P2: completion contract, loop guard, context truncation

README.md

@@ -1,6 +1,6 @@
 # GrokBoy
 
-Minimal local **GrokBot-like** CLI agent. Phase **P1**: tools + ReAct + sessions.
+Minimal local **GrokBot-like** CLI agent. Phase **P2**: completion contract, loop guard, context truncation.
 
 ## Status
 
@@ -8,7 +8,7 @@ Minimal local **GrokBot-like** CLI agent. Phase **P1**: tools + ReAct + sessions
 |-------|--------|
 | P0 streaming chat | done |
 | P1 shell / files + ReAct | done |
-| P2 completion / loop guard | planned |
+| P2 completion / loop guard / truncation | done |
 | P3 browser (Playwright) | later |
 
 No Docker desktop, no Codex/LazyBoy fork.
@@ -20,6 +20,7 @@ export GROKBOY_API_KEY=your_key   # or XAI_API_KEY
 # optional:
 # export GROKBOY_BASE_URL=https://api.x.ai/v1
 # export GROKBOY_MODEL=grok-4.6
+# export GROKBOY_CONTEXT_CHARS=100000
 
 cd ~/GrokBoy
 cargo run -p grokboy -- chat
@@ -28,16 +29,20 @@ cargo run -p grokboy -- chat
 ## Commands
 
 - `grokboy chat` — interactive streaming chat (no tools)
-- `grokboy run "<prompt>"` — one-shot agent with tools (`shell`, `list_dir`, `read_file`, `write_file`)
+- `grokboy run "<prompt>"` — one-shot agent with tools
 - `grokboy run --session <id> "<prompt>"` — continue a saved session
-- `grokboy smoke` — offline tool checks (no API key required)
+- `grokboy smoke` — offline checks (no API key required)
 - `grokboy help`
 
+Tools: `shell`, `list_dir`, `read_file`, `write_file`, `report_done`, `report_blocked`.
+
 Sessions are stored under `~/.grokboy/sessions/<id>.json`.
 
+The agent stops on `report_done` / `report_blocked`, blocks identical tool rounds (×3), and truncates old context when over budget.
+
 ## 繁體中文
 
-本機終端機 coding assistant。P1 已支援讀寫檔、列目錄、shell，以及多輪 tool-calling。
+本機終端機 coding assistant。P2 已支援完成合約（`report_done` / `report_blocked`）、迴圈守衛與上下文截斷。
 
 ```bash
 export GROKBOY_API_KEY=你的金鑰

crates/grokboy-core/src/agent.rs

@@ -1,63 +1,317 @@
-//! Multi-step OpenAI-compatible tool-calling ReAct loop.
+//! Multi-step OpenAI-compatible tool-calling ReAct loop (P2: completion, loop guard, truncation).
 
 use crate::config::Config;
-use crate::model::{ChatMessage, chat_completion};
-use crate::tools::{ToolContext, execute_tool, tool_definitions};
+use crate::model::{ChatMessage, Role, ToolCall, chat_completion};
+use crate::tools::{ToolContext, execute_tool, is_completion_tool, tool_definitions};
 use anyhow::Result;
+use serde_json::Value;
+use std::future::Future;
 
 pub const DEFAULT_MAX_ROUNDS: usize = 12;
+pub const DEFAULT_CONTEXT_CHARS: usize = 100_000;
+pub const LOOP_GUARD_REPEAT: usize = 3;
 
 pub const AGENT_SYSTEM: &str = "\
 You are GrokBoy, a concise local coding assistant with tools.
 Use tools when they help solve the task; otherwise answer directly.
 Prefer short, clear answers. Traditional Chinese is welcome when the user writes in Chinese.
-Available tools: shell, list_dir, read_file, write_file.
-Do not invent tool results — call the tools. Stop when you can give a final answer.";
+Available tools: shell, list_dir, read_file, write_file, report_done, report_blocked.
+When the task is finished, call report_done with a short summary.
+If you are stuck or cannot proceed, call report_blocked with the reason — do not invent results or loop.
+Do not invent tool results — call the tools.";
 
-/// Run the agent loop until the model returns text without tool_calls or max rounds.
-/// Appends all intermediate messages (assistant tool_calls + tool results + final) to `messages`.
-/// Returns the final assistant text (may be empty if stopped on round limit with only tools).
+/// Final verdict from the agent loop (fail-closed when stuck).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum AgentVerdict {
+    /// Model called `report_done`.
+    Done(String),
+    /// Model called `report_blocked`, loop guard, or max rounds.
+    Blocked(String),
+    /// Model returned final text without a completion tool.
+    Answer(String),
+}
+
+impl AgentVerdict {
+    pub fn message(&self) -> &str {
+        match self {
+            Self::Done(s) | Self::Blocked(s) | Self::Answer(s) => s,
+        }
+    }
+
+    pub fn kind(&self) -> &'static str {
+        match self {
+            Self::Done(_) => "done",
+            Self::Blocked(_) => "blocked",
+            Self::Answer(_) => "answer",
+        }
+    }
+}
+
+/// Resolve context budget from `GROKBOY_CONTEXT_CHARS` or default.
+pub fn context_char_budget() -> usize {
+    std::env::var("GROKBOY_CONTEXT_CHARS")
+        .ok()
+        .and_then(|s| s.parse().ok())
+        .filter(|&n| n > 0)
+        .unwrap_or(DEFAULT_CONTEXT_CHARS)
+}
+
+/// Stable signature for a single tool call (name + args).
+pub fn tool_call_signature(call: &ToolCall) -> String {
+    format!("{}:{}", call.function.name, call.function.arguments)
+}
+
+/// Signature for a whole round of tool calls (order-preserving).
+pub fn round_signature(calls: &[ToolCall]) -> String {
+    calls
+        .iter()
+        .map(tool_call_signature)
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+/// Approximate serialized size of one message (chars).
+pub fn message_char_len(msg: &ChatMessage) -> usize {
+    let mut n = 8; // role overhead
+    if let Some(c) = &msg.content {
+        n += c.len();
+    }
+    if let Some(id) = &msg.tool_call_id {
+        n += id.len();
+    }
+    if let Some(calls) = &msg.tool_calls {
+        for c in calls {
+            n += c.id.len() + c.function.name.len() + c.function.arguments.len() + 16;
+        }
+    }
+    n
+}
+
+pub fn messages_char_len(messages: &[ChatMessage]) -> usize {
+    messages.iter().map(message_char_len).sum()
+}
+
+/// Truncate older tool results / middle turns while keeping system + recent tail + last user goal.
+/// Mutates in place. Keeps it simple: shrink old tool contents first, then drop middle messages.
+pub fn truncate_messages(messages: &mut Vec<ChatMessage>, budget: usize) {
+    if budget == 0 || messages_char_len(messages) <= budget {
+        return;
+    }
+
+    // 1) Shrink older tool message contents (keep last few tool results intact).
+    let tool_indices: Vec<usize> = messages
+        .iter()
+        .enumerate()
+        .filter(|(_, m)| m.role == Role::Tool)
+        .map(|(i, _)| i)
+        .collect();
+    let keep_recent_tools = 4usize;
+    let shrink_until = tool_indices.len().saturating_sub(keep_recent_tools);
+    for &idx in tool_indices.iter().take(shrink_until) {
+        if let Some(content) = messages[idx].content.as_mut() {
+            if content.len() > 120 {
+                let omitted = content.len();
+                *content = format!("[truncated tool result; was {omitted} chars]");
+            }
+        }
+        if messages_char_len(messages) <= budget {
+            return;
+        }
+    }
+
+    // 2) Drop middle messages while preserving: leading system*, last user, and a recent tail.
+    while messages_char_len(messages) > budget && messages.len() > 4 {
+        let first_drop = messages
+            .iter()
+            .enumerate()
+            .find(|(i, m)| *i > 0 && m.role != Role::System)
+            .map(|(i, _)| i);
+        let Some(i) = first_drop else { break };
+
+        // Never drop the last user message or the last two messages.
+        let last_user = messages
+            .iter()
+            .rposition(|m| m.role == Role::User)
+            .unwrap_or(messages.len());
+        if i >= last_user || i + 2 >= messages.len() {
+            // Shrink remaining large contents instead.
+            for msg in messages.iter_mut() {
+                if let Some(content) = msg.content.as_mut() {
+                    if content.len() > 200 {
+                        let keep = 200.min(content.len());
+                        let omitted = content.len().saturating_sub(keep);
+                        *content = format!("{}…\n[truncated {omitted} chars]", &content[..keep]);
+                    }
+                }
+            }
+            break;
+        }
+
+        // If dropping an assistant with tool_calls, also drop following tool messages for those ids.
+        let drop_ids: Vec<String> = messages
+            .get(i)
+            .and_then(|m| m.tool_calls.as_ref())
+            .map(|calls| calls.iter().map(|c| c.id.clone()).collect())
+            .unwrap_or_default();
+
+        messages.remove(i);
+        let j = i;
+        while j < messages.len() {
+            let is_orphan_tool = messages[j].role == Role::Tool
+                && messages[j]
+                    .tool_call_id
+                    .as_ref()
+                    .is_some_and(|id| drop_ids.contains(id));
+            if is_orphan_tool {
+                messages.remove(j);
+            } else {
+                break;
+            }
+        }
+    }
+}
+
+/// Run the agent with a live HTTP model client.
 pub async fn run_agent(
     config: &Config,
     messages: &mut Vec<ChatMessage>,
     tool_ctx: &ToolContext,
     max_rounds: usize,
-) -> Result<String> {
-    let tools = tool_definitions();
-    let mut last_text = String::new();
+) -> Result<AgentVerdict> {
+    let budget = context_char_budget();
+    let config = config.clone();
+    run_agent_with(messages, tool_ctx, max_rounds, budget, move |msgs, tools| {
+        let config = config.clone();
+        async move { chat_completion(&config, &msgs, Some(&tools)).await }
+    })
+    .await
+}
 
-    for _round in 0..max_rounds {
-        let reply = chat_completion(config, messages, Some(&tools)).await?;
+/// Core loop with injectable completer (for offline tests).
+/// `complete` receives a snapshot of messages and tool defs each round.
+pub async fn run_agent_with<F, Fut>(
+    messages: &mut Vec<ChatMessage>,
+    tool_ctx: &ToolContext,
+    max_rounds: usize,
+    context_budget: usize,
+    mut complete: F,
+) -> Result<AgentVerdict>
+where
+    F: FnMut(Vec<ChatMessage>, Value) -> Fut,
+    Fut: Future<Output = Result<ChatMessage>>,
+{
+    let tools = tool_definitions();
+    let mut prev_round_sig: Option<String> = None;
+    let mut same_sig_streak: usize = 0;
+
+    for round in 0..max_rounds {
+        truncate_messages(messages, context_budget);
+
+        let reply = complete(messages.clone(), tools.clone()).await?;
         let tool_calls = reply.tool_calls.clone().unwrap_or_default();
 
         if tool_calls.is_empty() {
-            last_text = reply.text().to_string();
+            let last_text = reply.text().to_string();
             messages.push(reply);
-            return Ok(last_text);
+            if last_text.trim().is_empty() {
+                return Ok(AgentVerdict::Blocked(
+                    "model returned empty final answer".into(),
+                ));
+            }
+            return Ok(AgentVerdict::Answer(last_text));
         }
 
-        // Keep any content the model sent alongside tool_calls.
-        if let Some(c) = reply.content.as_ref().filter(|s| !s.is_empty()) {
-            last_text = c.clone();
+        // Loop guard: identical tool-call round repeated N times → fail closed.
+        let sig = round_signature(&tool_calls);
+        if prev_round_sig.as_deref() == Some(sig.as_str()) {
+            same_sig_streak += 1;
+        } else {
+            same_sig_streak = 1;
+            prev_round_sig = Some(sig);
         }
+        if same_sig_streak >= LOOP_GUARD_REPEAT {
+            messages.push(reply);
+            let reason = format!(
+                "loop guard: identical tool calls repeated {LOOP_GUARD_REPEAT} times (round {})",
+                round + 1
+            );
+            return Ok(AgentVerdict::Blocked(reason));
+        }
+
         messages.push(reply);
 
+        let mut completion: Option<AgentVerdict> = None;
         for call in &tool_calls {
-            let result = execute_tool(tool_ctx, &call.function.name, &call.function.arguments).await;
+            let result =
+                execute_tool(tool_ctx, &call.function.name, &call.function.arguments).await;
+            if completion.is_none() && is_completion_tool(&call.function.name) {
+                completion = parse_completion_verdict(&call.function.name, &result);
+            }
             messages.push(ChatMessage::tool(&call.id, result));
         }
+
+        if let Some(verdict) = completion {
+            return Ok(verdict);
+        }
     }
 
-    Ok(last_text)
+    Ok(AgentVerdict::Blocked(format!(
+        "blocked: reached max rounds ({max_rounds}) without completion"
+    )))
+}
+
+fn parse_completion_verdict(name: &str, result_json: &str) -> Option<AgentVerdict> {
+    let v: Value = serde_json::from_str(result_json).ok()?;
+    if v.get("error").is_some() {
+        return Some(AgentVerdict::Blocked(format!(
+            "completion tool {name} failed: {}",
+            v["error"]
+        )));
+    }
+    match name {
+        "report_done" => {
+            let msg = v
+                .get("message")
+                .and_then(|x| x.as_str())
+                .unwrap_or("done")
+                .to_string();
+            Some(AgentVerdict::Done(msg))
+        }
+        "report_blocked" => {
+            let reason = v
+                .get("reason")
+                .and_then(|x| x.as_str())
+                .unwrap_or("blocked")
+                .to_string();
+            Some(AgentVerdict::Blocked(reason))
+        }
+        _ => None,
+    }
 }
 
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::model::{FunctionCall, ToolCall};
+    use serde_json::json;
+    use std::sync::{Arc, Mutex};
+
+    fn tc(id: &str, name: &str, args: &str) -> ToolCall {
+        ToolCall {
+            id: id.into(),
+            kind: "function".into(),
+            function: FunctionCall {
+                name: name.into(),
+                arguments: args.into(),
+            },
+        }
+    }
 
     #[test]
-    fn system_prompt_mentions_tools() {
+    fn system_prompt_mentions_tools_and_chinese() {
         assert!(AGENT_SYSTEM.contains("shell"));
+        assert!(AGENT_SYSTEM.contains("report_done"));
+        assert!(AGENT_SYSTEM.contains("report_blocked"));
         assert!(AGENT_SYSTEM.contains("Traditional Chinese") || AGENT_SYSTEM.contains("Chinese"));
     }
 
@@ -65,4 +319,202 @@ mod tests {
     fn default_max_rounds_is_12() {
         assert_eq!(DEFAULT_MAX_ROUNDS, 12);
     }
+
+    #[test]
+    fn round_signature_stable() {
+        let a = vec![tc("1", "list_dir", r#"{"path":"."}"#)];
+        let b = vec![tc("2", "list_dir", r#"{"path":"."}"#)];
+        assert_eq!(round_signature(&a), round_signature(&b));
+        let c = vec![tc("1", "list_dir", r#"{"path":"src"}"#)];
+        assert_ne!(round_signature(&a), round_signature(&c));
+    }
+
+    #[test]
+    fn truncate_keeps_system_and_shrinks_old_tools() {
+        let mut msgs = vec![
+            ChatMessage::system("sys"),
+            ChatMessage::user("goal"),
+            ChatMessage::assistant_tool_calls(None, vec![tc("c1", "read_file", r#"{"path":"a"}"#)]),
+            ChatMessage::tool("c1", "X".repeat(5_000)),
+            ChatMessage::assistant_tool_calls(None, vec![tc("c2", "read_file", r#"{"path":"b"}"#)]),
+            ChatMessage::tool("c2", "Y".repeat(5_000)),
+            ChatMessage::assistant_tool_calls(None, vec![tc("c3", "read_file", r#"{"path":"c"}"#)]),
+            ChatMessage::tool("c3", "Z".repeat(5_000)),
+            ChatMessage::assistant_tool_calls(None, vec![tc("c4", "read_file", r#"{"path":"d"}"#)]),
+            ChatMessage::tool("c4", "W".repeat(5_000)),
+            ChatMessage::assistant_tool_calls(None, vec![tc("c5", "read_file", r#"{"path":"e"}"#)]),
+            ChatMessage::tool("c5", "V".repeat(5_000)),
+            ChatMessage::user("still the goal"),
+            ChatMessage::assistant("recent"),
+        ];
+        let before = messages_char_len(&msgs);
+        truncate_messages(&mut msgs, 8_000);
+        let after = messages_char_len(&msgs);
+        assert!(after < before);
+        assert_eq!(msgs[0].text(), "sys");
+        assert!(msgs.iter().any(|m| m.role == Role::User));
+        let old_tool = msgs
+            .iter()
+            .find(|m| m.role == Role::Tool && m.tool_call_id.as_deref() == Some("c1"));
+        if let Some(t) = old_tool {
+            assert!(
+                t.text().contains("truncated") || t.text().len() < 5_000,
+                "{}",
+                t.text()
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn loop_stops_on_report_done() {
+        let dir = std::env::temp_dir().join(format!("grokboy-agent-done-{}", std::process::id()));
+        let _ = std::fs::create_dir_all(&dir);
+        let ctx = ToolContext::new(dir.clone());
+        let mut messages = vec![
+            ChatMessage::system(AGENT_SYSTEM),
+            ChatMessage::user("finish"),
+        ];
+        let calls = Arc::new(Mutex::new(0usize));
+        let calls2 = calls.clone();
+
+        let verdict = run_agent_with(&mut messages, &ctx, 5, 100_000, move |_msgs, _tools| {
+            let n = {
+                let mut g = calls2.lock().unwrap();
+                *g += 1;
+                *g
+            };
+            async move {
+                if n == 1 {
+                    Ok(ChatMessage::assistant_tool_calls(
+                        None,
+                        vec![tc(
+                            "done1",
+                            "report_done",
+                            &json!({"message": "任務完成"}).to_string(),
+                        )],
+                    ))
+                } else {
+                    Ok(ChatMessage::assistant("should not reach"))
+                }
+            }
+        })
+        .await
+        .unwrap();
+
+        assert_eq!(verdict, AgentVerdict::Done("任務完成".into()));
+        assert_eq!(*calls.lock().unwrap(), 1);
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[tokio::test]
+    async fn loop_stops_on_report_blocked() {
+        let dir = std::env::temp_dir().join(format!("grokboy-agent-blk-{}", std::process::id()));
+        let _ = std::fs::create_dir_all(&dir);
+        let ctx = ToolContext::new(dir.clone());
+        let mut messages = vec![ChatMessage::user("x")];
+
+        let verdict = run_agent_with(&mut messages, &ctx, 5, 100_000, move |_msgs, _tools| async move {
+            Ok(ChatMessage::assistant_tool_calls(
+                None,
+                vec![tc(
+                    "b1",
+                    "report_blocked",
+                    &json!({"reason": "permission denied"}).to_string(),
+                )],
+            ))
+        })
+        .await
+        .unwrap();
+
+        assert_eq!(verdict, AgentVerdict::Blocked("permission denied".into()));
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[tokio::test]
+    async fn loop_guard_blocks_repeated_calls() {
+        let dir = std::env::temp_dir().join(format!("grokboy-agent-loop-{}", std::process::id()));
+        let _ = std::fs::create_dir_all(&dir);
+        let ctx = ToolContext::new(dir.clone());
+        std::fs::write(dir.join("f.txt"), "hi").unwrap();
+        let mut messages = vec![ChatMessage::user("loop")];
+        let args = json!({"path": "f.txt"}).to_string();
+
+        let verdict = run_agent_with(&mut messages, &ctx, 12, 100_000, move |_msgs, _tools| {
+            let args = args.clone();
+            async move {
+                Ok(ChatMessage::assistant_tool_calls(
+                    None,
+                    vec![tc("r1", "read_file", &args)],
+                ))
+            }
+        })
+        .await
+        .unwrap();
+
+        match verdict {
+            AgentVerdict::Blocked(reason) => {
+                assert!(reason.contains("loop guard"), "{reason}");
+            }
+            other => panic!("expected blocked, got {other:?}"),
+        }
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[tokio::test]
+    async fn max_rounds_fail_closed() {
+        let dir = std::env::temp_dir().join(format!("grokboy-agent-max-{}", std::process::id()));
+        let _ = std::fs::create_dir_all(&dir);
+        let ctx = ToolContext::new(dir.clone());
+        std::fs::write(dir.join("a.txt"), "1").unwrap();
+        std::fs::write(dir.join("b.txt"), "2").unwrap();
+        let mut messages = vec![ChatMessage::user("spin")];
+        let n = Arc::new(Mutex::new(0usize));
+        let n2 = n.clone();
+
+        let verdict = run_agent_with(&mut messages, &ctx, 2, 100_000, move |_msgs, _tools| {
+            let i = {
+                let mut g = n2.lock().unwrap();
+                *g += 1;
+                *g
+            };
+            async move {
+                let path = if i == 1 { "a.txt" } else { "b.txt" };
+                Ok(ChatMessage::assistant_tool_calls(
+                    None,
+                    vec![tc(
+                        &format!("c{i}"),
+                        "read_file",
+                        &json!({"path": path}).to_string(),
+                    )],
+                ))
+            }
+        })
+        .await
+        .unwrap();
+
+        match verdict {
+            AgentVerdict::Blocked(reason) => {
+                assert!(reason.contains("max rounds"), "{reason}");
+            }
+            other => panic!("expected blocked, got {other:?}"),
+        }
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
+    #[tokio::test]
+    async fn plain_answer_without_tools() {
+        let dir = std::env::temp_dir().join(format!("grokboy-agent-ans-{}", std::process::id()));
+        let _ = std::fs::create_dir_all(&dir);
+        let ctx = ToolContext::new(dir.clone());
+        let mut messages = vec![ChatMessage::user("hi")];
+
+        let verdict = run_agent_with(&mut messages, &ctx, 5, 100_000, move |_msgs, _tools| async move {
+            Ok(ChatMessage::assistant("hello there"))
+        })
+        .await
+        .unwrap();
+
+        assert_eq!(verdict, AgentVerdict::Answer("hello there".into()));
+        let _ = std::fs::remove_dir_all(&dir);
+    }
 }

crates/grokboy-core/src/lib.rs

@@ -6,8 +6,12 @@ mod model;
 mod session;
 mod tools;
 
-pub use agent::{AGENT_SYSTEM, DEFAULT_MAX_ROUNDS, run_agent};
+pub use agent::{
+    AGENT_SYSTEM, AgentVerdict, DEFAULT_CONTEXT_CHARS, DEFAULT_MAX_ROUNDS, LOOP_GUARD_REPEAT,
+    context_char_budget, message_char_len, messages_char_len, round_signature, run_agent,
+    run_agent_with, tool_call_signature, truncate_messages,
+};
 pub use config::Config;
 pub use model::{ChatMessage, FunctionCall, Role, ToolCall, chat_completion, stream_chat};
 pub use session::{Session, load_or_create, load_session, save_session, sessions_dir};
-pub use tools::{ToolContext, execute_tool, tool_definitions};
+pub use tools::{ToolContext, execute_tool, is_completion_tool, tool_definitions};

crates/grokboy-core/src/tools.rs

@@ -1,4 +1,4 @@
-//! Built-in tools: shell, list_dir, read_file, write_file.
+//! Built-in tools: shell, list_dir, read_file, write_file, report_done, report_blocked.
 
 use anyhow::{Context, Result, anyhow};
 use serde_json::{Value, json};
@@ -94,10 +94,43 @@ pub fn tool_definitions() -> Value {
                     "required": ["path", "content"]
                 }
             }
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "report_done",
+                "description": "Signal that the task is complete. Stops the agent loop with a done verdict. Call once when finished; include a short summary in message.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "message": { "type": "string", "description": "Final summary for the user" }
+                    },
+                    "required": ["message"]
+                }
+            }
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "report_blocked",
+                "description": "Signal that the task cannot proceed. Stops the agent loop with a blocked verdict and reason. Prefer this over spinning or inventing results.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "reason": { "type": "string", "description": "Why the agent is blocked" }
+                    },
+                    "required": ["reason"]
+                }
+            }
         }
     ])
 }
 
+/// True if the tool name ends the ReAct loop (completion contract).
+pub fn is_completion_tool(name: &str) -> bool {
+    matches!(name, "report_done" | "report_blocked")
+}
+
 /// Resolve a user-supplied path against cwd and optionally enforce workspace_root.
 pub fn resolve_path(ctx: &ToolContext, path: &str) -> Result<PathBuf> {
     let raw = Path::new(path);
@@ -156,10 +189,34 @@ async fn execute_tool_inner(ctx: &ToolContext, name: &str, arguments_json: &str)
         "list_dir" => tool_list_dir(ctx, &args).await,
         "read_file" => tool_read_file(ctx, &args).await,
         "write_file" => tool_write_file(ctx, &args).await,
+        "report_done" => tool_report_done(&args),
+        "report_blocked" => tool_report_blocked(&args),
         other => Err(anyhow!("unknown tool: {other}")),
     }
 }
 
+fn tool_report_done(args: &Value) -> Result<Value> {
+    let message = args
+        .get("message")
+        .and_then(|v| v.as_str())
+        .ok_or_else(|| anyhow!("report_done: missing 'message'"))?;
+    Ok(json!({
+        "status": "done",
+        "message": message,
+    }))
+}
+
+fn tool_report_blocked(args: &Value) -> Result<Value> {
+    let reason = args
+        .get("reason")
+        .and_then(|v| v.as_str())
+        .ok_or_else(|| anyhow!("report_blocked: missing 'reason'"))?;
+    Ok(json!({
+        "status": "blocked",
+        "reason": reason,
+    }))
+}
+
 async fn tool_shell(ctx: &ToolContext, args: &Value) -> Result<Value> {
     let command = args
         .get("command")
@@ -389,11 +446,40 @@ mod tests {
         let _ = std::fs::remove_dir_all(&dir);
     }
 
+    #[tokio::test]
+    async fn report_done_and_blocked() {
+        let (ctx, dir) = temp_ctx();
+        let done = execute_tool(
+            &ctx,
+            "report_done",
+            &json!({"message": "all good"}).to_string(),
+        )
+        .await;
+        let done: Value = serde_json::from_str(&done).unwrap();
+        assert_eq!(done["status"], "done");
+        assert_eq!(done["message"], "all good");
+
+        let blocked = execute_tool(
+            &ctx,
+            "report_blocked",
+            &json!({"reason": "no access"}).to_string(),
+        )
+        .await;
+        let blocked: Value = serde_json::from_str(&blocked).unwrap();
+        assert_eq!(blocked["status"], "blocked");
+        assert_eq!(blocked["reason"], "no access");
+
+        assert!(is_completion_tool("report_done"));
+        assert!(is_completion_tool("report_blocked"));
+        assert!(!is_completion_tool("shell"));
+        let _ = std::fs::remove_dir_all(&dir);
+    }
+
     #[test]
-    fn tool_defs_include_four() {
+    fn tool_defs_include_six() {
         let defs = tool_definitions();
         let arr = defs.as_array().unwrap();
-        assert_eq!(arr.len(), 4);
+        assert_eq!(arr.len(), 6);
         let names: Vec<&str> = arr
             .iter()
             .map(|t| t["function"]["name"].as_str().unwrap())
@@ -402,5 +488,7 @@ mod tests {
         assert!(names.contains(&"list_dir"));
         assert!(names.contains(&"read_file"));
         assert!(names.contains(&"write_file"));
+        assert!(names.contains(&"report_done"));
+        assert!(names.contains(&"report_blocked"));
     }
 }

crates/grokboy/src/main.rs

@@ -1,7 +1,8 @@
 use anyhow::{Context, Result, anyhow};
 use grokboy_core::{
-    AGENT_SYSTEM, ChatMessage, Config, DEFAULT_MAX_ROUNDS, Session, ToolContext, execute_tool,
-    load_or_create, run_agent, save_session, stream_chat, tool_definitions,
+    AGENT_SYSTEM, AgentVerdict, ChatMessage, Config, DEFAULT_MAX_ROUNDS, Session, ToolContext,
+    execute_tool, is_completion_tool, load_or_create, messages_char_len, run_agent, save_session,
+    stream_chat, tool_definitions, truncate_messages,
 };
 use serde_json::json;
 use std::io::{self, Write};
@@ -48,22 +49,24 @@ async fn run() -> Result<()> {
 fn print_help() {
     println!(
         "\
-GrokBoy — minimal local CLI agent (P1: tools + ReAct)
+GrokBoy — minimal local CLI agent (P2: completion + loop guard + truncation)
 
 USAGE:
   grokboy chat                      Interactive streaming chat (no tools)
   grokboy run \"<prompt>\"            One-shot agent with tools
   grokboy run --session <id> \"...\"  Continue a saved session
-  grokboy smoke                     Offline tool checks (no API key required)
+  grokboy smoke                     Offline checks (no API key required)
   grokboy version
   grokboy help
 
 ENV:
-  GROKBOY_API_KEY       API key (or XAI_API_KEY / OPENAI_API_KEY)
-  GROKBOY_BASE_URL      default https://api.x.ai/v1
-  GROKBOY_MODEL         default grok-4.6
+  GROKBOY_API_KEY          API key (or XAI_API_KEY / OPENAI_API_KEY)
+  GROKBOY_BASE_URL         default https://api.x.ai/v1
+  GROKBOY_MODEL            default grok-4.6
+  GROKBOY_CONTEXT_CHARS    context budget (default 100000)
 
-Sessions are stored under ~/.grokboy/sessions/<id>.json
+Tools: shell, list_dir, read_file, write_file, report_done, report_blocked
+Sessions: ~/.grokboy/sessions/<id>.json
 "
     );
 }
@@ -148,7 +151,7 @@ async fn cmd_run(args: &[String]) -> Result<()> {
     session.push(ChatMessage::user(&prompt));
 
     let tool_ctx = ToolContext::new(session.cwd.clone());
-    let answer = run_agent(
+    let verdict = run_agent(
         &config,
         &mut session.messages,
         &tool_ctx,
@@ -158,17 +161,22 @@ async fn cmd_run(args: &[String]) -> Result<()> {
 
     session.touch();
     let path = save_session(&session)?;
-    println!("{answer}");
+    println!("{}", verdict.message());
     eprintln!(
-        "\n[session {} saved → {}]",
+        "\n[verdict: {} | session {} → {}]",
+        verdict.kind(),
         session.id,
         path.display()
     );
+    // Non-zero exit on blocked so scripts can detect fail-closed.
+    if matches!(verdict, AgentVerdict::Blocked(_)) {
+        return Err(anyhow!("agent blocked: {}", verdict.message()));
+    }
     Ok(())
 }
 
 async fn cmd_smoke() -> Result<()> {
-    println!("GrokBoy smoke (offline tools)…");
+    println!("GrokBoy smoke (offline P2)…");
     let stamp = uuid_like();
     let dir = std::env::temp_dir().join(format!("grokboy-smoke-{stamp}"));
     std::fs::create_dir_all(&dir).context("temp dir")?;
@@ -183,7 +191,7 @@ async fn cmd_smoke() -> Result<()> {
     .await;
     let w: serde_json::Value = serde_json::from_str(&w)?;
     assert_ok(&w, "write_file")?;
-    println!("  write_file  ok");
+    println!("  write_file     ok");
 
     // read_file
     let r = execute_tool(
@@ -197,7 +205,7 @@ async fn cmd_smoke() -> Result<()> {
     if r["content"].as_str() != Some("smoke ok\n第二行") {
         return Err(anyhow!("read_file content mismatch: {r}"));
     }
-    println!("  read_file   ok");
+    println!("  read_file      ok");
 
     // list_dir
     let l = execute_tool(&ctx, "list_dir", &json!({"path": "."}).to_string()).await;
@@ -212,7 +220,7 @@ async fn cmd_smoke() -> Result<()> {
     if !names.contains(&"note.txt") {
         return Err(anyhow!("list_dir missing note.txt: {l}"));
     }
-    println!("  list_dir    ok");
+    println!("  list_dir       ok");
 
     // shell
     let s = execute_tool(
@@ -226,7 +234,7 @@ async fn cmd_smoke() -> Result<()> {
     if s["exit_code"] != 0 {
         return Err(anyhow!("shell exit_code != 0: {s}"));
     }
-    println!("  shell       ok");
+    println!("  shell          ok");
 
     // path traversal should fail
     let bad = execute_tool(
@@ -239,17 +247,65 @@ async fn cmd_smoke() -> Result<()> {
     if bad.get("error").is_none() {
         return Err(anyhow!("expected path traversal error, got {bad}"));
     }
-    println!("  sandbox     ok");
+    println!("  sandbox        ok");
 
-    // tool definitions present
-    let defs = tool_definitions();
-    if defs.as_array().map(|a| a.len()).unwrap_or(0) != 4 {
-        return Err(anyhow!("expected 4 tool defs"));
+    // completion tools
+    let done = execute_tool(
+        &ctx,
+        "report_done",
+        &json!({"message": "smoke done"}).to_string(),
+    )
+    .await;
+    let done: serde_json::Value = serde_json::from_str(&done)?;
+    if done["status"] != "done" || done["message"] != "smoke done" {
+        return Err(anyhow!("report_done mismatch: {done}"));
     }
-    println!("  tool defs   ok");
+    let blocked = execute_tool(
+        &ctx,
+        "report_blocked",
+        &json!({"reason": "smoke blocked"}).to_string(),
+    )
+    .await;
+    let blocked: serde_json::Value = serde_json::from_str(&blocked)?;
+    if blocked["status"] != "blocked" || blocked["reason"] != "smoke blocked" {
+        return Err(anyhow!("report_blocked mismatch: {blocked}"));
+    }
+    if !is_completion_tool("report_done") || !is_completion_tool("report_blocked") {
+        return Err(anyhow!("is_completion_tool failed"));
+    }
+    println!("  completion     ok");
 
-    // session roundtrip in temp (does not require ~/.grokboy for this check —
-    // we still exercise Session serialize via save into temp using core types)
+    // tool definitions present (6)
+    let defs = tool_definitions();
+    let n_tools = defs.as_array().map(|a| a.len()).unwrap_or(0);
+    if n_tools != 6 {
+        return Err(anyhow!("expected 6 tool defs, got {n_tools}"));
+    }
+    println!("  tool defs      ok");
+
+    // context truncation
+    let mut msgs = vec![
+        ChatMessage::system(AGENT_SYSTEM),
+        ChatMessage::user("goal"),
+        ChatMessage::tool("t1", "X".repeat(50_000)),
+        ChatMessage::tool("t2", "Y".repeat(50_000)),
+        ChatMessage::tool("t3", "Z".repeat(50_000)),
+        ChatMessage::tool("t4", "W".repeat(50_000)),
+        ChatMessage::tool("t5", "V".repeat(50_000)),
+        ChatMessage::user("goal again"),
+    ];
+    let before = messages_char_len(&msgs);
+    truncate_messages(&mut msgs, 10_000);
+    let after = messages_char_len(&msgs);
+    if after >= before {
+        return Err(anyhow!("truncate did not shrink ({before} -> {after})"));
+    }
+    if msgs[0].text() != AGENT_SYSTEM {
+        return Err(anyhow!("truncate dropped system prompt"));
+    }
+    println!("  truncation     ok");
+
+    // session roundtrip
     let mut sess = Session::new(dir.clone());
     sess.push(ChatMessage::system(AGENT_SYSTEM));
     sess.push(ChatMessage::user("smoke"));
@@ -260,11 +316,10 @@ async fn cmd_smoke() -> Result<()> {
     if loaded.messages.len() != 2 {
         return Err(anyhow!("session roundtrip failed"));
     }
-    println!("  session     ok");
+    println!("  session        ok");
 
     let _ = std::fs::remove_dir_all(&dir);
 
-    // Optional live ping if API key is present (does not fail smoke).
     if Config::from_env().is_ok() {
         println!("  (API key present — skipping live call in smoke; use `run` to exercise)");
     } else {

docs/ACCEPTANCE.md

@@ -15,9 +15,10 @@
 - [x] Default model `grok-4.6`
 
 ## P2 — completion contract
-- [ ] `report_done` / blocked
-- [ ] Loop guard
-- [ ] Context truncation
+- [x] `report_done` / `report_blocked` stop the loop with a clear verdict
+- [x] Loop guard (identical tool rounds ×3 → blocked)
+- [x] Context truncation (`GROKBOY_CONTEXT_CHARS`, default ~100k)
+- [x] Offline smoke/tests cover P2 without API key
 
 ## P3 — browser
 - [ ] Playwright DOM path (no screenshot-first)