[docwilde]

Codex now retrieves opted-in graph context at startup and on every prompt, including the first. Graph failures preserve the curated startup snapshot, and prompt retrieval does not schedule review or indexing.

For agents

• codex/hooks/hooks.json bundles native SessionStart and UserPromptSubmit handlers, including startup on fork.
• The native Codex prompt adapter returns graph context before review, live indexing, or memory-refresh scheduling. Startup uses the declared hook engine for freshness instructions and treats graph errors as optional context failures.
• Regression fixtures cover first-prompt retrieval without a session stamp, graph budgets, citation labels, opt-in and disable switches, unavailable graph context, unchanged curated memory and beliefs, and absent review/refresh schedules.
• Negative controls reproduce the missing prompt manifest, startup graph omission, first-prompt refusal, review scheduling, and incorrect engine instructions. Mutating graph error handling and the injection guard makes their regression tests fail.
• Validation: full native workspace suite; all 41 Python test files; 11 Python/native interoperability checks; native launcher smoke test; shell syntax and release-version consistency checks.
• Runs using shared /tmp encountered failures in existing pending-scope and dry-run fixtures. The final native gate uses a private temporary directory, matching CI.
• Version: 0.62.29 across the plugin, marketplace, native crate, and lockfile. README, native-core documentation, and changelog describe the bundled hooks and update requirements.