[Edo771977]

This imports upstream openai#840 as a merge commit, keeping upstream authorship. It also adapts one piece of upstream openai#837, which fixes fork issue #19.

openai/codex-plugin-cc#840: branch review pinned to commits

buildBranchComparison now resolves HEAD and the base to commits once (rev-parse --verify --end-of-options <ref>^{commit}) and builds both ranges from those SHAs. The adversarial-review guidance tells Codex to read surrounding files with git show <commit>:<path>. This fixes two problems that the fork also had:

• switching branches while a review runs could change what Codex sees;
• a base given as an annotated tag was not peeled to its commit.

--end-of-options also keeps a base starting with - from being parsed as a git option. It needs git 2.24 or newer.

• Conflict in lib/git.mjs: I took upstream's options.currentBranch and kept the fork's NUL-separated file list (listFilesZ, from fix: keep the status column and non-ASCII file names in review context openai/codex-plugin-cc#817).
• Conflict in tests/git.test.mjs: both sides' imports and tests are kept.
• Adapted tests: upstream's two snapshot tests switch branches when they see git diff --name-only as the first git arguments. This fork runs git as git -c core.quotePath=false …, so that trigger never fired. The trigger now skips that prefix; nothing else in the tests changed.
• Check: restoring the fork's previous lib/git.mjs makes three of the imported tests fail.

/codex:review --effort reaches Codex (adapted from openai/codex-plugin-cc#837)

Before this change, --effort was validated and then dropped (fork issue #19). A native review starts a thread and calls review/start. It never sends turn/start, which is the only request that carries a per-turn effort.

The effort now goes on thread/start as the config override model_reasoning_effort:

• It is merged into any config the access parameters already carry. The --read-root permission profile uses that same field, which is how we know thread/start accepts config.
• It is sent only when --effort is given, so otherwise the user's config.toml still decides.

Nothing else from openai#837 is taken. Its other changes are already in this fork in its own form: --sandbox, --model/--effort on reviews, atomic writes, gating detached workers, and interrupting a disconnected client's turns. Its max effort stays out, as with openai#761.

• Test: a new test checks that the override reaches thread/start and that nothing is sent without --effort. Removing the effort from the companion call makes it fail.
• Not verified against a real codex, which is not available here.

README now describes the effort as set on the review's thread, instead of saying it is dropped. The CHANGELOG is updated too.

Verification

• node --test tests/*.test.mjs: 451/451 on two full runs. No broker was left running afterwards.
• npx tsc -p tsconfig.app-server.json: clean.

Not imported

• #841 (do not launch Codex through SHELL on Windows): the fork never consults SHELL there, since fix: avoid shell:true deprecation on Windows openai/codex-plugin-cc#725.
• #838 (resolve the transferred thread from the completion results): already covered by fix(transfer): report already-imported sessions and add --refresh openai/codex-plugin-cc#835.
• #836 (idle shutdown of the broker, and a readiness probe on reuse): the fork already has both. Its idle timeout defaults to 10 minutes and is configurable.

---

Generated by Claude Code