macOS: live wrapper PID fails identity validation after exec, risking auto-reap and missing results

[zhoushoujianwork]

Summary

On macOS, getProcessIdentity() includes comm in the identity and describes the result as immutable. A legitimate exec changes comm while preserving the PID and process start time. Consequently, validateProcessIdentity() returns false for a live, non-reused PID.

This matters for the supported CC_PLUGIN_CODEX_CLAUDE_BIN override when the executable is a shell wrapper. runClaudeTurn() captures its identity immediately after spawning, and reapStaleJobs() interprets an identity mismatch as a dead process and marks the job failed.

Related: #79 and #83 (closed/fixed permission-denied cases). This report concerns an identity mismatch after exec, not EPERM/EACCES.

Environment

• cc@sendbird / cc-plugin-codex 1.5.0
• Codex CLI 0.153.4; Claude Code 2.1.223
• macOS; Node.js 24.5.0
• Custom Claude executable wrapper configured through CC_PLUGIN_CODEX_CLAUDE_BIN
• Wrapper reads literal gateway environment settings, then launches Claude; no credentials included here
• Current upstream main inspected at 19e5651 contains the relevant identity implementation

Deterministic minimal reproduction (no model call)

Run from a checkout of this repository on macOS with node on PATH:

node --input-type=module - <<'NODE'
import { spawn } from 'node:child_process';
import {
  getProcessIdentity,
  validateProcessIdentity,
  isProcessAlive,
} from './scripts/lib/process.mjs';

const child = spawn('/bin/zsh', [
  '-c',
  `read -r trigger; exec /usr/bin/env node -e 'process.stdout.write("ready");setTimeout(()=>{},10000)'`,
], { stdio: ['pipe', 'pipe', 'pipe'] });

try {
  await new Promise(resolve => setTimeout(resolve, 100));
  const before = getProcessIdentity(child.pid);
  const ready = new Promise((resolve, reject) => {
    child.stdout.once('data', resolve);
    child.once('error', reject);
  });
  child.stdin.end('go\n');
  await ready;
  const after = getProcessIdentity(child.pid);
  console.log({
    samePid: true,
    sameStartTime: before.slice(0, 24) === after.slice(0, 24),
    identityChanged: before !== after,
    alive: isProcessAlive(child.pid),
    identityValidated: validateProcessIdentity(child.pid, before),
  });
} finally {
  child.kill('SIGTERM'); // Only this reproduction's own child.
}
NODE

Observed against installed 1.5.0:

{"samePid":true,"sameStartTime":true,"identityChanged":true,"alive":true,"identityValidated":false}

The child remains the same process across exec; it has not exited and its PID has not been reused.

Real-world missing-result incident (supporting evidence, root cause not established)

A long, read-only Claude task launched through the Codex forwarding agent produced this timeline on 2026-09-10 (UTC):

Time	Observation
06:11:47	Tracked task started.
06:16:37	Claude tool activity continued.
06:16:44	Job receipt transitioned to failed: Process <pid> died without completing. Auto-reaped. No result field was saved.
06:18:08	The forwarding agent explicitly terminated the still-running companion with SIGTERM; companion exited 143.
06:19:39	The exact original Claude session wrote an end_turn assistant text of 13,077 characters. No additional user prompt was present.

The final answer was recovered from the original native Claude transcript; the tracked job remained failed without that answer.

Limitations: the receipt cleared the original pidIdentity during reaping, so we cannot prove the exec mismatch caused this specific incident. The explicit later termination is also a confounding factor for result delivery; we are not claiming that the plugin spontaneously killed the companion. The confirmed defect is the independently reproduced false identity verdict. The incident demonstrates that a failed tracker state does not by itself prove that underlying Claude work has stopped.

Short real calls succeed: a fresh Codex forwarding agent received an exact fixed-token response from Claude in approximately 8 seconds using the same wrapper, and the completed receipt contained that response. Authentication currently works.

Expected behavior / suggested regression coverage

• A legitimate exec should not invalidate the identity of the same live process.
• Preserve PID-reuse protection using an identity based on process birth/start information that remains stable across exec; do not simply discard identity checks.
• Add a macOS shell-to-executable regression test and retain the existing denied-probe and dead/PID-reuse coverage.
• Consider preserving diagnostic evidence for the reaping decision so the original expected/observed identity or probe failure can be investigated.
• Treat wrapper/supervisor liveness and underlying Claude completion carefully; surface missing-result/unknown outcomes truthfully without automatically rewriting historical receipts as successful.

Private prompts, checkout paths, credentials, native transcripts, and session identifiers have been omitted.