@wrongstack/core 0.148.0 → 0.236.0

package/dist/execution/index.js

@@ -9,7 +9,9 @@ import { EventEmitter } from 'events';
 // src/utils/expect-defined.ts
 function expectDefined(value, label) {
   if (value === null || value === void 0) {
-    throw new Error("Expected value to be defined");
+    const err = new Error("Expected value to be defined");
+    err.name = "ExpectDefinedError";
+    throw err;
   }
   return value;
 }
@@ -58,22 +60,31 @@ function estimateToolResultTokens(content) {
 function estimateTextTokens(text) {
   return RoughTokenEstimate(text);
 }
+function computeMessageTokens(msg) {
+  if (typeof msg.content === "string") return estimateTextTokens(msg.content);
+  let total = 0;
+  for (const b of msg.content) {
+    if (b.type === "text") total += estimateTextTokens(b.text);
+    else if (b.type === "tool_use") total += estimateToolInputTokens(b.input);
+    else if (b.type === "tool_result") total += estimateToolResultTokens(b.content);
+    else total += RoughTokenEstimate(JSON.stringify(b));
+  }
+  return total;
+}
 function estimateMessageTokens(messages) {
   let total = 0;
   for (const m of messages) {
-    if (typeof m.content === "string") {
-      total += estimateTextTokens(m.content);
-    } else {
-      for (const b of m.content) {
-        if (b.type === "text") total += estimateTextTokens(b.text);
-        else if (b.type === "tool_use") total += estimateToolInputTokens(b.input);
-        else if (b.type === "tool_result") total += estimateToolResultTokens(b.content);
-      }
+    if (typeof m._estTokens === "number" && m._estTokens > 0) {
+      total += m._estTokens;
+      continue;
     }
+    total += computeMessageTokens(m);
   }
   return total;
 }
 function estimateToolDefTokens(tool) {
+  const cached = tool._estDefTokens;
+  if (typeof cached === "number" && cached > 0) return cached;
   return RoughTokenEstimate(tool.name) + RoughTokenEstimate(tool.description ?? "") + RoughTokenEstimate(JSON.stringify(tool.inputSchema));
 }
 function estimateRequestTokens(messages, systemPrompt, tools, calibrationKey = CALIBRATION_GLOBAL_KEY) {
@@ -83,6 +94,11 @@ function estimateRequestTokens(messages, systemPrompt, tools, calibrationKey = C
   } else if (Array.isArray(messages)) {
     for (const m of messages) {
       if (typeof m === "object" && m !== null && "content" in m) {
+        const cached = m._estTokens;
+        if (typeof cached === "number" && cached > 0) {
+          messagesTokens += cached;
+          continue;
+        }
         const content = m.content;
         if (typeof content === "string") {
           messagesTokens += RoughTokenEstimate(content);
@@ -267,6 +283,18 @@ function findPreserveStart(messages, preserveK) {
 }
 function eliseOldToolResults(messages, opts) {
   const preserveStart = findPreserveStart(messages, opts.preserveK);
+  let hasOversized = false;
+  for (let i = 0; i < preserveStart && !hasOversized; i++) {
+    const msg = messages[i];
+    if (!msg || !Array.isArray(msg.content)) continue;
+    for (const b of msg.content) {
+      if (b.type === "tool_result" && estimateToolResultTokens(b.content) >= opts.eliseThreshold) {
+        hasOversized = true;
+        break;
+      }
+    }
+  }
+  if (!hasOversized) return { messages, saved: 0, changed: false };
   let saved = 0;
   let changed = false;
   const next = new Array(messages.length);
@@ -1245,6 +1273,15 @@ var AutoCompactionMiddleware = class _AutoCompactionMiddleware {
   static NOOP_RETRY_DELTA_TOKENS = 2e3;
   /** Tracks the most recent no-op attempt so we can avoid re-firing per turn. */
   lastNoopAttempt = null;
+  /**
+   * Cached token estimate from the last handler() invocation. When the
+   * message count and tool count haven't changed since the last estimate
+   * (autonomous idle loops), we skip the expensive O(n) token estimation
+   * and reuse this value. Reset to -1 when the context changes.
+   */
+  _cachedTokens = -1;
+  _cachedMsgCount = -1;
+  _cachedToolCount = -1;
   /**
    * @param compactor        Compactor to use for compaction.
    * @param maxContext Provider's max context window in tokens.
@@ -1280,12 +1317,24 @@ var AutoCompactionMiddleware = class _AutoCompactionMiddleware {
   }
   handler() {
     return async (ctx, next) => {
-      const tokens = this._estimator ? this._estimator(ctx) : estimateRequestTokensCalibrated(
-        ctx.messages,
-        ctx.systemPrompt,
-        ctx.tools ?? [],
-        `${ctx.provider?.id ?? "unknown"}/${ctx.model}`
-      ).total;
+      const msgCount = ctx.messages.length;
+      const toolCount = (ctx.tools ?? []).length;
+      let tokens;
+      if (this._estimator) {
+        tokens = this._estimator(ctx);
+      } else if (msgCount === this._cachedMsgCount && toolCount === this._cachedToolCount && this._cachedTokens >= 0) {
+        tokens = this._cachedTokens;
+      } else {
+        tokens = estimateRequestTokensCalibrated(
+          ctx.messages,
+          ctx.systemPrompt,
+          ctx.tools ?? [],
+          `${ctx.provider?.id ?? "unknown"}/${ctx.model}`
+        ).total;
+        this._cachedTokens = tokens;
+        this._cachedMsgCount = msgCount;
+        this._cachedToolCount = toolCount;
+      }
       const load = tokens / this._maxContext;
       const policy = this.policyProvider?.(ctx);
       const thresholds = policy?.thresholds ?? {
@@ -1566,7 +1615,7 @@ function createToolOutputSerializer(opts = {}) {
 }
 
 // src/execution/tool-executor.ts
-var ToolExecutor = class {
+var ToolExecutor = class _ToolExecutor {
   constructor(registry, opts) {
     this.registry = registry;
     this.opts = opts;
@@ -1578,6 +1627,10 @@ var ToolExecutor = class {
   }
   registry;
   opts;
+  /** Minimum gap between coalesced `partial_output` tool.progress emits. */
+  static PROGRESS_EMIT_INTERVAL_MS = 100;
+  /** Max chars of accumulated stream text carried per coalesced emit. */
+  static PROGRESS_TAIL_CHARS = 16384;
   serializer;
   iterationTimeoutMs;
   maxToolTimeoutMs;
@@ -1623,9 +1676,6 @@ Please call the tool again with arguments that match its inputSchema. You can us
         return { result, tool, durationMs: Date.now() - start };
       }
       const toolDangerousCaps = getDangerousCapabilities(tool);
-      if (toolDangerousCaps.length > 0) {
-        if (this.opts.events) ;
-      }
       if (hasMalformedArguments(use.input)) {
         const result = this.malformedInputResult(use, extractMalformedRaw(use.input));
         budget = this.decrementBudget(result, budget);
@@ -1863,17 +1913,48 @@ ${post.additionalContext}` };
       throw new Error(`Tool "${tool.name}" does not support streaming execution`);
     }
     const stream = tool.executeStream(input, ctx, { signal });
-    for await (const ev of stream) {
-      if (ev.type === "final") {
-        finalOutput = ev.output;
-        sawFinal = true;
-        break;
-      }
+    const iter = stream[Symbol.asyncIterator]();
+    let progressTail = "";
+    let lastProgressEmitAt = 0;
+    const emitProgress = (ev) => {
       this.opts.events?.emit("tool.progress", {
         name: tool.name,
         id: toolUseId ?? "<unknown>",
         event: ev
       });
+    };
+    const flushProgressTail = (force) => {
+      if (progressTail.length === 0) return;
+      const now = Date.now();
+      if (!force && now - lastProgressEmitAt < _ToolExecutor.PROGRESS_EMIT_INTERVAL_MS) return;
+      const text = progressTail;
+      progressTail = "";
+      lastProgressEmitAt = now;
+      emitProgress({ type: "partial_output", text });
+    };
+    try {
+      while (true) {
+        const { done, value: ev } = await iter.next();
+        if (done) break;
+        if (ev.type === "final") {
+          finalOutput = ev.output;
+          sawFinal = true;
+          break;
+        }
+        if (ev.type === "partial_output" && typeof ev.text === "string") {
+          progressTail += ev.text;
+          if (progressTail.length > _ToolExecutor.PROGRESS_TAIL_CHARS) {
+            progressTail = progressTail.slice(-_ToolExecutor.PROGRESS_TAIL_CHARS);
+          }
+          flushProgressTail(false);
+          continue;
+        }
+        flushProgressTail(true);
+        emitProgress(ev);
+      }
+      flushProgressTail(true);
+    } finally {
+      await iter.return?.(void 0);
     }
     if (!sawFinal) {
       throw new Error(`tool "${tool.name}" executeStream completed without a 'final' event`);
@@ -1984,9 +2065,11 @@ function extractMalformedRaw(input) {
 
 // src/utils/assert-never.ts
 function assertNever(x, message) {
-  throw new Error(
+  const err = new Error(
     `Unhandled case: ${JSON.stringify(x)}`
   );
+  err.name = "AssertNeverError";
+  throw err;
 }
 
 // src/utils/regex-guard.ts
@@ -2033,7 +2116,13 @@ var DoneConditionChecker = class {
       const result = compileUserRegex(condition.pattern, "");
       this.compiledRegex = result.ok ? result.regex : null;
       if (!result.ok) {
-        console.warn(`[DoneConditionChecker] Invalid regex pattern "${condition.pattern}": ${result.reason}`);
+        console.warn(JSON.stringify({
+          level: "warn",
+          event: "autonomous.done_condition_invalid_regex",
+          pattern: condition.pattern,
+          reason: result.reason,
+          timestamp: (/* @__PURE__ */ new Date()).toISOString()
+        }));
       }
     } else {
       this.compiledRegex = null;
@@ -2267,9 +2356,13 @@ function projectSlug(absRoot) {
 function slugify(name) {
   return name.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "").slice(0, 40) || "project";
 }
+function wstackGlobalRoot() {
+  const fromEnv = process.env["WRONGSTACK_HOME"];
+  if (fromEnv && fromEnv.trim().length > 0) return path2.resolve(fromEnv);
+  return path2.join(os.homedir(), ".wrongstack");
+}
 function resolveWstackPaths(opts) {
-  const home = opts.userHome ?? os.homedir();
-  const globalRoot = opts.globalRoot ?? path2.join(home, ".wrongstack");
+  const globalRoot = opts.globalRoot ?? (opts.userHome ? path2.join(opts.userHome, ".wrongstack") : wstackGlobalRoot());
   const hash = projectHash(opts.projectRoot);
   const slug = projectSlug(opts.projectRoot);
   const projectDir = path2.join(globalRoot, "projects", slug);
@@ -2326,12 +2419,24 @@ async function loadGoal(filePath) {
   try {
     const parsed = JSON.parse(raw);
     if (parsed?.version !== 1 || typeof parsed.goal !== "string" || !Array.isArray(parsed.journal)) {
-      console.warn(`[goal-store] Corrupt goal.json at ${filePath} \u2014 invalid schema. Consider deleting it and re-creating.`);
+      console.warn(JSON.stringify({
+        level: "warn",
+        event: "goal_store.invalid_schema",
+        path: filePath,
+        message: "invalid schema \u2014 consider deleting and re-creating",
+        timestamp: (/* @__PURE__ */ new Date()).toISOString()
+      }));
       return null;
     }
     return parsed;
   } catch {
-    console.warn(`[goal-store] Corrupt goal.json at ${filePath} \u2014 JSON parse failed. Consider deleting it and re-creating.`);
+    console.warn(JSON.stringify({
+      level: "warn",
+      event: "goal_store.parse_failed",
+      path: filePath,
+      message: "JSON parse failed \u2014 consider deleting and re-creating",
+      timestamp: (/* @__PURE__ */ new Date()).toISOString()
+    }));
     return null;
   }
 }
@@ -2406,6 +2511,25 @@ var RISK_LEVELS = {
   high: 2,
   critical: 3
 };
+function createTieredBrainArbiter(opts) {
+  return {
+    async decide(request) {
+      const policyDecision = await opts.policy.decide(request);
+      if (policyDecision.type !== "ask_human") return policyDecision;
+      const ceiling = opts.getMaxAutoRisk?.() ?? "medium";
+      if (!opts.autonomous || ceiling === "off") return policyDecision;
+      const ceilingLevel = ceiling === "all" ? 3 : RISK_LEVELS[ceiling] ?? 1;
+      const requestLevel = RISK_LEVELS[request.risk] ?? 2;
+      if (requestLevel > ceilingLevel) return policyDecision;
+      try {
+        const llmDecision = await opts.autonomous.decide(request);
+        if (llmDecision.type === "answer") return llmDecision;
+      } catch {
+      }
+      return policyDecision;
+    }
+  };
+}
 function createAutonomyBrain(opts) {
   const maxRisk = opts.maxAutoRisk ?? "high";
   const maxRiskLevel = RISK_LEVELS[maxRisk] ?? 2;
@@ -2612,7 +2736,14 @@ var EternalAutonomyEngine = class {
   stop() {
     this.stopRequested = true;
     this.currentCtrl?.abort();
-    void this.persistEngineState("stopped").catch(() => {
+    void this.persistEngineState("stopped").catch((err) => {
+      console.error(JSON.stringify({
+        level: "error",
+        event: "engine.persist_state_failed",
+        message: err instanceof Error ? err.message : String(err),
+        context: { expectedState: "stopped" },
+        timestamp: (/* @__PURE__ */ new Date()).toISOString()
+      }));
     });
     this.state = "stopped";
   }
@@ -3579,6 +3710,7 @@ var SubagentBudget = class _SubagentBudget {
 function makeAgentSubagentRunner(opts) {
   const format = opts.formatTaskInput ?? defaultFormatTaskInput;
   return async (task, ctx) => {
+    const taskStartedAt = Date.now();
     const factoryResult = await opts.factory(ctx.config);
     const { agent, events } = factoryResult;
     const detachFleet = opts.fleetBus?.attach(ctx.subagentId, events, task.id);
@@ -3675,7 +3807,7 @@ function makeAgentSubagentRunner(opts) {
       }),
       events.on("provider.text_delta", (e) => {
         ctx.budget.markActivity();
-        streamingTextAcc = (streamingTextAcc + e.text).slice(-200);
+        streamingTextAcc = (streamingTextAcc + e.text).slice(-2e3);
       })
     );
     const onParentAbort = () => aborter.abort();
@@ -3683,6 +3815,15 @@ function makeAgentSubagentRunner(opts) {
     let result;
     try {
       result = await agent.run(format(task, ctx.config), { signal: aborter.signal });
+      events.emit("subagent.task_completed", {
+        subagentId: ctx.subagentId,
+        taskId: task.id,
+        status: result.status === "done" ? "success" : "failed",
+        iterations: result.iterations,
+        toolCalls: ctx.budget.usage().toolCalls,
+        durationMs: Date.now() - taskStartedAt,
+        finalText: result.finalText?.trim() || void 0
+      });
     } finally {
       detachFleet?.();
       ctx.signal.removeEventListener("abort", onParentAbort);
@@ -3718,21 +3859,40 @@ function makeAgentSubagentRunner(opts) {
       if (budgetError) throw budgetError;
     }
     if (result.status === "failed") {
-      throw result.error instanceof Error ? result.error : new Error(String(result.error ?? "agent failed"));
+      throw result.error instanceof AgentError ? result.error : new AgentError({
+        message: result.error instanceof Error ? result.error.message : String(result.error ?? "agent failed"),
+        code: ERROR_CODES.AGENT_RUN_FAILED,
+        cause: result.error
+      });
     }
     if (result.status === "aborted") {
-      throw new Error("agent aborted");
+      throw new AgentError({
+        message: "agent aborted",
+        code: ERROR_CODES.AGENT_ABORTED
+      });
     }
     if (result.status === "max_iterations") {
-      throw new Error("agent exhausted iteration limit");
+      throw new AgentError({
+        message: "agent exhausted iteration limit",
+        code: ERROR_CODES.AGENT_ITERATION_LIMIT,
+        recoverable: true
+      });
     }
     const usage = ctx.budget.usage();
     const finalText = (result.finalText ?? "").trim();
     if (finalText.length === 0 && usage.toolCalls === 0) {
-      throw new Error("empty response");
+      throw new AgentError({
+        message: "empty response \u2014 agent produced no text and no tool calls",
+        code: ERROR_CODES.AGENT_RUN_FAILED,
+        context: { iterations: result.iterations }
+      });
     }
     if (finalText.length === 0 && lastToolFailed !== null) {
-      throw new Error(`tool failed: ${lastToolFailed}`);
+      throw new AgentError({
+        message: `unrecovered tool failure: ${lastToolFailed} \u2014 agent ended turn without acknowledging the error`,
+        code: ERROR_CODES.AGENT_RUN_FAILED,
+        context: { tool: lastToolFailed, iterations: result.iterations }
+      });
     }
     return {
       result: result.finalText,
@@ -3764,11 +3924,11 @@ var HEAVY_BUDGET = {
 };
 var TOOLS = {
   /** Pure read/inspect — safe for analysis and review agents. */
-  read: ["read", "grep", "glob", "search", "tree"],
+  read: ["read", "grep", "glob", "search", "tree", "mailbox"],
   /** Read + structured inspection (logs, diffs, json, dependency audit). */
-  inspect: ["read", "grep", "glob", "search", "tree", "json", "diff", "logs", "audit"],
+  inspect: ["read", "grep", "glob", "search", "tree", "json", "diff", "logs", "audit", "mailbox"],
   /** Read + edit (no shell). For agents that write code/docs but don't run it. */
-  write: ["read", "grep", "glob", "search", "tree", "write", "edit", "replace", "patch"],
+  write: ["read", "grep", "glob", "search", "tree", "write", "edit", "replace", "patch", "mailbox"],
   /** Full build loop: edit + run (lint/format/typecheck/test/bash). */
   build: [
     "read",
@@ -3785,16 +3945,17 @@ var TOOLS = {
     "lint",
     "format",
     "typecheck",
-    "test"
+    "test",
+    "mailbox"
   ],
   /** Version control. */
   vcs: ["read", "grep", "glob", "git", "diff"],
   /** Dependency management + CVE audit. */
-  deps: ["read", "grep", "glob", "install", "outdated", "audit", "json"],
+  deps: ["read", "grep", "glob", "install", "outdated", "audit", "json", "mailbox"],
   /** Documentation authoring. */
-  docs: ["read", "grep", "glob", "search", "tree", "write", "edit", "document"],
+  docs: ["read", "grep", "glob", "search", "tree", "write", "edit", "document", "mailbox"],
   /** Web research. */
-  research: ["read", "grep", "glob", "search", "fetch"]
+  research: ["read", "grep", "glob", "search", "fetch", "mailbox"]
 };
 
 // src/coordination/agents/phase1-discovery.ts
@@ -4592,15 +4753,44 @@ Working rules:
       id: "e2e",
       name: "E2E",
       role: "e2e",
-      tools: [...TOOLS.build, "fetch"],
+      tools: [
+        ...TOOLS.build,
+        "fetch",
+        "playwright_navigate",
+        "playwright_screenshot",
+        "playwright_click",
+        "playwright_type",
+        "playwright_evaluate",
+        "playwright_select_option",
+        "playwright_hover",
+        "playwright_fill_form",
+        "playwright_wait_for",
+        "playwright_press_key",
+        "playwright_drag"
+      ],
       prompt: `You are the E2E agent. Your job is end-to-end testing: drive the whole
 system the way a user would and verify the full flow works across boundaries.
 
 Scope:
 - Author end-to-end scenarios that exercise real user journeys
 - Drive UI/CLI/API across process and network boundaries
+- Use Playwright browser tools (navigate, click, type, screenshot, evaluate)
+  to automate web UI flows \u2014 open pages, interact with forms, capture evidence
 - Set up and tear down realistic test state
-- Capture failures with enough detail to reproduce (screenshots, logs)
+- Capture failures with enough detail to reproduce (screenshots, logs, page HTML)
+
+Playwright tools available (require the "playwright" MCP server to be enabled):
+  playwright_navigate(url)     \u2014 open a page at the given URL
+  playwright_screenshot()      \u2014 capture a full-page or viewport screenshot
+  playwright_click(selector)   \u2014 click on an element matching a CSS selector
+  playwright_type(selector, text) \u2014 type text into a focused input element
+  playwright_evaluate(script)  \u2014 run arbitrary JavaScript in the page context
+  playwright_select_option(selector, value) \u2014 pick a <select> dropdown option
+  playwright_hover(selector)   \u2014 hover the mouse over an element
+  playwright_fill_form(fields) \u2014 fill multiple form fields in one call
+  playwright_wait_for(selector) \u2014 block until an element appears on the page
+  playwright_press_key(key)    \u2014 press a keyboard key (Enter, Tab, Escape, \u2026)
+  playwright_drag(from, to)    \u2014 drag an element from one selector to another
 
 Input format you accept:
 { "task": "scenario | smoke | journey", "flow": "<user journey>", "surface": "ui | cli | api" }
@@ -4614,8 +4804,10 @@ Output: Markdown e2e report:
 Working rules:
 - Test the real flow end to end; don't stub the thing under test
 - Make scenarios deterministic \u2014 control time, randomness, and external state
-- On failure, capture artifacts (logs/screenshots) for reproduction
-- Keep scenarios independent so one failure doesn't cascade`
+- On failure, capture artifacts (screenshots, page HTML, logs) for reproduction
+- Keep scenarios independent so one failure doesn't cascade
+- For browser tests: playwright_navigate first, then interact, then playwright_screenshot as evidence
+- If playwright tools are unavailable, report it and fall back to API/CLI testing`
     },
     budget: HEAVY_BUDGET,
     capability: {
@@ -4628,10 +4820,106 @@ Working rules:
         "user journey",
         "smoke test",
         "playwright",
+        "browser",
+        "screenshot",
+        "web ui",
+        "headless",
         "cypress",
         "full flow",
         "browser test",
-        "acceptance test"
+        "acceptance test",
+        "navigate",
+        "click",
+        "form fill",
+        "dom",
+        "page load"
+      ]
+    }
+  },
+  {
+    config: {
+      id: "browser",
+      name: "Browser",
+      role: "browser",
+      tools: [
+        ...TOOLS.read,
+        "fetch",
+        "playwright_navigate",
+        "playwright_screenshot",
+        "playwright_click",
+        "playwright_type",
+        "playwright_evaluate",
+        "playwright_select_option",
+        "playwright_hover",
+        "playwright_fill_form",
+        "playwright_wait_for",
+        "playwright_press_key",
+        "playwright_drag"
+      ],
+      prompt: `You are the Browser agent. Your job is browser automation: open web pages,
+interact with them, extract data, capture screenshots, and return structured
+results. You are a read-focused agent \u2014 you drive the browser, not the filesystem.
+
+Scope:
+- Navigate to URLs and wait for pages to load
+- Take full-page or element screenshots as evidence
+- Click buttons, fill forms, select options, type text \u2014 full user simulation
+- Extract page content: text, HTML, element attributes, data tables
+- Evaluate JavaScript in the page context to extract structured data
+- Verify visual state (element visibility, text content, attribute values)
+
+Playwright tools available (require the "playwright" MCP server to be enabled):
+  playwright_navigate(url)          \u2014 open a page at the given URL
+  playwright_screenshot()           \u2014 capture a full-page or viewport screenshot
+  playwright_click(selector)        \u2014 click on an element matching a CSS selector
+  playwright_type(selector, text)   \u2014 type text into a focused input element
+  playwright_evaluate(script)       \u2014 run arbitrary JavaScript in the page context
+  playwright_select_option(selector, value) \u2014 pick a <select> dropdown option
+  playwright_hover(selector)        \u2014 hover the mouse over an element
+  playwright_fill_form(fields)      \u2014 fill multiple form fields in one call
+  playwright_wait_for(selector)     \u2014 block until an element appears on the page
+  playwright_press_key(key)         \u2014 press a keyboard key (Enter, Tab, Escape, \u2026)
+  playwright_drag(from, to)         \u2014 drag an element from one selector to another
+
+Input format you accept:
+{ "task": "navigate | screenshot | extract | interact | verify", "url": "<url>", "steps": ["step1", "step2"] }
+
+Output: Structured markdown report:
+- ## Page (URL, title, load status)
+- ## Actions Taken (step-by-step with timestamps)
+- ## Results (extracted data, element states, verification results)
+- ## Screenshots (list attached screenshot references)
+- ## Errors (any failures with stack traces)
+
+Working rules:
+- Always playwright_navigate first before any interaction
+- Always playwright_wait_for after navigation to ensure the page is ready
+- playwright_screenshot is your primary evidence \u2014 use it before and after interactions
+- Use playwright_evaluate for structured data extraction (JSON, text content)
+- If a selector fails, try alternative selectors before giving up
+- Report exact CSS selectors used \u2014 they're part of the evidence
+- If playwright tools are unavailable, report the error immediately \u2014 do not guess`
+    },
+    budget: MEDIUM_BUDGET,
+    capability: {
+      phase: "verify",
+      summary: "Browser automation: opens pages, clicks, types, screenshots, extracts data via Playwright headless Chromium.",
+      keywords: [
+        "browser",
+        "screenshot",
+        "navigate",
+        "web page",
+        "scrape",
+        "crawl",
+        "headless",
+        "chrome",
+        "open url",
+        "capture",
+        "page title",
+        "extract data",
+        "fill form",
+        "click button",
+        "take screenshot"
       ]
     }
   },
@@ -6080,7 +6368,7 @@ Working rules:
       id: "tech-stack",
       name: "Tech Stack Validator",
       role: "tech-stack",
-      tools: ["search", "fetch", "read", "grep", "glob", "outdated", "audit", "json"],
+      tools: ["search", "fetch", "read", "grep", "glob", "outdated", "audit", "json", "mailbox"],
       prompt: `You are the Tech Stack Validator \u2014 a single-shot validation agent that fires
 before any package, library, or framework choice is committed.
 
@@ -6088,6 +6376,16 @@ Your ONLY job: verify that a technology choice is current, real, and not obsolet
 You are the "this isn't code, this is 10-year-old technology" agent. Intervene
 hard when the LLM hallucinates a version number or suggests dead tech.
 
+## Before you begin
+
+Check the inter-agent mailbox for pending tasks. Other agents or the file-watcher
+may have left assign messages with dependency files to audit:
+- mailbox action=check
+
+If you find an assign message, use the specified file path and packages.
+When done, post results back:
+- mailbox action=send to=<sender> type=result subject="Tech stack audit results" body="..."
+
 ## Critical rules
 
 1. **Verify existence.** Search npm registry (fetch https://registry.npmjs.org/<pkg>/latest)
@@ -6146,11 +6444,11 @@ When APPROVED:
 **Install**: pnpm add <name>@^<major>.<minor>.0`
     },
     budget: {
-      timeoutMs: 6e4,
-      maxIterations: 5,
-      maxToolCalls: 20,
-      maxTokens: 4e4,
-      maxCostUsd: 0.1
+      timeoutMs: 12e4,
+      maxIterations: 10,
+      maxToolCalls: 40,
+      maxTokens: 6e4,
+      maxCostUsd: 0.25
     },
     capability: {
       phase: "meta",
@@ -6380,6 +6678,9 @@ function providerStatusToCode(status, type) {
 
 // src/coordination/coordinator/error-classifier.ts
 function classifySubagentError(err, hints = {}) {
+  if (err instanceof AgentError && err.cause) {
+    return classifySubagentError(err.cause, hints);
+  }
   const cause = err instanceof Error ? { name: err.name, message: err.message, stack: err.stack } : void 0;
   if (err instanceof ProviderError) {
     const baseMessage2 = err.describe();
@@ -6412,7 +6713,7 @@ function classifySubagentError(err, hints = {}) {
   if (/agent exhausted iteration limit$/i.test(baseMessage)) {
     return { kind: "budget_iterations", message: baseMessage, retryable: false, cause };
   }
-  if (/empty response$/i.test(baseMessage)) {
+  if (/empty response/i.test(baseMessage)) {
     return { kind: "empty_response", message: baseMessage, retryable: false, cause };
   }
   if (/^tool failed: /i.test(baseMessage)) {
@@ -7361,7 +7662,14 @@ var ParallelEternalEngine = class {
   }
   stop() {
     this.stopRequested = true;
-    void this.persistState("stopped").catch(() => {
+    void this.persistState("stopped").catch((err) => {
+      console.error(JSON.stringify({
+        level: "error",
+        event: "engine.persist_state_failed",
+        message: err instanceof Error ? err.message : String(err),
+        context: { expectedState: "stopped" },
+        timestamp: (/* @__PURE__ */ new Date()).toISOString()
+      }));
     });
     this.state = "stopped";
   }
@@ -8122,6 +8430,6 @@ function parseDescription(raw) {
   return { trigger, scope };
 }
 
-export { AutoCompactionMiddleware, AutonomousRunner, DefaultErrorHandler, DefaultRetryPolicy, DefaultSkillLoader, DoneConditionChecker, EternalAutonomyEngine, HybridCompactor, IntelligentCompactor, ParallelEternalEngine, SelectiveCompactor, ToolExecutor, buildGoalPreamble, createAutonomyBrain, createStrategyCompactor, formatDecisionSummary, makeAutonomyPromptContributor };
+export { AutoCompactionMiddleware, AutonomousRunner, DefaultErrorHandler, DefaultRetryPolicy, DefaultSkillLoader, DoneConditionChecker, EternalAutonomyEngine, HybridCompactor, IntelligentCompactor, ParallelEternalEngine, SelectiveCompactor, ToolExecutor, buildGoalPreamble, createAutonomyBrain, createStrategyCompactor, createTieredBrainArbiter, formatDecisionSummary, makeAutonomyPromptContributor };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map