@wrongstack/core 0.148.0 → 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-C75QuhAI.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-CYJLg0wk.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

@@ -3579,6 +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);
@@ -3675,7 +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();
@@ -3683,6 +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);
@@ -4592,15 +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" }
@@ -4614,8 +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: {
@@ -4628,10 +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"
       ]
     }
   },