PR #8143 demo readiness report

CopilotKit generative UI review agent · tested 26 August 2026 · head 134845e

Ready for the demo. The normal streamed review path, file opening, mobile approval, exact approval gate, and coding-agent handoff all passed. Keep this PR draft and do not merge it; the implementation intentionally favors a reliable local demo over production hardening.

What the demo shows

This is a custom review agent inside T3 Code, focused on CopilotKit generative UI rather than a chat-platform bot. It reads the active Git diff, renders a structured dashboard, opens a finding in the real file viewer, asks for explicit approval, and hands only the approved findings to the normal T3 coding agent.

What was fixed

Stream-safe dashboard

Partial tool arguments now show progress instead of crashing React while CopilotKit streams the dashboard payload.

Real local projects

The review service now accepts the active project root even when it is outside the server process directory, matching normal T3 Code usage.

Readable file opening

Opening a finding closes the review sidebar so the referenced source is visible instead of being covered.

Usable mobile approval

The approval card has enough trailing space to stay clear of the fixed composer at 390×844, with no horizontal overflow.

Visual evidence

CopilotKit review dashboard rendered in T3 Code — The normal streamed path renders the real 2-file diff, checks, two findings, and an approval card without a console error or error boundary.

Finding opened in the T3 Code file viewer — Clicking Open closes the sidebar and leaves `src/cart.js` readable in the file viewer, including the reviewed `eval` line.

Mobile approval card clear of the composer — At 390×844, the Approve button ends at y=476.5 and the composer starts at y=668: 191.5px of clear space, with 389px client and scroll widths.

Approved findings handed to the normal T3 coding agent — Approval sends the exact two findings and verification command into the active T3 thread and starts the normal coding agent.

End-to-end result

The seeded repository was deliberately outside the server runtime directory. Its branch changed 2 files by +3/−5 and contained two real defects: executable coupon input through eval and missing quantity validation. The review agent found both.

After approval, the normal T3 coding agent removed eval, validated coupon percentages, added positive-integer quantity validation, expanded the fixture tests, and left the disposable repository at 5/5 passing tests.

Checks run on the final patch

Check	Result	Evidence
Web and server typecheck	Passed	Both scoped workspace tasks completed under Node 24.18.1.
Focused web tests	11/11	Copilot review parsing and handoff logic.
Focused server tests	2/2	Arbitrary local project roots reach VCS detection for preview and file expansion.
Targeted formatting and diff check	Passed	All three changed files format cleanly; `git diff --check` is clean.
Browser assertions	17/17	Normal streamed dashboard, real diff, file open, mobile layout, approval, handoff, no console errors, and no failed requests.
Fixture after handoff	5/5	The ordinary coding agent made the approved edits and the expanded suite passed.

Demo runbook

Open a Git-backed T3 Code thread on a branch with a small, intentional diff.
Open the purple review-agent orb and ask: Review this branch and show me the review dashboard.
Show the checks and findings, then click Open on one finding.
Reopen the review agent, click Approve fixes, and show the exact findings appear in the normal T3 thread.
Let the coding agent make the edits and run the requested verification.

Deliberate draft-only caveats: GitHub still reports a merge conflict and production policy review issues. Those do not block this demo and were intentionally left alone because this branch must never merge. The deterministic browser run used a local OpenAI-compatible responder with dummy credentials for repeatable review output; the approval handoff used the normal installed coding provider.