fix(provider-acp): show the plan when an agent exits plan mode#25554

Approvability

Verdict: Approved at c791daf

Macroscope's review found this PR approvable — This focused ACP bug fix extracts plans from exit-plan tool calls and routes them through the existing proposed-plan rendering path. It adds only small, tested logic and does not change defaults, schemas, permissions, or deployment behavior.

^{You can add or adjust custom eligibility rules. Learn more.}

Review in Change Stack →

No actionable comments were generated in the recent review.

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

CodeRabbit Verify

• Run CodeRabbit Verify to test the changes in this pull request.

External image Recent review info

External image Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 1d626d0
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: b6cb2ad9-186e-48c2-a05b-3e34c03ff2ba

External image Commits

Reviewing files that changed from the base of the PR and between 6c1f2b5e52270f7cb6b173aabbfba7aa09efcd4d and 012b2b9996930b4af46356df12387b5a0e50c196.

External image Files selected for processing: 2

• apps/server/src/orchestration-v2/Adapters/AcpAdapterV2.test.ts
• packages/provider-acp/src/server/adapter.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

thanks @kareemsetl. the original correctly routes hidden exit-plan markdown into the existing proposed-plan card. i preserved your commit, merged current main, and pushed ede3807ca6152bec9e5caa2e3fc5db39dddf2def on top.

• kept capture after tool updates merge, so a title and plan arriving separately still work.
• tightened matching to the complete exit_plan_mode or exit plan mode title, fixing the existing review finding and rejecting prefix/suffix lookalikes.
• removed the separate helper and its new test file. the production fix is seven lines in the adapter; one regression in the existing adapter test file covers emitted plans, split updates, invalid inputs, both names, dedupe, and stable plan identity when the markdown changes. permissions stay unchanged.

verification: the adapter regression failed against current main with zero proposed plans, then passed with the fix. blacksmith vp run check:changed passes formatting, lint, both knip passes, graphql validation, typechecks for eight affected packages, and all 66 adapter tests. two independent code reviews approve this exact head. all seven reported head checks are green and the pr has no conflicts. coderabbit's latest automated review was rate-limited; its earlier finding is addressed.

in the running web app, a controlled acp wire peer reproduces main's missing plan and verifies the fixed plan card, repeated notifications, and the approve response through the real server/client path. matching screenshots show the same prompt and plan payload. the clip shows approval reaching the peer. uploaded images and video playback were checked.

unverified: authenticated cortex-code inference, other agents' title formats, electron desktop and native mobile. the fork update reported lightweight ci checks; the full ci test/build suites were not reported on this head. scoped checks ran in blacksmith as described above.

before: main asks for approval without a readable plan

after: the same acp payload renders one readable plan card

after: approve reaches the acp peer while the plan stays readable

2026-11-03 merge fixup at c39960705d28cb8e31cff421a7d59372808c65ad: merged current main with a merge commit. resolved the test-file conflict by retaining both the exit-plan regression and main's protocol-logging regression. the pr remains two files with 158 additions. no production changes beyond those already in the pr.

blacksmith vp run check:changed github/main passed formatting, lint, both knip passes, graphql validation, typechecks for 14 packages, and the focused two-file test run. the worker needed a temporary comparison parent because its synced source snapshot is shallow; the checked tree matches this head. two independent read-only reviews found no issues. resolved the already-addressed raw-title review thread after confirming the retained title lookup and execute-kind regression.

real inference through the affected acp provider/client path remains unverified. this pass changes only merge integration and does not replace the runtime limitations recorded below.

historical verification report follows:

Verified in a real dev environment

verification result: the specific ACP fix remains unverified. the real Codex checks below cover neighboring behavior only.

checked out github/main at 3ef16100796c77567ab8bdfeb05d77565ae02d53 in a separate worktree, installed dependencies, and ran its dev server. stopped that server before starting the exact PR head 1e07c9c49ccab3e1ee7623d654032dba6871d2ab, also after vp i, with isolated empty state. drove the actual web app with the repository's Playwright and /usr/bin/google-chrome. projects and threads were created through the app; turns used real Codex gpt-6.1-sol through the local proxy. no replay, mocks, fixtures, or controlled ACP peer was used for these runtime checks.

• main and the PR both displayed a readable Codex proposed-plan card for the same documentation-planning prompt.
• on the PR, the plan survived reload. switching from plan to build allowed a normal follow-up that read the real manifest and returned normal-turn-ok t3tools/monorepo.
• a second real follow-up returned capture-ok; the recording shows the build/plan controls, pending turn, and completed response.
• all 67 existing ACP adapter tests passed on this exact PR head in Blacksmith, including the exit-plan regression. this automated suite uses its existing test harness and is separate from the real-provider checks above.

short command transcript:

these screenshots document neighboring Codex behavior, not a before/after reproduction of the ACP bug. the prompt and viewport match; real model output and package metadata differ. the uploaded images were inspected, and the uploaded video played to its end and supported seeking. reopened both PR renderers: GitCafe displays external media as links; GitHub displays the screenshots but does not play this external MP4 inline. video playback was checked in the uploaded player directly.

main: native Codex already renders its proposed plan

PR: native Codex proposed plan remains readable

PR: mode controls and a real Codex follow-up complete

Not verified

could not reproduce the original missing-plan bug on main or exercise the fixed exit_plan_mode / rawInput.plan branch with a genuine ACP agent. native Codex uses CodexAdapterV2, while this PR changes makeAcpAdapterV2. this VM has working Codex proxy credentials but no configured Cortex/Snowflake or Grok credentials or executable. Claude is reported not authenticated. a native Codex plan therefore cannot establish that this ACP fix works. the earlier controlled-peer evidence is not being counted as real-provider verification.

ACP plan approval/denial, sparse tool updates, and duplicate-plan behavior remain unverified with real inference. Electron, native iOS/Android, macOS/Windows, and remote/relay/tunnel clients were not exercised. a 19-second CPU stall occurred during the PR's initial dev compilation; the app recovered and all recorded Codex turns completed well below ten minutes. no performance comparison is claimed.

GitHub reports its listed contexts passing or skipped, but full CI build/test workflows were not verified here. no source changes or pushes were made. this pass does not certify the PR's changed behavior for merge.