Repository navigation
Rebase fork on upstream #471 (RPC override) + cherry-pick shutdown hardening + #506/#480/#490 #3
Description
Context
The fork (1.0.6-fork.1) implemented --model/--effort override via spawn -c (PR #2, branch fix-app-server-overrides). Upstream PR openai/codex-plugin-cc#471 ("Support GPT-5.6 models and refresh stale brokers", +1302/-161) solves the same problem better and more broadly — via RPC thread/start.config + runtime-identity respawn (plugin/codex version, not just model/effort) + broker-lock (serialization) + broker-probe + model-catalog.
Decision: replace our spawn--c mechanism with openai#471, while preserving the fork's generic shutdown-hardening that openai#471 does not include.
Branch: rebase-on-471 (pushed). Target install version: 1.0.6-fork.2 (based on the last real upstream release v1.0.6, NOT on openai#471's claimed 1.0.7 — that PR may never merge).
What's in the branch (5 layers, all merged)
| Layer | Source | What |
|---|---|---|
| 1 | openai#471 (wholesale) | RPC thread/start.config override, broker-lock/probe/model-catalog, runtime-identity respawn. Replaces fork's spawn--c (ab75d49) + spawn-argv tests (eb3d78f, rewritten to RPC surface). |
| 2 | fork 8458741 |
Bounded sendBrokerShutdown (2s) — openai#471's rewrite is bool-returning but has NO timeout; re-applied the fork's settled/finish()/setTimeout-destroy on top. |
| 3 | fork 478be42 |
PID-reuse guard — existingReady gate in teardownExistingBroker (openai#471 unconditionally tree-kills by pid, risking killing a recycled pid on stale sessions). |
| 4 | openai#506 + openai#480 | resolved model/effort/sandbox in job records (fixes openai openai#481) + reasoning-start logging. Conflict in codex.mjs resolved by hand (preserve openai#471's validateReasoningSelection + openai#506's resolved). |
| 5 | openai#490 | Broker idle-timeout (15 min no clients → process.exit) + shutdown-path corrections. |
Dropped: fork 09b3a41 (graceful sendBrokerShutdown before teardown) — openai#471 already does graceful shutdown before teardown in loadReusableBrokerSessionUnlocked, so it's a duplicate.
Verification (codex-cli 0.144.4, macOS)
- Tests: 119 total, 114 pass, 5 fail — all 5 pre-existing (
status×2,result,resolveStateDir,task --background; reproduce identically on purepr-471). 0 regressions. Typecheck clean. - Empirical sweep (4/4 pass):
task --effort high→ rolloutturn_context reasoning_effort:"high"×3 (RPC override reaches the model — spawn--cis gone).- job record
resolved.reasoningEffort = "high"(Capture resolved model/effort/sandbox in job records openai/codex-plugin-cc#506). - SIGKILL worker mid-run → zombie
runningrecord retainsresolved(forensics, Job records never capture the resolved model/effort/sandbox a job ran with openai/codex-plugin-cc#481 scenario). - broker self-exits after idle timeout (Stop orphaned Codex companion brokers openai/codex-plugin-cc#490, tested with
CODEX_COMPANION_BROKER_IDLE_TIMEOUT_MS=5000).
Risks / notes
- Replacing the override mechanism drops the working spawn-
-cfallback. Rollback path:1.0.6-fork.1is preserved in git history (fix-app-server-overridesbranch + tag). - This is fork-internal — no upstream PR. Per [[upstream-effort-issue-map]], waiting for upstream reaction to Support GPT-5.6 models and refresh stale brokers openai/codex-plugin-cc#471.
- TODO before installing:
claude plugin uninstall codex@openai-codex && claude plugin install codex@openai-codex(per install-local-plugin-repoint —installalone skips if the version record exists).
Activity
axisrow commentedon Jul 21, 2026
Done — fork rebased on upstream openai#471, installed as 1.0.6-fork.2 (sha 06d5dc6). Shutdown hardening + openai#506/openai#480/openai#490 cherry-picked. See commit history for details.