# Review comments in pending question answers

Both runs: real web client in Chromium, 1440 × 1000 CSS pixels, light theme, separate checkout and fresh T3 database, T3CODE_DEV_SEED=0. OpenCode supplies the same typed-answer question. The selected model control and machine context control are hidden only for public capture privacy; answer handling and diff controls are unmodified.

Start the dev server in each checkout with T3CODE_DEV_SEED=0 T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0 vp run dev. Enable OpenCode in that checkout's dev settings. Wait for server readiness before running seed-public.py with EVIDENCE_MODEL set to an available question-capable model. Use the same selection on both sides.

The seed commits notes.md with line one, line two, line three. It then changes line 2 to line two revised and adds line four added as line 4. It adds the project and starts the question thread using the same prompt on both sides.

Browser interaction, performed using the shared preview browser tools:
1. Pair with the checkout's server and open Question review evidence. Close thread details.
2. Start recording when Which name should line two use? is pending, with alpha and beta options.
3. Click the question answer textbox and type See my note: followed by one space.
4. Toggle the right panel, open Diff, use Changes and stacked layout, expand notes.md.
5. Hover the added line 2 gutter and click its plus button. Type REVIEW-SENTINEL rename this to line two updated in Comment on lines +2, then click Comment.
6. Confirm the notes.md L2 chip is in the pending answer and click Submit.
7. Wait for the agent's final answer. It was instructed by the original seed prompt to report only the received comment text and diff, without reading files or inferring missing contents.
8. Stop recording. Capture the completed result beside the diff. For the final stills, two screenshots at least 500 ms apart compared with ImageMagick AE = 0 on both sides. The base still was recaptured after reopening its page; the recording contains the original uninterrupted interaction.

Observed base: the submitted answer has a bare t3-context review-comment link; the agent says The comment text and diff lines are missing.
Observed head: the submitted answer contains the review context with comment and +line two revised; the agent quotes both.

The built-in Replay question was not used because its multiple-choice ACP elicitation does not accept custom typed answers. This proof uses the real-provider path from triage instead of altering application code or the replay fixture.

AI-assisted; verified with the tests and tools listed above.