@archal/cli 0.9.1 → 0.9.6

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@archal/cli",
-  "version": "0.9.1",
+  "version": "0.9.6",
   "description": "Pre-deployment testing for AI agents",
   "type": "module",
   "main": "dist/index.cjs",
@@ -12,7 +12,8 @@
     ".": {
       "types": "./dist/index.d.cts",
       "default": "./dist/index.cjs"
-    }
+    },
+    "./seed/dynamic-generator": "./src/runner/seed/dynamic-generator.ts"
   },
   "license": "MIT",
   "repository": {
@@ -37,36 +38,44 @@
   "files": [
     "bin",
     "dist",
-    "harnesses",
-    "scenarios",
     "twin-assets"
   ],
-  "scripts": {
-    "sync:twin-assets": "node scripts/sync-twin-assets.mjs",
-    "sync:scenarios": "node scripts/sync-scenarios.mjs",
-    "build": "pnpm --filter @archal/twin-core run build && pnpm run sync:twin-assets && pnpm run sync:scenarios && tsup --config tsup.config.ts && node scripts/stage-runtime-assets.mjs",
-    "prepack": "pnpm run sync:twin-assets && pnpm run sync:scenarios",
-    "start": "tsx src/index.ts",
-    "test": "vitest run --exclude '__tests__/e2e/**'",
-    "test:e2e": "vitest run __tests__/e2e/",
-    "test:seed-stress": "vitest run __tests__/integration/gemini-seed-stress-live.test.ts",
-    "typecheck": "pnpm --filter @archal/twin-core run build && tsc --noEmit"
-  },
   "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.27.1",
+    "@modelcontextprotocol/sdk": "^1.29.0",
     "commander": "^14.0.3",
     "glob": "^11.0.3",
-    "posthog-node": "^5.26.2",
-    "zod": "^4.3.6"
+    "yaml": "^2.8.2",
+    "zod": "^4.3.6",
+    "@archal/sandbox-runtime": "0.1.0",
+    "@archal/seed-codegen-runtime": "0.1.0",
+    "@archal/openclaw-runtime": "0.1.0",
+    "@archal/node-auth": "0.1.0"
   },
   "devDependencies": {
-    "@archal/twin-core": "workspace:*",
-    "@archal/twin-github": "workspace:*",
-    "@archal/twin-slack": "workspace:*",
     "@types/node": "^25.3.3",
     "tsup": "^8.5.0",
     "tsx": "^4.19.0",
     "typescript": "^5.9.0",
-    "vitest": "^2.1.0"
+    "vitest": "^2.1.0",
+    "@archal/twin-core": "0.1.0",
+    "@archal/twin-jira": "0.1.0",
+    "@archal/twin-stripe": "0.1.0",
+    "@archal/twin-slack": "0.1.0",
+    "@archal/twin-github": "0.1.0"
+  },
+  "scripts": {
+    "sync:twin-assets": "node scripts/sync-twin-assets.mjs",
+    "build:base-deps": "node ../scripts/ensure-twin-core-build.mjs && node ../scripts/ensure-package-builds.mjs @archal/node-auth @archal/seed-codegen-runtime @archal/openclaw-runtime @archal/sandbox-runtime",
+    "build:test-deps": "pnpm run build:base-deps && node ../scripts/ensure-package-builds.mjs @archal/twin-github @archal/twin-jira @archal/twin-stripe",
+    "build:raw": "pnpm run sync:twin-assets && tsup --config tsup.config.ts && node scripts/stage-runtime-assets.mjs",
+    "build": "pnpm run build:base-deps && pnpm run build:raw",
+    "start": "tsx src/index.ts",
+    "test:openclaw-contract:raw": "vitest run __tests__/runner/openclaw-remote.test.ts __tests__/commands/run.test.ts __tests__/runner/run-executor.test.ts",
+    "test:targeted-regressions:raw": "vitest run __tests__/commands/run.test.ts",
+    "test:raw": "vitest run --exclude '__tests__/e2e/**'",
+    "test": "pnpm run build:test-deps && pnpm run test:raw",
+    "test:e2e": "vitest run __tests__/e2e/",
+    "typecheck:raw": "tsc -p tsconfig.typecheck.json --noEmit",
+    "typecheck": "pnpm run build:test-deps && pnpm run typecheck:raw"
   }
-}
+}

package/twin-assets/google-workspace/fidelity.json

@@ -0,0 +1,9 @@
+{
+  "twin": "google-workspace",
+  "version": "0.1",
+  "thresholds": {
+    "minReplaySteps": 0,
+    "minScenarioWindows": 0,
+    "maxValueDifferences": 0
+  }
+}

package/twin-assets/jira/fidelity.json

@@ -1,20 +1,19 @@
 {
   "twinName": "jira",
-  "version": "0.4.0",
+  "version": "0.6.0",
   "realServer": "sooperset/mcp-atlassian",
-  "realServerVersion": "0.13.1",
+  "realServerVersion": "0.21.0",
   "toolCount": {
-    "twin": 39,
-    "realServer": 39,
-    "implemented": 39,
+    "mcpVisible": 49,
+    "realServer": 49,
+    "totalHandlers": 210,
     "excluded": [
       "jira_get_issue_proforma_forms",
       "jira_get_proforma_form_details",
-      "jira_update_proforma_form_answers",
-      "jira_update_issue",
-      "jira_remove_issue_link"
+      "jira_update_proforma_form_answers"
     ]
   },
+  "restRouteCount": 818,
   "fixtureCount": {
     "realServer": 39,
     "replaySteps": 295,
@@ -26,15 +25,16 @@
     "replayPass": "295/295"
   },
   "seeds": ["empty", "small-project", "enterprise", "sprint-active", "large-backlog", "permissions-denied", "rate-limited", "conflict-states"],
-  "jqlSupported": ["=", "!=", "~", "IN", "NOT IN", "AND", "OR", "NOT", "IS EMPTY", "IS NOT EMPTY", "ORDER BY"],
-  "jqlNotSupported": ["WAS", "CHANGED", ">=", "<=", ">", "<", "date functions", "custom fields"],
-  "testCount": 207,
+  "jqlSupported": ["=", "!=", "~", ">", "<", ">=", "<=", "IN", "NOT IN", "AND", "OR", "NOT", "IS EMPTY", "IS NOT EMPTY", "ORDER BY", "WAS", "CHANGED", "CHANGED FROM x TO y", "customfield_XXXXX", "startOfDay()", "endOfDay()", "startOfWeek()", "endOfWeek()", "startOfMonth()", "endOfMonth()", "startOfYear()", "endOfYear()"],
+  "jqlNotSupported": [],
+  "stateEntities": 37,
+  "testCount": 1028,
+  "responseMappers": true,
+  "multipartUpload": true,
+  "fieldFiltering": true,
+  "expandSupport": ["changelog", "transitions"],
   "knownGaps": [
-    "5 tools are intentionally outside the audited publishable subset: ProForma forms (3) are out of scope, and jira_update_issue plus jira_remove_issue_link remain exposed but unaudited due upstream limitations.",
-    "jira_update_issue happy-path parity is blocked by the upstream server raising a backend error when called with its documented required fields payload.",
-    "jira_remove_issue_link happy-path parity is blocked because the upstream tool surface exposes no removable link id through create_issue_link output or get_issue reads.",
-    "JQL WAS/CHANGED operators — require historical state",
-    "JQL date functions (startOfDay, endOfWeek) — low priority",
-    "Custom field queries (customfield_XXXXX) — requires dynamic schema extension"
+    "3 ProForma tools are out of scope (SaaS customer portal feature, not commonly used in dev/test).",
+    "Multipart file upload handles base64-encoded content, not raw binary file streams."
   ]
 }

package/twin-assets/ramp/fidelity.json

@@ -0,0 +1,22 @@
+{
+  "twin": "ramp",
+  "api": "ramp-agent-tools",
+  "version": "0.1.0",
+  "capabilities": [
+    {
+      "name": "Published Ramp tool names mirrored into MCP",
+      "supported": true
+    },
+    {
+      "name": "Seed-backed cards, funds, expenses, approvals, travel, and receipts",
+      "supported": true
+    }
+  ],
+  "schemaCompatibilityExceptions": [
+    "approve-or-reject-bill"
+  ],
+  "thresholds": {
+    "minWorkflowScenarios": 1,
+    "minWorkflowSteps": 10
+  }
+}

package/twin-assets/slack/fidelity.json

@@ -3,12 +3,11 @@
   "api": "Web API",
   "version": "0.1.0",
   "capabilities": [
-    { "name": "Stateful CRUD (channels, messages, users)", "supported": true },
-    { "name": "Error responses (channel_not_found, not_authed)", "supported": true },
-    { "name": "Pagination", "supported": true },
-    { "name": "Thread replies", "supported": true },
-    { "name": "Reactions and pins", "supported": true },
-    { "name": "Real-time events (WebSocket)", "supported": false },
-    { "name": "File uploads", "supported": false }
+    { "name": "Visible MCP tools exactness (8-tool exposed surface)", "supported": true },
+    { "name": "Supported Web API flows: auth.test, conversations.create/info/list/history/replies, chat.postMessage/delete, reactions.add/remove, users.list/profile.get", "supported": true },
+    { "name": "State replay and reset exactness for supported flows", "supported": true },
+    { "name": "Hidden/internal-only Slack tools", "supported": false },
+    { "name": "files.* and other non-exposed Slack API areas", "supported": false },
+    { "name": "Real-time events (WebSocket)", "supported": false }
   ]
 }

package/dist/harnesses/_lib/agent-trace.mjs

@@ -1,57 +0,0 @@
-/**
- * Structured agent trace writer for bundled harnesses.
- *
- * Records per-step model thinking, text output, and tool calls as a structured
- * JSON trace. The orchestrator reads this file after the harness exits and flows
- * it into RunResult → artifacts → dashboard.
- *
- * Transport: writes to ARCHAL_AGENT_TRACE_FILE (set by orchestrator).
- * Safe no-op when the env var is not set.
- *
- * Trace format:
- *   { version: 1, steps: [ { step, thinking, text, toolCalls, durationMs } ] }
- */
-import { writeFileSync } from 'node:fs';
-
-/**
- * @typedef {Object} TraceStep
- * @property {number} step           - 1-indexed step number
- * @property {string|null} thinking  - Model's internal reasoning (extended thinking / reasoning_content)
- * @property {string|null} text      - Model's visible text output (reasoning "out loud")
- * @property {Array<{name: string, arguments: object}>} toolCalls - Tools called this step
- * @property {number} durationMs     - LLM call duration for this step
- */
-
-/**
- * Create a trace collector that accumulates steps and writes on flush.
- * @returns {{ addStep: (step: TraceStep) => void, flush: () => void }}
- */
-export function createAgentTrace() {
-  /** @type {TraceStep[]} */
-  const steps = [];
-
-  return {
-    /**
-     * Record a single agent step.
-     * @param {TraceStep} step
-     */
-    addStep(step) {
-      steps.push(step);
-    },
-
-    /**
-     * Write the accumulated trace to the file. Call once at the end.
-     */
-    flush() {
-      const tracePath = process.env['ARCHAL_AGENT_TRACE_FILE'];
-      if (!tracePath) return;
-
-      try {
-        const payload = { version: 1, steps };
-        writeFileSync(tracePath, JSON.stringify(payload));
-      } catch {
-        // Non-fatal — trace is best-effort
-      }
-    },
-  };
-}

package/dist/harnesses/_lib/env-utils.mjs

@@ -1,23 +0,0 @@
-/**
- * Shared environment variable parsing utilities for bundled harnesses.
- */
-
-/**
- * Parse an integer from an environment variable with validation and clamping.
- * Replaces the repeated IIFE pattern across agent files.
- *
- * @param {string} envVar - Environment variable name
- * @param {number} defaultValue - Default if env var is not set or invalid
- * @param {{ min?: number, max?: number }} [opts] - Optional min/max bounds
- * @returns {number}
- */
-export function parseEnvInt(envVar, defaultValue, { min, max } = {}) {
-  const raw = process.env[envVar]?.trim();
-  if (!raw) return defaultValue;
-  const parsed = parseInt(raw, 10);
-  if (Number.isNaN(parsed)) return defaultValue;
-  let value = parsed;
-  if (min !== undefined && value < min) value = min;
-  if (max !== undefined && value > max) value = max;
-  return value;
-}

package/dist/harnesses/_lib/harness-runner.mjs

@@ -1,373 +0,0 @@
-/**
- * Shared harness scaffolding for bundled agent files.
- *
- * Extracts the common init sequence and run-loop structure that all 4
- * bundled harnesses (naive, zero-shot, hardened, react) duplicate.
- *
- * Usage:
- *   const ctx = await createHarnessContext('react');
- *   await runAgentLoop(ctx, { ... });
- */
-import { collectTwinUrls, discoverAllTools } from './rest-client.mjs';
-import {
-  detectProvider,
-  resolveApiKey,
-  formatToolsForProvider,
-  buildInitialMessages,
-  appendAssistantResponse,
-  appendToolResults,
-  appendUserInstruction,
-  callLlmWithMessages,
-  parseToolCalls,
-  getResponseText,
-  getThinkingContent,
-  getStopReason,
-  withRetry,
-} from './providers.mjs';
-import { createLogger } from './logging.mjs';
-import { writeMetrics } from './metrics.mjs';
-import { createAgentTrace } from './agent-trace.mjs';
-
-// ── Context creation ──────────────────────────────────────────────────
-
-/**
- * @typedef {object} HarnessContext
- * @property {string} harnessName
- * @property {string} task
- * @property {string} model
- * @property {string} provider
- * @property {string} apiKey
- * @property {import('./logging.mjs').Logger} log
- * @property {Record<string, string>} twinUrls
- * @property {Array<{ name: string, description: string, inputSchema: object }>} allTools
- * @property {Record<string, { twinName: string, baseUrl: string, originalName: string }>} toolToTwin
- */
-
-/**
- * Create the full harness context: validate env vars, detect provider,
- * resolve API key, collect twin URLs, and discover tools.
- *
- * Exits with code 1 on missing env vars or unreachable twins.
- *
- * @param {string} harnessName
- * @returns {Promise<HarnessContext>}
- */
-export async function createHarnessContext(harnessName) {
-  let task = (process.env['ARCHAL_ENGINE_TASK'] || '').trim();
-  // If the task value is a file path inside the mounted config dir,
-  // read the actual task content from the file. This happens when the
-  // Docker harness writes multi-line task text to a file to avoid
-  // exposing it via docker -e flags (security: prevents secret leakage
-  // in docker ps / /proc/<pid>/cmdline).
-  if (task.startsWith('/archal-out/') || task.startsWith(process.env['ARCHAL_ENGINE_TASK_FILE'] ? '/' : '\0')) {
-    try {
-      const { readFileSync } = await import('node:fs');
-      task = readFileSync(task, 'utf-8').trim();
-    } catch { /* fall through to original value */ }
-  }
-  // Also check the _FILE convention: if ARCHAL_ENGINE_TASK is empty but
-  // ARCHAL_ENGINE_TASK_FILE points to a file, read from there.
-  if (!task && process.env['ARCHAL_ENGINE_TASK_FILE']) {
-    try {
-      const { readFileSync } = await import('node:fs');
-      task = readFileSync(process.env['ARCHAL_ENGINE_TASK_FILE'], 'utf-8').trim();
-    } catch { /* fall through */ }
-  }
-  const model = process.env['ARCHAL_ENGINE_MODEL'];
-
-  if (!task) { console.error('ARCHAL_ENGINE_TASK not set or empty'); process.exit(1); }
-  if (!model) { console.error('ARCHAL_ENGINE_MODEL not set'); process.exit(1); }
-
-  const provider = detectProvider(model);
-  const apiKey = resolveApiKey(provider);
-  const log = createLogger({ harness: harnessName, model, provider });
-
-  const twinUrls = collectTwinUrls();
-  if (Object.keys(twinUrls).length === 0) {
-    console.error(`[${harnessName}] No twin URLs found. Check ARCHAL_TWIN_NAMES and ARCHAL_<TWIN>_URL env vars.`);
-    process.exit(1);
-  }
-
-  const { tools: allTools, toolToTwin } = await discoverAllTools(twinUrls);
-  if (allTools.length === 0) {
-    console.error(`[${harnessName}] No tools discovered from twins. Twin endpoints may be unreachable.`);
-    process.exit(1);
-  }
-
-  return { harnessName, task, model, provider, apiKey, log, twinUrls, allTools, toolToTwin };
-}
-
-// ── Run loop ──────────────────────────────────────────────────────────
-
-/**
- * @typedef {object} RunLoopOptions
- * @property {string} systemPrompt - System prompt text (empty string for none)
- * @property {number} maxSteps - Maximum iteration count
- * @property {boolean} [useRetry=false] - Wrap LLM calls in withRetry
- * @property {number} [retryCount=4] - Max retries when useRetry is true
- * @property {boolean} [useTrace=false] - Record agent trace
- * @property {number} [maxConsecutiveErrors=0] - Bail threshold (0 = no limit)
- * @property {number} [maxInitialNoToolRecoveries=0] - Reprompt attempts when model doesn't call tools initially
- * @property {(ctx: HarnessContext, state: RunState) => Array} [selectTools] -
- *   Per-step tool selection function. Receives context and current state,
- *   returns the MCP tools array for this step. Default: use all tools.
- * @property {(ctx: HarnessContext, state: RunState, stepResult: StepResult) => 'continue' | 'break' | void} [onBeforeToolExecution] -
- *   Hook called after parsing tool calls but before executing them.
- *   Return 'continue' to skip tool execution and loop, 'break' to stop.
- * @property {(provider: string, messages: Array|object) => Array|object} [initMessages] -
- *   Optional post-init hook to modify the initial messages array before the
- *   run loop starts (e.g. to prepend a triage instruction).
- * @property {(ctx: HarnessContext, state: RunState, stepResult: StepResult) => void} [onAfterToolExecution] -
- *   Hook called after tool results are appended. Return value is ignored.
- * @property {(ctx: HarnessContext, state: RunState, stepResult: StepResult) => 'continue' | void} [onNoToolCalls] -
- *   Hook called when the model responds without tool calls. Return
- *   'continue' to add instructions and continue the loop.
- * @property {(tc: { name: string, arguments: object }) => void} [onToolSuccess] -
- *   Called after each successful tool call.
- */
-
-/**
- * @typedef {object} RunState
- * Mutable state tracked across loop iterations.
- * @property {Array|object} messages
- * @property {number} stepsCompleted
- * @property {number} totalInputTokens
- * @property {number} totalOutputTokens
- * @property {number} totalToolCalls
- * @property {number} totalToolErrors
- * @property {number} consecutiveErrors
- * @property {number} initialNoToolRecoveries
- * @property {string} exitReason
- * @property {import('./agent-trace.mjs').ReturnType<typeof createAgentTrace>|null} agentTrace
- */
-
-/**
- * @typedef {object} StepResult
- * @property {number} step - 1-indexed step number
- * @property {object} response - Raw LLM response wrapper
- * @property {Array|null} toolCalls - Parsed tool calls or null
- * @property {string|null} thinking - Model thinking content
- * @property {string|null} text - Model text content
- * @property {number} iterDurationMs
- * @property {string|null} stopReason
- */
-
-/**
- * Run the agent loop with shared metrics, logging, and tool execution.
- *
- * @param {HarnessContext} ctx
- * @param {RunLoopOptions} opts
- */
-export async function runAgentLoop(ctx, opts) {
-  const {
-    systemPrompt,
-    maxSteps,
-    useRetry = false,
-    retryCount = 4,
-    useTrace = false,
-    maxConsecutiveErrors = 0,
-    maxInitialNoToolRecoveries = 0,
-    selectTools,
-    onBeforeToolExecution,
-    onAfterToolExecution,
-    onNoToolCalls,
-    onToolSuccess,
-  } = opts;
-
-  const { harnessName, task, model, provider, apiKey, log, allTools, toolToTwin } = ctx;
-
-  let messages = buildInitialMessages(provider, systemPrompt, task, model);
-
-  // Allow callers to modify initial messages (e.g. react's triage instruction)
-  if (opts.initMessages) {
-    messages = opts.initMessages(provider, messages);
-  }
-
-  const state = {
-    messages,
-    stepsCompleted: 0,
-    totalInputTokens: 0,
-    totalOutputTokens: 0,
-    totalToolCalls: 0,
-    totalToolErrors: 0,
-    consecutiveErrors: 0,
-    initialNoToolRecoveries: 0,
-    exitReason: 'max_steps',
-    agentTrace: useTrace ? createAgentTrace() : null,
-  };
-
-  const runStart = Date.now();
-
-  log.info('run_start', { task: task.slice(0, 200), maxSteps });
-
-  try {
-    for (let step = 0; step < maxSteps; step++) {
-      state.stepsCompleted = step + 1;
-      const iterStart = Date.now();
-
-      // Select tools for this step (default: all tools)
-      const stepTools = selectTools ? selectTools(ctx, state) : allTools;
-      const providerTools = formatToolsForProvider(provider, stepTools);
-
-      // Call the LLM (optionally with retry)
-      log.llmCall(step + 1);
-      let response;
-      try {
-        const llmCall = () => callLlmWithMessages(provider, model, apiKey, state.messages, providerTools);
-        response = useRetry ? await withRetry(llmCall, retryCount) : await llmCall();
-      } catch (err) {
-        const msg = err?.message ?? String(err);
-        log.error('llm_call_failed', { step: step + 1, error: msg });
-        process.stderr.write(`[${harnessName}] LLM API error: ${msg.slice(0, 500)}\n`);
-        state.exitReason = 'llm_error';
-        break;
-      }
-
-      const iterDurationMs = Date.now() - iterStart;
-      state.totalInputTokens += response.usage.inputTokens;
-      state.totalOutputTokens += response.usage.outputTokens;
-
-      const toolCalls = parseToolCalls(provider, response);
-      const hasToolCalls = !!toolCalls;
-      const stopReason = getStopReason(provider, response);
-      log.llmResponse(step + 1, iterDurationMs, hasToolCalls, stopReason);
-      log.tokenUsage(step + 1, response.usage, {
-        inputTokens: state.totalInputTokens,
-        outputTokens: state.totalOutputTokens,
-      });
-
-      const thinking = getThinkingContent(provider, response);
-      const text = getResponseText(provider, response);
-
-      state.messages = appendAssistantResponse(provider, state.messages, response);
-
-      /** @type {StepResult} */
-      const stepResult = { step: step + 1, response, toolCalls, thinking, text, iterDurationMs, stopReason };
-
-      if (!toolCalls) {
-        // Record trace for no-tool-call steps
-        if (state.agentTrace) {
-          state.agentTrace.addStep({ step: step + 1, thinking, text, toolCalls: [], durationMs: iterDurationMs });
-        }
-        if (text) {
-          process.stderr.write(`[${harnessName}] Step ${step + 1}: ${text.slice(0, 200)}\n`);
-        }
-
-        // Initial no-tool recovery (reprompt)
-        const shouldRecoverInitial = state.totalToolCalls === 0
-          && maxInitialNoToolRecoveries > 0
-          && state.initialNoToolRecoveries < maxInitialNoToolRecoveries;
-        if (shouldRecoverInitial) {
-          state.initialNoToolRecoveries++;
-          state.messages = appendUserInstruction(
-            provider,
-            state.messages,
-            'You must use tools to make progress. ' +
-              'On your next response, call at least one relevant tool before giving any summary or conclusion. ' +
-              'Start by gathering concrete evidence from the systems, then execute the required actions.',
-          );
-          log.info('no_tool_calls_reprompt', {
-            step: step + 1,
-            attempt: state.initialNoToolRecoveries,
-          });
-          continue;
-        }
-
-        // Harness-specific no-tool-call handling
-        if (onNoToolCalls) {
-          const directive = onNoToolCalls(ctx, state, stepResult);
-          if (directive === 'continue') continue;
-        }
-
-        state.exitReason = state.totalToolCalls === 0 ? 'no_tool_calls' : 'completed';
-        break;
-      }
-
-      state.initialNoToolRecoveries = 0;
-
-      // Pre-execution hook (e.g. react's repo content guard)
-      if (onBeforeToolExecution) {
-        const directive = onBeforeToolExecution(ctx, state, stepResult);
-        if (directive === 'continue') continue;
-        if (directive === 'break') break;
-      }
-
-      // Execute tool calls
-      const { executeToolCalls } = await import('./tool-executor.mjs');
-      const { results, bailout } = await executeToolCalls(toolCalls, {
-        toolToTwin,
-        harnessName,
-        step: step + 1,
-        log,
-        counters: state,
-        maxConsecutiveErrors,
-        onSuccess: onToolSuccess,
-      });
-
-      // Record trace
-      if (state.agentTrace) {
-        state.agentTrace.addStep({
-          step: step + 1,
-          thinking,
-          text,
-          toolCalls: toolCalls.map((tc) => ({ name: tc.name, arguments: tc.arguments })),
-          durationMs: iterDurationMs,
-        });
-      }
-
-      if (bailout) {
-        state.exitReason = 'consecutive_errors';
-        break;
-      }
-
-      // Append tool results to conversation
-      state.messages = appendToolResults(provider, state.messages, toolCalls, results);
-
-      // Post-execution hook
-      if (onAfterToolExecution) {
-        onAfterToolExecution(ctx, state, stepResult);
-      }
-    }
-  } finally {
-    const totalTimeMs = Date.now() - runStart;
-
-    log.summary({
-      iterations: state.stepsCompleted,
-      totalInputTokens: state.totalInputTokens,
-      totalOutputTokens: state.totalOutputTokens,
-      totalTimeMs,
-      toolCallCount: state.totalToolCalls,
-      toolErrorCount: state.totalToolErrors,
-      exitReason: state.exitReason,
-    });
-
-    writeMetrics({
-      inputTokens: state.totalInputTokens,
-      outputTokens: state.totalOutputTokens,
-      llmCallCount: state.stepsCompleted,
-      toolCallCount: state.totalToolCalls,
-      toolErrorCount: state.totalToolErrors,
-      totalTimeMs,
-      exitReason: state.exitReason,
-      provider,
-      model,
-    });
-
-    if (state.agentTrace) {
-      state.agentTrace.flush();
-    }
-
-    process.stderr.write(
-      `\n[${harnessName}] Summary: ${state.stepsCompleted} iterations, ${state.totalToolCalls} tool calls ` +
-      `(${state.totalToolErrors} errors), ${state.totalInputTokens} input tokens, ` +
-      `${state.totalOutputTokens} output tokens, ${(totalTimeMs / 1000).toFixed(1)}s total\n`
-    );
-
-    if (state.exitReason === 'llm_error') {
-      process.exit(1);
-    }
-  }
-}
-
-// Re-export for convenience — harnesses that need to build custom initial messages
-export { appendUserInstruction };