@zibby/core 0.4.6 → 0.5.0

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zibby/core",
-  "version": "0.4.6",
+  "version": "0.5.0",
   "description": "Core test automation engine with multi-agent and multi-MCP support",
   "type": "module",
   "main": "dist/index.js",
@@ -17,7 +17,6 @@
     "./utils/parallel-config.js": "./dist/utils/parallel-config.js",
     "./utils/run-registry.js": "./dist/utils/run-registry.js",
     "./utils/run-index-merge.js": "./dist/utils/run-index-merge.js",
-    "./utils/run-index-post-cli.js": "./dist/utils/run-index-post-cli.js",
     "./utils/run-state-session.js": "./dist/utils/run-state-session.js",
     "./utils/session-state-live-runs.js": "./dist/utils/session-state-live-runs.js",
     "./utils/mission-control-from-run-states.js": "./dist/utils/mission-control-from-run-states.js",
@@ -26,8 +25,6 @@
     "./utils/run-capacity-queue.js": "./dist/utils/run-capacity-queue.js",
     "./utils/cursor-mcp-isolated-home.js": "./dist/utils/cursor-mcp-isolated-home.js",
     "./constants/zibby-scratch.js": "./dist/constants/zibby-scratch.js",
-    "./templates": "./templates/index.js",
-    "./templates/*": "./templates/*",
     "./package.json": "./package.json"
   },
   "scripts": {
@@ -59,10 +56,6 @@
   },
   "files": [
     "dist/",
-    "templates/",
-    "!templates/**/__tests__/",
-    "!templates/**/*.test.js",
-    "!templates/**/*.spec.js",
     "README.md",
     "LICENSE"
   ],

package/dist/utils/run-index-post-cli.js

@@ -1,4 +1 @@
-import{existsSync as P,readdirSync as F,statSync as Q}from"fs";import{join as a,relative as B,sep as x,resolve as m}from"path";import{appendFileSync as H,readFileSync as z,existsSync as _,mkdirSync as $}from"node:fs";import{join as h}from"node:path";import{DEFAULT_OUTPUT_BASE as E}from"@zibby/agent-workflow";var b="run-index.jsonl";function A(e,t=E){let n=h(e,t);return h(n,b)}function w(e){if(!e||!e.sessionId)return;let t=e.cwd||process.cwd(),n=e.outputBase||E,s=h(t,n);_(s)||$(s,{recursive:!0});let r=h(s,b),i=`${JSON.stringify(e)}
-`;H(r,i,"utf8")}function D(e){if(!e||!_(e))return[];let t;try{t=z(e,"utf8")}catch{return[]}let n=[];for(let s of t.split(`
-`)){let r=s.trim();if(r)try{n.push(JSON.parse(r))}catch{}}return n}import{existsSync as J,mkdirSync as K,readFileSync as W,readdirSync as de,statSync as fe,writeFileSync as G}from"node:fs";import{join as V}from"node:path";var X="zibby-run-state.json";function M(e){return V(e,X)}function R(e){if(!e||typeof e!="string")return null;let t=M(e);if(!J(t))return null;try{let n=W(t,"utf8"),s=JSON.parse(n);return s&&typeof s=="object"?s:null}catch{return null}}function v(e,t){if(!e||typeof e!="string")return;try{K(e,{recursive:!0})}catch{return}let s={...R(e)||{v:1},...t,v:1,updatedAt:Date.now()};try{G(M(e),`${JSON.stringify(s)}
-`,"utf8")}catch(r){console.warn(`[zibby run-state] ${r.message}`)}}function q(e){return e?.recordKind==="progress"}function O(e){let t=Number(e)||0;return t<=0?0:t<1e12?t*1e3:t}function C(e,t={}){let n=t.maxProgressAgeMs!=null&&Number.isFinite(t.maxProgressAgeMs)?Math.max(0,t.maxProgressAgeMs):21e5,s=typeof t.now=="number"?t.now:Date.now(),{summary:r,progress:i}=e||{};if(!i)return!1;let o=O(i.ts);if(n>0&&o>0&&s-o>n)return!1;if(!r)return o>0;let d=O(r.ts);return o>d}function k(e){let t=new Map;for(let n of e||[]){if(!n?.sessionId)continue;let s=t.get(n.sessionId);s||(s={summary:null,progress:null});let r=Number(n.ts)||0;q(n)?(!s.progress||r>=(Number(s.progress.ts)||0))&&(s.progress=n):(!s.summary||r>=(Number(s.summary.ts)||0))&&(s.summary=n),t.set(n.sessionId,s)}return t}import{DEFAULT_OUTPUT_BASE as I,SESSIONS_DIR as y}from"@zibby/agent-workflow";var N=Object.freeze(["preflight","execute_live","generate_script"]);function T(e){let t=process.env.ZIBBY_STUDIO_TEST_CASE_ID;return t!=null&&String(t).trim()!==""?String(t).trim():e!=null?String(e):""}var ee=[a("generate_script","generated-test.spec.js"),a("generate_script","generated-test.spec.ts"),a("generate_script","playwright.spec.ts"),a("generate_script","test.spec.ts")];function te(e){let t=[a(e,"execute_live","videos"),a(e,"execute_live"),e];for(let n of t){if(!P(n))continue;let s;try{s=F(n)}catch{continue}let r=s.find(i=>i.endsWith(".webm"));if(r)return a(n,r)}return null}function ne(e){let t=[a(e,"execute_live","events.json"),a(e,"events.json")];for(let n of t)if(P(n))return n;return null}function se(e){for(let t of ee){let n=a(e,t);if(P(n))return n}return null}function re(e){return!e||!P(e)?{videoPathAbs:null,eventsPathAbs:null,scriptPathAbs:null}:{videoPathAbs:te(e),eventsPathAbs:ne(e),scriptPathAbs:se(e)}}function j(e){let t=e.cwd||process.cwd(),n=e.outputBase||I,i=((e.result||{}).state||{}).sessionPath;if(!i||typeof i!="string")return null;let o=i.split(/[/\\]/).filter(Boolean).pop();if(!o)return null;let{videoPathAbs:d,eventsPathAbs:f,scriptPathAbs:l}=re(i),p=u=>{if(!u)return null;try{return B(t,u).split(x).join("/")}catch{return null}},c=null;if(e.specPath)try{let u=m(t,e.specPath);c=B(t,u).split(x).join("/")}catch{c=String(e.specPath).split(x).join("/")}return{v:1,recordKind:"summary",ts:Date.now(),sessionId:o,status:e.status??(e.success?"completed":"failed"),cwd:t,outputBase:n,sessionPathAbs:i,sessionDirRel:p(i),videoPathAbs:d||null,eventsPathAbs:f||null,scriptPathAbs:l||null,videoRel:p(d),eventsRel:p(f),scriptRel:p(l),specRel:c,source:process.env.ZIBBY_RUN_SOURCE||"cli",studioTestCaseId:T(o)||null,errorMessage:e.errorMessage||null}}function U({cwd:e,config:t,result:n,success:s,specPath:r,errorMessage:i}){try{let o=j({cwd:e||process.cwd(),result:n,success:s,outputBase:t?.paths?.output||I,specPath:r,errorMessage:i});o&&(w(o),o.sessionPathAbs&&v(o.sessionPathAbs,{sessionId:o.sessionId,studioTestCaseId:o.studioTestCaseId||o.sessionId,status:o.status,activeNode:null,activeStageIndex:null,errorMessage:o.errorMessage||null,runSource:o.source||"cli",cwd:o.cwd,outputBase:o.outputBase,sessionPathAbs:o.sessionPathAbs}))}catch(o){console.warn(`[zibby browser-test run-index] ${o.message}`)}}function oe({sessionPath:e,sessionId:t,cwd:n,outputBase:s=I}={}){let r=n||process.cwd(),i=s||I,o=t!=null&&String(t).trim()!==""?String(t).trim():null,d=process.env.ZIBBY_PIN_SESSION_PATH==="1"||process.env.ZIBBY_PIN_SESSION_PATH==="true",f=process.env.ZIBBY_SESSION_PATH&&String(process.env.ZIBBY_SESSION_PATH).trim();if(d&&f)return m(f);let l=e&&String(e).trim();if(l)return m(l);let p=process.env.ZIBBY_SESSIONS_ROOT&&String(process.env.ZIBBY_SESSIONS_ROOT).trim();return p&&o?m(a(p,o)):process.env.ZIBBY_SESSION_PATH&&String(process.env.ZIBBY_SESSION_PATH).trim()?m(String(process.env.ZIBBY_SESSION_PATH).trim()):m(a(r,i,y,o||"invalid"))}function ie(e){try{let t=e?.currentNode;if(!t||!N.includes(t))return;let n=e.sessionPath,s=e.sessionId||n&&String(n).split(/[/\\]/).filter(Boolean).pop()||null;if(!s)return;let r=e.cwd||process.cwd(),i=e.outputBase||I,o=N.indexOf(t),d=e?.specPath!=null?String(e.specPath).trim():"",f=e?.taskDescription!=null?String(e.taskDescription):"",l=null;if(d)try{let c=m(r,d);l=B(r,c).split(x).join("/")}catch{l=d.split(x).join("/")}let p=oe({sessionPath:n,sessionId:s,cwd:r,outputBase:i});w({v:1,recordKind:"progress",ts:Date.now(),sessionId:s,cwd:r,outputBase:i,sessionPathAbs:p,activeNode:t,activeStageIndex:o,specRel:l,taskDescription:f||null,studioTestCaseId:T(s)||null,source:process.env.ZIBBY_RUN_SOURCE||"cli"}),v(p,{sessionId:s,studioTestCaseId:T(s)||s,status:"running",activeNode:t,activeStageIndex:o,sessionPathAbs:p,cwd:r,outputBase:i,specPath:l||null,task:f||null,taskDescription:f||null,runSource:process.env.ZIBBY_RUN_SOURCE||"cli",pid:typeof process.pid=="number"?process.pid:null})}catch(t){console.warn(`[zibby browser-test run-index progress] ${t.message}`)}}function Y({cwd:e,config:t}={}){let n=e||process.cwd(),s=t?.paths?.output||I;return r=>{ie({cwd:r?.cwd||n,outputBase:r?.outputBase||s,sessionPath:r?.sessionPath,sessionId:r?.sessionId,currentNode:r?.currentNode,specPath:r?.specPath,taskDescription:r?.taskDescription})}}function Z(e={}){try{let t=e.cwd||process.cwd(),n=e.config?.paths?.output||e.outputBase||I,s=A(t,n),r=D(s),i=k(r),o=new Set,d=e.errorMessage||"Run stopped (SIGINT/SIGTERM) before a normal summary was written.",f=(c,u)=>{if(!c||!u||o.has(c))return;o.add(c);let S=j({cwd:t,outputBase:n,result:{state:{sessionPath:u}},success:!1,specPath:null,status:"interrupted",errorMessage:d});S&&(w(S),v(u,{sessionId:c,studioTestCaseId:S.studioTestCaseId||c,status:"interrupted",activeNode:null,activeStageIndex:null,errorMessage:S.errorMessage||null,runSource:S.source||"cli",cwd:t,outputBase:n,sessionPathAbs:u}))};for(let[c,u]of i){if(!C(u))continue;let S=u.progress;if(!S)continue;let g=String(c),L=S.sessionPathAbs&&String(S.sessionPathAbs)||a(t,n,y,g);f(g,L)}let l=a(t,n,y);if(!P(l))return;let p;try{p=F(l)}catch{return}for(let c of p){let u=a(l,c),S;try{S=Q(u)}catch{continue}if(!S.isDirectory())continue;let g=R(u);!g||g.status!=="running"||f(String(c),u)}}catch(t){console.warn(`[zibby browser-test run-index interrupt] ${t.message}`)}}function Te(e){U(e)}function _e(e){Z(e)}function Ee(e){return Y(e)}export{Ee as createCliRunIndexPipelineProgressAppender,_e as postCliInterruptedRunIndex,Te as postCliRunIndex};
+function e(n){}function t(n){}function o(n){return null}export{o as createCliRunIndexPipelineProgressAppender,t as postCliInterruptedRunIndex,e as postCliRunIndex};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zibby/core",
-  "version": "0.4.6",
+  "version": "0.5.0",
   "description": "Core test automation engine with multi-agent and multi-MCP support",
   "type": "module",
   "main": "dist/index.js",
@@ -17,7 +17,6 @@
     "./utils/parallel-config.js": "./dist/utils/parallel-config.js",
     "./utils/run-registry.js": "./dist/utils/run-registry.js",
     "./utils/run-index-merge.js": "./dist/utils/run-index-merge.js",
-    "./utils/run-index-post-cli.js": "./dist/utils/run-index-post-cli.js",
     "./utils/run-state-session.js": "./dist/utils/run-state-session.js",
     "./utils/session-state-live-runs.js": "./dist/utils/session-state-live-runs.js",
     "./utils/mission-control-from-run-states.js": "./dist/utils/mission-control-from-run-states.js",
@@ -26,8 +25,6 @@
     "./utils/run-capacity-queue.js": "./dist/utils/run-capacity-queue.js",
     "./utils/cursor-mcp-isolated-home.js": "./dist/utils/cursor-mcp-isolated-home.js",
     "./constants/zibby-scratch.js": "./dist/constants/zibby-scratch.js",
-    "./templates": "./templates/index.js",
-    "./templates/*": "./templates/*",
     "./package.json": "./package.json"
   },
   "scripts": {
@@ -59,10 +56,6 @@
   },
   "files": [
     "dist/",
-    "templates/",
-    "!templates/**/__tests__/",
-    "!templates/**/*.test.js",
-    "!templates/**/*.spec.js",
     "README.md",
     "LICENSE"
   ],

package/dist/templates/browser-test-automation/README.md

@@ -1,136 +0,0 @@
-# Browser Test Automation Workflow
-
-This is YOUR workflow graph. You can customize it however you want!
-
-Works with **Claude** or **Cursor** agents (configured in `.zibby.config.mjs`).
-
-## Default Flow
-
-```
-preflight → execute_live → generate_script
-```
-
-The workflow generates a test title, executes the test live in a **browser** with AI assistance, and generates a Playwright script with stable selectors.
-
-## Customization
-
-### Add Custom Nodes
-
-Create a new file in `nodes/`:
-
-```javascript
-// nodes/send-slack.js
-export const sendSlackNode = {
-  name: 'send_slack',
-  agent: { type: 'openai', model: 'gpt-4o-mini' },
-  prompt: (state) => `Send Slack notification...`,
-  outputSchema: { success: { type: 'boolean', required: true } }
-};
-```
-
-Then add it to your graph in `graph.js`:
-
-```javascript
-import { sendSlackNode } from './nodes/send-slack.js';
-
-buildGraph() {
-  const graph = new WorkflowGraph();
-  // ... existing nodes
-  graph.addNode('send_slack', sendSlackNode);
-  graph.addEdge('verify_script', 'send_slack');
-  return graph;
-}
-```
-
-### Multi-Agent Configuration
-
-Each node can use a different LLM:
-
-```javascript
-graph.addNode('generate_title', {
-  agent: { type: 'claude', model: 'claude-sonnet-4' },
-  prompt: (state) => `Generate title...`
-});
-
-graph.addNode('verify_script', {
-  agent: { type: 'deepseek', model: 'deepseek-coder' }, // Cheap & fast
-  prompt: (state) => `Run test...`
-});
-
-graph.addNode('update_jira', {
-  agent: { type: 'ollama', model: 'llama3' }, // Local for privacy
-  prompt: (state) => `Update Jira...`
-});
-```
-
-### Skip Nodes
-
-Comment out nodes you don't need:
-
-```javascript
-// graph.addNode('verify_script', verifyScriptNode);
-graph.addEdge('generate_script', 'update_jira'); // Skip verification
-```
-
-### Parallel Execution
-
-Run multiple nodes in parallel:
-
-```javascript
-graph.addParallelEdges('verify_script', [
-  'send_slack',
-  'update_jira',
-  'log_datadog'
-]);
-```
-
-## Configuration
-
-Edit `.zibby.config.mjs` to set your default agent and optional per-node model overrides:
-
-```javascript
-export default {
-  agent: {
-    cursor: { model: 'auto' }, // or claude: { model: 'auto' }
-    strictMode: false,
-  },
-  models: {
-    default: 'auto',
-    execute_live: 'auto',
-    generate_script: 'auto',
-  },
-};
-```
-
-## Studio / Scripts tab (code discovery)
-
-Runs write `generate_script/result.json` with a `scriptPath` (often under your repo `tests/`). After the graph finishes, **`BrowserTestResultHandler.ensureStudioCodegenMirror`** copies that file into the session folder under stable names so tools don’t need Studio running at generation time:
-
-| File (under `.zibby/output/sessions/<sessionId>/generate_script/`) | Role |
-|---------------------------------------------------------------------|------|
-| `generated-test.spec.js` | Playwright (`.js`) |
-| `playwright.spec.ts` | Playwright (`.ts` / `.tsx` source) |
-| `test.selenium.py` | Selenium |
-
-**Electron Studio** resolves these via `discoverCodegenArtifactsElectron` (after `session/codegen/`).
-
-**Web Studio** (`VITE_STUDIO_API_ORIGIN`, e.g. `:3847`) should implement `GET /api/sessions/:id/codegen/playwright` (and `/selenium`) by reading, in order:
-
-1. `sessions/<id>/codegen/` legacy JIT names (`test.spec.ts`, `generated-test.spec.js`, …)
-2. **`sessions/<id>/generate_script/`** canonical names above
-3. `scriptPath` from `generate_script/result.json` (resolve relative to session / `cwd` from session meta)
-
-## Documentation
-
-- [Full Graph Framework Design](../../docs/GRAPH_FRAMEWORK_DESIGN.md)
-- [Multi-Agent Patterns](../../docs/FRAMEWORK_CONVERSATION_SUMMARY.md)
-
-## Updates
-
-To get latest template updates:
-
-```bash
-zibby update-graph --merge
-```
-
-This will merge bug fixes while preserving your customizations.

package/dist/templates/browser-test-automation/chat.mjs

@@ -1,36 +0,0 @@
-/**
- * Zibby Chat Agent
- *
- * Interactive conversational node that acts as the default entry point
- * when users type `zibby` with no subcommand.
- *
- * This is a plain chat bot — no MCP servers, no middleware, no structured output.
- * Just streamed text conversation with the AI agent.
- *
- * The skill-installer skill injects its promptFragment so the LLM knows which
- * skills are available and can install/uninstall them via natural conversation.
- * Users can customize this file after `zibby init` copies it to .zibby/chat.mjs
- */
-
-import { SKILLS } from '@zibby/core';
-
-export const CHAT_CONFIG = {
-  name: 'zibby_chat',
-  skills: [SKILLS.CORE_TOOLS, SKILLS.SKILL_INSTALLER, SKILLS.CHAT_MEMORY, SKILLS.WORKFLOW_BUILDER],
-  timeout: 0,
-
-  systemPrompt: `You are Zibby, a helpful AI assistant. Capabilities come from installed skills.
-
-## How you work
-1. When you need data, call tools. You can chain up to 5 calls per turn.
-2. After each tool result, decide: "Would I be embarrassed to give this answer to a coworker?" If yes, call another tool.
-3. Only respond once you have something genuinely useful.
-4. Never claim you did something without actually calling the tool.
-5. After EVERY response, self-evaluate: is the user's goal fully achieved? Is anything still pending or running? If yes, DO NOT ASK — autonomously poll: call wait (you decide how long), then check status, then respond with an update. Repeat until done or the user interrupts.
-
-## How you talk
-- Talk like a teammate in Slack, not a report generator.
-- Summarize and paraphrase. Never copy-paste field values or list raw steps verbatim.
-- Short paragraphs, not numbered lists (unless the user specifically asks for steps).
-- Match the user's tone and energy. Be concise.`,
-};

package/dist/templates/browser-test-automation/graph.mjs

@@ -1,80 +0,0 @@
-/**
- * Test Automation Workflow Graph
- * 
- * buildGraph() - define nodes, edges, routing
- * onComplete(result) - post-processing after graph finishes (save artifacts, etc.)
- */
-
-import { WorkflowAgent, WorkflowGraph } from '@zibby/core';
-import {
-  preflightNode,
-  cacheReplayNode,
-  executeLiveNode,
-  generateScriptNode,
-} from './nodes/index.mjs';
-import { BrowserTestResultHandler } from './result-handler.mjs';
-import { browserTestAutomationStateSchema } from './state.js';
-
-export class BrowserTestAutomationAgent extends WorkflowAgent {
-  buildGraph() {
-    const graph = new WorkflowGraph();
-    graph.setStateSchema(browserTestAutomationStateSchema);
-
-    graph.addNode('preflight', preflightNode);
-    graph.addNode('cache_replay', cacheReplayNode);
-    graph.addNode('execute_live', executeLiveNode);
-    graph.addNode('generate_script', generateScriptNode);
-
-    graph.setEntryPoint('preflight');
-
-    // Short-circuit when preflight produced nothing usable. Triggered when:
-    //   - the user invoked `zibby workflow run browser-tests` with no spec
-    //     (state.input is undefined / empty), so preflight had nothing to
-    //     analyze and the LLM came back with `assertions: []`
-    //   - the spec is so vague the LLM can't extract any assertions
-    // Without this gate the graph would barrel into execute_live, fire up
-    // a real browser session + a second expensive LLM call, then waste
-    // ~30s before failing — bad UX and bad bill.
-    graph.addConditionalEdges('preflight', (state) => {
-      const assertions = state.preflight?.assertions || [];
-      return assertions.length > 0 ? 'cache_replay' : 'END';
-    });
-
-    // Lever-#2 fork: cache_replay attempted a Playwright-only replay of
-    // a prior successful action sequence. On hit it side-wrote
-    // state.execute_live with synthesized output, so we can skip
-    // execute_live and jump straight to generate_script — zero LLM
-    // tokens. On miss / replay failure / cold cache, fall through to
-    // the normal LLM-driven execute_live path.
-    graph.addConditionalEdges('cache_replay', (state) => {
-      return state.cache_replay?.hit === true ? 'generate_script' : 'execute_live';
-    });
-
-    graph.addConditionalEdges('execute_live', (state) => {
-      const result = state.execute_live;
-      const hasExecution = (result?.steps?.length > 0) || (result?.actions?.length > 0);
-      return hasExecution ? 'generate_script' : 'END';
-    });
-
-    graph.addEdge('generate_script', 'END');
-    return graph;
-  }
-
-  async onComplete(result) {
-    const cwd = result.state.cwd || process.cwd();
-    BrowserTestResultHandler.saveTitle(result, cwd);
-    await BrowserTestResultHandler.saveExecutionData(result);
-    BrowserTestResultHandler.ensureStudioCodegenMirror(
-      result.state?.sessionPath,
-      result.state?.cwd || cwd,
-    );
-
-    // Memory end-run hook (if @zibby/ui-memory is installed)
-    try {
-      const { memoryEndRun, memorySyncPush } = await import('@zibby/ui-memory');
-      const sessionId = result.state.sessionPath?.split('/').pop();
-      memoryEndRun(cwd, { sessionId, passed: result.success !== false });
-      memorySyncPush(cwd);
-    } catch { /* @zibby/ui-memory not available */ }
-  }
-}

package/dist/templates/browser-test-automation/nodes/cache-replay.mjs

@@ -1,213 +0,0 @@
-/**
- * cache_replay node — lever-#2 read path inside the workflow.
- *
- * Sits between `preflight` and `execute_live` in the graph. Tries to
- * replay a prior successful run's action sequence via Playwright
- * directly, completely skipping the LLM. On a clean cache hit it
- * populates `state.execute_live` with the result so downstream
- * `generate_script` works exactly as if execute_live had run.
- *
- * Conditional edge after this node:
- *   - state.cache_replay.hit === true → skip execute_live → generate_script
- *   - state.cache_replay.hit === false → execute_live (LLM-driven path)
- *
- * Not user-configurable per-spec — the cache key derivation handles
- * staleness (page fingerprint drift invalidates) and replay failures
- * fall through cleanly to the LLM path.
- */
-
-import { z } from '@zibby/core';
-import { chromium } from 'playwright';
-import { spawn } from 'child_process';
-import { extractDomain, replayActions } from '@zibby/ui-memory';
-import { join } from 'path';
-
-const REPLAY_TIMEOUT_MS = 60_000;
-
-export const cacheReplayNode = {
-  name: 'cache_replay',
-  skills: [],
-  timeout: 90000,
-  outputSchema: z.object({
-    hit: z.boolean(),
-    elapsed_ms: z.number().nullish(),
-    executed: z.number().nullish(),
-    total: z.number().nullish(),
-    cache_key: z.string().nullish(),
-    error: z.string().nullish(),
-    // When hit, we also write a synthesized execute_live block so the
-    // downstream generate_script node sees what it expects.
-    execute_live_synthesized: z.boolean().nullish(),
-  }),
-
-  execute: async (context) => {
-    // graph.js builds nodeContext as `{ state, invokeAgent, _coreInvokeAgent,
-    // ...state.getAll() }`. So `context.testSpec` works (spread) AND
-    // `context.state.get('testSpec')` works (instance). Reading from the
-    // spread is the natural shape — `context.state` is reserved for the
-    // .set(key, value) side-write below.
-    const cwd = context.cwd || context.workspace || process.cwd();
-    const testSpec = context.testSpec || '';
-    const specPath = context.specPath || '';
-
-    // Derive domain from the spec text (no DOM access yet — pure parse).
-    const domain = extractDomainFromSpec(testSpec);
-    if (!domain) {
-      return { hit: false, error: 'cannot derive domain from spec' };
-    }
-
-    // Cache key requires page_fingerprint, which is page-state-dependent
-    // and only available AFTER navigation. We compute a key WITHOUT
-    // fingerprint first and look up by (domain, spec_path) prefix —
-    // the persister wrote spec_path too. If we find a candidate, we
-    // use its stored fingerprint to compute the full key and verify.
-    //
-    // Lookup order:
-    //   1. Exact (domain, spec_path) match in action_cache.
-    //   2. If found, use its actions for replay attempt.
-    //   3. On replay success: signal hit, populate state.execute_live.
-    //   4. On replay failure (or cache miss): hit=false, fall back to LLM.
-    const cached = await lookupCacheByDomainAndSpec({ cwd, domain, specPath });
-    if (!cached) {
-      return { hit: false, error: 'no cached actions for this spec' };
-    }
-
-    // Run the replay in a freshly-launched Playwright browser. Cleanly
-    // independent from the @zibby/mcp-browser path execute_live uses.
-    const t0 = Date.now();
-    const browser = await chromium.launch({ headless: true });
-    const page = await browser.newPage();
-    let replayResult;
-    try {
-      replayResult = await Promise.race([
-        replayActions({
-          actions: cached.actions,
-          page,
-          log: (m) => console.log(`[cache_replay] ${m}`),
-        }),
-        new Promise((_, reject) =>
-          setTimeout(() => reject(new Error('replay timeout')), REPLAY_TIMEOUT_MS),
-        ),
-      ]);
-    } catch (err) {
-      replayResult = { success: false, error: err.message, executed: 0, total: cached.actions.length };
-    }
-    const finalUrl = page.url();
-    await browser.close().catch(() => {});
-    const elapsedMs = Date.now() - t0;
-
-    if (!replayResult.success) {
-      // Increment failure_count so we can drop chronic misses later.
-      await incrementCacheFailure({ cwd, cacheKey: cached.cache_key });
-      return {
-        hit: false,
-        elapsed_ms: elapsedMs,
-        executed: replayResult.executed,
-        total: replayResult.total,
-        cache_key: cached.cache_key,
-        error: replayResult.error,
-      };
-    }
-
-    // HIT path. Side-write the synthesized execute_live output via
-    // context.state.set so downstream generate_script reads the same
-    // shape it expects (actions[], finalUrl, …). The customExecute
-    // return-value lands in state.cache_replay; the execute_live slot
-    // has to be populated separately.
-    if (typeof context.state?.set === 'function') {
-      context.state.set('execute_live', {
-        success: true,
-        steps: cached.actions.map((a) => a.description),
-        actions: cached.actions,
-        assertions: [],
-        finalUrl,
-        browserClosed: true,
-        notes: 'cache_replay hit — actions replayed via Playwright, no LLM',
-      });
-    }
-
-    return {
-      hit: true,
-      elapsed_ms: elapsedMs,
-      executed: replayResult.executed,
-      total: replayResult.total,
-      cache_key: cached.cache_key,
-      execute_live_synthesized: true,
-    };
-  },
-};
-
-// ─── helpers ────────────────────────────────────────────────────────────
-
-function extractDomainFromSpec(spec) {
-  if (!spec) return null;
-  // Find the first http(s) URL in the spec and run it through the
-  // SAME `extractDomain` the persister uses, so the cache-key lookup
-  // matches what was actually written (notably: `www.` is stripped).
-  const m = String(spec).match(/https?:\/\/[^\s"'<>]+/);
-  if (!m) return null;
-  return extractDomain(m[0]);
-}
-
-/**
- * Find a cached row by (domain, spec_path). Picks the row with
- * highest success_count if multiple match.
- * Uses dolt via subprocess (matching the rest of the codebase's
- * Dolt-access pattern).
- */
-async function lookupCacheByDomainAndSpec({ cwd, domain, specPath }) {
-  const dbDir = join(cwd, '.zibby', 'memory');
-  const safeDomain = escapeSql(domain);
-  const safeSpec = escapeSql(specPath);
-  const sql = `SELECT cache_key, actions_json, page_fingerprint
-    FROM action_cache
-    WHERE domain = ${safeDomain} AND spec_path = ${safeSpec}
-    ORDER BY success_count DESC, last_used_at DESC
-    LIMIT 1`;
-  const rows = await runDoltJson(dbDir, sql);
-  if (!rows || rows.length === 0) return null;
-  try {
-    const actions = JSON.parse(rows[0].actions_json);
-    return { cache_key: rows[0].cache_key, actions, fingerprint: rows[0].page_fingerprint };
-  } catch {
-    return null;
-  }
-}
-
-async function incrementCacheFailure({ cwd, cacheKey }) {
-  const dbDir = join(cwd, '.zibby', 'memory');
-  const sql = `UPDATE action_cache
-    SET failure_count = failure_count + 1, last_replay_status = 'replay-failed'
-    WHERE cache_key = ${escapeSql(cacheKey)}`;
-  await runDoltExec(dbDir, sql).catch(() => { /* non-fatal */ });
-}
-
-function escapeSql(v) {
-  if (v == null) return 'NULL';
-  return `'${String(v).replace(/'/g, "''")}'`;
-}
-
-function runDoltJson(dir, sql) {
-  return new Promise((resolve) => {
-    const child = spawn('dolt', ['sql', '-r', 'json', '-q', sql], { cwd: dir });
-    let out = '';
-    child.stdout.on('data', (d) => { out += d; });
-    child.on('close', () => {
-      try {
-        const parsed = JSON.parse(out);
-        resolve(parsed.rows || []);
-      } catch {
-        resolve([]);
-      }
-    });
-    child.on('error', () => resolve([]));
-  });
-}
-
-function runDoltExec(dir, sql) {
-  return new Promise((resolve, reject) => {
-    const child = spawn('dolt', ['sql', '-q', sql], { cwd: dir });
-    child.on('close', (code) => (code === 0 ? resolve() : reject(new Error(`dolt exit ${code}`))));
-    child.on('error', reject);
-  });
-}