[abruption]

Summary

Investigates #268 against fixed native Codex sources and adds bounded pre-effect storage guards. The investigation in #268 remains open: effective storage of a running shared server and native managed/project configuration remains an evidence gate.

• Refuse nonempty inherited CODEX_SQLITE_HOME and root sqlite_home configuration before discovery, dry-run, queue and wake preflight. Refusal is structured and does not reflect configuration contents or environment values.
• Bound selected-home configuration reads to a 1 MiB regular, non-symlink UTF-8 file with descriptor/path identity checks. Recognize quoted/escaped root keys, dotted keys and tables while distinguishing nested keys, comments and string bodies. The dependency-free classifier is not a complete TOML/schema parser.
• Recheck before queue/app-server child creation, strip empty SQLite overrides, normalize case-insensitive Windows environment keys, and preserve the existing private native-home translation.
• Add 13 isolated storage tests and an actual fake-child environment recorder. Update the CLI reference in all four locales and regenerate the standalone artifact.

Type

• fix
• docs
• chore / refactor / test / perf

Fixed source findings

• Official configuration reference: https://developers.openai.com/codex/config-reference/
• Native 0.154.0 peeled commit 6b9826e3aa83b1a5947db50f4332cb9c65f1b340 and 0.159.0 peeled commit 687a119f0fcaace47e1f1abcc77cec6c813fd6da: codex-rs/core/src/config/mod.rs resolves configured sqlite_home, then trimmed nonempty CODEX_SQLITE_HOME resolved against native working directory, then the Codex home.
• https://github.com/openai/codex/blob/687a119f0fcaace47e1f1abcc77cec6c813fd6da/codex-rs/core/src/config/requirements.rs applies managed storage requirements before final construction. Command-line -c alone is not a universal selected-storage fence; requirements can replace it with warnings.
• https://github.com/openai/codex/blob/687a119f0fcaace47e1f1abcc77cec6c813fd6da/codex-rs/tui/src/session_queue_commands.rs can reuse an existing shared app-server. Client environment/config does not attest that server's storage. A config override can exclude daemon reuse and the queue rejects an embedded server while a daemon is running.
• https://github.com/openai/codex/blob/687a119f0fcaace47e1f1abcc77cec6c813fd6da/codex-rs/state/src/sqlite.rs derives both the state and queue database paths from resolved SQLite storage.

These are source findings, not a fresh native routing experiment or proof of an actual wrong-target submission/security exploit. No live accounts, transcripts, native message submissions or operational services were inspected or modified.

Checklist

• Real local CLI listing of live sessions (not performed; isolated fixtures only)
• Real SSH listing (not performed)
• Dry-run refusal occurs without queue or app-server child execution in isolated fixtures
• Commits follow Conventional Commits
• CLI reference and translations updated
• No secrets, tokens, session keys, private paths or account identities in the diff

Verified against

Local macOS 27.0 arm64, Python 3.14.8. Native Codex versions were reviewed from immutable upstream source only; they were not executed. The live wake support matrix remains unchanged.

Local checks:

• python3 -X dev -W error::ResourceWarning -m unittest tests.codex.test_storage -q: 13 PASS.
• python3 -m unittest discover -s tests/codex -t . -q: 120 tests PASS, 1 Windows-only skip.
• python3 -m unittest discover -s tests/core -t . -q: 345 tests PASS, 2 skips.
• python3 -m unittest discover -s tests/docs -t . -q: 32 PASS.
• Generated artifact, compilation and git diff --check: PASS.
• Initial system-Python full discovery ran 615 tests with one missing-test-dependency error (packaging) and 29 skips. This environment failure is not hidden; a private test venv with the repository-pinned packaging==26.3 is used for the final rerun.
• Final pinned-test-venv python -X dev -W error::ResourceWarning -m unittest discover -s tests -t . -q: 615 tests PASS, 29 explicit skips. This is default discovery, not a claim that all optional Relay/Control integrations ran.
• The initial Windows CI run exposed a fixture-only case-simulation error: real Windows os.environ normalized the lowercase fixture key even while the test simulated POSIX. The follow-up uses a plain mocked mapping for cross-platform simulation; runtime behavior is unchanged. Current-head CI remains a required gate.

Notes and remaining gates

This guard checks inherited overrides and each inspected home's configuration, not the complete native configuration stack. Unselected nested configuration is not treated as a root storage setting, but a profile/project/system/managed layer or an already-running daemon can still determine effective storage. Configuration can also change after the last check. Passing this guard is not storage attestation, and fake-child fixtures cannot prove real routing or consumption.

Before #268 can be closed, establish an authoritative body-free effective-storage contract for supported native versions and daemon/layer cases, then verify it on the relevant platforms. This PR adds no relocated-storage support, does not relax writer ownership/inactive opt-in, and does not expand the wake version matrix. Native full TOML/schema validation remains native Codex's responsibility; lexically unclassifiable input fails closed here.

No merge, tag, package publication, fleet installation or operational changes are part of this work.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-09T00:09:34.155276Z	0a388fc	Draft marked ready
🔒 Security Review	✅ Completed 2026-10-09T00:10:41.772641Z	0a388fc	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.