitleNetwork: "連不到模型",
+  errorTitleComputerGone: "電腦不見了",
+  errorTitleLeaseLost: "工作被另一邊接手",
+  errorTitleUnknown: "出錯了",
+  errorChip: "任務失敗",
+  errorRetry: "重試",
+  errorRetryHint: "從中斷的地方接著做",
+  errorOpenSettings: "開模型設定",
+  errorDetails: "看詳細記錄",
+  errorRetried: "已重新排進佇列",
+  errorRetryFailed: "現在重試排不進去，可能還有別的工作在跑。",
+
 } as const;
diff --git a/apps/web/src/main.tsx b/apps/web/src/main.tsx
index 6c22dd5..7736aac 100644
--- a/apps/web/src/main.tsx
+++ b/apps/web/src/main.tsx
@@ -8,6 +8,7 @@ import "./avatar.css";
 import "./chat.css";
 import "./computer.css";
 import "./schedule.css";
+import "./monitor.css";
 import "./call.css";
 import "./responsive.css";
 import { App } from "./App";
diff --git a/apps/web/src/monitor.css b/apps/web/src/monitor.css
new file mode 100644
index 0000000..297c71c
--- /dev/null
+++ b/apps/web/src/monitor.css
@@ -0,0 +1,39 @@
+/* The live run bubble. It floats above whatever opened it, so it never has to
+   fit inside the composer or a message bubble. */
+.run-probe{position:relative;display:inline-flex;align-items:center;border-radius:10px;outline:none}
+.run-probe:focus-visible{box-shadow:0 0 0 2px var(--focus)}
+.run-probe.inactive{cursor:default}
+.run-monitor{position:fixed;z-index:60;display:flex;flex-direction:column;gap:8px;max-height:min(64vh,460px);padding:12px 13px 10px;border:1px solid var(--border);border-radius:14px;background:var(--menu);box-shadow:0 18px 44px rgba(0,0,0,.55);color:var(--ink);font-size:12px;line-height:1.45;text-align:left;cursor:default}
+.run-monitor-head{display:flex;flex-wrap:wrap;align-items:baseline;justify-content:space-between;gap:8px}
+.run-monitor-head b{font-size:13px}
+.run-monitor-turn{color:var(--accent);font-variant-numeric:tabular-nums}
+.run-monitor-step{color:var(--muted);overflow-wrap:anywhere}
+.run-monitor-step.bad{color:var(--error-fg)}
+.run-monitor-list{display:flex;flex-direction:column;gap:5px;min-height:44px;max-height:min(38vh,300px);overflow-y:auto;padding-right:2px;border-top:1px solid var(--line);border-bottom:1px solid var(--line);padding-top:7px;padding-bottom:7px}
+.run-monitor-empty{color:var(--faint)}
+.run-line{display:grid;grid-template-columns:58px 40px 1fr;gap:7px;align-items:baseline}
+.run-line-clock{color:var(--faint);font-variant-numeric:tabular-nums}
+.run-line-kind{color:var(--muted)}
+.run-line-text{overflow-wrap:anywhere}
+.run-line.is-error .run-line-text,.run-line.is-timed_out .run-line-text{color:var(--error-fg)}
+.run-line.is-paused .run-line-text{color:var(--warn-fg)}
+.run-line.kind-run .run-line-text{color:var(--ink)}
+.run-line.kind-retry .run-line-text{color:var(--warn-fg)}
+.run-monitor-error{display:flex;flex-direction:column;gap:4px;padding:9px 10px;border:1px solid var(--error-border);border-radius:11px;background:var(--error-bg);color:var(--error-fg)}
+.run-monitor-error code{display:block;max-height:110px;overflow:auto;white-space:pre-wrap;overflow-wrap:anywhere;color:var(--error-fg);font-size:11px;opacity:.85}
+.run-monitor-foot{display:flex;align-items:center;gap:9px}
+.run-monitor-foot button{display:inline-flex;align-items:center;gap:5px;padding:5px 9px;border:1px solid var(--border);border-radius:9px;background:var(--elevated);color:var(--ink);font-size:12px;cursor:pointer}
+.run-monitor-foot button:hover{background:var(--surface-hover)}
+.run-monitor-foot small{flex:1;color:var(--faint)}
+.run-monitor-close{padding:3px 8px !important;color:var(--muted)}
+
+/* Failure card in the chat: what broke, and the one thing to do next. */
+.error-chip{border-color:var(--error-border);background:var(--error-bg)}
+.error-chip .login-label{display:flex;align-items:center;gap:7px;color:var(--error-fg);font-weight:600}
+.error-chip .error-body{color:var(--ink);overflow-wrap:anywhere}
+.error-chip .error-actions{display:flex;flex-wrap:wrap;align-items:center;gap:8px;margin-top:2px}
+.error-chip .error-actions button{padding:6px 12px;border:1px solid var(--border);border-radius:9px;background:var(--elevated);color:var(--ink);font-size:13px;cursor:pointer}
+.error-chip .error-actions button.primary{border-color:transparent;background:var(--cream);color:var(--on-cream)}
+.error-chip .error-actions button:disabled{opacity:.55;cursor:default}
+.error-chip .error-probe{margin-left:auto;color:var(--error-fg);cursor:pointer}
+.error-chip .error-said{color:var(--error-fg);font-size:12px}
diff --git a/apps/web/src/run-monitor.tsx b/apps/web/src/run-monitor.tsx
new file mode 100644
index 0000000..56506b6
--- /dev/null
+++ b/apps/web/src/run-monitor.tsx
@@ -0,0 +1,372 @@
+import { useCallback, useEffect, useRef, useState, type KeyboardEvent as ReactKeyboardEvent, type ReactNode } from "react";
+import { api } from "./api";
+import { t, type MessageKey } from "./i18n";
+import type { RunActivity, RunActivityEntry } from "./types";
+
+/**
+ * The observability bubble: hover (or tap, or focus) an avatar to watch the run
+ * that avatar is doing, turn by turn. The trail is polled rather than streamed
+ * because it is only interesting while somebody is looking at it.
+ */
+const POLL_MS = 1500;
+const CLOSE_MS = 300;
+const KEEP = 200;
+const PANEL_WIDTH = 380;
+const DONE = ["completed", "failed", "cancelled"];
+
+export type RunActionId = "retry" | "screen" | "settings";
+
+/**
+ * Which buttons actually help for a failure code, best first. A bad API key is
+ * fixed in settings and a lost computer is checked on screen — retrying alone
+ * would only burn another turn on the same wall.
+ */
+const ACTIONS: Record<string, RunActionId[]> = {
+  interrupted: ["screen", "retry"],
+  tool_timeout: ["screen", "retry"],
+  model_key: ["settings", "retry"],
+  model_unknown: ["settings", "retry"],
+  model_quota: ["retry", "settings"],
+  model_timeout: ["retry", "settings"],
+  network: ["retry", "settings"],
+  computer_gone: ["screen", "retry"],
+  lease_lost: ["retry"],
+  unknown: ["retry"],
+};
+
+const TITLES: Record<string, MessageKey> = {
+  interrupted: "errorTitleInterrupted",
+  tool_timeout: "errorTitleToolTimeout",
+  model_key: "errorTitleModelKey",
+  model_quota: "errorTitleModelQuota",
+  model_unknown: "errorTitleModelUnknown",
+  model_timeout: "errorTitleModelTimeout",
+  network: "errorTitleNetwork",
+  computer_gone: "errorTitleComputerGone",
+  lease_lost: "errorTitleLeaseLost",
+  unknown: "errorTitleUnknown",
+};
+
+export function errorActions(code?: string | null): RunActionId[] {
+  return ACTIONS[code || ""] || ACTIONS.unknown;
+}
+
+/** Short label for a failure code; the long sentence lives in the message body. */
+export function errorTitle(code?: string | null): string {
+  return t(TITLES[code || ""] || TITLES.unknown);
+}
+
+/** 1:05 / 1:02:05 — how long the run has been at it. */
+export function formatElapsed(ms: number): string {
+  const seconds = Number.isFinite(ms) ? Math.max(0, Math.floor(ms / 1000)) : 0;
+  const pad = (value: number) => String(value).padStart(2, "0");
+  const hours = Math.floor(seconds / 3600);
+  const minutes = Math.floor((seconds % 3600) / 60);
+  return hours > 0 ? `${hours}:${pad(minutes)}:${pad(seconds % 60)}` : `${minutes}:${pad(seconds % 60)}`;
+}
+
+/** 340ms / 6.4s / 1:05 — durations inside the trail stay glanceable. */
+export function shortDuration(ms?: number | null): string {
+  if (typeof ms !== "number" || !Number.isFinite(ms) || ms < 0) return "";
+  if (ms < 1000) return `${Math.round(ms)}ms`;
+  if (ms < 10_000) return `${(ms / 1000).toFixed(1)}s`;
+  return formatElapsed(ms);
+}
+
+function kindLabel(kind: string): string {
+  if (kind === "model") return t("monitorKindModel");
+  if (kind === "tool") return t("monitorKindTool");
+  if (kind === "retry") return t("monitorKindRetry");
+  if (kind === "notice") return t("monitorKindNotice");
+  return t("monitorKindRun");
+}
+
+function statusLabel(status?: string | null): string {
+  if (status === "ok") return t("monitorStatusOk");
+  if (status === "timed_out") return t("monitorStatusTimedOut");
+  if (status === "paused") return t("monitorStatusPaused");
+  if (status === "error") return t("monitorStatusError");
+  return "";
+}
+
+/** One trail line as plain text — what the panel shows and what gets copied. */
+export function trailText(entry: RunActivityEntry): string {
+  if (entry.kind === "run") {
+    if (entry.event === "started") return t("monitorRunStarted", { task: entry.task || "" });
+    if (entry.event === "completed") return t("monitorRunCompleted", { turns: entry.turns ?? 0 });
+    if (entry.event === "failed") return t("monitorRunFailed");
+    if (entry.event === "waiting_input") return t("monitorRunWaiting", { reason: entry.reason || "" });
+    if (entry.event === "paused") return t("monitorRunPaused", { reason: entry.reason || "" });
+    if (entry.event === "retry") return t("monitorRunRetry");
+    return entry.text || entry.reason || t("monitorKindRun");
+  }
+  if (entry.kind === "model") {
+    const head = t("monitorModelTurn", { turn: entry.turn ?? 0, time: shortDuration(entry.elapsedMs) });
+    return entry.text ? `${head} — ${entry.text}` : head;
+  }
+  if (entry.kind === "tool") {
+    const parts = [entry.step || entry.name || t("monitorKindTool")];
+    const status = statusLabel(entry.status);
+    if (status) parts.push(status);
+    const time = shortDuration(entry.elapsedMs);
+    if (time) parts.push(time);
+    const detail = entry.snippet ? ` — ${entry.snippet}` : "";
+    return `${parts.join(" · ")}${detail}`;
+  }
+  if (entry.kind === "retry") {
+    const head = `${t("monitorRetryAttempt", { attempt: entry.attempt ?? 0 })}${entry.gaveUp ? ` · ${t("monitorGaveUp")}` : ""}`;
+    return entry.error ? `${head} — ${entry.error}` : head;
+  }
+  return entry.text || t("monitorKindNotice");
+}
+
+function clockOf(createdAt: string): string {
+  const date = new Date(createdAt);
+  if (Number.isNaN(date.getTime())) return "";
+  const pad = (value: number) => String(value).padStart(2, "0");
+  return `${pad(date.getHours())}:${pad(date.getMinutes())}:${pad(date.getSeconds())}`;
+}
+
+function logDump(snapshot: RunActivity | null, entries: RunActivityEntry[]): string {
+  const turn = snapshot?.turnLimit
+    ? t("monitorTurnOf", { turn: snapshot.turn ?? 0, limit: snapshot.turnLimit })
+    : t("monitorTurn", { turn: snapshot?.turn ?? 0 });
+  const head = [
+    snapshot?.runId ?? "",
+    snapshot?.status ?? "",
+    snapshot?.turn != null ? turn : "",
+    snapshot?.step ?? "",
+    snapshot?.elapsedMs != null ? t("monitorElapsed", { time: formatElapsed(snapshot.elapsedMs) }) : "",
+  ].filter(Boolean).join(" · ");
+  const trail = entries.map(entry => `${clockOf(entry.createdAt)} ${trailText(entry)}`);
+  const error = snapshot?.error ? `\n[${snapshot.error.code}] ${snapshot.error.raw}` : "";
+  return [head, ...trail].join("\n") + error;
+}
+
+async function copyLog(text: string): Promise<boolean> {
+  try {
+    await navigator.clipboard.writeText(text);
+    return true;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Wraps an avatar (or any small trigger) in a live readout of the run behind
+ * it. Hover opens it and it closes ~300ms after the pointer leaves; a click
+ * pins it so the panel survives reaching for the copy button; Esc lets go.
+ */
+export function RunProbe({ runId, align = "start", label, children }: { runId?: string | null; align?: "start" | "end"; label?: string; children: ReactNode }) {
+  const [open, setOpen] = useState(false);
+  const [pinned, setPinned] = useState(false);
+  const [snapshot, setSnapshot] = useState<RunActivity | null>(null);
+  const [entries, setEntries] = useState<RunActivityEntry[]>([]);
+  const [stale, setStale] = useState(false);
+  const [copied, setCopied] = useState<boolean | null>(null);
+  const [box, setBox] = useState<{ left: number; bottom: number } | null>(null);
+  const anchorRef = useRef<HTMLSpanElement | null>(null);
+  const listRef = useRef<HTMLSpanElement | null>(null);
+  const lastId = useRef(0);
+  const followTail = useRef(true);
+  const linger = useRef<number | null>(null);
+  const pinnedRef = useRef(false);
+
+  useEffect(() => {
+    pinnedRef.current = pinned;
+  }, [pinned]);
+
+  useEffect(() => {
+    lastId.current = 0;
+    setSnapshot(null);
+    setEntries([]);
+    setStale(false);
+  }, [runId]);
+
+  const place = useCallback(() => {
+    const node = anchorRef.current;
+    if (!node) return;
+    const rect = node.getBoundingClientRect();
+    const width = Math.min(PANEL_WIDTH, window.innerWidth - 24);
+    const edge = align === "end" ? rect.right - width : rect.left;
+    setBox({
+      left: Math.max(12, Math.min(edge, window.innerWidth - 12 - width)),
+      bottom: Math.max(12, window.innerHeight - rect.top + 8),
+    });
+  }, [align]);
+
+  useEffect(() => {
+    if (!open || !runId) return;
+    let stopped = false;
+    let timer = 0;
+    const load = async () => {
+      try {
+        const after = lastId.current > 0 ? `?after=${lastId.current}` : "";
+        const next = await api<RunActivity>(`/api/runs/${runId}/activity${after}`);
+        if (stopped) return;
+        setStale(false);
+        setSnapshot(next);
+        setEntries(current => {
+          if (lastId.current <= 0) return next.activity.slice(-KEEP);
+          const seen = new Set(current.map(entry => entry.id));
+          return [...current, ...next.activity.filter(entry => !seen.has(entry.id))].slice(-KEEP);
+        });
+        for (const entry of next.activity) lastId.current = Math.max(lastId.current, entry.id);
+        if (DONE.includes(next.status)) window.clearInterval(timer);
+      } catch {
+        if (!stopped) setStale(true);
+      }
+    };
+    timer = window.setInterval(() => void load(), POLL_MS);
+    void load();
+    return () => {
+      stopped = true;
+      window.clearInterval(timer);
+    };
+  }, [open, runId]);
+
+  useEffect(() => {
+    if (!open) return;
+    place();
+    window.addEventListener("resize", place);
+    return () => window.removeEventListener("resize", place);
+  }, [open, place]);
+
+  useEffect(() => {
+    const list = listRef.current;
+    if (open && followTail.current && list) list.scrollTop = list.scrollHeight;
+  }, [entries, open]);
+
+  const hoverIn = useCallback(() => {
+    if (linger.current) window.clearTimeout(linger.current);
+    linger.current = null;
+    setOpen(true);
+  }, []);
+
+  const hoverOut = useCallback(() => {
+    if (pinnedRef.current) return;
+    if (linger.current) window.clearTimeout(linger.current);
+    linger.current = window.setTimeout(() => setOpen(false), CLOSE_MS);
+  }, []);
+
+  useEffect(
+    () => () => {
+      if (linger.current) window.clearTimeout(linger.current);
+    },
+    [],
+  );
+
+  const onKeyDown = (event: ReactKeyboardEvent<HTMLSpanElement>) => {
+    if (event.key === "Escape") {
+      setPinned(false);
+      setOpen(false);
+      return;
+    }
+    if (event.key === "Enter" || event.key === " ") {
+      event.preventDefault();
+      setPinned(current => !current);
+      setOpen(true);
+    }
+  };
+
+  const turnText =
+    snapshot?.status === "queued"
+      ? t("monitorQueued")
+      : snapshot?.turn != null
+        ? snapshot.turnLimit
+          ? t("monitorTurnOf", { turn: snapshot.turn, limit: snapshot.turnLimit })
+          : t("monitorTurn", { turn: snapshot.turn })
+        : "—";
+
+  return (
+    <span
+      className={`run-probe ${runId ? "" : "inactive"}`}
+      ref={anchorRef}
+      tabIndex={runId ? 0 : -1}
+      aria-expanded={open}
+      aria-label={label || t("monitorTitle")}
+      onMouseEnter={runId ? hoverIn : undefined}
+      onMouseLeave={runId ? hoverOut : undefined}
+      onClick={
+        runId
+          ? () => {
+              setOpen(true);
+              setPinned(current => !current);
+            }
+          : undefined
+      }
+      onKeyDown={runId ? onKeyDown : undefined}
+    >
+      {children}
+      {open ? (
+        <span
+          className={`run-monitor ${pinned ? "pinned" : ""}`}
+          style={box ? { left: box.left, bottom: box.bottom, width: Math.min(PANEL_WIDTH, window.innerWidth - 24) } : { display: "none" }}
+          role="dialog"
+          aria-label={t("monitorTitle")}
+          onMouseEnter={hoverIn}
+          onMouseLeave={hoverOut}
+        >
+          <span className="run-monitor-head">
+            <b>{t("monitorTitle")}</b>
+            <span className="run-monitor-turn">
+              {turnText}
+              {snapshot?.elapsedMs != null ? ` · ${t("monitorElapsed", { time: formatElapsed(snapshot.elapsedMs) })}` : ""}
+            </span>
+          </span>
+          <span className={`run-monitor-step ${snapshot?.error ? "bad" : ""}`}>
+            {snapshot?.step ? t("monitorNow", { step: snapshot.step }) : snapshot?.error?.headline || t("monitorEmpty")}
+          </span>
+          <span
+            className="run-monitor-list"
+            ref={listRef}
+            onScroll={event => {
+              const node = event.currentTarget;
+              followTail.current = node.scrollHeight - node.scrollTop - node.clientHeight < 32;
+            }}
+          >
+            {entries.length === 0 ? <span className="run-monitor-empty">{stale ? t("monitorLoadFailed") : t("monitorEmpty")}</span> : null}
+            {entries.map(entry => (
+              <span className={`run-line kind-${entry.kind} ${entry.status && entry.status !== "ok" ? `is-${entry.status}` : ""}`} key={entry.id}>
+                <span className="run-line-clock">{clockOf(entry.createdAt)}</span>
+                <span className="run-line-kind">{kindLabel(entry.kind)}</span>
+                <span className="run-line-text">{trailText(entry)}</span>
+              </span>
+            ))}
+          </span>
+          {snapshot?.error ? (
+            <span className="run-monitor-error">
+              <b>{`[${snapshot.error.code}] ${errorTitle(snapshot.error.code)}`}</b>
+              <code>{snapshot.error.raw}</code>
+            </span>
+          ) : null}
+          <span className="run-monitor-foot">
+            <button
+              type="button"
+              onClick={() =>
+                void copyLog(logDump(snapshot, entries)).then(ok => {
+                  setCopied(ok);
+                  window.setTimeout(() => setCopied(null), 1800);
+                })
+              }
+            >
+              {copied === null ? t("monitorCopy") : copied ? t("monitorCopied") : t("clipboardWriteBlocked")}
+            </button>
+            <small>{pinned ? t("monitorPinned") : t("monitorPinHint")}</small>
+            <button
+              type="button"
+              className="run-monitor-close"
+              aria-label={t("monitorClose")}
+              onClick={() => {
+                setPinned(false);
+                setOpen(false);
+              }}
+            >
+              ×
+            </button>
+          </span>
+        </span>
+      ) : null}
+    </span>
+  );
+}
diff --git a/apps/web/src/types.ts b/apps/web/src/types.ts
index 5ba57b9..d870c2e 100644
--- a/apps/web/src/types.ts
+++ b/apps/web/src/types.ts
@@ -10,6 +10,12 @@ export interface MessageFile { kind:"image"|"file"; name:string; mimeType?:strin
 export interface RoomMember { id:string; name:string; avatarColor:string; avatarShape:AvatarShape }
 export interface Room { id:string; name:string; members:RoomMember[]; lastMessageAt:string|null; lastPreview:string|null; unreadCount:number }
 export interface ComputerStatus { botId:string; mode:ComputerMode; state:ComputerState; controlHolder:"none"|"bot"|"user"; takeoverRequested:boolean; busyBotName:string|null; busySessionId:string|null; busyRunId:string|null; busyStep?:string|null; usingComputer?:boolean; waitingRunId?:string|null; waitingSessionId?:string|null; queuedRuns?:number; display:string|null; profileMode:string; screenAvailable:boolean }
+/** One line of the live trail a run writes while it works. */
+export type RunActivityKind = "run"|"model"|"tool"|"retry"|"notice";
+export interface RunActivityEntry { id:number; kind:RunActivityKind; createdAt:string; turn?:number|null; event?:string|null; task?:string|null; reason?:string|null; turns?:number|null; limit?:number|null; error?:string|null; name?:string|null; step?:string|null; status?:string|null; elapsedMs?:number|null; toolCalls?:number|null; text?:string|null; snippet?:string|null; attempt?:number|null; gaveUp?:boolean|null }
+export interface RunActivityError { code:string; headline:string; raw:string }
+export interface RunActivity { runId:string; status:string; turn:number|null; turnLimit:number|null; step:string|null; elapsedMs:number|null; error:RunActivityError|null; activity:RunActivityEntry[] }
+
 export interface PlaybookStep { do:string; expect?:string; note?:string }
 export interface PlaybookInput { name:string; description?:string; example?:string }
 export interface Playbook { name?:string; whenToUse?:string; intent?:string; inputs?:PlaybookInput[]; preconditions?:string[]; steps?:(PlaybookStep|string)[]; howToCheck?:string; whatToReturn?:string; cautions?:string[] }
diff --git a/clippy.toml b/clippy.toml
new file mode 100644
index 0000000..18e7b04
--- /dev/null
+++ b/clippy.toml
@@ -0,0 +1,9 @@
+# Clippy configuration for the LazyBoy workspace.
+#
+# cargo-clippy looks for this file next to the working directory it was started
+# in, so always run clippy from the workspace root (`make clippy`).
+#
+# Handlers and run/vault/voice helpers legitimately thread 8-10 parameters
+# (state + actor + ids). The default of 7 is too tight for this codebase, so we
+# keep the guard for genuinely oversized signatures instead of disabling it.
+too-many-arguments-threshold = 10
diff --git a/crates/api/Cargo.toml b/crates/api/Cargo.toml
index f02f732..9d6ecbe 100644
--- a/crates/api/Cargo.toml
+++ b/crates/api/Cargo.toml
@@ -2,6 +2,7 @@
 name = "lazyboy-api"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
@@ -14,23 +15,19 @@ axum.workspace = true
 tokio.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-thiserror.workspace = true
 tracing.workspace = true
 tracing-subscriber.workspace = true
 sqlx.workspace = true
 uuid.workspace = true
 chrono.workspace = true
 tower-http.workspace = true
-async-trait = "0.1"
 rig-core.workspace = true
 base64.workspace = true
-hmac.workspace = true
 sha2.workspace = true
 hex.workspace = true
 reqwest.workspace = true
 tokio-tungstenite.workspace = true
 futures-util = "0.3"
-http-body-util = "0.1"
 dotenvy = "0.15"
 fastembed = { version = "6.0.2", default-features = false, features = ["hf-hub-rustls-tls", "ort-load-dynamic"] }
 rmcp = { version = "3.2", default-features = false, features = ["client", "transport-child-process", "transport-streamable-http-client-reqwest"] }
@@ -41,3 +38,6 @@ chrono-tz = "0.10"
 rand = "0.8"
 
 cap-std = "3"
+
+[lints]
+workspace = true
diff --git a/crates/api/src/attachments.rs b/crates/api/src/attachments.rs
index 3c7c651..5bdf4b0 100644
--- a/crates/api/src/attachments.rs
+++ b/crates/api/src/attachments.rs
@@ -185,13 +185,13 @@ pub async fn llm_parts(
                 }
             }
         } else if is_text_mime(&stored.mime_type) {
-            if let Some(bytes) = bytes {
-                if let Some(text) = utf8_preview(&bytes) {
-                    parts.push(UserContent::text(format!(
-                        "Contents of {}:\n```\n{text}\n```",
-                        stored.name
-                    )));
-                }
+            if let Some(bytes) = bytes
+                && let Some(text) = utf8_preview(&bytes)
+            {
+                parts.push(UserContent::text(format!(
+                    "Contents of {}:\n```\n{text}\n```",
+                    stored.name
+                )));
             }
         } else {
             notes.push(
@@ -220,12 +220,12 @@ pub async fn sweep_all_inboxes(data_dir: &str) {
                 continue;
             };
             sweep_cap_inbox(&home, "inbox");
-            if let Ok(bots) = home.open_dir("bots") {
-                if let Ok(entries) = bots.entries() {
-                    for bot in entries.flatten() {
-                        if let Ok(dir) = bots.open_dir(bot.file_name()) {
-                            sweep_cap_inbox(&dir, "inbox");
-                        }
+            if let Ok(bots) = home.open_dir("bots")
+                && let Ok(entries) = bots.entries()
+            {
+                for bot in entries.flatten() {
+                    if let Ok(dir) = bots.open_dir(bot.file_name()) {
+                        sweep_cap_inbox(&dir, "inbox");
                     }
                 }
             }
@@ -351,9 +351,7 @@ fn parse_stored(value: &Value) -> Option<StoredAttachment> {
     }
     let path = value.get("path").and_then(Value::as_str).unwrap_or("");
     if !path.is_empty() {
-        let Some(stored) = path.strip_prefix("inbox/") else {
-            return None;
-        };
+        let stored = path.strip_prefix("inbox/")?;
         if stored.is_empty() || stored == "." || stored == ".." || stored.contains(['/', '\\']) {
             return None;
         }
diff --git a/crates/api/src/auth.rs b/crates/api/src/auth.rs
index 97bc36c..9099095 100644
--- a/crates/api/src/auth.rs
+++ b/crates/api/src/auth.rs
@@ -45,7 +45,7 @@ impl AuthConfig {
     pub fn strong_enough_for_network(&self) -> bool {
         self.token
             .as_ref()
-            .map(|token| token.as_bytes().len() >= 32 && token != "dev-token")
+            .map(|token| token.len() >= 32 && token != "dev-token")
             .unwrap_or(false)
     }
 
@@ -83,14 +83,13 @@ impl AuthConfig {
         let key = hex::encode(Sha256::digest(value.as_bytes()));
         let mut sessions = self.sessions.lock().unwrap();
         sessions.retain(|_, expires| *expires > Instant::now());
-        if sessions.len() >= 4096 {
-            if let Some(oldest) = sessions
+        if sessions.len() >= 4096
+            && let Some(oldest) = sessions
                 .iter()
                 .min_by_key(|(_, t)| **t)
                 .map(|(k, _)| k.clone())
-            {
-                sessions.remove(&oldest);
-            }
+        {
+            sessions.remove(&oldest);
         }
         sessions.insert(key, Instant::now() + Duration::from_secs(604800));
         Some(format!(
diff --git a/crates/api/src/computer.rs b/crates/api/src/computer.rs
index e3e69cd..047e576 100644
--- a/crates/api/src/computer.rs
+++ b/crates/api/src/computer.rs
@@ -1076,15 +1076,14 @@ async fn pause_idle_computers(state: &AppState) {
         if computer_has_active_work(state, &computer.id).await {
             continue;
         }
-        if let Some(computer_ref) = computer_ref(&computer) {
-            if state
+        if let Some(computer_ref) = computer_ref(&computer)
+            && state
                 .sandbox
                 .suspend(&computer_ref, &idle_adapter(&computer, "idle"))
                 .await
                 .is_err()
-            {
-                continue;
-            }
+        {
+            continue;
         }
         let _ = sqlx::query(
             "UPDATE computers SET state = 'suspended', updated_at = now()
diff --git a/crates/api/src/file_skills.rs b/crates/api/src/file_skills.rs
index 726a59c..915fb0b 100644
--- a/crates/api/src/file_skills.rs
+++ b/crates/api/src/file_skills.rs
@@ -33,9 +33,7 @@ fn parse_file(path: PathBuf, name: String) -> Option<FileSkill> {
         let (header, body) = rest.split_once("\n---")?;
         (
             header,
-            body.trim_start_matches(|ch| ch == '\n' || ch == '\r')
-                .trim()
-                .to_string(),
+            body.trim_start_matches(['\n', '\r']).trim().to_string(),
         )
     } else {
         ("", text.trim().to_string())
diff --git a/crates/api/src/main.rs b/crates/api/src/main.rs
index 0c701cb..250a12b 100644
--- a/crates/api/src/main.rs
+++ b/crates/api/src/main.rs
@@ -6,6 +6,7 @@ mod file_skills;
 mod mcp;
 mod mcp_catalog;
 mod memory;
+mod monitor;
 mod retention;
 mod rooms;
 mod routes;
diff --git a/crates/api/src/mcp.rs b/crates/api/src/mcp.rs
index 40c2ce2..99052c7 100644
--- a/crates/api/src/mcp.rs
+++ b/crates/api/src/mcp.rs
@@ -570,7 +570,7 @@ async fn create_server(
             internal(error.to_string())
         }
     })?;
-    let mut row = load_row(state.pool(), &actor, &id)
+    let row = load_row(state.pool(), &actor, &id)
         .await
         .map_err(|error| internal(error.to_string()))?
         .ok_or_else(|| internal("missing row".into()))?;
diff --git a/crates/api/src/memory.rs b/crates/api/src/memory.rs
index ed85b71..c5f77cf 100644
--- a/crates/api/src/memory.rs
+++ b/crates/api/src/memory.rs
@@ -28,7 +28,9 @@ pub struct MemoryService {
 
 enum ModelState {
     Uninitialized,
-    Ready(TextEmbedding),
+    // Boxed: the embedding model is far larger than the other two variants and
+    // this enum lives inside an Arc<Mutex<..>> shared by every request.
+    Ready(Box<TextEmbedding>),
     Unavailable,
 }
 
@@ -102,7 +104,7 @@ impl MemoryService {
                     .with_cache_dir(cache_dir)
                     .with_show_download_progress(false);
                 match TextEmbedding::try_new(options) {
-                    Ok(embedding) => *state = ModelState::Ready(embedding),
+                    Ok(embedding) => *state = ModelState::Ready(Box::new(embedding)),
                     Err(error) => {
                         *state = ModelState::Unavailable;
                         return Err(format!("FastEmbed unavailable: {error}"));
@@ -151,7 +153,7 @@ impl MemoryService {
         let content = validate_content(&input.content)?;
         validate_importance(input.importance)?;
         let embedding = self.embed(content.clone()).await;
-        let vector = embedding.as_ref().map(vector_literal);
+        let vector = embedding.as_deref().map(vector_literal);
         let id = Uuid::new_v4();
         let mut tx = pool.begin().await.map_err(|error| error.to_string())?;
         let item: MemoryItem = sqlx::query_as(
@@ -212,7 +214,7 @@ impl MemoryService {
         }
         let limit = limit.unwrap_or(self.top_k).clamp(1, 50);
         let embedding = self.embed(query.to_string()).await;
-        let rows = if let Some(vector) = embedding.as_ref().map(vector_literal) {
+        let rows = if let Some(vector) = embedding.as_deref().map(vector_literal) {
             sqlx::query_as(
                 "SELECT id,session_id,source_run_id,source_message_id,content,importance,revision,
                         created_at,updated_at,deleted_at
@@ -269,7 +271,7 @@ impl MemoryService {
         let vector = self
             .embed(content.clone())
             .await
-            .as_ref()
+            .as_deref()
             .map(vector_literal);
         let mut tx = pool.begin().await.map_err(|error| error.to_string())?;
         let item: Option<MemoryItem> = sqlx::query_as(
@@ -377,7 +379,7 @@ async fn insert_revision(
     Ok(())
 }
 
-fn vector_literal(vector: &Vec<f32>) -> String {
+fn vector_literal(vector: &[f32]) -> String {
     format!(
         "[{}]",
         vector
diff --git a/crates/api/src/monitor.rs b/crates/api/src/monitor.rs
new file mode 100644
index 0000000..3913df3
--- /dev/null
+++ b/crates/api/src/monitor.rs
@@ -0,0 +1,538 @@
+//! Observability surface for a run: the activity trail the chat bubble reads,
+//! and the plain-language reading of a failure that turns into an action item.
+use axum::extract::{Path, Query, State};
+use axum::http::StatusCode;
+use axum::routing::{get, post};
+use axum::{Json, Router};
+use chrono::{DateTime, Utc};
+use serde_json::{Value, json};
+
+use crate::db::Actor;
+use crate::state::AppState;
+
+type ApiError = (StatusCode, Json<Value>);
+
+/// Longest a single string inside an activity payload may get. The trail is a
+/// glanceable summary, not a second copy of the transcript.
+const MAX_STRING_CHARS: usize = 512;
+const DEFAULT_ACTIVITY_LIMIT: i32 = 60;
+const MAX_ACTIVITY_LIMIT: i32 = 200;
+
+pub fn router() -> Router<AppState> {
+    Router::new()
+        .route("/api/runs/{id}/activity", get(activity))
+        .route("/api/runs/{id}/retry", post(retry))
+}
+
+/// Append one line to the run's trail. Diagnostics never fail a run: a write
+/// that cannot land is logged and dropped.
+pub async fn record(state: &AppState, run_id: &str, kind: &str, payload: Value) {
+    let result = sqlx::query("INSERT INTO run_activity (run_id,kind,payload) VALUES ($1,$2,$3)")
+        .bind(run_id)
+        .bind(kind)
+        .bind(clamp_strings(&payload, MAX_STRING_CHARS))
+        .execute(state.pool())
+        .await;
+    if let Err(error) = result {
+        tracing::warn!(run_id, kind, "failed to record run activity: {error}");
+    }
+}
+
+/// One line of text for the trail: single line, bounded, never an image.
+pub fn snippet(text: &str, max_chars: usize) -> String {
+    let flat = text.split_whitespace().collect::<Vec<_>>().join(" ");
+    if flat.chars().count() <= max_chars {
+        return flat;
+    }
+    let mut head: String = flat.chars().take(max_chars).collect();
+    head.push('…');
+    head
+}
+
+/// Bound every string leaf so one chatty tool result cannot bloat the trail.
+pub fn clamp_strings(value: &Value, max_chars: usize) -> Value {
+    match value {
+        Value::String(text) if text.chars().count() > max_chars => {
+            let mut head: String = text.chars().take(max_chars).collect();
+            head.push('…');
+            Value::String(head)
+        }
+        Value::Array(items) => Value::Array(
+            items
+                .iter()
+                .map(|item| clamp_strings(item, max_chars))
+                .collect(),
+        ),
+        Value::Object(map) => Value::Object(
+            map.iter()
+                .map(|(key, item)| (key.clone(), clamp_strings(item, max_chars)))
+                .collect(),
+        ),
+        other => other.clone(),
+    }
+}
+
+/// What the user is told, and what they can do about it. `code` is the stable
+/// contract the chat card renders from; the Chinese strings are the fallback
+/// body that also lives in the message history.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct RunFailure {
+    pub code: &'static str,
+    pub headline: &'static str,
+    pub action: &'static str,
+    pub retryable: bool,
+    pub needs_settings: bool,
+    pub needs_computer: bool,
+}
+
+impl RunFailure {
+    fn new(
+        code: &'static str,
+        headline: &'static str,
+        action: &'static str,
+        retryable: bool,
+        needs_settings: bool,
+        needs_computer: bool,
+    ) -> Self {
+        Self {
+            code,
+            headline,
+            action,
+            retryable,
+            needs_settings,
+            needs_computer,
+        }
+    }
+}
+
+/// Turn a raw run error into a classified, human-readable failure. Ordering is
+/// the policy: the most specific and most actionable cause wins, so a 429 that
+/// also mentions a timeout is reported as a quota problem, not a slow network.
+pub fn classify_run_error(error: &str) -> RunFailure {
+    let text = error.to_lowercase();
+    let has = |needles: &[&str]| needles.iter().any(|needle| text.contains(needle));
+    if has(&["worker interrupted"]) {
+        return RunFailure::new(
+            "interrupted",
+            "我剛剛被打斷了，最後那個動作可能已經做下去了。",
+            "先按重試；不確定的話，打開它的畫面確認現狀再繼續。",
+            true,
+            false,
+            false,
+        );
+    }
+    if text.contains("tool ") && text.contains("timed out") {
+        return RunFailure::new(
+            "tool_timeout",
+            "有一個動作卡超過 150 秒，我無法確定它有沒有做完。",
+            "先打開它的畫面看一眼，再按重試；我不會重複已經成功的步驟。",
+            true,
+            false,
+            false,
+        );
+    }
+    if has(&[
+        "invalid api key",
+        "incorrect api key",
+        "unauthorized",
+        "authentication",
+        "forbidden",
+        "401",
+        "403",
+    ]) {
+        return RunFailure::new(
+            "model_key",
+            "模型不認得這個 API 金鑰。",
+            "到「設定 → 模型」重新貼一次金鑰，再按重試。",
+            false,
+            true,
+            false,
+        );
+    }
+    if has(&[
+        "429",
+        "rate limit",
+        "ratelimit",
+        "too many requests",
+        "insufficient",
+        "quota",
+        "credit",
+        "billing",
+    ]) {
+        return RunFailure::new(
+            "model_quota",
+            "模型那邊限流了，或是額度已經用完。",
+            "等一下再按重試；如果是額度用完，到「設定 → 模型」換一個或加值。",
+            true,
+            true,
+            false,
+        );
+    }
+    if has(&[
+        "model not found",
+        "unknown model",
+        "invalid model",
+        "does not exist",
+        "404",
+    ]) {
+        return RunFailure::new(
+            "model_unknown",
+            "這個模型名稱找不到，可能已下架或打錯了。",
+            "到「設定 → 模型」選一個存在的模型，再按重試。",
+            false,
+            true,
+            false,
+        );
+    }
+    if has(&["timed out", "timeout", "逾時"]) {
+        return RunFailure::new(
+            "model_timeout",
+            "模型太久沒有回話（超過 165 秒）。",
+            "先按重試；每次都發生的話，到「設定 → 模型」換一個快一點的模型。",
+            true,
+            true,
+            false,
+        );
+    }
+    if has(&[
+        "error sending request",
+        "failed to connect",
+        "connection refused",
+        "dns",
+        "host not found",
+        "certificate",
+        "tls",
+        "network",
+    ]) {
+        return RunFailure::new(
+            "network",
+            "連不到模型服務，網路或位址不對。",
+            "確認網路與「設定 → 模型」的 Base URL，再按重試。",
+            true,
+            true,
+            false,
+        );
+    }
+    if has(&[
+        "no such container",
+        "container",
+        "docker",
+        "computer not found",
+        "computer is not running",
+    ]) {
+        return RunFailure::new(
+            "computer_gone",
+            "它的電腦不見了或已經停止。",
+            "先把電腦啟動，再按重試。",
+            true,
+            false,
+            true,
+        );
+    }
+    if text.contains("lease") {
+        return RunFailure::new(
+            "lease_lost",
+            "這份工作被另一邊接手過，我手上這份已經失效。",
+            "按重試重新開始這一步就好。",
+            true,
+            false,
+            false,
+        );
+    }
+    RunFailure::new(
+        "unknown",
+        "我卡住了，這一輪沒有完成。",
+        "按重試看看；需要的話把下面的記錄複製起來給我。",
+        true,
+        false,
+        false,
+    )
+}
+
+/// The sentence that goes into the chat: what happened, how far the run got,
+/// and the one thing the human can do next.
+pub fn failure_message(failure: &RunFailure, last_step: Option<&str>, turn: Option<i64>) -> String {
+    let progress = match (last_step, turn) {
+        (Some(step), Some(n)) if n > 0 && !step.trim().is_empty() => {
+            format!("我最後在做：{}（第 {n} 輪）。", snippet(step, 80))
+        }
+        (Some(step), _) if !step.trim().is_empty() => {
+            format!("我最後在做：{}。", snippet(step, 80))
+        }
+        _ => "我還沒有開始動手。".to_string(),
+    };
+    format!("{} {} {}", failure.headline, progress, failure.action)
+}
+
+#[derive(Debug, serde::Deserialize)]
+struct ActivityQuery {
+    after: Option<i64>,
+    limit: Option<i32>,
+}
+
+#[derive(Debug, sqlx::FromRow)]
+struct RunRow {
+    status: String,
+    step: Option<String>,
+    turn: Option<i64>,
+    turn_limit: Option<i64>,
+    error: Option<String>,
+    started_at: Option<DateTime<Utc>>,
+    completed_at: Option<DateTime<Utc>>,
+}
+
+async fn actor(state: &AppState) -> Result<Actor, ApiError> {
+    state.bootstrap().await.map_err(|error| {
+        (
+            StatusCode::INTERNAL_SERVER_ERROR,
+            Json(json!({"message": error.to_string()})),
+        )
+    })
+}
+
+/// The bubble's whole payload: where the run stands right now plus its trail.
+/// `after` pages forward; the first read returns the newest window, oldest
+/// first, so the UI renders it and then tails from the last id.
+async fn activity(
+    State(state): State<AppState>,
+    Path(id): Path<String>,
+    Query(query): Query<ActivityQuery>,
+) -> Result<Json<Value>, ApiError> {
+    let actor = actor(&state).await?;
+    let run = sqlx::query_as::<_, RunRow>(
+        "SELECT status, checkpoint->>'step' AS step, (checkpoint->>'turn')::bigint AS turn,
+                (checkpoint->>'turnLimit')::bigint AS turn_limit, error, started_at, completed_at
+         FROM runs WHERE id=$1 AND space_id=$2 AND user_id=$3",
+    )
+    .bind(&id)
+    .bind(&actor.space_id)
+    .bind(&actor.user_id)
+    .fetch_optional(state.pool())
+    .await
+    .map_err(internal)?;
+    let Some(run) = run else {
+        return Err(not_found("run not found"));
+    };
+    let limit = query
+        .limit
+        .unwrap_or(DEFAULT_ACTIVITY_LIMIT)
+        .clamp(1, MAX_ACTIVITY_LIMIT);
+    let rows = sqlx::query_as::<_, (i64, String, Value, DateTime<Utc>)>(
+        "SELECT id, kind, payload, created_at FROM (
+             SELECT id, kind, payload, created_at FROM run_activity
+             WHERE run_id=$1 AND ($2::bigint IS NULL OR id>$2)
+             ORDER BY id DESC LIMIT $3
+         ) recent ORDER BY id ASC",
+    )
+    .bind(&id)
+    .bind(query.after)
+    .bind(limit)
+    .fetch_all(state.pool())
+    .await
+    .map_err(internal)?;
+    let error = run
+        .error
+        .as_deref()
+        .filter(|raw| !raw.trim().is_empty())
+        .map(|raw| {
+            let failure = classify_run_error(raw);
+            json!({"code": failure.code, "headline": failure.headline, "raw": raw})
+        });
+    let elapsed = run.started_at.map(|started| {
+        (run.completed_at.unwrap_or_else(Utc::now) - started)
+            .num_milliseconds()
+            .max(0) as u64
+    });
+    Ok(Json(json!({
+        "runId": id,
+        "status": run.status,
+        "turn": run.turn,
+        "turnLimit": run.turn_limit,
+        "step": run.step,
+        "elapsedMs": elapsed,
+        "error": error,
+        "activity": rows.into_iter().map(|(row_id, kind, payload, created_at)| {
+            let mut entry = payload;
+            if let Some(map) = entry.as_object_mut() {
+                map.insert("id".to_string(), json!(row_id));
+                map.insert("kind".to_string(), json!(kind));
+                map.insert("createdAt".to_string(), json!(created_at.to_rfc3339()));
+            }
+            entry
+        }).collect::<Vec<_>>(),
+    })))
+}
+
+/// Re-queue a run that stopped on an error, keeping its harness checkpoint so
+/// it continues where it stopped instead of starting the task over.
+async fn retry(
+    State(state): State<AppState>,
+    Path(id): Path<String>,
+) -> Result<Json<Value>, ApiError> {
+    let actor = actor(&state).await?;
+    let exists: Option<String> =
+        sqlx::query_scalar("SELECT status FROM runs WHERE id=$1 AND space_id=$2 AND user_id=$3")
+            .bind(&id)
+            .bind(&actor.space_id)
+            .bind(&actor.user_id)
+            .fetch_optional(state.pool())
+            .await
+            .map_err(internal)?;
+    if exists.is_none() {
+        return Err(not_found("run not found"));
+    }
+    let queued: Option<String> = sqlx::query_scalar(
+        "UPDATE runs SET status='queued', retry_count=0, error=NULL, completed_at=NULL,
+                lease_owner=NULL, lease_expires_at=NULL, updated_at=now()
+         WHERE id=$1 AND space_id=$2 AND user_id=$3
+           AND status IN ('failed','cancelled')
+           AND NOT EXISTS (
+               SELECT 1 FROM runs a
+               WHERE a.bot_id=runs.bot_id AND a.id<>runs.id
+                 AND a.status IN ('leased','running','waiting_input','waiting_takeover')
+                 AND (a.lease_expires_at IS NULL OR a.lease_expires_at >= now())
+           )
+         RETURNING thread_id",
+    )
+    .bind(&id)
+    .bind(&actor.space_id)
+    .bind(&actor.user_id)
+    .fetch_optional(state.pool())
+    .await
+    .map_err(internal)?;
+    if queued.is_none() {
+        return Err((
+            StatusCode::CONFLICT,
+            Json(json!({"message": "run is not retryable"})),
+        ));
+    }
+    record(&state, &id, "run", json!({"event": "retry"})).await;
+    tracing::info!(run_id = id.as_str(), "run re-queued from the chat");
+    Ok(Json(json!({"ok": true, "runId": id})))
+}
+
+fn not_found(message: &str) -> ApiError {
+    (StatusCode::NOT_FOUND, Json(json!({"message": message})))
+}
+
+fn internal(error: sqlx::Error) -> ApiError {
+    tracing::error!("monitor: {error}");
+    (
+        StatusCode::INTERNAL_SERVER_ERROR,
+        Json(json!({"message": "internal error"})),
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{RunFailure, clamp_strings, classify_run_error, failure_message, snippet};
+    use serde_json::json;
+
+    fn code(error: &str) -> String {
+        classify_run_error(error).code.to_string()
+    }
+
+    #[test]
+    fn every_failure_offers_an_action_and_a_code() {
+        for error in [
+            "boom",
+            "HTTP 401",
+            "model request timed out after 165 seconds",
+            "no such container",
+        ] {
+            let failure = classify_run_error(error);
+            assert!(!failure.code.is_empty(), "{error}");
+            assert!(!failure.action.is_empty(), "{error}");
+            assert!(!failure.headline.is_empty(), "{error}");
+        }
+    }
+
+    #[test]
+    fn auth_quota_and_unknown_model_are_told_apart() {
+        assert_eq!(
+            code("Api error 401 Unauthorized: invalid api key"),
+            "model_key"
+        );
+        assert_eq!(code("HTTP status 403 Forbidden"), "model_key");
+        assert_eq!(code("status 429 Too Many Requests"), "model_quota");
+        assert_eq!(code("insufficient quota for this project"), "model_quota");
+        assert_eq!(code("The model gpt-9 does not exist"), "model_unknown");
+        assert_eq!(code("HTTP status 404 Not Found"), "model_unknown");
+    }
+
+    #[test]
+    fn the_specific_timeouts_win_over_the_generic_one() {
+        assert_eq!(
+            code("tool browser timed out after 150 seconds"),
+            "tool_timeout"
+        );
+        assert_eq!(
+            code("model request timed out after 165 seconds"),
+            "model_timeout"
+        );
+        assert_eq!(code("AI 回應逾時（150 秒）"), "model_timeout");
+    }
+
+    #[test]
+    fn infrastructure_and_network_failures_are_separate_from_the_model() {
+        assert_eq!(code("no such container: lazyboy-bot-1"), "computer_gone");
+        assert_eq!(code("computer not found"), "computer_gone");
+        assert_eq!(code("error sending request: dns failure"), "network");
+        assert_eq!(code("run lease was lost"), "lease_lost");
+        assert_eq!(
+            code("Worker interrupted after tool execution; inspect current state"),
+            "interrupted"
+        );
+        assert_eq!(code("something odd happened"), "unknown");
+    }
+
+    #[test]
+    fn a_retryable_answer_never_promises_a_setting_that_cannot_help() {
+        assert!(!classify_run_error("401 invalid api key").retryable);
+        assert!(classify_run_error("401 invalid api key").needs_settings);
+        assert!(classify_run_error("no such container").needs_computer);
+        assert!(!classify_run_error("no such container").needs_settings);
+        assert!(classify_run_error("status 429 rate limit").retryable);
+    }
+
+    #[test]
+    fn a_failure_sentence_names_the_last_step_and_turn() {
+        let failure = RunFailure::new("tool_timeout", "卡住了。", "按重試。", true, false, false);
+        let said = failure_message(&failure, Some("browser: click #12"), Some(7));
+        assert!(said.contains("卡住了。"), "{said}");
+        assert!(said.contains("browser: click #12"), "{said}");
+        assert!(said.contains("第 7 輪"), "{said}");
+        assert!(said.contains("按重試。"), "{said}");
+    }
+
+    #[test]
+    fn a_failure_before_any_work_says_so_instead_of_inventing_a_step() {
+        let failure = RunFailure::new("model_key", "金鑰不對。", "改金鑰。", false, true, false);
+        assert!(failure_message(&failure, None, None).contains("我還沒有開始動手。"));
+        assert!(failure_message(&failure, Some("  "), Some(0)).contains("我還沒有開始動手。"));
+        assert!(failure_message(&failure, Some("思考中"), None).contains("思考中"));
+    }
+
+    #[test]
+    fn snippets_stay_single_line_and_bounded() {
+        assert_eq!(snippet("  多行\n文字\t在這裡  ", 40), "多行 文字 在這裡");
+        let long = "字".repeat(500);
+        let cut = snippet(&long, 200);
+        assert_eq!(cut.chars().count(), 201);
+        assert!(cut.ends_with('…'));
+    }
+
+    #[test]
+    fn payloads_are_clamped_at_every_string_leaf() {
+        let clamped = clamp_strings(
+            &json!({"error": "x".repeat(900), "nested": {"raw": "y".repeat(900)}, "n": 7}),
+            50,
+        );
+        assert_eq!(clamped["error"].as_str().unwrap().chars().count(), 51);
+        assert_eq!(
+            clamped["nested"]["raw"].as_str().unwrap().chars().count(),
+            51
+        );
+        assert_eq!(clamped["n"].as_i64(), Some(7));
+    }
+}
diff --git a/crates/api/src/retention.rs b/crates/api/src/retention.rs
index b0a8c06..c12fde0 100644
--- a/crates/api/src/retention.rs
+++ b/crates/api/src/retention.rs
@@ -9,11 +9,13 @@ fn days(name: &str, default: i32) -> i32 {
 }
 
 pub async fn retention_loop(state: AppState) {
+    let recording_days = days("LAZYBOY_RECORDING_RETENTION_DAYS", 30);
     let rules = [
         ("events", include_str!("retention/events.sql"), days("LAZYBOY_EVENT_RETENTION_DAYS", 30)),
         ("checkpoints", include_str!("retention/checkpoints.sql"), days("LAZYBOY_CHECKPOINT_RETENTION_DAYS", 7)),
         ("runs", include_str!("retention/runs.sql"), days("LAZYBOY_RUN_RETENTION_DAYS", 90)),
-        ("recordings", include_str!("retention/recordings.sql"), days("LAZYBOY_RECORDING_RETENTION_DAYS", 30)),
+        ("run_activity", include_str!("retention/run_activity.sql"), days("LAZYBOY_RUN_ACTIVITY_RETENTION_DAYS", 7)),
+        ("recordings", include_str!("retention/recordings.sql"), recording_days),
         ("revisions", include_str!("retention/revisions.sql"), days("LAZYBOY_MEMORY_HISTORY_RETENTION_DAYS", 90)),
         ("deleted_memories", include_str!("retention/deleted_memories.sql"), days("LAZYBOY_MEMORY_HISTORY_RETENTION_DAYS", 90)),
         ("leases", include_str!("retention/leases.sql"), 7),
@@ -32,7 +34,7 @@ pub async fn retention_loop(state: AppState) {
             }
             if removed > 0 { tracing::info!(name, rows=removed, "retention cleaned expired diagnostics"); }
         }
-        if let Err(error) = clean_frames(&state, rules[3].2).await {
+        if let Err(error) = clean_frames(&state, recording_days).await {
             tracing::warn!(%error, "recording file retention failed; retry next hour");
         }
         if let Ok(bytes) = sqlx::query_scalar::<_, i64>("SELECT pg_database_size(current_database())").fetch_one(state.pool()).await {
diff --git a/crates/api/src/retention/run_activity.sql b/crates/api/src/retention/run_activity.sql
new file mode 100644
index 0000000..85066d8
--- /dev/null
+++ b/crates/api/src/retention/run_activity.sql
@@ -0,0 +1,4 @@
+DELETE FROM run_activity WHERE id IN (
+ SELECT id FROM run_activity WHERE created_at < now() - make_interval(days => $1)
+ ORDER BY created_at LIMIT $2 FOR UPDATE SKIP LOCKED
+)
diff --git a/crates/api/src/rooms.rs b/crates/api/src/rooms.rs
index 2214e5c..149f99a 100644
--- a/crates/api/src/rooms.rs
+++ b/crates/api/src/rooms.rs
@@ -64,13 +64,21 @@ async fn members_for(state: &AppState, room_id: &str) -> Result<Vec<RoomMember>,
     Ok(rows.into_iter().map(member_from_row).collect())
 }
 
+/// `room_from_id` projection: id, name, last message time, preview, unread count.
+type RoomSummaryRow = (
+    String,
+    String,
+    Option<chrono::DateTime<chrono::Utc>>,
+    Option<String>,
+    i64,
+);
+
 async fn room_from_id(
     state: &AppState,
     actor: &Actor,
     id: &str,
 ) -> Result<Option<Room>, sqlx::Error> {
-    let row: Option<(String, String, Option<chrono::DateTime<chrono::Utc>>, Option<String>, i64)> =
-        sqlx::query_as(
+    let row: Option<RoomSummaryRow> = sqlx::query_as(
             "SELECT r.id, r.name,
                     (SELECT MAX(m.created_at) FROM messages m JOIN threads t ON t.id=m.thread_id WHERE t.room_id=r.id),
                     (SELECT m.body FROM messages m JOIN threads t ON t.id=m.thread_id
diff --git a/crates/api/src/routes.rs b/crates/api/src/routes.rs
index 031cbfe..e450b69 100644
--- a/crates/api/src/routes.rs
+++ b/crates/api/src/routes.rs
@@ -15,6 +15,7 @@ pub fn router(state: AppState) -> Router {
     Router::new()
         .merge(crate::sessions::router())
         .merge(crate::memory::router())
+        .merge(crate::monitor::router())
         .merge(crate::rooms::router())
         .merge(crate::mcp::router())
         .merge(crate::workspace::router())
@@ -347,17 +348,16 @@ async fn delete_bot(
     if let Some(computer) = computer
         .as_ref()
         .filter(|row| parse_mode(&row.scope) == ComputerMode::Dedicated)
+        && let Some(computer_ref) = computer::computer_ref(computer)
     {
-        if let Some(computer_ref) = computer::computer_ref(computer) {
-            state
-                .sandbox
-                .destroy(
-                    &computer_ref,
-                    &computer::adapter_context(&actor, &id, "delete-bot"),
-                )
-                .await
-                .map_err(|error| bad_gateway(error.to_string()))?;
-        }
+        state
+            .sandbox
+            .destroy(
+                &computer_ref,
+                &computer::adapter_context(&actor, &id, "delete-bot"),
+            )
+            .await
+            .map_err(|error| bad_gateway(error.to_string()))?;
     }
 
     let mut tx = state.pool().begin().await.map_err(internal_error)?;
diff --git a/crates/api/src/runs.rs b/crates/api/src/runs.rs
index 1574b74..22ba4b7 100644
--- a/crates/api/src/runs.rs
+++ b/crates/api/src/runs.rs
@@ -236,7 +236,7 @@ pub async fn send(
         .bind(&member_run)
         .bind(&actor.space_id)
         .bind(member_id)
-        .bind(&thread_id)
+        .bind(thread_id)
         .bind(&actor.user_id)
         .bind(&stored_body)
         .bind(json!({"messageSeq":seq}))
@@ -296,19 +296,43 @@ pub async fn send(
     }))
 }
 
+/// `worker_loop` claim projection: run id, bot id, thread id, prompt, user id, space id.
+type RetryCandidateRow = (String, String, String, String, String, String);
+
 pub async fn worker_loop(state: AppState) {
     let inflight = Arc::new(tokio::sync::Semaphore::new(16));
     let lease_owner = format!("api-{}", Uuid::new_v4());
     loop {
         tokio::time::sleep(Duration::from_millis(200)).await;
-        let _ = sqlx::query("UPDATE runs SET status='failed',error='Worker interrupted after tool execution; inspect current state before continuing.',completed_at=now(),lease_owner=NULL,lease_expires_at=NULL WHERE status IN ('leased','running') AND lease_expires_at<now() AND COALESCE((checkpoint->>'toolsStarted')::boolean,false)")
-            .execute(state.pool()).await;
+        let interrupted: Vec<(String, String, String)> = sqlx::query_as(
+            "WITH doomed AS (
+                 UPDATE runs SET status='failed',error=$1,completed_at=now(),
+                     lease_owner=NULL, lease_expires_at=NULL
+                 WHERE status IN ('leased','running') AND lease_expires_at<now()
+                   AND COALESCE((checkpoint->>'toolsStarted')::boolean,false)
+                 RETURNING id, thread_id, bot_id
+             )
+             SELECT id, thread_id, bot_id FROM doomed LIMIT 50",
+        )
+        .bind(INTERRUPTED_ERROR)
+        .fetch_all(state.pool())
+        .await
+        .unwrap_or_default();
+        for (orphan_run, orphan_thread, orphan_bot) in interrupted {
+            report_run_failure(
+                &state,
+                &orphan_run,
+                &orphan_thread,
+                &orphan_bot,
+                INTERRUPTED_ERROR,
+            )
+            .await;
+        }
         let Ok(permit) = inflight.clone().try_acquire_owned() else {
             continue;
         };
-        let queued: Result<Option<(String, String, String, String, String, String)>, _> =
-            sqlx::query_as(
-                "WITH candidate AS (
+        let queued: Result<Option<RetryCandidateRow>, _> = sqlx::query_as(
+            "WITH candidate AS (
                SELECT r.id
                FROM runs r
                WHERE r.retry_count < r.max_retries
@@ -335,10 +359,10 @@ pub async fn worker_loop(state: AppState) {
                  lease_fence=lease_fence+1, retry_count=retry_count+1, updated_at=now()
              FROM candidate c WHERE r.id=c.id
              RETURNING r.id,r.bot_id,r.thread_id,r.prompt,r.user_id,r.space_id",
-            )
-            .bind(&lease_owner)
-            .fetch_optional(state.pool())
-            .await;
+        )
+        .bind(&lease_owner)
+        .fetch_optional(state.pool())
+        .await;
         let Ok(Some((run_id, bot_id, thread_id, prompt, user_id, space_id))) = queued else {
             drop(permit);
             continue;
@@ -371,14 +395,7 @@ pub async fn worker_loop(state: AppState) {
                     .ok()
                     .flatten();
                 if next_status.as_deref() == Some("failed") {
-                    let _ = append_bot_message(
-                        &state,
-                        &thread_id,
-                        &run_id,
-                        &bot_id,
-                        &format!("Run failed after retries: {error}"),
-                    )
-                    .await;
+                    report_run_failure(&state, &run_id, &thread_id, &bot_id, &error).await;
                     let _ = crate::sessions::append_event(
                         &state,
                         &thread_id,
@@ -432,6 +449,13 @@ async fn execute_run(
     }
     let _ = crate::sessions::append_event(state, thread_id, "run.started", json!({"runId":run_id}))
         .await;
+    crate::monitor::record(
+        state,
+        run_id,
+        "run",
+        json!({"event": "started", "task": crate::monitor::snippet(prompt, 160)}),
+    )
+        .await;
 
     let bot = state
         .db
@@ -656,6 +680,13 @@ async fn execute_run(
         // model call with no chance of accidentally invoking any capability.
         defs.clear();
         tracing::info!(run_id, "chat-only turn: all tools withheld");
+        crate::monitor::record(
+            state,
+            run_id,
+            "notice",
+            json!({"text": "這一輪是對白，工具先收起來（不會動到電腦）。"}),
+        )
+        .await;
     }
     // Taught skills run long (a 24-page course is 24 clicks); plain chats stay
     // bounded tighter so a confused model cannot burn budget for as long.
@@ -668,6 +699,10 @@ async fn execute_run(
     } else {
         ExecutionMode::Bounded(40)
     };
+    let turn_limit = match execution_mode {
+        ExecutionMode::Goal => None,
+        ExecutionMode::Bounded(limit) => Some(limit as i64),
+    };
     let mut nudges: u32 = 0;
     let mut screenshots: u32 = 0;
     let mut screenshot_bytes: u64 = 0;
@@ -685,22 +720,21 @@ async fn execute_run(
         }
     }
     let mut pending = Message::User { content: first };
-    if !resume_after_takeover {
-        if let (Some(saved_history), Some(saved_pending)) = (
+    if !resume_after_takeover
+        && let (Some(saved_history), Some(saved_pending)) = (
             checkpoint.get("harnessHistory"),
             checkpoint.get("harnessPending"),
-        ) {
-            if let (Ok(restored), Ok(mut next)) = (
-                serde_json::from_value::<Vec<Message>>(saved_history.clone()),
-                serde_json::from_value::<Message>(saved_pending.clone()),
-            ) {
-                history = restored;
-                if let Message::User { content } = &mut next {
-                    content.push(UserContent::text("Resumed after a completed tool batch. Do not repeat completed actions. Observe current browser/desktop before any new mutation; prior element references may be stale."));
-                }
-                pending = next;
-            }
+        )
+        && let (Ok(restored), Ok(mut next)) = (
+            serde_json::from_value::<Vec<Message>>(saved_history.clone()),
+            serde_json::from_value::<Message>(saved_pending.clone()),
+        )
+    {
+        history = restored;
+        if let Message::User { content } = &mut next {
+            content.push(UserContent::text("Resumed after a completed tool batch. Do not repeat completed actions. Observe current browser/desktop before any new mutation; prior element references may be stale."));
         }
+        pending = next;
     }
 
     let mut final_text = String::new();
@@ -727,6 +761,13 @@ async fn execute_run(
         {
             Ok(items) => state.memory.durable_block(&items),
             Err(error) => {
+                crate::monitor::record(
+                    state,
+                    run_id,
+                    "notice",
+                    json!({"text": format!("記憶讀取失敗，這輪不用記憶：{error}")}),
+                )
+                .await;
                 tracing::warn!("memory retrieval failed for run {run_id}: {error}");
                 String::new()
             }
@@ -791,17 +832,17 @@ async fn execute_run(
             preamble.push_str("\n\n");
             preamble.push_str(&file_skill_index);
         }
-        if let Ok(accounts) = crate::vault::list_on(state.pool(), actor, bot_id).await {
-            if !accounts.is_empty() {
-                let names = accounts
-                    .iter()
-                    .map(|item| format!("{} ({})", item.site, item.username))
-                    .collect::<Vec<_>>()
-                    .join(", ");
-                preamble.push_str(&format!(
+        if let Ok(accounts) = crate::vault::list_on(state.pool(), actor, bot_id).await
+            && !accounts.is_empty()
+        {
+            let names = accounts
+                .iter()
+                .map(|item| format!("{} ({})", item.site, item.username))
+                .collect::<Vec<_>>()
+                .join(", ");
+            preamble.push_str(&format!(
                     "\n\nSaved logins (passwords are not shown): {names}. At a login wall call use_saved_login with the matching accountId from list_accounts."
                 ));
-            }
         }
     }
 
@@ -837,6 +878,16 @@ async fn execute_run(
                             guidance.join("\n")
                         )
                     };
+                    crate::monitor::record(
+                        state,
+                        run_id,
+                        "notice",
+                        json!({
+                            "turn": turns,
+                            "text": format!("收到新指示：{}", crate::monitor::snippet(&guidance.join(" / "), 160)),
+                        }),
+                    )
+                    .await;
                     if let Message::User { content } = &mut pending {
                         content.push(UserContent::text(text));
                     }
@@ -858,7 +909,7 @@ async fn execute_run(
             .await;
         }
         drop_history_screenshots(&mut history, &pending);
-        set_run_step(state, run_id, MODEL_STEP).await;
+        set_run_progress(state, run_id, MODEL_STEP, turns, turn_limit).await;
         let model_started = std::time::Instant::now();
         let content = tokio::select! {
             halt = wait_for_halt(state, run_id) => {
@@ -875,7 +926,14 @@ async fn execute_run(
                 )
                 .await;
             }
-            result = complete_with_retry(&model, pending.clone(), &preamble, &history, &defs) => result
+            result = complete_with_retry(
+                &model,
+                pending.clone(),
+                &preamble,
+                &history,
+                &defs,
+                Trace { state, run_id, turn: turns },
+            ) => result
         }?;
         let assistant = Message::Assistant {
             id: None,
@@ -885,20 +943,37 @@ async fn execute_run(
         history.push(assistant);
 
         let mut calls = Vec::new();
+        let mut turn_text = String::new();
         for item in &content {
             match item {
-                AssistantContent::Text(text) => final_text.push_str(&text.text),
+                AssistantContent::Text(text) => {
+                    final_text.push_str(&text.text);
+                    turn_text.push_str(&text.text);
+                }
                 AssistantContent::ToolCall(call) => calls.push(call.clone()),
                 _ => {}
             }
         }
+        let model_elapsed = model_started.elapsed().as_millis() as u64;
         tracing::info!(
             run_id,
             turn = turns,
-            elapsed_ms = model_started.elapsed().as_millis() as u64,
+            elapsed_ms = model_elapsed,
             tool_calls = calls.len(),
             "model turn"
         );
+        crate::monitor::record(
+            state,
+            run_id,
+            "model",
+            json!({
+                "turn": turns,
+                "elapsedMs": model_elapsed,
+                "toolCalls": calls.len(),
+                "text": crate::monitor::snippet(&turn_text, 200),
+            }),
+        )
+        .await;
         if calls.is_empty() {
             // A model that quits a playbook early, or parrots an earlier reply
             // instead of describing the current screen, gets pushed back to
@@ -946,6 +1021,22 @@ async fn execute_run(
                     verify = verify_chosen,
                     "nudging model back to tools"
                 );
+                crate::monitor::record(
+                    state,
+                    run_id,
+                    "notice",
+                    json!({
+                        "turn": turns,
+                        "text": if verify_chosen {
+                            format!("要求模型先證明做完了才准結束（第 {nudges} 次攔下）。")
+                        } else if parroted {
+                            format!("模型重複了舊的回答，叫它看現在畫面繼續做（第 {nudges} 次）。")
+                        } else {
+                            format!("模型想用一句話收尾，叫它用工具繼續（第 {nudges} 次）。")
+                        },
+                    }),
+                )
+                .await;
                 earlier_replies.push(final_text.trim().to_string());
                 final_text.clear();
                 let mut content = vec![UserContent::text(text)];
@@ -1047,13 +1138,14 @@ async fn execute_run(
             );
             did_work = true;
             let step = describe_step(&name, &call.function.arguments);
-            set_run_step(state, run_id, &step).await;
+            set_run_progress(state, run_id, &step, turns, turn_limit).await;
             let fence=sqlx::query("UPDATE runs SET checkpoint=COALESCE(checkpoint,'{}'::jsonb)||jsonb_build_object('toolsStarted',true) WHERE id=$1 AND lease_owner=$2 AND status='running'")
                 .bind(run_id).bind(lease_owner).execute(state.pool()).await.map_err(|e| e.to_string())?;
             if fence.rows_affected() != 1 {
                 return Err("run lease lost before tool dispatch".into());
             }
             let tool_started = std::time::Instant::now();
+            let mut tool_timed_out = false;
             let outcome = tokio::select! {
                 halt = wait_for_halt(state, run_id) => {
                     return finish_halt(
@@ -1076,12 +1168,15 @@ async fn execute_run(
                     Ok(outcome) => outcome,
                     // A slow tool is a normal turn, not a dead run: say the
                     // effects are unknown and let the model re-observe.
-                    Err(_) => ToolOutcome {
+                    Err(_) => {
+                        tool_timed_out = true;
+                        ToolOutcome {
                         text: format!("tool {name} timed out after 150 seconds. Its effects are unknown: observe the current screen or files before anything else, and never repeat a step that already worked."),
                         image: None,
                         pause: false,
                         blocks: Vec::new(),
-                    },
+                        }
+                    }
                 }
             };
             tracing::info!(
@@ -1094,6 +1189,20 @@ async fn execute_run(
                 pause = outcome.pause,
                 "tool call"
             );
+            crate::monitor::record(
+                state,
+                run_id,
+                "tool",
+                json!({
+                    "turn": turns,
+                    "name": name.clone(),
+                    "step": step.clone(),
+                    "status": tool_status(tool_timed_out, outcome.pause, &outcome.text),
+                    "elapsedMs": tool_started.elapsed().as_millis() as u64,
+                    "snippet": crate::monitor::snippet(&outcome.text, 200),
+                }),
+            )
+            .await;
             // xAI rejects images inside tool results. Attach a changed
             // screenshot as a following user image instead.
             if let Some(image) = outcome.image {
@@ -1161,7 +1270,16 @@ async fn execute_run(
             results.extend(screenshot_parts(png));
         }
         pending = Message::User { content: results };
-        save_harness_checkpoint(state, run_id, lease_owner, &history, &pending, turns, steering_seq).await?;
+        save_harness_checkpoint(
+            state,
+            run_id,
+            lease_owner,
+            &history,
+            &pending,
+            turns,
+            steering_seq,
+        )
+        .await?;
     }
 
     let status: Option<String> = sqlx::query_scalar("SELECT status FROM runs WHERE id = $1")
@@ -1214,8 +1332,21 @@ async fn execute_run(
     }
     let needs_input = goal_mode && goal_outcome(&final_text) == GoalOutcome::NeedsInput;
     if needs_input {
-        let next = Message::User { content: vec![UserContent::text("The goal was paused for required user input. Read the user's new information and continue from completed work.")] };
-        save_harness_checkpoint(state, run_id, lease_owner, &history, &next, turns, steering_seq).await?;
+        let next = Message::User {
+            content: vec![UserContent::text(
+                "The goal was paused for required user input. Read the user's new information and continue from completed work.",
+            )],
+        };
+        save_harness_checkpoint(
+            state,
+            run_id,
+            lease_owner,
+            &history,
+            &next,
+            turns,
+            steering_seq,
+        )
+        .await?;
     }
     let final_text = final_text
         .replace("[GOAL_COMPLETE]", "")
@@ -1238,6 +1369,16 @@ async fn execute_run(
     if completed.rows_affected() != 1 {
         return Err("run lease was lost before completion".into());
     }
+    crate::monitor::record(
+        state,
+        run_id,
+        "run",
+        json!({
+            "event": if needs_input { "waiting_input" } else { "completed" },
+            "turns": turns,
+        }),
+    )
+    .await;
     let click_misses = *ctx.click_misses.lock().unwrap();
     let takeover = *ctx.takeover_requested.lock().unwrap();
     record_run_metrics(
@@ -1331,12 +1472,21 @@ fn retryable_run_error(error: &str) -> bool {
     !matches!(code, 400..=499 if !matches!(code, 408 | 409 | 425 | 429))
 }
 
+/// Where a helper that is not the run loop itself writes its diagnostics.
+#[derive(Clone, Copy)]
+struct Trace<'a> {
+    state: &'a AppState,
+    run_id: &'a str,
+    turn: u32,
+}
+
 async fn complete_with_retry(
     model: &DynModel,
     pending: Message,
     preamble: &str,
     history: &[Message],
     defs: &[ToolDefinition],
+    trace: Trace<'_>,
 ) -> Result<Vec<AssistantContent>, String> {
     let mut last = String::new();
     for attempt in 0..3 {
@@ -1349,6 +1499,18 @@ async fn complete_with_retry(
         match result {
             Ok(Ok(content)) => return Ok(content),
             Ok(Err(error)) => {
+                crate::monitor::record(
+                    trace.state,
+                    trace.run_id,
+                    "retry",
+                    json!({
+                        "turn": trace.turn,
+                        "attempt": attempt + 1,
+                        "error": error,
+                        "gaveUp": !retryable_run_error(&error),
+                    }),
+                )
+                .await;
                 if !retryable_run_error(&error) {
                     return Err(error);
                 }
@@ -1360,8 +1522,19 @@ async fn complete_with_retry(
                 last = error;
             }
             Err(_) => {
-                tracing::warn!(attempt = attempt + 1, "model attempt exceeded its 165 second budget");
-                last = "model request timed out after 165 seconds".into();
+                let error = "model request timed out after 165 seconds";
+                crate::monitor::record(
+                    trace.state,
+                    trace.run_id,
+                    "retry",
+                    json!({"turn": trace.turn, "attempt": attempt + 1, "error": error}),
+                )
+                .await;
+                tracing::warn!(
+                    attempt = attempt + 1,
+                    "model attempt exceeded its 165 second budget"
+                );
+                last = error.into();
             }
         }
         if attempt < 2 {
@@ -1702,6 +1875,13 @@ async fn finish_halt(
             .bind(run_id)
             .execute(state.pool())
             .await;
+            crate::monitor::record(
+                state,
+                run_id,
+                "run",
+                json!({"event": "paused", "reason": "takeover", "turns": turns}),
+            )
+            .await;
             let click_misses = *ctx.click_misses.lock().unwrap();
             record_run_metrics(
                 state,
@@ -1792,6 +1972,18 @@ async fn pause_for_answer(
     if paused.rows_affected() != 1 {
         return Err("run lease was lost before pausing".into());
     }
+    crate::monitor::record(
+        state,
+        run_id,
+        "run",
+        json!({
+            "event": "paused",
+            "reason": req.reason.as_str(),
+            "turns": req.turns,
+            "limit": req.limit,
+        }),
+    )
+    .await;
     append_bot_message_with(
         state,
         thread_id,
@@ -2369,7 +2561,7 @@ fn is_greeting(normalized: &str) -> bool {
         "🙂",
         "😀",
     ];
-    if EXACT.iter().any(|item| normalized == *item) {
+    if EXACT.contains(&normalized) {
         return true;
     }
     const PREFIXES: &[&str] = &["hi ", "hey ", "hello ", "嗨", "你好"];
@@ -2641,6 +2833,113 @@ fn describe_step(name: &str, args: &Value) -> String {
     }
 }
 
+/// The error the claim sweep writes when a worker died holding a tool open.
+const INTERRUPTED_ERROR: &str =
+    "Worker interrupted after tool execution; inspect current state before continuing.";
+
+/// Step plus turn counter: what the bubble header reads out of the checkpoint.
+pub(crate) async fn set_run_progress(
+    state: &AppState,
+    run_id: &str,
+    step: &str,
+    turn: u32,
+    limit: Option<i64>,
+) {
+    let result = sqlx::query(
+        "UPDATE runs SET checkpoint = COALESCE(checkpoint, '{}'::jsonb)
+                || jsonb_build_object('step', $2::text, 'stepAt', now(),
+                                      'turn', $3::bigint, 'turnLimit', $4::bigint),
+                updated_at = now()
+         WHERE id = $1",
+    )
+    .bind(run_id)
+    .bind(step)
+    .bind(turn as i64)
+    .bind(limit)
+    .execute(state.pool())
+    .await;
+    if let Err(error) = result {
+        tracing::warn!(run_id, "failed to record run progress: {error}");
+    }
+}
+
+/// Tell the human a run died, in their language, with the one next action. The
+/// raw error stays available through the run activity endpoint.
+async fn report_run_failure(
+    state: &AppState,
+    run_id: &str,
+    thread_id: &str,
+    bot_id: &str,
+    error: &str,
+) {
+    let progress: Option<(Option<String>, Option<i64>)> = sqlx::query_as(
+        "SELECT checkpoint->>'step', (checkpoint->>'turn')::bigint FROM runs WHERE id=$1",
+    )
+    .bind(run_id)
+    .fetch_optional(state.pool())
+    .await
+    .unwrap_or(None);
+    let (last_step, turn) = progress.unwrap_or((None, None));
+    let failure = crate::monitor::classify_run_error(error);
+    let body = crate::monitor::failure_message(&failure, last_step.as_deref(), turn);
+    let _ = append_bot_message_with(
+        state,
+        thread_id,
+        run_id,
+        bot_id,
+        &body,
+        json!([{
+            "kind": "error",
+            "code": failure.code,
+            "retryable": failure.retryable,
+            "runId": run_id,
+            "turn": turn,
+            "step": last_step,
+        }]),
+    )
+    .await;
+    crate::monitor::record(
+        state,
+        run_id,
+        "run",
+        json!({"event": "failed", "error": error}),
+    )
+    .await;
+}
+
+/// Display-only reading of a tool result. Tools report failure in prose, so the
+/// trail marks the shapes it recognises rather than pretending to know more.
+fn tool_status(timed_out: bool, pause: bool, text: &str) -> &'static str {
+    if timed_out {
+        return "timed_out";
+    }
+    if pause {
+        return "paused";
+    }
+    const FAILURES: [&str; 14] = [
+        "error",
+        "failed",
+        "failure",
+        "unable to",
+        "cannot ",
+        "can't ",
+        "timed out",
+        "denied",
+        "no such",
+        "exception",
+        "not available",
+        "失敗",
+        "錯誤",
+        "無法",
+    ];
+    let head: String = text.to_lowercase().chars().take(160).collect();
+    if FAILURES.iter().any(|needle| head.contains(needle)) {
+        "error"
+    } else {
+        "ok"
+    }
+}
+
 pub(crate) async fn set_run_step(state: &AppState, run_id: &str, step: &str) {
     let result = sqlx::query(
         "UPDATE runs SET checkpoint = COALESCE(checkpoint, '{}'::jsonb)
@@ -2696,11 +2995,23 @@ mod tests {
     use super::{
         RunHalt, SCREENSHOT_CAPTION, describe_step, drop_history_screenshots, halt_from_status,
         history_window_start, is_plain_chat, retryable_run_error, screenshot_parts, tool_needs_gui,
-        tool_needs_sandbox,
+        tool_needs_sandbox, tool_status,
     };
     use rig_core::completion::message::{Message, UserContent};
     use serde_json::json;
 
+    #[test]
+    fn the_trail_reads_a_tool_result_as_ok_error_timeout_or_pause() {
+        assert_eq!(tool_status(false, false, "clicked element #12"), "ok");
+        assert_eq!(
+            tool_status(false, false, "Error: selector not found"),
+            "error"
+        );
+        assert_eq!(tool_status(false, false, "操作失敗：視窗關閉"), "error");
+        assert_eq!(tool_status(true, false, "tool timed out"), "timed_out");
+        assert_eq!(tool_status(false, true, "需要人先登入"), "paused");
+    }
+
     #[test]
     fn chat_tools_do_not_need_the_desktop() {
         assert!(!tool_needs_sandbox("remember"));
diff --git a/crates/api/src/schedules.rs b/crates/api/src/schedules.rs
index abd6732..84045e8 100644
--- a/crates/api/src/schedules.rs
+++ b/crates/api/src/schedules.rs
@@ -617,7 +617,7 @@ pub fn describe_cron(expr: &str) -> String {
     if let Ok(Some(seconds)) = interval_seconds(expr) {
         return format!("每隔 {} 分鐘（固定間隔）", seconds / 60);
     }
-    let parts: Vec<&str> = expr.trim().split_whitespace().collect();
+    let parts: Vec<&str> = expr.split_whitespace().collect();
     if parts.len() != 5 {
         return expr.to_string();
     }
@@ -625,36 +625,40 @@ pub fn describe_cron(expr: &str) -> String {
     if expr.trim() == "* * * * *" {
         return "每分鐘".into();
     }
-    if let Some(rest) = min.strip_prefix("*/") {
-        if hour == "*" && dom == "*" && month == "*" && dow == "*" {
-            return if rest
-                .parse::<u32>()
-                .ok()
-                .is_some_and(|n| n > 0 && 60 % n == 0)
-            {
-                format!("每 {rest} 分鐘")
-            } else {
-                format!("日曆排程：{expr}")
-            };
-        }
+    if let Some(rest) = min.strip_prefix("*/")
+        && hour == "*"
+        && dom == "*"
+        && month == "*"
+        && dow == "*"
+    {
+        return if rest
+            .parse::<u32>()
+            .ok()
+            .is_some_and(|n| n > 0 && 60 % n == 0)
+        {
+            format!("每 {rest} 分鐘")
+        } else {
+            format!("日曆排程：{expr}")
+        };
     }
     if min == "0" && hour == "*" && dom == "*" && month == "*" && dow == "*" {
         return "每小時".into();
     }
-    if min == "0" {
-        if let Some(rest) = hour.strip_prefix("*/") {
-            if dom == "*" && month == "*" && dow == "*" {
-                return if rest
-                    .parse::<u32>()
-                    .ok()
-                    .is_some_and(|n| n > 0 && 24 % n == 0)
-                {
-                    format!("每 {rest} 小時")
-                } else {
-                    format!("日曆排程：{expr}")
-                };
-            }
-        }
+    if min == "0"
+        && let Some(rest) = hour.strip_prefix("*/")
+        && dom == "*"
+        && month == "*"
+        && dow == "*"
+    {
+        return if rest
+            .parse::<u32>()
+            .ok()
+            .is_some_and(|n| n > 0 && 24 % n == 0)
+        {
+            format!("每 {rest} 小時")
+        } else {
+            format!("日曆排程：{expr}")
+        };
     }
     if min.parse::<u32>().is_ok() && hour.parse::<u32>().is_ok() && month == "*" {
         let at = format!("{hour:0>2}:{min:0>2}");
diff --git a/crates/api/src/screen_proxy.rs b/crates/api/src/screen_proxy.rs
index 8c7201b..fd09359 100644
--- a/crates/api/src/screen_proxy.rs
+++ b/crates/api/src/screen_proxy.rs
@@ -241,10 +241,10 @@ async fn proxy_socket(mut client: WebSocket, host: String, port: u16, rest: Stri
                         if client.send(AxumMessage::Text(text.to_string().into())).await.is_err() { break; }
                     }
                     Some(Ok(WsMessage::Ping(data))) => {
-                        if client.send(AxumMessage::Ping(data.into())).await.is_err() { break; }
+                        if client.send(AxumMessage::Ping(data)).await.is_err() { break; }
                     }
                     Some(Ok(WsMessage::Pong(data))) => {
-                        if client.send(AxumMessage::Pong(data.into())).await.is_err() { break; }
+                        if client.send(AxumMessage::Pong(data)).await.is_err() { break; }
                     }
                     Some(Ok(WsMessage::Close(_))) | Some(Ok(WsMessage::Frame(_))) | None => break,
                     Some(Err(_)) => break,
diff --git a/crates/api/src/sessions.rs b/crates/api/src/sessions.rs
index b496b09..b13d2b3 100644
--- a/crates/api/src/sessions.rs
+++ b/crates/api/src/sessions.rs
@@ -220,9 +220,7 @@ async fn delete_session(
             Json(json!({"message":"session not found"})),
         ));
     }
-    cancel_session_runs(&state, &id)
-        .await
-        .map_err(|error| internal(error))?;
+    cancel_session_runs(&state, &id).await.map_err(internal)?;
     let mut tx = state
         .pool()
         .begin()
@@ -286,9 +284,7 @@ async fn clear_messages(
             Json(json!({"message":"session not found"})),
         ));
     }
-    cancel_session_runs(&state, &id)
-        .await
-        .map_err(|error| internal(error))?;
+    cancel_session_runs(&state, &id).await.map_err(internal)?;
     let mut tx = state
         .pool()
         .begin()
@@ -364,9 +360,7 @@ async fn stop_session(
             Json(json!({"message":"session not found"})),
         ));
     }
-    cancel_session_runs(&state, &id)
-        .await
-        .map_err(|error| internal(error))?;
+    cancel_session_runs(&state, &id).await.map_err(internal)?;
     Ok(Json(json!({"ok":true})))
 }
 
@@ -447,26 +441,29 @@ pub async fn default_session_for_bot(
     .await
 }
 
+/// `messages_for_session` projection: message columns joined with the speaking bot.
+type MessageWithSpeakerRow = (
+    String,
+    String,
+    i32,
+    String,
+    String,
+    Value,
+    Option<String>,
+    Option<String>,
+    chrono::DateTime<chrono::Utc>,
+    Option<String>,
+    Option<String>,
+    Option<String>,
+    Option<String>,
+);
+
 pub async fn messages_for_session(
     state: &AppState,
     actor: &Actor,
     id: &str,
 ) -> Result<Vec<SessionMessage>, ApiError> {
-    let rows: Vec<(
-        String,
-        String,
-        i32,
-        String,
-        String,
-        Value,
-        Option<String>,
-        Option<String>,
-        chrono::DateTime<chrono::Utc>,
-        Option<String>,
-        Option<String>,
-        Option<String>,
-        Option<String>,
-    )> = sqlx::query_as(
+    let rows: Vec<MessageWithSpeakerRow> = sqlx::query_as(
         "SELECT m.id, m.thread_id, m.seq, m.role, m.body, m.blocks, m.run_id,
                 m.client_nonce, m.created_at, m.speaker_bot_id, b.name, b.avatar_color, b.avatar_shape
          FROM messages m
diff --git a/crates/api/src/skills.rs b/crates/api/src/skills.rs
index 93c7715..5bbe53a 100644
--- a/crates/api/src/skills.rs
+++ b/crates/api/src/skills.rs
@@ -1394,7 +1394,11 @@ pub fn fallback_playbook(goal: &str, events: &[Value]) -> Value {
         })
         .filter_map(|event| describe_event(event, t0))
         .map(|line| {
-            let text = line.splitn(2, ' ').nth(1).unwrap_or(&line).to_string();
+            let text = line
+                .split_once(' ')
+                .map(|x| x.1)
+                .unwrap_or(&line)
+                .to_string();
             json!({ "do": text, "expect": "", "note": "" })
         })
         .take(40)
diff --git a/crates/api/src/tools.rs b/crates/api/src/tools.rs
index 2ba6d8c..4f6e481 100644
--- a/crates/api/src/tools.rs
+++ b/crates/api/src/tools.rs
@@ -728,20 +728,20 @@ async fn browser(ctx: &ToolCtx, args: &Value) -> ToolOutcome {
             request[key] = value.clone();
         }
     }
-    if request.get("selector").and_then(Value::as_str).is_none() {
-        if let Some(id) = element_id(args.get("element").or_else(|| args.get("id"))) {
-            let elements = ctx.elements.lock().unwrap().clone();
-            match elements
-                .iter()
-                .find(|element| u64::from(element.id) == id)
-                .and_then(|element| element.selector.clone())
-            {
-                Some(selector) => request["selector"] = json!(selector),
-                None if matches!(action, "click" | "type") => {
-                    return pause_unknown_element(ctx, id as u32, &elements);
-                }
-                None => {}
+    if request.get("selector").and_then(Value::as_str).is_none()
+        && let Some(id) = element_id(args.get("element").or_else(|| args.get("id")))
+    {
+        let elements = ctx.elements.lock().unwrap().clone();
+        match elements
+            .iter()
+            .find(|element| u64::from(element.id) == id)
+            .and_then(|element| element.selector.clone())
+        {
+            Some(selector) => request["selector"] = json!(selector),
+            None if matches!(action, "click" | "type") => {
+                return pause_unknown_element(ctx, id as u32, &elements);
             }
+            None => {}
         }
     }
     if matches!(action, "click") && request.get("selector").and_then(Value::as_str).is_none() {
diff --git a/crates/api/src/vault.rs b/crates/api/src/vault.rs
index c0dd5ac..a482139 100644
--- a/crates/api/src/vault.rs
+++ b/crates/api/src/vault.rs
@@ -4,7 +4,7 @@ use aes_gcm::aead::{Aead, KeyInit};
 use aes_gcm::{Aes256Gcm, Nonce};
 use axum::extract::{Path, State};
 use axum::http::StatusCode;
-use axum::routing::{delete, get, patch, post};
+use axum::routing::{get, patch};
 use axum::{Json, Router};
 use chrono::{DateTime, Utc};
 use rand::RngCore;
@@ -161,14 +161,18 @@ pub async fn list_on(
     .map_err(|error| error.to_string())
 }
 
-pub async fn get_secret(
-    state: &AppState,
-    actor: &Actor,
-    bot_id: &str,
-    account_id: &str,
-) -> Result<Option<(VaultAccount, String, String)>, String> {
-    get_secret_on(state.pool(), actor, bot_id, account_id).await
-}
+/// `get_secret_on` projection: vault columns plus the encrypted password.
+type SecretRow = (
+    String,
+    String,
+    String,
+    String,
+    String,
+    String,
+    DateTime<Utc>,
+    DateTime<Utc>,
+    String,
+);
 
 pub async fn get_secret_on(
     pool: &sqlx::PgPool,
@@ -176,8 +180,7 @@ pub async fn get_secret_on(
     bot_id: &str,
     account_id: &str,
 ) -> Result<Option<(VaultAccount, String, String)>, String> {
-    let row: Option<(String, String, String, String, String, String, DateTime<Utc>, DateTime<Utc>, String)> =
-        sqlx::query_as(
+    let row: Option<SecretRow> = sqlx::query_as(
             "SELECT id, bot_id, site, host, username, notes, created_at, updated_at, password_ciphertext
              FROM vault_accounts
              WHERE id=$1 AND bot_id=$2 AND space_id=$3 AND user_id=$4",
@@ -368,6 +371,9 @@ mod tests {
     use super::{decrypt, encrypt, normalize_host};
 
     #[test]
+    // Rust 2024 marks `set_var` unsafe; this only seeds a test-only key and no
+    // other test in this binary reads it.
+    #[allow(unsafe_code)]
     fn round_trips_a_password() {
         unsafe { std::env::set_var("LAZYBOY_VAULT_KEY", "test-vault-key-for-unit-tests") };
         let packed = encrypt("s3cret!").unwrap();
diff --git a/crates/contracts/Cargo.toml b/crates/contracts/Cargo.toml
index acaef4c..33e5e27 100644
--- a/crates/contracts/Cargo.toml
+++ b/crates/contracts/Cargo.toml
@@ -2,6 +2,7 @@
 name = "lazyboy-contracts"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
@@ -11,3 +12,6 @@ serde_json.workspace = true
 thiserror.workspace = true
 chrono.workspace = true
 base64.workspace = true
+
+[lints]
+workspace = true
diff --git a/crates/contracts/src/computer.rs b/crates/contracts/src/computer.rs
index 03297e3..81bb9bf 100644
--- a/crates/contracts/src/computer.rs
+++ b/crates/contracts/src/computer.rs
@@ -113,8 +113,10 @@ pub fn computer_home_key(
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize)]
 #[serde(rename_all = "kebab-case")]
+#[derive(Default)]
 pub enum BrowserProfileMode {
     Shared,
+    #[default]
     PerBot,
     PerTask,
 }
@@ -129,12 +131,6 @@ impl BrowserProfileMode {
     }
 }
 
-impl Default for BrowserProfileMode {
-    fn default() -> Self {
-        Self::PerBot
-    }
-}
-
 impl std::str::FromStr for BrowserProfileMode {
     type Err = UnknownProfileMode;
 
diff --git a/crates/control/Cargo.toml b/crates/control/Cargo.toml
index 75c6318..68bd0d3 100644
--- a/crates/control/Cargo.toml
+++ b/crates/control/Cargo.toml
@@ -2,6 +2,7 @@
 name = "lazyboy-control"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
@@ -15,3 +16,6 @@ hex.workspace = true
 chrono.workspace = true
 async-trait = "0.1"
 image.workspace = true
+
+[lints]
+workspace = true
diff --git a/crates/control/src/a11y.rs b/crates/control/src/a11y.rs
index ebcdc26..e13c5ac 100644
--- a/crates/control/src/a11y.rs
+++ b/crates/control/src/a11y.rs
@@ -130,7 +130,6 @@ mod tests {
             kind: Some("a11y".into()),
             selector: Some(format!("0/{id}")),
             role: Some("push button".into()),
-            ..UiElement::default()
         }
     }
 
diff --git a/crates/control/src/actions.rs b/crates/control/src/actions.rs
index 08508e5..e5d42b3 100644
--- a/crates/control/src/actions.rs
+++ b/crates/control/src/actions.rs
@@ -200,22 +200,22 @@ pub fn parse_computer_actions(value: &Value) -> Result<Vec<ComputerAction>, Acti
             .unwrap_or_default();
         match kind {
             "click" | "move" | "down" | "up" => {
-                if kind == "click" {
-                    if let Some(target) = ref_target(action) {
-                        let pointer = ComputerAction::Ref {
-                            verb: RefVerb::Click,
-                            target,
-                            ref_kind: ref_kind(action),
-                            text: None,
-                        };
-                        let doubled = action.get("double").and_then(Value::as_bool) == Some(true);
-                        actions.push(pointer.clone());
-                        if doubled {
-                            actions.push(ComputerAction::Wait { ms: 70 });
-                            actions.push(pointer);
-                        }
-                        continue;
+                if kind == "click"
+                    && let Some(target) = ref_target(action)
+                {
+                    let pointer = ComputerAction::Ref {
+                        verb: RefVerb::Click,
+                        target,
+                        ref_kind: ref_kind(action),
+                        text: None,
+                    };
+                    let doubled = action.get("double").and_then(Value::as_bool) == Some(true);
+                    actions.push(pointer.clone());
+                    if doubled {
+                        actions.push(ComputerAction::Wait { ms: 70 });
+                        actions.push(pointer);
                     }
+                    continue;
                 }
                 let x = coordinate(action.get("x"), "x")?;
                 let y = coordinate(action.get("y"), "y")?;
@@ -558,7 +558,6 @@ mod tests {
             y: 20,
             w: 80,
             h: 24,
-            ..UiElement::default()
         }
     }
 
diff --git a/crates/control/src/overlay.rs b/crates/control/src/overlay.rs
index 481f720..c4b5b72 100644
--- a/crates/control/src/overlay.rs
+++ b/crates/control/src/overlay.rs
@@ -150,7 +150,7 @@ fn draw_badge(img: &mut RgbImage, x: u32, y: u32, id: u32) {
         2,
     );
     for (i, ch) in label.bytes().enumerate() {
-        if !(b'0'..=b'9').contains(&ch) {
+        if !ch.is_ascii_digit() {
             continue;
         }
         let dx = bx + pad + i as u32 * digit_w;
diff --git a/crates/control/src/sandbox.rs b/crates/control/src/sandbox.rs
index 8c85613..64aee79 100644
--- a/crates/control/src/sandbox.rs
+++ b/crates/control/src/sandbox.rs
@@ -2,7 +2,7 @@ use lazyboy_contracts::{ComputerAction, ComputerCapabilities, ComputerObservatio
 use serde::{Deserialize, Serialize};
 use thiserror::Error;
 
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
 pub struct AdapterContext {
     pub operation_id: String,
     pub space_id: String,
@@ -20,23 +20,6 @@ pub struct AdapterContext {
     pub profile_path: Option<String>,
 }
 
-impl Default for AdapterContext {
-    fn default() -> Self {
-        Self {
-            operation_id: String::new(),
-            space_id: String::new(),
-            user_id: String::new(),
-            bot_id: None,
-            run_id: None,
-            screen_lease_id: None,
-            screen_id: None,
-            screen_slot: None,
-            display: None,
-            profile_path: None,
-        }
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct ComputerRef {
     pub id: String,
@@ -53,7 +36,7 @@ pub struct ProvisionRequest {
     pub provider_ref: Option<String>,
 }
 
-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize, Default)]
 pub struct CommandRequest {
     pub argv: Vec<String>,
     pub cwd: Option<String>,
@@ -64,17 +47,6 @@ pub struct CommandRequest {
     pub stdin: Option<String>,
 }
 
-impl Default for CommandRequest {
-    fn default() -> Self {
-        Self {
-            argv: Vec::new(),
-            cwd: None,
-            timeout_ms: None,
-            stdin: None,
-        }
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct CommandResult {
     pub stdout: String,
diff --git a/crates/control/src/x11.rs b/crates/control/src/x11.rs
index d588e9b..65d6980 100644
--- a/crates/control/src/x11.rs
+++ b/crates/control/src/x11.rs
@@ -386,6 +386,17 @@ pub fn screenshot_command_on(display: &str) -> Vec<String> {
     ]
 }
 
+/// Native clipboard path; text is supplied on stdin, never process arguments.
+pub fn paste_command_on(display: &str) -> Vec<String> {
+    vec![
+        "env".into(),
+        display_env(display),
+        "python3".into(),
+        "-c".into(),
+        include_str!("clipboard.py").into(),
+    ]
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -490,14 +501,3 @@ mod tests {
         assert!(parse_ui_elements("not-json").is_empty());
     }
 }
-
-/// Native clipboard path; text is supplied on stdin, never process arguments.
-pub fn paste_command_on(display: &str) -> Vec<String> {
-    vec![
-        "env".into(),
-        display_env(display),
-        "python3".into(),
-        "-c".into(),
-        include_str!("clipboard.py").into(),
-    ]
-}
diff --git a/crates/controld/Cargo.toml b/crates/controld/Cargo.toml
index f5420f7..06ddbe8 100644
--- a/crates/controld/Cargo.toml
+++ b/crates/controld/Cargo.toml
@@ -2,6 +2,7 @@
 name = "lazyboy-controld"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
@@ -10,9 +11,10 @@ lazyboy-contracts.workspace = true
 lazyboy-control.workspace = true
 axum.workspace = true
 tokio.workspace = true
-serde.workspace = true
 serde_json.workspace = true
-thiserror.workspace = true
 tracing.workspace = true
 tracing-subscriber.workspace = true
 base64.workspace = true
+
+[lints]
+workspace = true
diff --git a/crates/harness/Cargo.toml b/crates/harness/Cargo.toml
index d591737..479d371 100644
--- a/crates/harness/Cargo.toml
+++ b/crates/harness/Cargo.toml
@@ -2,12 +2,12 @@
 name = "lazyboy-harness"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
 [dependencies]
 lazyboy-contracts.workspace = true
-serde.workspace = true
 serde_json.workspace = true
 thiserror.workspace = true
 rig-core.workspace = true
@@ -21,3 +21,6 @@ http = "1"
 reqwest.workspace = true
 rustls.workspace = true
 rustls-native-certs = "0.8"
+
+[lints]
+workspace = true
diff --git a/crates/sandbox/Cargo.toml b/crates/sandbox/Cargo.toml
index 9705648..f7496c9 100644
--- a/crates/sandbox/Cargo.toml
+++ b/crates/sandbox/Cargo.toml
@@ -2,6 +2,7 @@
 name = "lazyboy-sandbox"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
@@ -10,10 +11,8 @@ lazyboy-contracts.workspace = true
 lazyboy-control.workspace = true
 async-trait = "0.1"
 reqwest.workspace = true
-serde.workspace = true
 serde_json.workspace = true
-tokio.workspace = true
 base64.workspace = true
-chrono.workspace = true
-sha2.workspace = true
-hex.workspace = true
+
+[lints]
+workspace = true
diff --git a/crates/sandbox/src/docker.rs b/crates/sandbox/src/docker.rs
index 1ef6f5d..773c62f 100644
--- a/crates/sandbox/src/docker.rs
+++ b/crates/sandbox/src/docker.rs
@@ -36,20 +36,20 @@ impl DockerSandbox {
         if let Some(bot_id) = &context.bot_id {
             headers.insert("x-lazyboy-bot-id", bot_id.parse().unwrap());
         }
-        if let Some(display) = &context.display {
-            if let Ok(value) = display.parse() {
-                headers.insert("x-lazyboy-display", value);
-            }
+        if let Some(display) = &context.display
+            && let Ok(value) = display.parse()
+        {
+            headers.insert("x-lazyboy-display", value);
         }
-        if let Some(profile) = &context.profile_path {
-            if let Ok(value) = profile.parse() {
-                headers.insert("x-lazyboy-profile", value);
-            }
+        if let Some(profile) = &context.profile_path
+            && let Ok(value) = profile.parse()
+        {
+            headers.insert("x-lazyboy-profile", value);
         }
-        if let Some(slot) = context.screen_slot {
-            if let Ok(value) = slot.to_string().parse() {
-                headers.insert("x-lazyboy-screen-slot", value);
-            }
+        if let Some(slot) = context.screen_slot
+            && let Ok(value) = slot.to_string().parse()
+        {
+            headers.insert("x-lazyboy-screen-slot", value);
         }
         headers
     }
diff --git a/crates/sandbox/src/fake.rs b/crates/sandbox/src/fake.rs
index cea3356..627df9c 100644
--- a/crates/sandbox/src/fake.rs
+++ b/crates/sandbox/src/fake.rs
@@ -55,22 +55,22 @@ impl SandboxProvider for FakeSandbox {
         request: CommandRequest,
         _context: &AdapterContext,
     ) -> Result<CommandResult, SandboxError> {
-        if request.argv.get(0).map(String::as_str) == Some("mkdir") {
+        if request.argv.first().map(String::as_str) == Some("mkdir") {
             return Ok(CommandResult {
                 stdout: String::new(),
                 stderr: String::new(),
                 code: 0,
             });
         }
-        if request.argv.get(0).map(String::as_str) == Some("touch") {
-            if let Some(path) = request.argv.get(1) {
-                self.files
-                    .lock()
-                    .unwrap()
-                    .entry(computer.home_key.clone())
-                    .or_default()
-                    .insert(path.clone(), Vec::new());
-            }
+        if request.argv.first().map(String::as_str) == Some("touch")
+            && let Some(path) = request.argv.get(1)
+        {
+            self.files
+                .lock()
+                .unwrap()
+                .entry(computer.home_key.clone())
+                .or_default()
+                .insert(path.clone(), Vec::new());
         }
         Ok(CommandResult {
             stdout: request.argv.join(" "),
diff --git a/crates/supervisor/Cargo.toml b/crates/supervisor/Cargo.toml
index 5ed107f..1f33109 100644
--- a/crates/supervisor/Cargo.toml
+++ b/crates/supervisor/Cargo.toml
@@ -2,26 +2,26 @@
 name = "lazyboy-supervisor"
 version.workspace = true
 edition.workspace = true
+rust-version.workspace = true
 license.workspace = true
 publish.workspace = true
 
 [dependencies]
-lazyboy-contracts.workspace = true
 lazyboy-control.workspace = true
 axum.workspace = true
 tokio.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-thiserror.workspace = true
 tracing.workspace = true
 tracing-subscriber.workspace = true
 bollard.workspace = true
-uuid.workspace = true
 base64.workspace = true
 reqwest.workspace = true
-async-trait = "0.1"
 futures-util = "0.3"
 
 hmac.workspace = true
 sha2.workspace = true
 hex.workspace = true
+
+[lints]
+workspace = true
diff --git a/crates/supervisor/src/docker.rs b/crates/supervisor/src/docker.rs
index 5563bdf..692d647 100644
--- a/crates/supervisor/src/docker.rs
+++ b/crates/supervisor/src/docker.rs
@@ -34,20 +34,20 @@ pub struct Provisioned {
 }
 
 pub struct ObservePayload {
-    pub png: Vec<u8>,
     pub json: serde_json::Value,
 }
 
 impl ObservePayload {
     fn from_json(value: serde_json::Value) -> Result<Self, String> {
-        let encoded = value
+        // The screenshot travels inside `json` as base64; only its presence is
+        // checked here so a broken capture fails fast instead of returning an
+        // empty observation.
+        value
             .get("png_base64")
             .and_then(serde_json::Value::as_str)
+            .filter(|encoded| !encoded.is_empty())
             .ok_or_else(|| "missing png".to_string())?;
-        let png = base64::engine::general_purpose::STANDARD
-            .decode(encoded)
-            .map_err(|error| error.to_string())?;
-        Ok(Self { png, json: value })
+        Ok(Self { json: value })
     }
 }
 
@@ -76,7 +76,7 @@ impl DockerHost {
             return Err("invalid home key".into());
         }
         let expected = PathBuf::from(&data_dir).join("homes").join(home_key);
-        if PathBuf::from(home_path) != expected {
+        if *home_path != expected {
             return Err("home path is outside managed homes".into());
         }
         // Work on the container-local path, not the host daemon's bind path.
@@ -251,42 +251,6 @@ impl DockerHost {
         })
     }
 
-    pub async fn exec(&self, id: &str, request: CommandRequest) -> Result<CommandResult, String> {
-        let cwd = match request.cwd {
-            Some(cwd) if PathBuf::from(&cwd).is_absolute() => cwd,
-            Some(cwd) => {
-                let relative = normalize_workspace_path(&cwd).map_err(|error| error.to_string())?;
-                if relative.is_empty() {
-                    HOME.to_string()
-                } else {
-                    format!("{HOME}/{relative}")
-                }
-            }
-            None => HOME.to_string(),
-        };
-        let timeout_ms = request.timeout_ms.unwrap_or(30_000).clamp(100, 120_000);
-        let argv = if request.argv.is_empty() {
-            vec!["/bin/echo".into(), "ready".into()]
-        } else {
-            request.argv
-        };
-        let mut bounded = vec![
-            "timeout".into(),
-            "--signal=TERM".into(),
-            "--kill-after=2s".into(),
-            format!("{}s", timeout_ms as f64 / 1000.0),
-        ];
-        bounded.extend(argv);
-        self.exec_raw_cmd(
-            id,
-            &bounded,
-            Some(&cwd),
-            &ScreenTarget::default(),
-            request.stdin,
-        )
-        .await
-    }
-
     pub async fn exec_on(
         &self,
         id: &str,
@@ -322,13 +286,6 @@ impl DockerHost {
             .await
     }
 
-    pub async fn observe(&self, id: &str) -> Result<Vec<u8>, String> {
-        Ok(self
-            .observe_payload(id, &ScreenTarget::default())
-            .await?
-            .png)
-    }
-
     pub async fn observe_payload(
         &self,
         id: &str,
@@ -352,23 +309,22 @@ impl DockerHost {
             let mut body = serde_json::json!({
                 "png_base64": base64::engine::general_purpose::STANDARD.encode(&stdout)
             });
-            if let Ok(meta) = self.pointer_state(id, target).await {
-                if let serde_json::Value::Object(map) = meta {
-                    if let Some(obj) = body.as_object_mut() {
-                        if let (Some(x), Some(y)) = (map.get("x"), map.get("y")) {
-                            obj.insert("cursor".into(), serde_json::json!({ "x": x, "y": y }));
-                        }
-                        if map
-                            .get("id")
-                            .and_then(serde_json::Value::as_str)
-                            .is_some_and(|id| !id.is_empty())
-                        {
-                            obj.insert(
-                                "activeWindow".into(),
-                                serde_json::json!({ "id": map.get("id"), "title": map.get("title") }),
-                            );
-                        }
-                    }
+            if let Ok(meta) = self.pointer_state(id, target).await
+                && let serde_json::Value::Object(map) = meta
+                && let Some(obj) = body.as_object_mut()
+            {
+                if let (Some(x), Some(y)) = (map.get("x"), map.get("y")) {
+                    obj.insert("cursor".into(), serde_json::json!({ "x": x, "y": y }));
+                }
+                if map
+                    .get("id")
+                    .and_then(serde_json::Value::as_str)
+                    .is_some_and(|id| !id.is_empty())
+                {
+                    obj.insert(
+                        "activeWindow".into(),
+                        serde_json::json!({ "id": map.get("id"), "title": map.get("title") }),
+                    );
                 }
             }
             body
diff --git a/crates/supervisor/src/main.rs b/crates/supervisor/src/main.rs
index 82f3bf1..a23f344 100644
--- a/crates/supervisor/src/main.rs
+++ b/crates/supervisor/src/main.rs
@@ -387,10 +387,9 @@ async fn managed_boundary(
             .path()
             .strip_prefix("/computers/")
             .and_then(|p| p.split('/').next())
+            && app.docker.container_control_token(id).await.is_err()
         {
-            if app.docker.container_control_token(id).await.is_err() {
-                return StatusCode::NOT_FOUND.into_response();
-            }
+            return StatusCode::NOT_FOUND.into_response();
         }
     }
     next.run(req).await
diff --git a/deny.toml b/deny.toml
new file mode 100644
index 0000000..67b3c26
--- /dev/null
+++ b/deny.toml
@@ -0,0 +1,58 @@
+# Supply-chain policy for the LazyBoy workspace, checked by `make audit`.
+# cargo-deny is not part of the Rust toolchain:
+#   cargo install --locked cargo-deny    (or grab a release binary)
+#
+# Three things are enforced: known vulnerabilities (RustSec), the license set we
+# are willing to ship, and where dependencies may come from.
+
+[graph]
+targets = [
+    { triple = "x86_64-unknown-linux-gnu" },
+    { triple = "aarch64-unknown-linux-gnu" },
+]
+
+[advisories]
+yanked = "deny"
+ignore = [
+    # `paste` is an eager-macro helper pulled in by fastembed -> tokenizers. It is
+    # archived upstream (no safe upgrade), only used while compiling macros, and we
+    # keep embeddings optional. Re-check on the next fastembed/tokenizers bump.
+    { id = "RUSTSEC-2024-0436", reason = "transitive proc-macro helper behind fastembed; unmaintained, no upgrade available" },
+]
+
+[licenses]
+# Permissive set only; copyleft candidates (MPL-2.0) stay file-level.
+confidence-threshold = 0.93
+allow = [
+    "0BSD",
+    "Apache-2.0",
+    "Apache-2.0 WITH LLVM-exception",
+    "BSD-2-Clause",
+    "BSD-3-Clause",
+    "BSL-1.0",
+    "CC0-1.0",
+    "CDLA-Permissive-2.0",
+    "ISC",
+    "MIT",
+    "MIT-0",
+    "MPL-2.0",
+    "Unicode-3.0",
+    "Unlicense",
+    "Zlib",
+]
+
+[bans]
+# Duplicate major/minor versions are normal in an application graph (they are
+# compiled once each); we only want to know when a dependency disappears.
+multiple-versions = "allow"
+wildcards = "deny"
+# Workspace members are path dependencies and carry no version requirement.
+allow-workspace = true
+allow-wildcard-paths = true
+highlight = "all"
+
+[sources]
+unknown-registry = "deny"
+unknown-git = "deny"
+allow-registry = ["https://github.com/rust-lang/crates.io-index"]
+allow-git = []
diff --git a/docker-compose.dev.yml b/docker-compose.dev.yml
index d274c57..fc8f019 100644
--- a/docker-compose.dev.yml
+++ b/docker-compose.dev.yml
@@ -1,5 +1,14 @@
 # Opt-in host access for local Rust development only.
+# The project `database` network is internal, and Docker silently ignores published
+# ports for containers attached only to internal networks. Postgres therefore also
+# joins this host bridge, so 127.0.0.1:5434 really listens for `cargo test`.
 services:
   postgres:
     ports:
       - "127.0.0.1:5434:5432"
+    networks:
+      - database
+      - dev_host
+networks:
+  dev_host:
+    name: lazyboy_dev_host
diff --git a/docs/development.md b/docs/development.md
index 3853fad..8264e60 100644
--- a/docs/development.md
+++ b/docs/development.md
@@ -21,9 +21,12 @@ npm run dev           # http://127.0.0.1:5173
 ### 檢查與測試
 
 ```bash
-make fmt
-make clippy
-make test
+# 有資料庫的 Rust 測試需要 127.0.0.1:5434，先用 dev overlay 啟動 Postgres
+make postgres
+
+make lint               # clippy（-D warnings），政策見下方
+make test               # Rust 測試（需要 127.0.0.1:5434）
+make audit              # 依賴漏洞、授權、來源掃描（需要 cargo-deny）
 
 cd apps/web
 npm run typecheck
@@ -34,11 +37,46 @@ node --test tests/frontend.test.mjs
 python3 tests/control.test.py
 python3 tests/log-rotation.test.py
 
-# 需要已啟動的 PostgreSQL Compose service
-docker compose up -d postgres
+# Python 整合測試用 docker compose exec 連進 Postgres，自己建一次性資料庫後清掉
 python3 tests/retention.test.py
+python3 tests/run-resume.test.py
+python3 tests/run-activity.test.py
 ```
 
+`make postgres` 才會發布 `127.0.0.1:5434`（`docker-compose.dev.yml`）；整套堆疊已經用
+基礎設定跑著時，請改用 `export COMPOSE_FILE=docker-compose.yml:docker-compose.dev.yml`
+再 `docker compose up -d`，否則後續的 `docker compose` 指令會拿基礎設定重建 Postgres、
+把發布埠拿掉。
+
+### Rust 品質檢查
+
+工具鏈以穩定版為準：workspace 宣告 `rust-version = "1.98"`（MSRV，代表 let-chains
+等語法下限），容器用 `rust:1-*` 映像會自動跟最新小版。本機更新只需：
+
+```bash
+rustup update stable && rustc --version
+```
+
+Lint 政策集中在三處，新增 crate 時只要補 `[lints] workspace = true` 就會繼承：
+
+- 根 `Cargo.toml` 的 `[workspace.lints]`：不安全程式碼預設警告、未使用的
+  `Result`/`Future` 直接拒絕，並用 `unused_crate_dependencies` 抓多餘依賴
+  （不需要另外裝 `cargo-machete`）。
+- `clippy.toml`：閾值類設定。`too-many-arguments-threshold = 10` 是因為 handler
+  與 run/vault/voice 協助函式本來就要帶 state + actor + 多個 id，超過 10 個參數才會警告。
+  `cargo-clippy` 只看得到啟動目錄下的 `clippy.toml`，請一律在 repo 根目錄跑 `make lint`。
+- 各 crate 的 `[lints] workspace = true`。
+
+要放寬一條 lint 時，請在**呼叫點**加 `#[allow(...)]` 並附一句理由（例：
+`crates/api/src/vault.rs` 測試裡的 `#[allow(unsafe_code)]`），不要在工作區層級關掉規則。
+
+`make fmt` 會重排整個 workspace；目前倉庫仍有歷史格式偏差，`make fmt-check` 會列出來，
+等到一次性重整後再併入 CI。CI 目前只需 `make lint` + `make test`。
+
+供應鏈檢查用 `cargo-deny`（`make audit`，設定見 `deny.toml`）：RustSec 漏洞、授權白名單、
+依賴來源。它不是內建工具，第一次要先 `cargo install --locked cargo-deny`。
+已知無法升級的項目會寫進 `deny.toml` 的 `ignore` 並附原因與重檢時機。
+
 ### 專案結構
 
 ```text
@@ -59,6 +97,8 @@ LazyBoy/
 ├── docs/hero.html            README 首頁視覺原稿
 ├── docs/workflow.html        可互動產品流程圖
 ├── docker-compose.yml        正式堆疊
+├── clippy.toml               Clippy 閾值（要在 repo 根目錄執行才讀得到）
+├── deny.toml                 cargo-deny 供應鏈政策（make audit）
 └── Makefile                  常用開發與部署指令
 ```
 
diff --git a/docs/operations.md b/docs/operations.md
index 0a621fe..3207b36 100644
--- a/docs/operations.md
+++ b/docs/operations.md
@@ -77,6 +77,37 @@ docker compose up -d api
 - 暫時性跨網存取建議維持 `127.0.0.1` 綁定改用隧道：`ssh -L 3101:127.0.0.1:3101 <host>`。
 - 桌面 noVNC 走 `LAZYBOY_SCREEN_NETWORK` 這條 internal 網路，不佔主機埠；同機跑多組 LazyBoy 時請為每組取不同名稱，compose 與 supervisor 會共用同一個值。
 
+## 執行記錄與錯誤診斷
+
+任務執行中的每一步會寫進 `run_activity`：模型每一輪在想什麼、哪個動作成功或
+失敗、花了幾秒、現在是第幾輪。這屬於診斷資料，不是對話內容，因此可以被清理。
+
+**看即時記錄。** 在對話框把滑鼠移到「思考中」的頭像上（或點一下釘住、用 `Tab`
+聚焦），會浮出即時記錄面板：
+
+- 標題列顯示「第 {n}/{limit} 輪」、已運行時間與現在在做什麼；goal 模式沒有輪次
+  上限時只顯示目前輪次，不會假裝有上限。
+- 記錄由舊到新排列並自動捲到底，包含模型回合、動作成敗與逾時、重試第幾次，以及
+  停下來等你的原因，最後停在哪裡一目了然。
+- 「複製記錄」可把整份純文字記錄貼給別人除錯；錯誤原文以等寬字型原樣顯示。
+- 面板只在開啟時每 1.5 秒增量抓取 `GET /api/runs/{id}/activity`，run 結束
+  （完成／失敗／取消）後停止輪詢，不開就完全不打 API。
+
+**出錯時說什麼。** 任務失敗會在對話框出現錯誤卡，而不是一行紅字：
+
+- 一句人話講清楚原因與下一步，例如「模型不認得這個 API 金鑰。到「設定 →
+  模型」重新貼一次金鑰，再按重試。」
+- 按鈕依錯誤類型給：`重試`（從中斷的地方接著做，保留 checkpoint，不重複已成功的
+  步驟）、`開模型設定`、`打開它的畫面`。金鑰錯誤不會只給重試，電腦消失也不會叫
+  你去看設定。
+- `ⓘ` 直接展開同一個即時記錄面板看詳細 log；錯誤代碼與原文都在裡面。
+- 重試走 `POST /api/runs/{id}/retry`，只對失敗或已取消的 run 生效；同一台電腦還有
+  別的工作在跑時會被擋下，並明確告訴你是因為還有別的工作。
+
+**保留多久。** `LAZYBOY_RUN_ACTIVITY_RETENTION_DAYS` 控制記錄保留天數，預設 7 天，
+每小時清理一次。想留更久的除錯軌跡就調大，在意資料庫體積就調小；run 本身被
+`LAZYBOY_RUN_RETENTION_DAYS` 清掉時，它的記錄一併消失。
+
 ## 網站連線驗證
 
 AI 遇到可辨識的 Cloudflare 連線驗證頁時，會先取得最新桌面截圖，再使用
@@ -95,3 +126,31 @@ AI 遇到可辨識的 Cloudflare 連線驗證頁時，會先取得最新桌面
 這會重建桌面容器，讓 `lazyboy` 使用者能執行免密碼 `sudo`；可用
 `sudo -n id -u` 檢查，預期輸出 `0`。既有容器只暫停／恢復不會套用此設定。
 重新啟動前請儲存工作；家目錄保留，容器系統層自行安裝的套件需重新安裝。
+
+## Postgres collation 版本不符
+
+`pgvector/pgvector:pg16` 是浮動 tag。重新 `docker compose pull` 之後，容器內的 glibc
+版本可能和 `pgdata` 卷建立時不同，Postgres 會在使用每個資料庫時抱怨：
+
+```text
+WARNING:  database "template1" has a collation version mismatch
+DETAIL:   The database was created using collation version 2.41, but the operating system provides version 2.36.
+```
+
+此時 `CREATE DATABASE` 直接被拒（`ERROR: template database "template1" has a collation
+version mismatch`），凡是需要建立資料庫的測試都會失敗：`cargo test` 裡的 `#[sqlx::test]`
+表現在連線逾時（`PoolTimedOut`），`tests/retention.test.py`、`tests/run-resume.test.py`
+則在 `CREATE DATABASE` 就掛掉。
+
+原地修好，資料不遺失（會重建索引，資料量大時安排在離峰）：
+
+```bash
+docker compose stop api
+make pg-collation
+docker compose up -d api
+```
+
+它對 `template1`、`postgres`、`lazyboy` 各跑一次 `REINDEX DATABASE` 與
+`ALTER DATABASE ... REFRESH COLLATION VERSION`。collation provider 是 libc，排序規則
+真的可能跟著 glibc 變，所以先重建索引再更新記錄的版號，不要只改版號。想避免重複發生，
+把 Postgres 映像改成固定 digest，讓同一個資料卷不會被不同 glibc 開起來。
diff --git a/migrations/015_run_activity.sql b/migrations/015_run_activity.sql
new file mode 100644
index 0000000..5bfb132
--- /dev/null
+++ b/migrations/015_run_activity.sql
@@ -0,0 +1,12 @@
+-- Per-run activity trail for the chat's live monitor bubble. Diagnostics only:
+-- never user-authored content, and cheap enough to write once per model turn.
+CREATE TABLE IF NOT EXISTS run_activity (
+    id BIGSERIAL PRIMARY KEY,
+    run_id TEXT NOT NULL REFERENCES runs (id) ON DELETE CASCADE,
+    kind TEXT NOT NULL,
+    payload JSONB NOT NULL DEFAULT '{}'::jsonb,
+    created_at TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+
+CREATE INDEX IF NOT EXISTS run_activity_run_idx ON run_activity (run_id, id);
+CREATE INDEX IF NOT EXISTS run_activity_retention_idx ON run_activity (created_at);
diff --git a/tests/frontend.test.mjs b/tests/frontend.test.mjs
index 9211f6b..81d2029 100644
--- a/tests/frontend.test.mjs
+++ b/tests/frontend.test.mjs
@@ -241,3 +241,90 @@ test('view-only pointer controls never send mouse input and trackpad asks for co
  assert.equal(v.pointerEvents.length,0);
  assert.equal(v.sent.at(-1).type,'lazyboy-request-control');
 });
+
+const monitorJs=ts.transpileModule(fs.readFileSync('apps/web/src/run-monitor.tsx','utf8'),{compilerOptions:{module:ts.ModuleKind.CommonJS,jsx:ts.JsxEmit.ReactJSX}}).outputText;
+const monitorBox={exports:{},require:name=>{
+ if(name==='react')return{useCallback:fn=>fn,useEffect:()=>{},useRef:()=>({current:null}),useState:()=>[null,()=>{}]};
+ if(name==='react/jsx-runtime')return{jsx:()=>null,jsxs:()=>null,Fragment:'Fragment'};
+ if(name==='./api')return{api:async()=>({activity:[]})};
+ if(name==='./i18n')return{t:i18n.t,getLocale:()=>i18n.locale};
+ throw new Error('unexpected import '+name);
+}};
+vm.runInNewContext(monitorJs,monitorBox);
+const {formatElapsed,shortDuration,errorActions,errorTitle,trailText}=monitorBox.exports;
+const FAILURE_CODES=['interrupted','tool_timeout','model_key','model_quota','model_unknown','model_timeout','network','computer_gone','lease_lost','unknown'];
+
+test('run timings stay on one glanceable line',()=>{
+ assert.equal(formatElapsed(0),'0:00');
+ assert.equal(formatElapsed(9_500),'0:09');
+ assert.equal(formatElapsed(65_000),'1:05');
+ assert.equal(formatElapsed(3_725_000),'1:02:05');
+ assert.equal(formatElapsed(-5),'0:00');
+ assert.equal(formatElapsed(NaN),'0:00');
+ assert.equal(shortDuration(340),'340ms');
+ assert.equal(shortDuration(6_400),'6.4s');
+ assert.equal(shortDuration(65_000),'1:05');
+ assert.equal(shortDuration(null),'');
+ assert.equal(shortDuration(-1),'');
+});
+
+test('every failure code offers at least one action, in the right order',()=>{
+ const buttons=code=>errorActions(code).join(',');
+ assert.equal(buttons('model_key'),'settings,retry');
+ assert.equal(buttons('model_unknown'),'settings,retry');
+ assert.equal(buttons('model_quota'),'retry,settings');
+ assert.equal(buttons('model_timeout'),'retry,settings');
+ assert.equal(buttons('network'),'retry,settings');
+ assert.equal(buttons('computer_gone'),'screen,retry');
+ assert.equal(buttons('tool_timeout'),'screen,retry');
+ assert.equal(buttons('interrupted'),'screen,retry');
+ assert.equal(buttons('lease_lost'),'retry');
+ for(const code of [...FAILURE_CODES,'made_up',undefined,null]){
+  const actions=errorActions(code);
+  assert.ok(actions.length>0,String(code));
+  assert.ok(actions.every(action=>['retry','screen','settings'].includes(action)),code);
+  assert.equal(new Set(actions).size,actions.length,'no duplicate buttons for '+code);
+ }
+});
+
+test('failure titles resolve in both catalogs and never repeat themselves',()=>{
+ for(const locale of Object.keys(catalogs)){
+  i18n.locale=locale;
+  const titles=FAILURE_CODES.map(code=>errorTitle(code));
+  for(const [index,title] of titles.entries()){
+   assert.notEqual(title,title.startsWith('errorTitle')?title:`${title}-missing`,FAILURE_CODES[index]);
+   assert.ok(!title.startsWith('errorTitle'),`${locale} ${FAILURE_CODES[index]} has no copy`);
+   assert.ok(title.trim().length>0);
+  }
+  assert.equal(new Set(titles).size,titles.length,locale);
+  assert.ok(errorTitle('made_up').length>0);
+ }
+ i18n.locale='en';
+});
+
+test('trail lines read as sentences with the detail a stuck run needs',()=>{
+ assert.equal(trailText({id:1,kind:'model',createdAt:'',turn:3,elapsedMs:6400,text:'先打開網頁'}),'turn 3, thought for 6.4s — 先打開網頁');
+ assert.match(trailText({id:2,kind:'tool',createdAt:'',step:'browser: click #12',status:'ok',elapsedMs:400}),/browser: click #12 · ok · 400ms/);
+ assert.match(trailText({id:3,kind:'tool',createdAt:'',step:'shell: npm test',status:'timed_out',snippet:'killed after 150s'}),/timed out — killed after 150s/);
+ assert.match(trailText({id:4,kind:'run',createdAt:'',event:'started',task:'整理下載資料'}),/整理下載資料/);
+ assert.match(trailText({id:5,kind:'run',createdAt:'',event:'completed',turns:9}),/Done in 9 turns/);
+ assert.match(trailText({id:6,kind:'run',createdAt:'',event:'waiting_input',reason:'登入'}),/Waiting for you: 登入/);
+ assert.match(trailText({id:7,kind:'run',createdAt:'',event:'retry'}),/Re-queued/);
+ assert.match(trailText({id:8,kind:'retry',createdAt:'',attempt:2,gaveUp:true,error:'429 rate limit'}),/attempt 2 failed, retrying · gave up — 429 rate limit/);
+ assert.match(trailText({id:9,kind:'notice',createdAt:'',text:'這輪不需要電腦'}),/這輪不需要電腦/);
+});
+
+test('the bubble reads the run activity endpoint the API actually mounts',()=>{
+ const monitor=fs.readFileSync('crates/api/src/monitor.rs','utf8');
+ assert.match(monitor,/\.route\("\/api\/runs\/\{id\}\/activity", get\(activity\)\)/);
+ assert.match(monitor,/\.route\("\/api\/runs\/\{id\}\/retry", post\(retry\)\)/);
+ const app=fs.readFileSync('apps/web/src/App.tsx','utf8');
+ assert.match(app,/chip\.kind==="error"\?/);
+ assert.match(app,/errorActions\(chip\.code\)/);
+ assert.match(app,/`\/api\/runs\/\$\{runId\}\/retry`,\{method:"POST",body:"\{\}"\}/);
+ assert.match(app,/<RunProbe runId=\{chip\.runId\|\|null\} align="end" label=\{t\("errorDetails"\)\}/);
+ assert.match(app,/<RunProbe runId=\{member\.id===computer\.botId\?computer\.busyRunId:null\}><Avatar/);
+ const probe=fs.readFileSync('apps/web/src/run-monitor.tsx','utf8');
+ assert.match(probe,/`\/api\/runs\/\$\{runId\}\/activity\$\{after\}`/);
+ assert.match(fs.readFileSync('apps/web/src/main.tsx','utf8'),/import "\.\/monitor\.css";/);
+});
diff --git a/tests/retention.test.py b/tests/retention.test.py
index bfbece7..f67e568 100644
--- a/tests/retention.test.py
+++ b/tests/retention.test.py
@@ -44,6 +44,11 @@ try:
       FROM unnest(ARRAY['completed','failed','cancelled','running','queued','leased','waiting_input','waiting_takeover']) status;
     INSERT INTO runs(id,space_id,bot_id,thread_id,user_id,status,checkpoint,updated_at,completed_at)
       VALUES ('recent-completed','s','b','t','u','completed','{"keep":true}',now(),now());
+    INSERT INTO run_activity(run_id,kind,payload,created_at)
+      SELECT run_id,'notice','{"text":"step from an old run"}',now()-interval '10 days'
+      FROM unnest(ARRAY['completed','failed','running','recent-completed']) run_id;
+    INSERT INTO run_activity(run_id,kind,payload,created_at)
+      VALUES ('recent-completed','notice','{"text":"fresh step"}',now());
     INSERT INTO memory_items(id,space_id,user_id,bot_id,content,revision,source_run_id)
       VALUES ('00000000-0000-0000-0000-000000000001','s','u','b','keep current memory',15,'completed');
     INSERT INTO memory_revisions(memory_id,revision,space_id,user_id,bot_id,content,importance,action)
@@ -69,6 +74,8 @@ try:
     clean('runs',90)
     check('(SELECT count(*) FROM runs)=6')
     check("EXISTS(SELECT 1 FROM memory_items WHERE content='keep current memory' AND source_run_id IS NULL)")
+    clean('run_activity',7)
+    check("(SELECT count(*) FROM run_activity)=1 AND EXISTS(SELECT 1 FROM run_activity WHERE run_id='recent-completed' AND payload->>'text'='fresh step')")
     clean('recordings',30)
     check("(SELECT count(*) FROM taught_skills WHERE recording='{}')=3 AND (SELECT count(*) FROM taught_skills WHERE playbook->>'keep'='true')=5")
     clean('revisions',90)
@@ -80,9 +87,9 @@ try:
     check("(SELECT count(*) FROM computer_execution_leases)=1 AND EXISTS(SELECT 1 FROM computer_execution_leases WHERE id='active')")
     check("(SELECT count(*) FROM computer_profile_locks)=1 AND EXISTS(SELECT 1 FROM computer_profile_locks WHERE profile_key='active')")
     check("EXISTS(SELECT 1 FROM messages WHERE body='keep my conversation')")
-    for name,age in [('events',30),('checkpoints',7),('runs',90),('recordings',30),('revisions',90),('deleted_memories',90),('leases',7),('profile_locks',7)]:
+    for name,age in [('events',30),('checkpoints',7),('runs',90),('run_activity',7),('recordings',30),('revisions',90),('deleted_memories',90),('leases',7),('profile_locks',7)]:
         clean(name,age)
     check('(SELECT count(*) FROM runs)=6 AND (SELECT count(*) FROM memory_revisions)=9')
-    print('PASS: all 8 retention rules; bounded batches; live work and user content preserved; repeat-safe')
+    print('PASS: all 9 retention rules; bounded batches; live work and user content preserved; repeat-safe')
 finally:
     sql(f'DROP DATABASE {DB};','postgres')
diff --git a/tests/run-activity.test.py b/tests/run-activity.test.py
new file mode 100644
index 0000000..076f993
--- /dev/null
+++ b/tests/run-activity.test.py
@@ -0,0 +1,210 @@
+"""Integration regression: the run trail the hover bubble reads, and the manual
+retry that re-queues a dead run without losing where it stopped.
+Requires the project's Postgres Compose service; never modifies the app database.
+"""
+from pathlib import Path
+import re
+import subprocess
+import uuid
+
+ROOT = Path(__file__).resolve().parents[1]
+DB = 'run_activity_test_' + uuid.uuid4().hex
+MONITOR_RS = (ROOT/'crates/api/src/monitor.rs').read_text()
+RUNS_RS = (ROOT/'crates/api/src/runs.rs').read_text()
+ACTIVITY_SQL = (ROOT/'crates/api/src/retention/run_activity.sql').read_text()
+
+
+def sql(text, database=DB):
+    result = subprocess.run(['docker','compose','exec','-T','postgres','psql','-X','-q','-v','ON_ERROR_STOP=1','-U','lazyboy','-d',database], input=text, text=True, capture_output=True, cwd=ROOT)
+    if result.returncode:
+        raise AssertionError(result.stderr)
+    return result.stdout.strip()
+
+
+def query(text):
+    result = subprocess.run(['docker','compose','exec','-T','postgres','psql','-X','-q','-t','-A','-U','lazyboy','-d',DB], input=text, text=True, capture_output=True, cwd=ROOT)
+    if result.returncode:
+        raise AssertionError(result.stderr)
+    return result.stdout.strip()
+
+
+def check(condition, label):
+    if query(f"SELECT ({condition});") != 't':
+        raise AssertionError(f'run-activity assertion failed: {label}')
+
+
+def flatten(text):
+    return re.sub(r'\s+',' ',text).strip()
+
+
+def trail(run, after=None, limit=60):
+    """The handler's paging query: newest window first read, then forward by id."""
+    cursor = 'NULL' if after is None else str(after)
+    statement = """SELECT id FROM ( SELECT id FROM run_activity WHERE run_id='%s' AND (%s::bigint IS NULL OR id>%s) ORDER BY id DESC LIMIT %s ) recent ORDER BY id ASC""" % (run, cursor, cursor, limit)
+    return query(statement)
+
+
+def row_ids(run):
+    return query(f"SELECT id FROM run_activity WHERE run_id='{run}' ORDER BY id;").split()
+
+
+def retry(run):
+    """The handler's retry UPDATE, verbatim apart from the bind parameters."""
+    statement = """
+    UPDATE runs SET status='queued', retry_count=0, error=NULL, completed_at=NULL,
+            lease_owner=NULL, lease_expires_at=NULL, updated_at=now()
+     WHERE id='%s' AND space_id='s' AND user_id='u'
+       AND status IN ('failed','cancelled')
+       AND NOT EXISTS (
+           SELECT 1 FROM runs a
+           WHERE a.bot_id=runs.bot_id AND a.id<>runs.id
+             AND a.status IN ('leased','running','waiting_input','waiting_takeover')
+             AND (a.lease_expires_at IS NULL OR a.lease_expires_at >= now())
+       )
+     RETURNING thread_id""" % run
+    return query(statement) or '<null>'
+
+
+# The bubble is only as honest as what the run loop writes. Pin the shapes so a
+# silent rename shows up here instead of as a blank panel.
+monitor = flatten(MONITOR_RS)
+for fragment in [
+    '.route("/api/runs/{id}/activity", get(activity))',
+    '.route("/api/runs/{id}/retry", post(retry))',
+    'INSERT INTO run_activity (run_id,kind,payload)',
+    'SELECT id, kind, payload, created_at FROM run_activity WHERE run_id=$1 AND ($2::bigint IS NULL OR id>$2) ORDER BY id DESC LIMIT $3',
+    "(checkpoint->>'turn')::bigint AS turn",
+    "(checkpoint->>'turnLimit')::bigint AS turn_limit",
+    "AND status IN ('failed','cancelled')",
+    'a.bot_id=runs.bot_id AND a.id<>runs.id',
+    "a.status IN ('leased','running','waiting_input','waiting_takeover')",
+    '"run is not retryable"',
+]:
+    assert flatten(fragment) in monitor, f'monitor SQL changed, update tests/run-activity.test.py: {fragment}'
+
+runs_source = flatten(RUNS_RS)
+for fragment in [
+    '"kind": "error"',
+    '"code": failure.code',
+    '"Worker interrupted after tool execution',
+    "'step', $2::text, 'stepAt', now(), 'turn', $3::bigint, 'turnLimit', $4::bigint",
+    '"event": "started"',
+    '"event": "failed"',
+    '"status": tool_status(',
+    '"elapsedMs": model_elapsed',
+]:
+    assert flatten(fragment) in runs_source, f'run loop instrumentation changed, update tests/run-activity.test.py: {fragment}'
+
+# template0, not template1: hosts whose Postgres collation version drifted refuse to
+# copy template1, and C collation keeps this test's ordering deterministic anyway.
+sql(f"CREATE DATABASE {DB} TEMPLATE template0 LC_COLLATE 'C' LC_CTYPE 'C';",'postgres')
+try:
+    for migration in sorted((ROOT/'migrations').glob('*.sql')):
+        sql(migration.read_text())
+
+    # 015 has to survive being applied twice, like every other migration.
+    sql((ROOT/'migrations/015_run_activity.sql').read_text())
+    check("""(SELECT count(*) FROM pg_indexes WHERE tablename='run_activity' AND indexname IN ('run_activity_run_idx','run_activity_retention_idx'))=2""",
+          'the trail needs its run index and its retention index')
+
+    sql("""
+    INSERT INTO users(id,name) VALUES ('u','test');
+    INSERT INTO spaces(id,user_id,name) VALUES ('s','u','test');
+    INSERT INTO computers(id,space_id,user_id,scope,scope_key,home_key) VALUES ('c','s','u','bot','c','c');
+    INSERT INTO bots(id,space_id,user_id,name,computer_id) VALUES
+      ('b_live','s','u','live','c'),('b_idle','s','u','idle','c'),
+      ('b_busy','s','u','busy','c'),('b_other','s','u','other','c');
+    INSERT INTO threads(id,space_id,bot_id,user_id) VALUES
+      ('t_live','s','b_live','u'),('t_idle','s','b_idle','u'),
+      ('t_busy','s','b_busy','u'),('t_other','s','b_other','u');
+    INSERT INTO runs(id,space_id,bot_id,thread_id,user_id,status,prompt,checkpoint) VALUES
+      ('r_live','s','b_live','t_live','u','running','整理報價','{"step":"browser: click #12","turn":3,"turnLimit":40}'),
+      ('r_goal','s','b_live','t_live','u','running','/goal 每天彙整日報','{"step":"shell: npm test","turn":7,"turnLimit":null}'),
+      ('r_failed','s','b_idle','t_idle','u','failed','整理報價','{"step":"browser: click #12","turn":7}'),
+      ('r_cancelled','s','b_idle','t_idle','u','cancelled','整理報價','{"keep":true}'),
+      ('r_busy_running','s','b_busy','t_busy','u','running','還沒結束','{}'),
+      ('r_busy_failed','s','b_busy','t_busy','u','failed','整理報價','{}'),
+      ('r_other','s','b_other','t_other','u','running','別人的工作正在跑','{}');
+    UPDATE runs SET error='HTTP status 401 Unauthorized: invalid api key' WHERE id='r_failed';
+    UPDATE runs SET retry_count=3, lease_owner='stale-worker', lease_expires_at=now()+interval '5 minutes' WHERE id='r_failed';
+    """)
+
+    # The header the bubble reads: turn, limit and step must come back typed.
+    check("""(SELECT (checkpoint->>'turn')::bigint FROM runs WHERE id='r_live')=3
+             AND (SELECT (checkpoint->>'turnLimit')::bigint FROM runs WHERE id='r_live')=40""",
+          'turn and turn limit must be readable as numbers')
+    check("""(SELECT (checkpoint->>'turnLimit')::bigint IS NULL FROM runs WHERE id='r_goal')""",
+          'a goal run has no limit instead of a broken one')
+
+    # A trail, oldest first, with the payload the panel renders.
+    sql("""
+    INSERT INTO run_activity(run_id,kind,payload) VALUES
+      ('r_live','run','{"event":"started","task":"整理報價"}'),
+      ('r_live','model','{"turn":1,"elapsedMs":6400,"toolCalls":1,"text":"先打開網頁"}'),
+      ('r_live','tool','{"turn":1,"name":"browser","step":"browser: click #12","status":"ok","elapsedMs":400}'),
+      ('r_live','retry','{"turn":2,"attempt":1,"error":"429 rate limit","gaveUp":false}'),
+      ('r_live','tool','{"turn":2,"name":"browser","step":"browser: type #13","status":"timed_out","elapsedMs":150000}');
+    INSERT INTO run_activity(run_id,kind,payload) VALUES ('r_failed','run','{"event":"failed","error":"HTTP status 401"}');
+    """)
+    assert len(row_ids('r_live')) == 5, 'the trail collected every turn'
+    check("""(SELECT payload->>'step' FROM run_activity WHERE kind='tool' ORDER BY id LIMIT 1)='browser: click #12'""",
+          'a tool line carries the step it is on')
+
+    # Paging: the first read is the newest window (oldest first so it renders in
+    # order), later reads only carry what is newer than the row already shown.
+    live = row_ids('r_live')
+    assert len(live) == 5, f'expected five trail rows, got {live}'
+    assert trail('r_live', None, 60).split() == live, 'a full window is the whole trail, oldest first'
+    assert trail('r_live', None, 2).split() == live[3:], f'the first read must be the newest window, got {trail("r_live", None, 2).split()}'
+    assert trail('r_live', live[2], 60).split() == live[3:], 'a cursor must never serve a row twice'
+    assert trail('r_live', live[4], 60) == '', 'a tail read on a quiet run stays empty'
+    assert trail('r_nothing', None, 60) == '', 'a run with no trail shows an empty panel, not an error'
+    sql("""
+    INSERT INTO run_activity(run_id,kind,payload) VALUES
+      ('r_live','tool','{"turn":3,"name":"shell","step":"shell: npm run build","status":"ok"}'),
+      ('r_live','notice','{"turn":3,"text":"這輪不需要電腦"}');
+    """)
+    fresh = row_ids('r_live')[5:]
+    assert len(fresh) == 2, f'two new rows expected, got {fresh}'
+    assert trail('r_live', live[4], 60).split() == fresh, 'new rows extend the panel in arrival order'
+    assert trail('r_live', live[4], 1) == fresh[1], 'a small window keeps the freshest line, not the oldest'
+    assert len(set(trail('r_live', None, 60).split())) == 7, 'the window never repeats a row'
+
+    # Manual retry: back to queued, counters reset, checkpoint kept for the resume.
+    assert retry('r_failed') == 't_idle', 'a failed run must be re-queueable from the chat'
+    check("""(SELECT status FROM runs WHERE id='r_failed')='queued'""", 'a retry goes back to the queue')
+    check("""(SELECT retry_count FROM runs WHERE id='r_failed')=0""", 'a human retry is not the automatic retry')
+    check("""(SELECT error FROM runs WHERE id='r_failed') IS NULL""", 'the stale failure must stop being shown')
+    check("""(SELECT lease_owner FROM runs WHERE id='r_failed') IS NULL""", 'a re-queued run must not keep the dead worker lease')
+    check("""(SELECT checkpoint->>'step' FROM runs WHERE id='r_failed')='browser: click #12'""",
+          'the checkpoint survives so the run continues where it stopped')
+    assert retry('r_failed') == '<null>', 'a queued run cannot be re-queued twice'
+    assert retry('r_cancelled') == 't_idle', 'a cancelled run can be restarted too'
+
+    # Only the bot that is actually idle may be restarted, and only for itself.
+    assert retry('r_busy_running') == '<null>', 'a live run is never hijacked by a retry'
+    assert retry('r_busy_failed') == '<null>', 'a bot with work in flight cannot be double-booked'
+    sql("UPDATE runs SET status='cancelled' WHERE id='r_busy_running'")
+    assert retry('r_busy_failed') == 't_busy', 'once its other run is gone the retry works'
+    assert retry('r_other') == '<null>', 'another bot is none of this retry\'s business'
+
+    # Diagnostics are expendable: old trail lines expire, fresh ones stay.
+    sql("""
+    INSERT INTO run_activity(run_id,kind,payload,created_at)
+      SELECT 'r_live','notice','{"text":"old"}',now()-interval '30 days' FROM generate_series(1,3) n;
+    UPDATE run_activity SET created_at=now()-interval '30 days' WHERE payload->>'text'='old';
+    """)
+    stale = query("SELECT count(*) FROM run_activity WHERE created_at < now()-interval '7 days';")
+    sql(ACTIVITY_SQL.replace('$1','7').replace('$2','1000')+';')
+    removed = int(stale) - int(query("SELECT count(*) FROM run_activity WHERE created_at < now()-interval '7 days';"))
+    assert removed == int(stale), f'retention must remove every expired row ({stale}), removed {removed}'
+    check("""(SELECT count(*) FROM run_activity WHERE run_id='r_live')=7""", 'fresh trail rows are never collected')
+    sql(ACTIVITY_SQL.replace('$1','7').replace('$2','1000')+';')
+    check("""(SELECT count(*) FROM run_activity WHERE run_id='r_live')=7""", 'repeat retention batches stay safe')
+
+    # And the trail dies with its run, never as an orphan holding the table open.
+    sql("DELETE FROM runs WHERE id='r_live';")
+    check("""(SELECT count(*) FROM run_activity WHERE run_id='r_live')=0""", 'deleting a run must delete its trail')
+    print('PASS: trail paging, typed turn/limit header, guarded manual retry that keeps the checkpoint, expiring diagnostics, cascading cleanup')
+finally:
+    sql(f'DROP DATABASE {DB};','postgres')