[ScriptedAlchemy]

T3's Codex threads could miss native Computer Use even when it was configured. This change checks the project's effective Codex configuration before each turn on macOS and supplies the installed skill path. If discovery fails or the plugin is disabled, T3 tells Codex that Computer Use is unavailable for that turn. Remote clients control the server environment's Mac, with native permissions and confirmations.

Discovery also respects the effective node_repl MCP tool filters. An allowlist without js, including an empty allowlist, makes native Computer Use unavailable. A denylist containing js does the same. A denylist takes precedence over an allowlist.

The change includes resolver and adapter tests, synthetic eval cases, and user guidance. It does not install Computer Use or modify shared skill directories.

Verification

The original author completed real Calculator tasks through the production adapter, including after thread resume. Disabling the plugin prevented native calls on the next turn. This review did not repeat those macOS interactions.

The follow-up review used Node 26.12.1 and pnpm 12.14.0, with these results:

• 697 tests passed across the six provider, adapter, instruction, context-budget, and eval test files that the scoped checks selected.
• Six added MCP-filter cases cover accepted and rejected configurations. Four rejection cases failed before the fix.
• All 157 CodexAdapterV2 tests passed, including the previously reported capacity-retry failure and two background-stop failures.
• Formatting, lint, both Knip passes, GraphQL validation, and direct typechecks for server, scripts, and evals passed.
• The local Vite+ aggregate typecheck launcher could not initialize its seccomp supervisor, so the review ran the affected typechecks and selected tests directly.

Remaining validation

The review ran on Linux, and platform and filesystem fixtures covered the macOS discovery paths. The review ran no new live macOS interaction and no paid synthetic model eval. Native autonomous continuations and Codex-spawned children remain untested. The review did not rebuild or restart the installed T3 app.

GPT-6.1 Sol through Codex did the original implementation and verification. GPT-6 Astra through Codex prepared the original PR. GPT-6 Astra Pro in ChatGPT did the follow-up review and fixes.

[changeset-bot]

⚠️ No Changeset found

Latest commit: a820c4f

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR