[ScriptedAlchemy]

T3 repeatedly fetched full PR details for checks, duplicated reads across worktrees, and invalidated caches after every agent turn. Most providers could also bypass the Claude-only CLI gate, while reserve protection was reported as an exhausted GitHub quota.

This change:

• Uses a dedicated checks query, shares credential-scoped observations across worktrees, and coalesces turn-triggered refreshes.
• Stops mobile polling when the app or route is inactive.
• Installs a managed gh launcher for local providers, preserves Codex startup paths, and redirects supported reads to T3 tools while preserving writes and explicit bypasses.
• Separates REST and GraphQL primary cooldowns, retains shared secondary throttling, and distinguishes reserve protection from exhausted quota.
• Makes the one-PR limit explicit for feedback and failure tools, with eval cases for splitting requests. Canonicalizes the subprocess test directory on macOS. Stabilizes the pre-existing hung-stop persistence regression by exercising one stop timeout and awaiting the actual turn-start signal. Runs the real ACP archive-installation fixture with a live clock so external I/O can complete.

Validation: CARGO_NET_OFFLINE=true vp run check:changed passes: formatting, lint, both knip passes, generated GraphQL checks, 35 package typechecks, 1,213 tests in 27 files, and eight selected React compiler checks. Full local validation on pinned Node 26.12.1 passes: 6,516 server tests (475 files), 5,482 web tests, 3,240 mobile tests, 923 desktop tests, 1,596 shared-client tests, and all remaining package suites. Repository scripts and both native test suites also pass. Existing platform/fixture skips remain. The maintainer requested merging after local validation without waiting for CI. The hung-stop test failed on the base branch before the test-only correction. The subprocess regression verifies intercepted reads never execute the underlying CLI and that writes preserve literal arguments and exit status. The two additional feedback/failure eval cases generate valid requests; paid model evals remain unrun because the local vibeproxy configuration is unavailable.

The launcher covers normal local PATH-based calls. Absolute executables, replaced PATHs, explicit direct-read bypasses, and remote provider processes remain outside it. Existing provider processes must restart to receive it. Native Windows and live web/mobile verification remain unperformed; traffic reductions are demonstrated by regression fixtures rather than production measurements.

Updated and verified with GPT-6 Astra through the Codex harness in T3 Code.

[changeset-bot]

⚠️ No Changeset found

Latest commit: 8b9b826

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