You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
feat(pr-management): add interactive review findings pane Claude Code mod (part of #1497) #1887
pr-management-code-review should display drafted code review findings in an interactive terminal review pane (ui.renderPane) when running under Claude Code, allowing maintainers to inspect findings and toggle individual inline comments ([Accept] / [Skip] / [Edit]) directly rather than stepping through sequential text prompts.
Why
During deep code review (Step 6 / Step 7a in review-flow.md), reviewing inline comment candidates currently requires text-prompt interactions ([A]ll, [N]one, 1,3, -2, [E 2]). As identified in #1497, moving this presentation to an interactive terminal pane (ui.render) provides a zero-token visual review surface, removes interactive prompt roundtrips, and accelerates review triage while maintaining strict vendor-neutral Markdown fallbacks across other harnesses.
Which layer
Other / unsure
Boundary conditions (optional)
Vendor Neutrality: When running on non-Claude harnesses (Codex, Antigravity, Cursor, CI) or when mods are disabled/unsupported, the skill must cleanly fall back to the existing Markdown prompt flow in review-flow.md without throwing errors or halting execution.
Empty Findings: If a review yields zero findings (clean approval), the pane hook must cleanly pass through (next(e)) without rendering empty UI frames.
Anchor Handling: Findings anchored to file:line should be clearly distinguished from unanchored general observations (which fold into the body's Smaller observations).
Graceful Fallback: If UI resolution fails or runs in a constrained/narrow terminal, it should fall back cleanly to standard text blocks.
Out of scope (optional)
Autonomous review submission: maintainers retain full control and must explicitly confirm the final review submission.
Modifying backend capability tools in tools/pr-management/.
Replacing agent reasoning: diff analysis, severity classification, and finding generation remain in the skill.
To keep review cycles swift and easy to review (avoiding the large diff we saw in #1676, that was absolutely nuts), I’m planning to break the implementation into 3 small, incremental PRs:
PR 1 (Data Model & Parser): Plugin hooks wiring (hooks.json), findings parser, and claude plugin test/validate unit tests.
PR 2 (UI Rendering):ui.renderPane component, severity badges, and terminal fallback handling.
PR 3 (Actions & Skill Wiring): Action state toggling and updating review-flow.md Step 7a with the mod enhancement.
Does this incremental breakdown sound good to you, @potiuk? I'll start working on PR-1
Activity
Kaap10 commentedon Oct 28, 2026
To keep review cycles swift and easy to review (avoiding the large diff we saw in #1676, that was absolutely nuts), I’m planning to break the implementation into 3 small, incremental PRs:
hooks.json), findings parser, andclaude plugin test/validateunit tests.ui.renderPanecomponent, severity badges, and terminal fallback handling.review-flow.mdStep 7a with the mod enhancement.Does this incremental breakdown sound good to you, @potiuk? I'll start working on PR-1
feat(pr-management): add review findings data model and parser (part of
feat(pr-management): add review findings data model and parser (part of
feat(pr-management): add review findings data model and parser (part of