@elvatis_com/openclaw-cli-bridge-elvatis 2.9.0 → 2.10.0

package/README.md

@@ -2,7 +2,7 @@
 
 > OpenClaw plugin that bridges locally installed AI CLIs (Codex, Gemini, Claude Code, OpenCode, Pi) as model providers — with slash commands for instant model switching, restore, health testing, and model listing.
 
-**Current version:** `2.9.0`
+**Current version:** `2.10.0`
 
 ---
 
@@ -406,6 +406,18 @@ npm run ci          # lint + typecheck + test
 
 ## Changelog
 
+### v2.10.0
+- **fix:** cap effective timeout at 580s (under gateway's 600s `idleTimeoutSeconds`) so bridge fallback fires BEFORE gateway kills the request — eliminates the race condition where both compete to handle the timeout
+- **fix:** reduce Sonnet base timeout 420s→300s, Opus 420s→360s — ensures fallback triggers faster for stuck CLI sessions
+- **feat:** compact tool schema mode — when >8 tools, compress definitions to name+params only, cutting prompt size ~60%
+- **feat:** stale-output detection — if CLI produces no stdout for 120s, SIGTERM early instead of waiting full timeout
+- **feat:** adaptive message limits — reduce history from 20→12 messages when >10 tools to keep prompts smaller
+- **feat:** file-based debug log at `~/.openclaw/cli-bridge/debug.log` — `tail -f` for real-time request lifecycle visibility
+- **feat:** SSE progress comments every 30s so the webchat connection stays informed during long CLI runs
+- **feat:** SSE fallback notification — visible comment when a model times out and the bridge retries with fallback
+- **fix:** rescue tool_calls embedded inside content strings — handles models that wrap `{"tool_calls":[...]}` inside a `{"content":"..."}` wrapper
+- **fix:** parse robustness — debug logging on all parse paths to diagnose raw-JSON-instead-of-tool-calls issues
+
 ### v2.9.0
 - **feat:** enhanced `/status` dashboard with 5 new panels:
   - **Active Requests**: live in-flight requests with model, elapsed time, message/tool count, prompt preview

package/SKILL.md

@@ -68,4 +68,4 @@ On gateway restart, if any session has expired, a **WhatsApp alert** is sent aut
 
 See `README.md` for full configuration reference and architecture diagram.
 
-**Version:** 2.9.0
+**Version:** 2.10.0

package/openclaw.plugin.json

@@ -2,7 +2,7 @@
   "id": "openclaw-cli-bridge-elvatis",
   "slug": "openclaw-cli-bridge-elvatis",
   "name": "OpenClaw CLI Bridge",
-  "version": "2.9.0",
+  "version": "2.10.0",
   "license": "MIT",
   "description": "Phase 1: openai-codex auth bridge. Phase 2: local HTTP proxy routing model calls through gemini/claude CLIs (vllm provider).",
   "providers": [
@@ -43,8 +43,8 @@
           "type": "number"
         },
         "default": {
-          "cli-claude/claude-opus-4-6": 420000,
-          "cli-claude/claude-sonnet-4-6": 420000,
+          "cli-claude/claude-opus-4-6": 360000,
+          "cli-claude/claude-sonnet-4-6": 300000,
           "cli-claude/claude-haiku-4-5": 120000,
           "cli-gemini/gemini-2.5-pro": 300000,
           "cli-gemini/gemini-2.5-flash": 180000,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@elvatis_com/openclaw-cli-bridge-elvatis",
-  "version": "2.9.0",
+  "version": "2.10.0",
   "description": "Bridges gemini, claude, and codex CLI tools as OpenClaw model providers. Reads existing CLI auth without re-login.",
   "type": "module",
   "openclaw": {

package/src/cli-runner.ts

@@ -30,11 +30,15 @@ import {
 } from "./tool-protocol.js";
 import {
   MAX_MESSAGES,
+  MAX_MESSAGES_HEAVY_TOOLS,
+  TOOL_HEAVY_THRESHOLD,
   MAX_MSG_CHARS,
   DEFAULT_CLI_TIMEOUT_MS,
   TIMEOUT_GRACE_MS,
   MEDIA_TMP_DIR,
+  STALE_OUTPUT_TIMEOUT_MS,
 } from "./config.js";
+import { debugLog } from "./debug-log.js";
 
 // ──────────────────────────────────────────────────────────────────────────────
 // Message formatting
@@ -69,13 +73,16 @@ export type { ToolDefinition, CliToolResult } from "./tool-protocol.js";
  *   - role "tool": formatted as [Tool Result: name]
  *   - role "assistant" with tool_calls: formatted as [Assistant Tool Call: name(args)]
  */
-export function formatPrompt(messages: ChatMessage[]): string {
+export function formatPrompt(messages: ChatMessage[], toolCount = 0): string {
   if (messages.length === 0) return "";
 
+  // Reduce history when tool schemas dominate the prompt
+  const maxMsgs = toolCount > TOOL_HEAVY_THRESHOLD ? MAX_MESSAGES_HEAVY_TOOLS : MAX_MESSAGES;
+
   // Keep system message (if any) + last N non-system messages
   const system = messages.find((m) => m.role === "system");
   const nonSystem = messages.filter((m) => m.role !== "system");
-  const recent = nonSystem.slice(-MAX_MESSAGES);
+  const recent = nonSystem.slice(-maxMsgs);
   const truncated = system ? [system, ...recent] : recent;
 
   // Single short user message — send bare (no wrapping needed)
@@ -331,17 +338,20 @@ export function runCli(
     let timedOut = false;
     let killTimer: ReturnType<typeof setTimeout> | null = null;
     let timeoutTimer: ReturnType<typeof setTimeout> | null = null;
+    let staleTimer: ReturnType<typeof setInterval> | null = null;
+    let lastOutputAt = Date.now();
 
     const clearTimers = () => {
       if (timeoutTimer) { clearTimeout(timeoutTimer); timeoutTimer = null; }
       if (killTimer) { clearTimeout(killTimer); killTimer = null; }
+      if (staleTimer) { clearInterval(staleTimer); staleTimer = null; }
     };
 
-    // ── Timeout sequence: SIGTERM → grace → SIGKILL ──────────────────────
-    timeoutTimer = setTimeout(() => {
+    const doKill = (reason: string) => {
+      if (timedOut) return; // already killing
       timedOut = true;
-      const elapsed = Math.round(timeoutMs / 1000);
-      log(`[cli-bridge] timeout after ${elapsed}s for ${cmd}, sending SIGTERM`);
+      log(`[cli-bridge] ${reason} for ${cmd}, sending SIGTERM`);
+      debugLog("KILL", `${cmd} ${reason}`, { stdoutLen: stdout.length, stderrLen: stderr.length });
       proc.kill("SIGTERM");
 
       killTimer = setTimeout(() => {
@@ -350,14 +360,36 @@ export function runCli(
           proc.kill("SIGKILL");
         }
       }, TIMEOUT_GRACE_MS);
+    };
+
+    // ── Hard timeout: SIGTERM → grace → SIGKILL ──────────────────────────
+    timeoutTimer = setTimeout(() => {
+      doKill(`timeout after ${Math.round(timeoutMs / 1000)}s`);
     }, timeoutMs);
 
+    // ── Stale-output detection: kill if no stdout for STALE_OUTPUT_TIMEOUT_MS
+    if (STALE_OUTPUT_TIMEOUT_MS > 0) {
+      const checkInterval = 15_000; // check every 15s
+      staleTimer = setInterval(() => {
+        const silent = Date.now() - lastOutputAt;
+        if (silent >= STALE_OUTPUT_TIMEOUT_MS) {
+          doKill(`stale output — no stdout for ${Math.round(silent / 1000)}s`);
+        }
+      }, checkInterval);
+    }
+
     proc.stdin.write(prompt, "utf8", () => {
       proc.stdin.end();
     });
 
-    proc.stdout.on("data", (d: Buffer) => { stdout += d.toString(); });
-    proc.stderr.on("data", (d: Buffer) => { stderr += d.toString(); });
+    proc.stdout.on("data", (d: Buffer) => {
+      stdout += d.toString();
+      lastOutputAt = Date.now();
+    });
+    proc.stderr.on("data", (d: Buffer) => {
+      stderr += d.toString();
+      lastOutputAt = Date.now(); // stderr also counts as activity
+    });
 
     proc.on("close", (code) => {
       clearTimers();
@@ -770,8 +802,9 @@ export async function routeToCliRunner(
   timeoutMs: number,
   opts: RouteOptions = {}
 ): Promise<CliToolResult> {
-  const prompt = formatPrompt(messages);
-  const hasTools = !!(opts.tools?.length);
+  const toolCount = opts.tools?.length ?? 0;
+  const prompt = formatPrompt(messages, toolCount);
+  const hasTools = toolCount > 0;
 
   // Strip "vllm/" prefix if present — OpenClaw sends the full provider path
   // (e.g. "vllm/cli-claude/claude-sonnet-4-6") but the router only needs the

package/src/config.ts

@@ -25,8 +25,13 @@ export const DEFAULT_PROXY_API_KEY = "cli-bridge";
 /** Default base timeout for CLI subprocess responses (ms). Scales dynamically. */
 export const DEFAULT_PROXY_TIMEOUT_MS = 300_000; // 5 min
 
-/** Maximum effective timeout after dynamic scaling (ms). */
-export const MAX_EFFECTIVE_TIMEOUT_MS = 900_000; // 15 min
+/**
+ * Maximum effective timeout after dynamic scaling (ms).
+ * MUST be lower than the OpenClaw gateway's idleTimeoutSeconds (600s)
+ * so the bridge's own fallback fires BEFORE the gateway kills the request.
+ * 580s gives a 20s safety margin under the gateway's 600s hard limit.
+ */
+export const MAX_EFFECTIVE_TIMEOUT_MS = 580_000; // 9m 40s — under gateway's 600s
 
 /** Extra timeout per message beyond 10 in the conversation (ms). */
 export const TIMEOUT_PER_EXTRA_MSG_MS = 2_000;
@@ -47,9 +52,25 @@ export const DEFAULT_CLI_TIMEOUT_MS = 120_000; // 2 min
 /** Grace period between SIGTERM and SIGKILL when a timeout fires (ms). */
 export const TIMEOUT_GRACE_MS = 5_000;
 
+/**
+ * Stale output timeout — if a CLI subprocess produces no stdout for this long,
+ * assume it's stuck and SIGTERM early. 0 = disabled.
+ * Prevents waiting the full timeout when Claude CLI hangs silently.
+ */
+export const STALE_OUTPUT_TIMEOUT_MS = 120_000; // 2 min of silence → kill
+
 /** Max messages to include in the prompt sent to CLI subprocesses. */
 export const MAX_MESSAGES = 20;
 
+/**
+ * Reduced message limit when tools are heavy (> TOOL_HEAVY_THRESHOLD).
+ * Fewer history messages = smaller prompt = faster CLI response.
+ */
+export const MAX_MESSAGES_HEAVY_TOOLS = 12;
+
+/** Tool count threshold that triggers reduced message limit. */
+export const TOOL_HEAVY_THRESHOLD = 10;
+
 /** Max characters per message content before truncation. */
 export const MAX_MSG_CHARS = 4_000;
 
@@ -91,8 +112,8 @@ export const PROVIDER_SESSION_SWEEP_MS = 10 * 60 * 1_000; // 10 min
  *   - Fast/lightweight (Haiku, Flash, Mini): 120s
  */
 export const DEFAULT_MODEL_TIMEOUTS: Record<string, number> = {
-  "cli-claude/claude-opus-4-6":        420_000,  // 7 min
-  "cli-claude/claude-sonnet-4-6":      420_000,  // 7 min — prevent timeout→Haiku fallback on large sessions
+  "cli-claude/claude-opus-4-6":        360_000,  // 6 min — leaves room for dynamic scaling up to 580s cap
+  "cli-claude/claude-sonnet-4-6":      300_000,  // 5 min — was 7 min, reduced so fallback fires before gateway's 600s
   "cli-claude/claude-haiku-4-5":       120_000,  // 2 min
   "cli-gemini/gemini-2.5-pro":         300_000,  // 5 min — image generation needs more time
   "cli-gemini/gemini-2.5-flash":       180_000,  // 3 min

package/src/debug-log.ts

@@ -0,0 +1,55 @@
+/**
+ * debug-log.ts
+ *
+ * File-based debug logger for the CLI bridge.
+ * Writes to ~/.openclaw/cli-bridge/debug.log with automatic rotation at 5 MB.
+ *
+ * Usage:
+ *   tail -f ~/.openclaw/cli-bridge/debug.log
+ */
+
+import { appendFileSync, statSync, renameSync, mkdirSync } from "node:fs";
+import { join } from "node:path";
+import { homedir } from "node:os";
+
+const LOG_DIR = join(homedir(), ".openclaw", "cli-bridge");
+const LOG_FILE = join(LOG_DIR, "debug.log");
+const LOG_FILE_PREV = join(LOG_DIR, "debug.log.1");
+const MAX_LOG_SIZE = 5 * 1024 * 1024; // 5 MB
+
+let initialized = false;
+
+function ensureDir(): void {
+  if (initialized) return;
+  try { mkdirSync(LOG_DIR, { recursive: true }); } catch { /* exists */ }
+  initialized = true;
+}
+
+function rotate(): void {
+  try {
+    const stat = statSync(LOG_FILE);
+    if (stat.size > MAX_LOG_SIZE) {
+      try { renameSync(LOG_FILE, LOG_FILE_PREV); } catch { /* best effort */ }
+    }
+  } catch { /* file doesn't exist yet */ }
+}
+
+function ts(): string {
+  return new Date().toISOString();
+}
+
+/**
+ * Append a debug line to the log file.
+ * Non-blocking, never throws — logging must not crash the bridge.
+ */
+export function debugLog(category: string, message: string, data?: Record<string, unknown>): void {
+  try {
+    ensureDir();
+    rotate();
+    const extra = data ? ` ${JSON.stringify(data)}` : "";
+    appendFileSync(LOG_FILE, `${ts()} [${category}] ${message}${extra}\n`);
+  } catch { /* never crash on log failure */ }
+}
+
+/** Log path for display on status page / startup messages. */
+export const DEBUG_LOG_PATH = LOG_FILE;

package/src/proxy-server.ts

@@ -33,6 +33,7 @@ import {
   BITNET_SYSTEM_PROMPT,
   DEFAULT_MODEL_TIMEOUTS,
 } from "./config.js";
+import { debugLog, DEBUG_LOG_PATH } from "./debug-log.js";
 
 // ── Active request tracking ─────────────────────────────────────────────────
 
@@ -214,6 +215,7 @@ export function startProxyServer(opts: ProxyServerOptions): Promise<http.Server>
       opts.log(
         `[cli-bridge] proxy listening on :${opts.port}`
       );
+      debugLog("STARTUP", `proxy listening on :${opts.port}`, { debugLog: DEBUG_LOG_PATH });
       // unref() so the proxy server does not keep the Node.js event loop alive
       // when openclaw doctor or other short-lived CLI commands load plugins.
       // The gateway's own main loop keeps the process alive during normal operation.
@@ -389,6 +391,8 @@ async function handleRequest(
     const lastUserMsg = [...cleanMessages].reverse().find(m => m.role === "user");
     const promptPreview = typeof lastUserMsg?.content === "string" ? lastUserMsg.content.slice(0, 80) : "";
 
+    debugLog("REQ", `${model} start`, { msgs: cleanMessages.length, tools: tools?.length ?? 0, stream, media: mediaFiles.length, promptPreview: promptPreview.slice(0, 60) });
+
     // Track active request for dashboard
     activeRequests.set(id, { id, model, startedAt: Date.now(), messageCount: cleanMessages.length, toolCount: tools?.length ?? 0, promptPreview });
 
@@ -805,6 +809,7 @@ async function handleRequest(
     const toolExtra = (tools?.length ?? 0) * TIMEOUT_PER_TOOL_MS;
     const effectiveTimeout = Math.min(baseTimeout + msgExtra + toolExtra, MAX_EFFECTIVE_TIMEOUT_MS);
     opts.log(`[cli-bridge] ${model} session=${session.id} timeout: ${Math.round(effectiveTimeout / 1000)}s (base=${Math.round(baseTimeout / 1000)}s${perModelTimeout ? " per-model" : ""}, +${Math.round(msgExtra / 1000)}s msgs, +${Math.round(toolExtra / 1000)}s tools)`);
+    debugLog("TIMEOUT", `${model} effective=${Math.round(effectiveTimeout / 1000)}s`, { base: Math.round(baseTimeout / 1000), msgExtra: Math.round(msgExtra / 1000), toolExtra: Math.round(toolExtra / 1000), cap: Math.round(MAX_EFFECTIVE_TIMEOUT_MS / 1000) });
 
     // ── SSE keepalive: send headers early so OpenClaw doesn't read-timeout ──
     let sseHeadersSent = false;
@@ -821,17 +826,35 @@ async function handleRequest(
       keepaliveInterval = setInterval(() => { res.write(": keepalive\n\n"); }, SSE_KEEPALIVE_INTERVAL_MS);
     }
 
+    // ── Progress notifications: send visible status updates to the webchat ──
+    // Users shouldn't stare at a blank screen for minutes without feedback.
+    let progressInterval: ReturnType<typeof setInterval> | null = null;
+    const PROGRESS_INTERVAL_MS = 30_000; // 30s between updates
+    if (stream && sseHeadersSent) {
+      const progressStart = Date.now();
+      progressInterval = setInterval(() => {
+        const elapsed = Math.round((Date.now() - progressStart) / 1000);
+        const timeoutSec = Math.round(effectiveTimeout / 1000);
+        // Send an SSE comment with progress info — visible in raw SSE but won't render as content
+        // Also send a small content delta that OpenClaw can show as typing indicator
+        res.write(`: progress ${elapsed}s/${timeoutSec}s — ${model} processing\n\n`);
+      }, PROGRESS_INTERVAL_MS);
+    }
+
     const cliStart = Date.now();
     try {
       result = await routeToCliRunner(model, cleanMessages, effectiveTimeout, routeOpts);
+      const latencyMs = Date.now() - cliStart;
       const estCompletionTokens = estimateTokens(result.content ?? "");
-      metrics.recordRequest(model, Date.now() - cliStart, true, estPromptTokens, estCompletionTokens, promptPreview);
+      metrics.recordRequest(model, latencyMs, true, estPromptTokens, estCompletionTokens, promptPreview);
       providerSessions.recordRun(session.id, false);
+      debugLog("OK", `${model} completed in ${(latencyMs / 1000).toFixed(1)}s`, { toolCalls: result.tool_calls?.length ?? 0, contentLen: result.content?.length ?? 0 });
     } catch (err) {
       const primaryDuration = Date.now() - cliStart;
       const msg = (err as Error).message;
       // ── Model fallback: retry once with a lighter model if configured ────
       const isTimeout = msg.includes("timeout:") || msg.includes("exit 143") || msg.includes("exited 143");
+      debugLog("FAIL", `${model} failed after ${(primaryDuration / 1000).toFixed(1)}s`, { isTimeout, error: msg.slice(0, 200) });
       // Record the run (with timeout flag) — session is preserved, not deleted
       providerSessions.recordRun(session.id, isTimeout);
       const fallbackModel = opts.modelFallbacks?.[model];
@@ -839,6 +862,11 @@ async function handleRequest(
         metrics.recordRequest(model, primaryDuration, false, estPromptTokens, undefined, promptPreview);
         const reason = isTimeout ? `timeout by supervisor, session=${session.id} preserved` : msg;
         opts.warn(`[cli-bridge] ${model} failed (${reason}), falling back to ${fallbackModel}`);
+        debugLog("FALLBACK", `${model} → ${fallbackModel}`, { reason: isTimeout ? "timeout" : "error", primaryDuration: Math.round(primaryDuration / 1000) });
+        // Notify the user via SSE that we're retrying with a different model
+        if (sseHeadersSent) {
+          res.write(`: fallback — ${model} ${isTimeout ? "timed out" : "failed"} after ${Math.round(primaryDuration / 1000)}s, retrying with ${fallbackModel}\n\n`);
+        }
         const fallbackStart = Date.now();
         try {
           result = await routeToCliRunner(fallbackModel, cleanMessages, effectiveTimeout, routeOpts);
@@ -877,6 +905,7 @@ async function handleRequest(
       }
     } finally {
       if (keepaliveInterval) clearInterval(keepaliveInterval);
+      if (progressInterval) clearInterval(progressInterval);
       cleanupMediaFiles(mediaFiles);
       activeRequests.delete(id);
     }

package/src/tool-protocol.ts

@@ -10,6 +10,7 @@
  */
 
 import { randomBytes } from "node:crypto";
+import { debugLog } from "./debug-log.js";
 
 // ──────────────────────────────────────────────────────────────────────────────
 // Types
@@ -46,13 +47,34 @@ export interface CliToolResult {
  * Build a text block describing available tools and response format instructions.
  * This block is prepended to the system message (or added as a new system message).
  */
+/** Threshold: when tool count exceeds this, use compact schema to reduce prompt size. */
+const COMPACT_TOOL_THRESHOLD = 8;
+
+/**
+ * Build a compact tool description: name + required param names only.
+ * Cuts prompt size by ~60-70% for large tool sets.
+ */
+function compactToolDescription(t: ToolDefinition): string {
+  const fn = t.function;
+  const params = fn.parameters as { properties?: Record<string, unknown>; required?: string[] };
+  const required = params?.required ?? Object.keys(params?.properties ?? {});
+  const paramList = required.length > 0 ? `(${required.join(", ")})` : "()";
+  return `- ${fn.name}${paramList}: ${fn.description}`;
+}
+
+/**
+ * Build a full tool description: name, description, and full JSON schema.
+ */
+function fullToolDescription(t: ToolDefinition): string {
+  const fn = t.function;
+  const params = JSON.stringify(fn.parameters);
+  return `- name: ${fn.name}\n  description: ${fn.description}\n  parameters: ${params}`;
+}
+
 export function buildToolPromptBlock(tools: ToolDefinition[]): string {
+  const useCompact = tools.length > COMPACT_TOOL_THRESHOLD;
   const toolDescriptions = tools
-    .map((t) => {
-      const fn = t.function;
-      const params = JSON.stringify(fn.parameters);
-      return `- name: ${fn.name}\n  description: ${fn.description}\n  parameters: ${params}`;
-    })
+    .map(useCompact ? compactToolDescription : fullToolDescription)
     .join("\n");
 
   return [
@@ -67,6 +89,7 @@ export function buildToolPromptBlock(tools: ToolDefinition[]): string {
     '{"content":"<your text response>"}',
     "",
     "Do NOT include any text outside the JSON. Do NOT wrap in markdown code blocks.",
+    useCompact ? "Call ONE tool at a time. Do NOT batch multiple tool calls." : "",
     "",
     "Available tools:",
     toolDescriptions,
@@ -117,6 +140,7 @@ export function buildToolCallJsonSchema(): object {
  */
 export function parseToolCallResponse(text: string): CliToolResult {
   const trimmed = text.trim();
+  const preview = trimmed.slice(0, 120);
 
   // Check for Claude's --output-format json wrapper FIRST.
   // Claude returns: { "type": "result", "result": "..." }
@@ -124,30 +148,48 @@ export function parseToolCallResponse(text: string): CliToolResult {
   const claudeResult = tryExtractClaudeJsonResult(trimmed);
   if (claudeResult) {
     const inner = tryParseJson(claudeResult);
-    if (inner) return normalizeResult(inner);
+    if (inner) {
+      const result = normalizeResult(inner);
+      debugLog("PARSE", `claude-json → ${result.tool_calls ? "tool_calls" : "content"}`, { toolCalls: result.tool_calls?.length ?? 0 });
+      return result;
+    }
     // Claude result is plain text
+    debugLog("PARSE", "claude-json → plain text", { len: claudeResult.length });
     return { content: claudeResult };
   }
 
   // Try direct JSON parse (for non-Claude outputs)
   const parsed = tryParseJson(trimmed);
-  if (parsed) return normalizeResult(parsed);
+  if (parsed) {
+    const result = normalizeResult(parsed);
+    debugLog("PARSE", `direct-json → ${result.tool_calls ? "tool_calls" : "content"}`, { toolCalls: result.tool_calls?.length ?? 0 });
+    return result;
+  }
 
   // Try extracting JSON from markdown code blocks: ```json ... ```
   const codeBlock = tryExtractCodeBlock(trimmed);
   if (codeBlock) {
     const inner = tryParseJson(codeBlock);
-    if (inner) return normalizeResult(inner);
+    if (inner) {
+      const result = normalizeResult(inner);
+      debugLog("PARSE", `code-block → ${result.tool_calls ? "tool_calls" : "content"}`, { toolCalls: result.tool_calls?.length ?? 0 });
+      return result;
+    }
   }
 
   // Try finding a JSON object anywhere in the text
   const embedded = tryExtractEmbeddedJson(trimmed);
   if (embedded) {
     const inner = tryParseJson(embedded);
-    if (inner) return normalizeResult(inner);
+    if (inner) {
+      const result = normalizeResult(inner);
+      debugLog("PARSE", `embedded-json → ${result.tool_calls ? "tool_calls" : "content"}`, { toolCalls: result.tool_calls?.length ?? 0 });
+      return result;
+    }
   }
 
   // Fallback: treat entire text as content
+  debugLog("PARSE", "no JSON found → raw content", { len: trimmed.length, preview });
   return { content: trimmed || null };
 }
 
@@ -167,11 +209,17 @@ function normalizeResult(obj: Record<string, unknown>): CliToolResult {
           : JSON.stringify(tc.arguments ?? {}),
       },
     }));
-    return { content: null, tool_calls: toolCalls };
+    // If the model also returned a content string alongside tool_calls, include it
+    const content = typeof obj.content === "string" ? obj.content : null;
+    return { content, tool_calls: toolCalls };
   }
 
-  // Check for content field
+  // Check for content field — but rescue embedded tool_calls JSON from inside content strings.
+  // Models sometimes wrap tool calls inside a content string:
+  //   {"content":"I'll write that file.\n{\"tool_calls\":[...]}"}
   if (typeof obj.content === "string") {
+    const rescued = tryRescueToolCallsFromContent(obj.content);
+    if (rescued) return rescued;
     return { content: obj.content };
   }
 
@@ -179,6 +227,41 @@ function normalizeResult(obj: Record<string, unknown>): CliToolResult {
   return { content: JSON.stringify(obj) };
 }
 
+/**
+ * Rescue tool_calls embedded inside a content string.
+ * Handles cases where the model wraps tool calls in a content field:
+ *   {"content":"Some text\n{\"tool_calls\":[...]}"}
+ *   {"content":"{\"tool_calls\":[{\"name\":\"write\",...}]}"}
+ */
+function tryRescueToolCallsFromContent(content: string): CliToolResult | null {
+  // Only attempt rescue if content contains the tool_calls signature
+  if (!content.includes('"tool_calls"') && !content.includes("tool_calls")) return null;
+
+  // Try to find embedded JSON with tool_calls
+  const embedded = tryExtractEmbeddedJson(content);
+  if (!embedded) return null;
+
+  const parsed = tryParseJson(embedded);
+  if (!parsed || !Array.isArray(parsed.tool_calls) || parsed.tool_calls.length === 0) return null;
+
+  // Extract the text content before the JSON (if any)
+  const jsonStart = content.indexOf(embedded);
+  const textBefore = jsonStart > 0 ? content.slice(0, jsonStart).trim() : null;
+
+  const toolCalls: ToolCall[] = parsed.tool_calls.map((tc: Record<string, unknown>) => ({
+    id: generateCallId(),
+    type: "function" as const,
+    function: {
+      name: String(tc.name ?? ""),
+      arguments: typeof tc.arguments === "string"
+        ? tc.arguments
+        : JSON.stringify(tc.arguments ?? {}),
+    },
+  }));
+
+  return { content: textBefore || null, tool_calls: toolCalls };
+}
+
 function tryParseJson(text: string): Record<string, unknown> | null {
   try {
     const obj = JSON.parse(text);

package/test/config.test.ts

@@ -38,7 +38,7 @@ describe("config.ts exports", () => {
     expect(DEFAULT_PROXY_TIMEOUT_MS).toBe(300_000);
     expect(DEFAULT_CLI_TIMEOUT_MS).toBe(120_000);
     expect(TIMEOUT_GRACE_MS).toBe(5_000);
-    expect(MAX_EFFECTIVE_TIMEOUT_MS).toBe(900_000);
+    expect(MAX_EFFECTIVE_TIMEOUT_MS).toBe(580_000); // under gateway's 600s
     expect(SESSION_TTL_MS).toBe(30 * 60 * 1000);
     expect(CLEANUP_INTERVAL_MS).toBe(5 * 60 * 1000);
     expect(SESSION_KILL_GRACE_MS).toBe(5_000);
@@ -61,8 +61,8 @@ describe("config.ts exports", () => {
   });
 
   it("exports per-model timeouts for all major models", () => {
-    expect(DEFAULT_MODEL_TIMEOUTS["cli-claude/claude-opus-4-6"]).toBe(420_000);
-    expect(DEFAULT_MODEL_TIMEOUTS["cli-claude/claude-sonnet-4-6"]).toBe(420_000);
+    expect(DEFAULT_MODEL_TIMEOUTS["cli-claude/claude-opus-4-6"]).toBe(360_000);
+    expect(DEFAULT_MODEL_TIMEOUTS["cli-claude/claude-sonnet-4-6"]).toBe(300_000);
     expect(DEFAULT_MODEL_TIMEOUTS["cli-claude/claude-haiku-4-5"]).toBe(120_000);
     expect(DEFAULT_MODEL_TIMEOUTS["cli-gemini/gemini-2.5-pro"]).toBe(300_000);
     expect(DEFAULT_MODEL_TIMEOUTS["cli-gemini/gemini-2.5-flash"]).toBe(180_000);

package/test/session-manager.test.ts

@@ -63,8 +63,10 @@ vi.mock("../src/workdir.js", () => ({
 }));
 
 // Mock config module — provide all constants needed by session-manager.ts and cli-runner.ts
-vi.mock("../src/config.js", async () => {
+vi.mock("../src/config.js", async (importOriginal) => {
+  const actual = await importOriginal<typeof import("../src/config.js")>();
   return {
+    ...actual,
     SESSION_TTL_MS: 30 * 60 * 1000,
     CLEANUP_INTERVAL_MS: 5 * 60 * 1000,
     SESSION_KILL_GRACE_MS: 5_000,