Record Codex native review trial and credential-store research - #916

Scope

• Record a read-only Codex 0.162.0 native-review trial against codex-plugin-cc 1.0.6 on real PR882. Both direct paths returned no actionable regressions; the verdict is retain-bridge because automatic dedicated-review routing and the Claude integration surface remain unqualified. Add source-only OS credential-store research and unapplied owner/command-center proposal fragments.
• Base commit: b7dfe638fc825a52ff4e7d1a9ccf2fde7de889fd.
• Lane: lane:foundation.
• Owned paths touched: two documentation records, evidence/artifacts/codex-review-trial-20261009/{receipt.json,owner-proposals.md}, and their four manifests/evidence.json hash rows. The shared manifest is the last commit.

SOTA sources

• openai/codex rust-v0.162.0, commit c1382380de69521303b416720a52f42d51af6248: codex-rs/exec/src/cli.rs, codex-rs/core/src/tasks/review.rs:105–141, codex-rs/config/src/types.rs:133–183, codex-rs/login/src/auth/storage.rs:308–346,478–527,596–626, codex-rs/core/src/config/auth_keyring.rs, and codex-rs/keyring-store/Cargo.toml. Release run 37812662295 passed; the official package and extracted binary checksums are retained in the receipt.
• openai/codex-plugin-cc v1.0.6, commit db52e28f4d9ded852ab3942cea316258ae4ef346: plugins/codex/commands/review.md, plugins/codex/scripts/lib/codex.mjs:1002–1054, plugins/codex/scripts/lib/app-server.mjs:189–195, package.json, .github/workflows/pull-request-ci.yml, and upstream Node tests. Its native review command already delegates to Codex review/start.
• Official OpenAI CLI reference and authentication / credential storage, fetched October 9; keyring 3.6.3 persistent Linux backend, the dependency locked by the pinned Codex source.
• Corpus: PR882, base a484989a43f70ff2575b6fb054cf00c9834e7679, head f4965224b569325a8ab41c2a97b8d71157d630c6, five-file extraction-host refusal delta. Its source PR and worktree were not changed.

Evidence-class table

Local commands run

nice -n 10 ionice -c2 -n7 tar -xzf <retained official full-package archive> -C <lane-owned runtime/package>
<extracted codex> --version
  rc 0, codex-cli 0.162.0; publisher archive and binary SHA256 verified
npm run check-version
  rc 0, all version metadata matches 1.0.6
env -u CLAUDE_PLUGIN_DATA nice -n 10 ionice -c2 -n7 timeout 600 npm test
  rc 0, 91 tests pass, zero failures/skips
claude plugin validate <pinned source>/plugins/codex
  rc 0
timeout 600 python3 -B -m unittest tests.test_research_dispatch_exact_queries
  rc 0, 32 cases / 30 pass / 2 skipped, on the pinned corpus head
<extracted codex> exec -s read-only --ephemeral --json -o <native-last.md> review --base a484989a43f70ff2575b6fb054cf00c9834e7679
  rc 0, no actionable regressions
node <pinned plugin>/scripts/codex-companion.mjs review --wait --scope branch --base a484989a43f70ff2575b6fb054cf00c9834e7679 --model cx/gpt-6.1-sol --json
  outer rc 0 / nested Codex status 0, no actionable regression
<extracted codex> exec -s read-only --ephemeral --json -o <fresh-last.md> <unnamed review intent>
  rc 0, general review succeeded; dedicated-review automatic routing unproven
rm -r -- <exact lane-owned trial runtime>
kill -TERM -- <this trial's dedicated broker PID>
  inverse verified: runtime, broker process and socket absent; prior install preserved
nice -n 10 ionice -c2 -n7 timeout 600 python3 -B -m unittest tests.test_validate
  rc 0, 68 tests pass
nice -n 10 ionice -c2 -n7 timeout 600 python3 -B -m unittest tests.test_evidence_manifest
  rc 0, 5 tests pass
nice -n 10 ionice -c2 -n7 timeout 600 python3 scripts/validate.py
  rc 0: 70 components / 10925 hashed files / 4 profiles / 239 receipts
git diff --check
  rc 0

Decision record

docs/decisions/2026-10-09-codex-native-review-trial.md explains the identical reviewer backend, scope difference, retained bridge verdict and overturn condition. docs/codex-credential-store-research-20261009.md distinguishes setting semantics, migration, cleanup, Linux/WSL prerequisites and source-only boundaries. Owner proposal fragments remain unapplied; no client-configuration adoption follows.

Private native receipts and measurement script: $HOME/.local/state/native-agent-stack/research/api-surfaces-trials/codex-review/20261009. The registered public receipt supplies the retained output hashes. Current sweep FINAL.md SHA256 is 77e17817d8ecc8e53329ac32303ece8dc4e86eaf67cba1e3ae36196aa328a109, superseding its handoff digest.

Host evidence

No files under evidence/hosts/ or platform-status claims change. No host-receipt or platform-status promotion is asserted.

Checklist

• Existing workflows remain unchanged; no new Action or permission grant.
• No secrets are printed, logged or committed; no new required secret.
• No new hosting, subscription or billing surface.
• Peer-owned files/worktrees and pinned PR901 work are preserved.
• Vendor distribution, one version check and installation inverse recorded; isolated runtime removed.
• No user-level client configuration, plugin registration or credential-store change.
• Both designated reads at the published full head.
• Required CI at that head.
• Designated pre-cue tool at that head.
• Explicit command-center cue before landing through 5f.

Keep this PR draft. Publication requests the co-op to schedule the two designated reads and the pre-cue tool; no push after a read snapshot/cue or LANDING-READY. Rebase the shared evidence manifest once at landing and refresh the affected head gates.