Summary

• Problem: direct codex-cli runs still finalize the user-visible reply from stdout even though Codex can write a dedicated final-message artifact
• Why it matters: stdout is useful for transport and telemetry, but it is a weaker source of truth for the final assistant answer than Codex's explicit --output-last-message output
• What changed: codex-cli runs now allocate a temp --output-last-message file, prefer that artifact when it contains text, and keep stdout parsing as the fallback path
• What did NOT change (scope boundary): other CLI backends still use their existing stdout/json/jsonl parsing paths, and this PR does not change failure classification or streaming behavior

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes Enhancement: finalize codex-cli replies from --output-last-message artifact instead of relying on stdout openclaw/openclaw#65074
• Related [Bug]: OpenClaw treats codex exec sandbox-denied runs as success because codex exits 0 openclaw/openclaw#39317
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: the direct codex-cli runner only treated stdout parsing as the final reply source even though Codex already exposes a more stable final-message artifact contract
• Missing detection / guardrail: there was no runner-level test that forced stdout and the final-message artifact to disagree and verified that OpenClaw finalized from the artifact
• Contributing context (if known): Codex CLI already supports --output-last-message, but OpenClaw was not wiring that contract into the codex-cli backend path

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: src/agents/cli-runner.spawn.test.ts, src/agents/cli-output.test.ts
• Scenario the test should lock in: when Codex writes both stdout JSONL and an --output-last-message file, OpenClaw should prefer the artifact when it is non-empty and fall back to parsed stdout when the artifact is empty
• Why this is the smallest reliable guardrail: it exercises the exact runner boundary where args are built, the artifact is read, stdout is still parsed, and the final reply is selected
• Existing test that already covers this (if any): none covered the artifact-vs-stdout precedence or Codex message-content JSONL fallback
• If no new test is added, why not:

User-visible / Behavior Changes

• codex-cli replies now finalize from Codex's explicit final-message artifact when available, which makes delegated Codex runs less dependent on stdout formatting details.

Diagram (if applicable)

Before:
[codex exec --json] -> [stdout parsed] -> [reply text selected from stdout]

After:
[codex exec --json --output-last-message <file>]
  -> [stdout parsed for session/usage/fallback]
  -> [artifact read]
  -> [artifact text wins when present]

Security Impact (required)

• New permissions/capabilities? (No)
• Secrets/tokens handling changed? (No)
• New/changed network calls? (No)
• Command/tool execution surface changed? (Yes)
• Data access scope changed? (No)
• If any Yes, explain risk + mitigation:
  • This only adds Codex's own --output-last-message flag plus a temp file under OpenClaw's preferred temp dir. The file is per-run, read back locally, and deleted in cleanup even when parsing falls back to stdout.

Repro + Verification

Environment

• OS: macOS
• Runtime/container: local pnpm/vitest
• Model/provider: codex-cli/gpt-5.4
• Integration/channel (if any): direct CLI runner
• Relevant config (redacted): default codex-cli backend with --json

Steps

1. Run a direct codex-cli execution through the CLI runner.
2. Have Codex emit stdout JSONL and write a final-message artifact.
3. Compare the final reply selection when the artifact is populated vs empty.

Expected

• OpenClaw should prefer the artifact text when it exists and keep stdout parsing as the fallback.

Actual

• Before this change, the runner always finalized from stdout because it never requested or read Codex's final-message artifact.

Evidence

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Human Verification (required)

• Verified scenarios: targeted vitest run for src/agents/cli-output.test.ts, src/agents/cli-runner.spawn.test.ts, and src/agents/cli-runner.reliability.test.ts
• Edge cases checked: non-empty artifact overriding stdout; empty artifact falling back to stdout; Codex JSONL message content still parsing to plain assistant text
• What you did not verify: full repo-wide pnpm check / pnpm test lanes and a live end-to-end Codex CLI run against the actual binary in this branch

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? (Yes)
• Config/env changes? (No)
• Migration needed? (No)
• If yes, exact upgrade steps:

Risks and Mitigations

• Risk: temp artifact cleanup could leak files if the runner exits early
  • Mitigation: cleanup runs in the outer finally, and the unit test asserts the artifact path is removed after success
• Risk: Codex stdout fallback could regress if the artifact is empty and stdout uses the normal message-content JSONL shape
  • Mitigation: parseCliJsonl now extracts text from Codex message content records, and the fallback path is covered by a dedicated unit test

AI-assisted: Yes (Codex). Testing: targeted vitest coverage listed above. I attempted codex review --base origin/main, but the local Codex CLI is currently rate-limited.