@magnusekdahl/parallix 1.2.0 → 1.3.0

package/lib/agents/opencode.ts

@@ -0,0 +1,370 @@
+import { spawnAndTee } from '../core/spawn-tee.js';
+import { extractOpencodeTelemetryFromExport } from './opencode-telemetry.js';
+import { captureOpencodeExport } from './opencode-export.js';
+import { detectLimitHit } from './limit-hit.js';
+// tools/sessions and core/subagent-limit are still CJS (not converted in this
+// wave); require keeps them untyped (any) without pulling non-included .js into
+// the typecheck program.
+const sessions = require('../tools/sessions');
+const { buildSubagentLimitPrefix } = require('../core/subagent-limit');
+
+interface BuildOpencodeInvocationOptions {
+  prompt: string;
+  worktree: string;
+  env?: object;
+  resume?: boolean;
+  sessionId?: string | null;
+  model?: string | null;
+  preferJson?: boolean;
+}
+
+interface StartOpencodeAgentOptions {
+  prompt: string;
+  worktree: string;
+  env?: object;
+  resume?: boolean;
+  sessionId?: string | null;
+  model?: string | null;
+  teeOptions?: object;
+  slug?: string | null;
+  role?: string | null;
+  maxTransientRetries?: number;
+}
+
+// Injectable I/O for tests. Production uses the real spawn-tee / export capture.
+let _spawnAndTee: any = spawnAndTee;
+let _captureExport: any = captureOpencodeExport;
+let _sessions: any = sessions;
+
+// Test hooks: override the launcher's I/O without touching the public signature.
+function __setSpawnAndTeeForTest(fn: any) { _spawnAndTee = fn || spawnAndTee; }
+function __setExportCaptureForTest(fn: any) { _captureExport = fn || captureOpencodeExport; }
+function __setSessionsForTest(mod: any) { _sessions = mod || sessions; }
+
+// Test hook: override the cached feature-detect for `--format json` support.
+// Set to true/false to force inclusion/exclusion of the flag regardless of
+// the real binary.  Pass null to reset to the live detect.
+function __setJsonFormatSupportForTest(val: boolean | null) {
+  _jsonFormatSupported = val;
+  _jsonFormatDetectFn = null;
+}
+
+// Test hook: inject a feature-detect function for tests.  This replaces the
+// live shell-out with a stubbed function, making tests hermetic.
+function __setJsonFormatDetectForTest(fn: any) {
+  _jsonFormatDetectFn = fn;
+  _jsonFormatSupported = null;  // invalidate cache so the detect fn runs
+}
+
+// Session-id recovery for telemetry. After session completion `opencode export`
+// is invoked to extract token-usage data (see opencode-telemetry.js), so the
+// session id MUST be recoverable from the launcher's stdout. Stats hooks in
+// active.js and review-loop.js call recordStageStats which defaults to '0' for
+// tokens when telemetry is null.
+//
+// Two shapes are recognized:
+//   1. Legacy footer:  "Continue  opencode -s ses_<id>"  (older opencode TUI).
+//   2. JSON stream:     `"sessionID":"ses_<id>"`           (opencode v2.0.0
+//      `run --format json`). v2.0.0's plain `run` no longer prints the footer,
+//      so without the JSON field the id is lost and custom telemetry drops to all
+//      zeros (task-1339). We request `--format json` in buildOpencodeInvocation
+//      so this field is always present.
+const OPENCODE_SESSION_ID_RE = /opencode\s+-s\s+(ses_\S+)/i;
+const OPENCODE_JSON_SESSION_ID_RE = /"sessionID"\s*:\s*"(ses_[^"]+)"/;
+
+function extractOpencodeSessionId(stdout: string) {
+  if (!stdout) {return null;}
+  const footer = OPENCODE_SESSION_ID_RE.exec(stdout);
+  if (footer) {return footer[1];}
+  const json = OPENCODE_JSON_SESSION_ID_RE.exec(stdout);
+  return json ? json[1] : null;
+}
+
+// Cached feature-detect for `--format json` support. Set to `false` only when
+// a real invocation proves the flag is rejected (older opencode versions).
+// Tests can inject a canned result via __setJsonFormatSupportForTest.
+let _jsonFormatSupported: boolean | null = null;
+
+// Injectable feature-detect function for tests.  When set, checkJsonFormatSupport()
+// calls this function instead of shelling out, making tests hermetic.
+let _jsonFormatDetectFn: (() => any) | null = null;
+
+function resolveOpencodeCommand() {
+  return 'opencode';
+}
+
+// Lightweight feature-detect: try `opencode --format json --help`.
+// Returns true if the flag is accepted (exit 0 or non-zero for unrelated reasons),
+// false if the flag itself is rejected ("unrecognized" / "unknown option").
+// Results are cached so the check runs at most once in production.
+// Tests can inject a canned result via __setJsonFormatSupportForTest.
+function checkJsonFormatSupport() {
+  if (_jsonFormatSupported !== null) {return _jsonFormatSupported;}
+  if (_jsonFormatDetectFn) {
+    _jsonFormatSupported = _jsonFormatDetectFn();
+    return _jsonFormatSupported;
+  }
+  try {
+    const { spawnSync } = require('node:child_process');
+    const result = spawnSync('opencode', ['--format', 'json', '--help'], {
+      timeout: 3000,
+      stdio: ['ignore', 'pipe', 'pipe'],
+    });
+    const stderr = (result.stderr && result.stderr.toString()) || '';
+    const combined = stderr;
+    // Flag rejected → fallback to legacy invocation.
+    if (/unrecognized|unknown option|no such option|invalid option/i.test(combined)) {
+      _jsonFormatSupported = false;
+      return false;
+    }
+    _jsonFormatSupported = true;
+    return true;
+  } catch (_) {
+    // ENOENT or other spawn errors → assume legacy (no JSON flag).
+    _jsonFormatSupported = false;
+    return false;
+  }
+}
+
+function buildOpencodeInvocation({ prompt, worktree, env, resume = false, sessionId = null, model = null, preferJson = true }: BuildOpencodeInvocationOptions) {
+  // `--format json` makes opencode stream NDJSON events that each carry a
+  // "sessionID":"ses_..." field. opencode v2.0.0's default `run` output no
+  // longer prints the "Continue  opencode -s ses_..." footer, so JSON is the
+  // only reliable way to recover the session id needed for `opencode export`
+  // telemetry capture (task-1339).
+  //
+  // Compatibility guard: feature-detect the flag on first call; older opencode
+  // versions that reject `--format json` will silently fall back to the legacy
+  // invocation, preserving launch behaviour at the cost of telemetry.
+  const useJson = preferJson && checkJsonFormatSupport();
+  const args = ['run', '--pure', '--dangerously-skip-permissions'];
+  if (useJson) {args.push('--format', 'json');}
+  if (model) {args.push('-m', model);}
+  // Only resume when the marker explicitly says so (resume=true).
+  // sessionId from a stale cross-family marker must be ignored when resume=false.
+  if (resume && sessionId) {
+    args.push('-s', sessionId);
+  } else if (resume) {
+    args.push('--continue');
+  }
+  args.push(prompt);
+  return {
+    command: resolveOpencodeCommand(),
+    args,
+    options: {
+      stdio: 'inherit',
+      cwd: worktree,
+      env: { ...process.env, ...env }
+    }
+  };
+}
+
+// Transient custom/vLLM backend failures that opencode surfaces as a non-zero
+// exit. These are recoverable: the same prompt frequently succeeds on a second
+// attempt, so the launcher retries once in-family before agents.js reroutes to a
+// different agent family. Limit-hit phrases are intentionally NOT listed here —
+// detectLimitHit() in agents.js owns that accounting (lib/agents/limit-hit.js).
+// Hard errors (model-not-found, auth, ENOENT/EACCES) are never retried.
+const TRANSIENT_OPENCODE_PATTERNS = [
+  /\bECONNRESET\b/i,
+  /\bECONNREFUSED\b/i,
+  /\bETIMEDOUT\b/i,
+  /\bENETUNREACH\b/i,
+  /\bENOTFOUND\b/i,
+  /\bEAI_AGAIN\b/i,
+  /\bEPIPE\b/i,
+  /socket hang ?up/i,
+  /\bfetch failed\b/i,
+  /network (?:error|timeout)/i,
+  /connection (?:error|reset|closed|refused|timed? ?out)/i,
+  /\b(?:502|503|504|529)\b/,
+  /\bbad gateway\b/i,
+  /\bservice (?:is )?unavailable\b/i,
+  /\bgateway time-?out\b/i,
+  /\b(?:server |api |model )?overloaded\b/i,
+  /\btemporarily unavailable\b/i,
+  /\bplease (?:try|retry) again\b/i,
+  /\bretry[- ]after\b/i,
+];
+
+// Hard, non-retryable failures: retrying cannot help and only wastes time/quota.
+const HARD_OPENCODE_PATTERNS = [
+  /\bmodel not found\b/i,
+  /\bno such model\b/i,
+  /\bunknown model\b/i,
+  /\binvalid (?:api )?key\b/i,
+  /\b401\b[^\n]*\bunauthorized\b/i,
+  /\bauthentication (?:failed|error)\b/i,
+];
+
+function failureText(result: any) {
+  return `${(result && result.stdout) || ''}\n${(result && result.stderr) || ''}`;
+}
+
+function isHardOpencodeFailure(result: any) {
+  if (!result) {return false;}
+  if (result.error && (result.error.code === 'ENOENT' || result.error.code === 'EACCES')) {return true;}
+  const text = failureText(result);
+  return HARD_OPENCODE_PATTERNS.some((re) => re.test(text));
+}
+
+// Opencode v2.0.0 JSON-mode sometimes exits with code 1 during post-run
+// cleanup (race closing the stdout pipe, EPIPE, or draining pending events).
+// When the transcript contains a valid "reason":"stop" completion event and
+// no "type":"error" event, the exit is spurious – the agent actually
+// completed successfully. See opencode issues #31446 and #33653.
+function isSpuriousOpencodeExit(result: any) {
+  if (!result || result.status !== 1) {return false;}
+  const text = failureText(result);
+  const hasStop = /"reason"\s*:\s*"stop"/.test(text);
+  const hasError = /"type"\s*:\s*"error"/.test(text);
+  return hasStop && !hasError;
+}
+
+function isTransientOpencodeFailure(result: any) {
+  if (!result) {return false;}
+  const text = failureText(result);
+  return TRANSIENT_OPENCODE_PATTERNS.some((re) => re.test(text));
+}
+
+// Decide whether a failed opencode run should be retried once in-family.
+// Only genuine non-zero exits carrying a transient backend signature qualify;
+// clean exits, spawn errors (ENOENT/EACCES), killing signals, limit-hits, and
+// hard errors never retry — so real failures still surface to the reroute loop.
+function shouldRetryOpencodeFailure(result: any) {
+  if (!result) {return false;}
+  if (result.error && (result.error.code === 'ENOENT' || result.error.code === 'EACCES')) {return false;}
+  if (result.signal) {return false;}
+  if (!(typeof result.status === 'number' && result.status !== 0)) {return false;}
+  if (detectLimitHit({
+    agent: 'custom',
+    stdout: result.stdout,
+    stderr: result.stderr,
+    status: result.status,
+    signal: result.signal,
+    error: result.error,
+  })) {return false;}
+  if (isHardOpencodeFailure(result)) {return false;}
+  return isTransientOpencodeFailure(result);
+}
+
+function startOpencodeAgent({
+  prompt,
+  worktree,
+  env,
+  resume = false,
+  sessionId = null,
+  model = null,
+  teeOptions = {},
+  slug = null,
+  role = null,
+  maxTransientRetries = 1
+}: StartOpencodeAgentOptions) {
+  // Prepend the subagent-limit advisory prefix to the prompt.
+  // When maxParallel is unset/null/zero the prefix is empty (no-op).
+  const subagentPrefix = buildSubagentLimitPrefix();
+  const injectedPrompt = subagentPrefix + prompt;
+
+  function isStaleSessionResult(result: any) {
+    if (!result) {return false;}
+    const stderr = result.stderr || '';
+    const stdout = result.stdout || '';
+    return stderr.includes('Session not found') || stdout.includes('Session not found');
+  }
+
+  function runInvocation(invocation: any) {
+    const spawnOptions = { ...invocation.options, ...teeOptions };
+    return _spawnAndTee(invocation.command, invocation.args, spawnOptions);
+  }
+
+  async function runInvocationWithRetry(invocation: any) {
+    let attempts = 0;
+    let result = await runInvocation(invocation);
+    while (attempts < maxTransientRetries && shouldRetryOpencodeFailure(result)) {
+      attempts += 1;
+      result = await runInvocation(invocation);
+    }
+    if (result) {result.transientRetries = attempts;}
+    return result;
+  }
+
+  // Detect whether `--format json` was rejected at runtime (e.g. older opencode
+  // that wasn't caught by the feature-detect).  If so, retry the invocation
+  // without the flag so the prompt still runs (telemetry will be lost but the
+  // agent won't fail outright).
+  function isJsonFlagError(result: any) {
+    if (!result) {return false;}
+    const text = failureText(result);
+    return (/\bunrecognized option\b|\bunknown option\b|\bno such option\b|\binvalid option\b|\bunrecognized flag\b|\bunknown flag\b/i).test(text);
+  }
+
+  async function processResult(result: any) {
+    if (result && result.stdout) {
+      result.sessionId = extractOpencodeSessionId(result.stdout) || undefined;
+    }
+    if (result && result.sessionId) {
+      try {
+        const exportJson = await _captureExport(result.sessionId, { worktree, env });
+        if (exportJson) {
+          const telemetry = extractOpencodeTelemetryFromExport(exportJson, model || undefined);
+          if (telemetry) {
+            result.telemetry = telemetry;
+            if ((telemetry as any).model) {result.model = (telemetry as any).model;}
+            if ((telemetry as any).provider) {result.provider = (telemetry as any).provider;}
+          }
+        }
+      } catch (_) {
+        return result;
+      }
+    }
+    return result;
+  }
+
+  async function runWithJsonFallback(invocation: any) {
+    let result = await runInvocationWithRetry(invocation);
+    // Runtime fallback: if the first invocation used --format json but the
+    // binary rejected it, retry without the flag so the prompt still executes.
+    if (isJsonFlagError(result) && invocation.args.includes('--format')) {
+      const legacyInv = buildOpencodeInvocation({
+        prompt: injectedPrompt, worktree, env, resume, sessionId, model, preferJson: false,
+      });
+      result = await runInvocationWithRetry(legacyInv);
+      // Mark that we fell back so callers know telemetry may be absent.
+      if (result) {result._jsonFallback = true;}
+    }
+    return result;
+  }
+
+  async function staleSessionHandler(invocation: any) {
+    let result = await runWithJsonFallback(invocation);
+    if (isStaleSessionResult(result) && worktree && resume) {
+      try {
+        _sessions.clearSession(worktree, slug || '', role || '');
+      } catch (_) { /* best-effort */ }
+      const freshInv = buildOpencodeInvocation({ prompt: injectedPrompt, worktree, env, resume: false, sessionId: null, model });
+      result = await runWithJsonFallback(freshInv);
+    }
+    return processResult(result);
+  }
+
+  const invocation = buildOpencodeInvocation({ prompt: injectedPrompt, worktree, env, resume, sessionId, model });
+  const resultPromise = staleSessionHandler(invocation);
+
+  return { invocation, resultPromise };
+}
+
+export {
+  buildOpencodeInvocation,
+  extractOpencodeSessionId,
+  resolveOpencodeCommand,
+  startOpencodeAgent,
+  isSpuriousOpencodeExit,
+  isTransientOpencodeFailure,
+  isHardOpencodeFailure,
+  shouldRetryOpencodeFailure,
+  __setSpawnAndTeeForTest,
+  __setExportCaptureForTest,
+  __setSessionsForTest,
+  __setJsonFormatSupportForTest,
+  __setJsonFormatDetectForTest,
+};

package/lib/agents/stage-telemetry.js

@@ -1,5 +1,6 @@
 'use strict';
-
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.resolveStageTelemetry = resolveStageTelemetry;
 // Shared resolution of the most faithful telemetry for a recorded workflow
 // stage (execute/review). Codex's authoritative usage is the windowed sum read
 // from its on-disk rollout — `sinceMs` bounds the read to just this stage — so
@@ -13,30 +14,28 @@
 // fixes was active.js computing `sinceMs` and then discarding it, recording
 // cumulative Codex usage for the execute phase instead of the windowed delta
 // (task-1285 review F1/F6).
-
-const { codexHomeRoot, extractCodexTelemetry } = require('./codex');
-
+const codex_js_1 = require("./codex.js");
 /**
- * @param {{ worktree: string, result: { telemetry?: { provider?: string } & object }, sinceMs?: number }} opts
- * @returns {object|null} normalized telemetry, or null when none is available.
+ * @returns normalized telemetry, or null when none is available.
  */
-function resolveStageTelemetry({ worktree, result, sinceMs = 0 } = /** @type {{ worktree: string, result: { telemetry?: { provider?: string } & object }, sinceMs?: number }} */({})) {
-  if (!result || !result.telemetry) {return null;}
-  try {
-    const codexT = extractCodexTelemetry(codexHomeRoot(worktree), { sinceMs: sinceMs || 0 });
-    // Guard against stale Codex rollouts bleeding into non-Codex stages:
-    // if the launcher attached telemetry with a provider that differs from
-    // what Codex would produce (e.g. anthropic vs openai), discard the Codex
-    // read and use the launcher-attached telemetry instead.
-    if (codexT && codexT.provider && result.telemetry.provider
-        && codexT.provider.toLowerCase() !== result.telemetry.provider.toLowerCase()) {
-      return result.telemetry;
+function resolveStageTelemetry({ worktree, result, sinceMs = 0 } = {}) {
+    if (!result || !result.telemetry) {
+        return null;
+    }
+    try {
+        const codexT = (0, codex_js_1.extractCodexTelemetry)((0, codex_js_1.codexHomeRoot)(worktree), { sinceMs: sinceMs || 0 });
+        // Guard against stale Codex rollouts bleeding into non-Codex stages:
+        // if the launcher attached telemetry with a provider that differs from
+        // what Codex would produce (e.g. anthropic vs openai), discard the Codex
+        // read and use the launcher-attached telemetry instead.
+        if (codexT && codexT.provider && result.telemetry.provider
+            && codexT.provider.toLowerCase() !== result.telemetry.provider.toLowerCase()) {
+            return result.telemetry;
+        }
+        return codexT || result.telemetry;
+    }
+    catch (_) {
+        // Codex rollout unreadable: fall back to this stage's launcher telemetry.
+        return result.telemetry;
     }
-    return codexT || result.telemetry;
-  } catch (_) {
-    // Codex rollout unreadable: fall back to this stage's launcher telemetry.
-    return result.telemetry;
-  }
 }
-
-module.exports = { resolveStageTelemetry };

package/lib/agents/stage-telemetry.ts

@@ -0,0 +1,47 @@
+'use strict';
+
+// Shared resolution of the most faithful telemetry for a recorded workflow
+// stage (execute/review). Codex's authoritative usage is the windowed sum read
+// from its on-disk rollout — `sinceMs` bounds the read to just this stage — so
+// prefer it when present. For non-Codex agents (e.g. Claude, whose usage the
+// launcher parses from stdout via claude-telemetry.js) no codex rollout exists
+// and extractCodexTelemetry returns null *without throwing*, so we fall back to
+// the telemetry the launcher already attached.
+//
+// This lives in one place so the execute hook (commands/active.js) and the
+// review-loop hook (review/review-loop.js) cannot drift apart — the bug this
+// fixes was active.js computing `sinceMs` and then discarding it, recording
+// cumulative Codex usage for the execute phase instead of the windowed delta
+// (task-1285 review F1/F6).
+
+import { codexHomeRoot, extractCodexTelemetry } from './codex.js';
+
+interface StageTelemetryOptions {
+  worktree: string;
+  result: { telemetry?: { provider?: string } & Record<string, any> };
+  sinceMs?: number;
+}
+
+/**
+ * @returns normalized telemetry, or null when none is available.
+ */
+function resolveStageTelemetry({ worktree, result, sinceMs = 0 }: StageTelemetryOptions = {} as StageTelemetryOptions) {
+  if (!result || !result.telemetry) {return null;}
+  try {
+    const codexT = extractCodexTelemetry(codexHomeRoot(worktree), { sinceMs: sinceMs || 0 });
+    // Guard against stale Codex rollouts bleeding into non-Codex stages:
+    // if the launcher attached telemetry with a provider that differs from
+    // what Codex would produce (e.g. anthropic vs openai), discard the Codex
+    // read and use the launcher-attached telemetry instead.
+    if (codexT && codexT.provider && result.telemetry.provider
+        && codexT.provider.toLowerCase() !== result.telemetry.provider.toLowerCase()) {
+      return result.telemetry;
+    }
+    return codexT || result.telemetry;
+  } catch (_) {
+    // Codex rollout unreadable: fall back to this stage's launcher telemetry.
+    return result.telemetry;
+  }
+}
+
+export { resolveStageTelemetry };