@wrongstack/core 0.141.0 → 0.155.0

package/dist/execution/index.d.ts

@@ -1,18 +1,18 @@
-export { C as CompactorOptions, D as DefaultErrorHandler, a as DefaultRetryPolicy, E as EternalAutonomyEngine, b as EternalAutonomyOptions, c as EternalEngineState, H as HybridCompactor, I as IterationStage, P as ParallelEngineState, d as ParallelEternalEngine, e as ParallelEternalOptions, f as ParallelIterationStage, T as ToolExecutor } from '../parallel-eternal-engine-Dj2SYzha.js';
-export { A as AutoCompactionMiddleware, a as AutonomousRunner, b as AutonomousRunnerOptions, c as AutonomyPromptContributorOptions, C as CompactorStrategy, D as DefaultSkillLoader, d as DoneCheckResult, e as DoneConditionChecker, I as IntelligentCompactor, f as IntelligentCompactorOptions, S as SelectiveCompactor, g as SelectiveCompactorOptions, h as SkillLoaderOptions, i as StrategyCompactorOptions, j as buildGoalPreamble, k as createStrategyCompactor, m as makeAutonomyPromptContributor } from '../goal-preamble-iuIUTQwk.js';
+export { C as CompactorOptions, D as DefaultErrorHandler, a as DefaultRetryPolicy, E as EternalAutonomyEngine, b as EternalAutonomyOptions, c as EternalEngineState, H as HybridCompactor, I as IterationStage, P as ParallelEngineState, d as ParallelEternalEngine, e as ParallelEternalOptions, f as ParallelIterationStage, T as ToolExecutor } from '../parallel-eternal-engine-CYoTKjsz.js';
+export { A as AutoCompactionMiddleware, a as AutonomousRunner, b as AutonomousRunnerOptions, c as AutonomyPromptContributorOptions, C as CompactorStrategy, D as DefaultSkillLoader, d as DoneCheckResult, e as DoneConditionChecker, I as IntelligentCompactor, f as IntelligentCompactorOptions, S as SelectiveCompactor, g as SelectiveCompactorOptions, h as SkillLoaderOptions, i as StrategyCompactorOptions, j as buildGoalPreamble, k as createStrategyCompactor, m as makeAutonomyPromptContributor } from '../goal-preamble-CbV8pXLD.js';
 import { P as Provider } from '../context-C7G_MtLV.js';
-import { B as BrainDecision, a as BrainDecisionRequest, b as BrainArbiter } from '../brain-Cp_3GIS2.js';
+import { B as BrainDecision, a as BrainDecisionRequest, b as BrainArbiter } from '../brain-CS_B0vIE.js';
 import '../retry-policy-rutAfVeR.js';
 import '../compactor-BueGt7LG.js';
 import '../config-BaVThgnT.js';
-import '../index-CPweVoFM.js';
+import '../index-B5wz-GXm.js';
 import '../logger-B63L5bTg.js';
 import '../pipeline-BG7UgbDc.js';
 import '../observability-D-HZN_mF.js';
 import '../permission-B7nKnEvQ.js';
-import '../agent-subagent-runner-1GeQE_L0.js';
+import '../agent-subagent-runner-Bsueu0J2.js';
 import '../goal-store-CV9Yz2X_.js';
-import '../multi-agent-coordinator-QWEzJDlm.js';
+import '../multi-agent-coordinator-BSKSFNhv.js';
 import 'node:events';
 import '../skill-Bj6Ezqb8.js';
 import '../wstack-paths-DD50Omgn.js';

package/dist/execution/index.js

@@ -255,8 +255,8 @@ function findPreserveStart(messages, preserveK) {
   for (let i = preserveStart; i < messages.length; i++) {
     const m = messages[i];
     if (!m || typeof m.content === "string" || !Array.isArray(m.content)) continue;
-    const hasToolUse2 = m.content.some((b) => b.type === "tool_use");
-    if (hasToolUse2 && i + 1 < messages.length) {
+    const hasToolUse3 = m.content.some((b) => b.type === "tool_use");
+    if (hasToolUse3 && i + 1 < messages.length) {
       const next = messages[i + 1];
       if (next && next.role === "user" && typeof next.content !== "string" && Array.isArray(next.content) && next.content.some((b) => b.type === "tool_result")) {
         preserveStart = i + 1;
@@ -320,6 +320,127 @@ function buildLosslessDigest(messages) {
   }
   return lines.join("\n");
 }
+function extractText(m) {
+  if (typeof m.content === "string") return m.content;
+  return m.content.filter(isTextBlock).map((b) => b.text).join(" ");
+}
+function hasToolUse2(m) {
+  if (typeof m.content === "string") return false;
+  return m.content.some((b) => b.type === "tool_use");
+}
+function hasLargeToolResult(m, threshold = 3e3) {
+  if (typeof m.content === "string") return false;
+  return m.content.some(
+    (b) => b.type === "tool_result" && b.content && (typeof b.content === "string" ? b.content.length : JSON.stringify(b.content).length) > threshold
+  );
+}
+function scoreMessage(m, context) {
+  const text = extractText(m).toLowerCase();
+  if (text.trim().length === 0 && (hasToolUse2(m) || typeof m.content !== "string")) {
+    const hasResult = typeof m.content !== "string" && m.content.some((b) => b.type === "tool_result");
+    if (hasToolUse2(m) || hasResult) return 0;
+  }
+  if (context?.failureCounts && m.role === "user" && hasToolUse2(m) === false) {
+    const isFailure = /error|fail|exception|timeout|enonet|eacces|eperm|enoent|abort/i.test(text);
+    if (isFailure) {
+      const errKey = /(error|fail|exception|timeout|enonet|eacces|eperm|enoent|abort)/i.exec(text)?.[0]?.toLowerCase() ?? "error";
+      const count = (context.failureCounts.get(errKey) ?? 0) + 1;
+      context.failureCounts.set(errKey, count);
+      if (count >= 5) return 0;
+      if (count >= 3) return 1;
+    }
+  }
+  if (m.role === "user") {
+    if (/\b(wrong|no\b|stop\b|don'?t\b|actually|fix that|undo|revert|forget|ignore|skip)\b/i.test(
+      text
+    )) {
+      return 5;
+    }
+  }
+  if (/\b(error|exception|fatal|critical|crash|panic|abort|segfault|core dump|undefined is not|null pointer|typeerror|referenceerror|syntaxerror)\b/i.test(
+    text
+  )) {
+    return 5;
+  }
+  if (/\b(security|vulnerability|injection|xss|csrf|secret|apikey|api.key|hardcoded|leak|exploit|cve)\b/i.test(
+    text
+  )) {
+    return 5;
+  }
+  if (m.role === "assistant" && /\b(architecture|design|approach|strategy|pattern|refactor|migrate|restructure|decision|trade.?off)\b/i.test(
+    text
+  )) {
+    return 5;
+  }
+  if (hasLargeToolResult(m)) return 1;
+  if (m.role === "user" && !hasToolUse2(m) && /\b(files_with_matches|count|found \d+ match|directory tree|\.\.\. and \d+ more)\b/i.test(text)) {
+    return 1;
+  }
+  return 3;
+}
+function buildSmartDigest(messages) {
+  const lines = [];
+  const failureCounts = /* @__PURE__ */ new Map();
+  let noiseCount = 0;
+  for (const m of messages) {
+    const score = scoreMessage(m, { failureCounts });
+    const text = extractText(m);
+    const toolCount = countToolBlocks(m);
+    if (score === 0) {
+      noiseCount++;
+      continue;
+    }
+    const marker = toolCount > 0 ? ` [${toolCount} tool call(s)]` : "";
+    let display;
+    switch (score) {
+      case 5:
+        display = text.trim();
+        break;
+      case 3:
+        display = firstSentence(text);
+        break;
+      case 1:
+        display = oneLineSummary(m, text);
+        break;
+      default:
+        display = firstSentence(text);
+    }
+    if (display.length === 0 && toolCount === 0) continue;
+    lines.push(`[${m.role}]: ${display}${marker}`);
+  }
+  if (noiseCount > 0) {
+    lines.push(`[system]: ${noiseCount} low-importance turn(s) collapsed (repeated failures / pure tool I/O)`);
+  }
+  return lines.join("\n");
+}
+function countToolBlocks(m) {
+  if (typeof m.content === "string") return 0;
+  return m.content.filter(
+    (b) => b.type === "tool_use" || b.type === "tool_result"
+  ).length;
+}
+function firstSentence(text) {
+  const trimmed = text.trim();
+  if (trimmed.length === 0) return "";
+  const dot = trimmed.indexOf(". ");
+  if (dot === -1) return trimmed.length > 150 ? `${trimmed.slice(0, 147)}\u2026` : trimmed;
+  const sentence = trimmed.slice(0, dot + 1);
+  return sentence.length > 150 ? `${sentence.slice(0, 147)}\u2026` : sentence;
+}
+function oneLineSummary(m, text) {
+  const trimmed = text.trim();
+  if (trimmed.length === 0) {
+    if (typeof m.content !== "string") {
+      const results = m.content.filter((b) => b.type === "tool_result");
+      if (results.length > 0) {
+        return `[${results.length} tool result(s) \u2014 see session log]`;
+      }
+    }
+    return "[no text content]";
+  }
+  const firstLine = trimmed.split("\n")[0] ?? "";
+  return firstLine.length > 100 ? `${firstLine.slice(0, 97)}\u2026` : firstLine;
+}
 function findSafeBoundary(messages, from, to) {
   for (let i = to; i >= from; i--) {
     const m = messages[i];
@@ -335,8 +456,8 @@ function findExchangeStart(messages, userIndex) {
     const m = messages[i];
     if (!m) continue;
     if (m.role === "assistant") {
-      const hasToolUse2 = Array.isArray(m.content) ? m.content.some((b) => b.type === "tool_use") : false;
-      if (!hasToolUse2) return i + 1;
+      const hasToolUse3 = Array.isArray(m.content) ? m.content.some((b) => b.type === "tool_use") : false;
+      if (!hasToolUse3) return i + 1;
     } else if (m.role === "user") {
       return i;
     }
@@ -348,9 +469,11 @@ function findExchangeStart(messages, userIndex) {
 var HybridCompactor = class {
   preserveK;
   eliseThreshold;
+  smart;
   constructor(opts = {}) {
     this.preserveK = opts.preserveK ?? 5;
     this.eliseThreshold = opts.eliseThreshold ?? 2e3;
+    this.smart = opts.smart ?? false;
   }
   async compact(ctx, opts = {}) {
     const beforeTokens = estimateMessages(ctx.messages);
@@ -422,7 +545,7 @@ var HybridCompactor = class {
     if (boundary <= 0) return { saved: 0 };
     const removed = messages.slice(0, boundary);
     const removedTokens = estimateMessages(removed);
-    const digest = buildLosslessDigest(removed) || `${removed.length} earlier turns (no textual content; tool I/O omitted \u2014 see session log)`;
+    const digest = this.smart ? buildSmartDigest(removed) || `${removed.length} earlier turns (no textual content; tool I/O omitted \u2014 see session log)` : buildLosslessDigest(removed) || `${removed.length} earlier turns (no textual content; tool I/O omitted \u2014 see session log)`;
     const summaryMsg = {
       role: "system",
       content: `[prior_turns_digest: ${digest}]`
@@ -946,7 +1069,8 @@ function createStrategyCompactor(opts = {}) {
   }
   return new HybridCompactor({
     preserveK: opts.preserveK,
-    eliseThreshold: opts.eliseThreshold
+    eliseThreshold: opts.eliseThreshold,
+    smart: opts.smart
   });
 }
 var ProviderBackedCompactor = class {
@@ -1807,7 +1931,7 @@ ${excerpt}`;
   subjectFor(toolName, input, subjectKey) {
     if (!input || typeof input !== "object") return void 0;
     const obj = input;
-    const globChars = /[*?\[\]]/g;
+    const globChars = /[*?[\]]/g;
     const escapeGlob = (s) => s.replace(globChars, (c) => `\\${c}`);
     const normalizePath = (s) => escapeGlob(s.replace(/\\/g, "/"));
     if (subjectKey) {
@@ -2237,7 +2361,7 @@ function appendJournal(goal, entry) {
   };
 }
 function parseProgressFromText(text) {
-  const re = /\[progress:\s*(\d{1,3})%\]\s*(?:[—\-]\s*(.+))?/i;
+  const re = /\[progress:\s*(\d{1,3})%\]\s*(?:[—-]\s*(.+))?/i;
   const m = text.match(re);
   if (!m) return null;
   const progress = Math.min(100, Math.max(0, Number.parseInt(m[1] ?? "0", 10)));
@@ -2415,7 +2539,7 @@ ${request.context}` : "",
       },
       { signal }
     );
-    const text = extractText(response).trim();
+    const text = extractText2(response).trim();
     if (request.options?.length) {
       for (const opt of request.options) {
         if (text.toLowerCase().includes(opt.id.toLowerCase())) {
@@ -2443,7 +2567,7 @@ ${request.context}` : "",
     return { type: "deny", reason: "Autonomy Brain LLM unavailable for decision." };
   }
 }
-function extractText(result) {
+function extractText2(result) {
   if (!result || typeof result !== "object") return "";
   const r = result;
   if (Array.isArray(r.content)) {
@@ -3455,6 +3579,7 @@ 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);
@@ -3551,7 +3676,7 @@ 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();
@@ -3559,6 +3684,15 @@ 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);
@@ -4468,15 +4602,44 @@ 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" }
@@ -4490,8 +4653,10 @@ 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: {
@@ -4504,10 +4669,106 @@ 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"
       ]
     }
   },