FEA-372: Add post-loop code review and fix skill by wongk · Pull Request #47 · closedloop-ai/claude-plugins

wongk · 2026-04-09T21:59:43Z

Summary

Adds comprehensive code review as the final step in the run-loop pipeline, replacing the lightweight in-loop code:code-reviewer with the full code-review:start pipeline after implementation completes
Creates a new code-review:fix skill that verifies findings, fixes confirmed issues serially, runs build validation, and supports configurable re-review cycles
Writes CLOSEDLOOP_START_SHA to config.env so the review can scope its diff to only this run's changes

Implementation plan: PLN-233

Changes

`plugins/code/prompts/prompt.md`

Remove Phase 5 Step 2 (lightweight single-agent code reviewer)
Renumber Step 3 → Step 2
Rename phase from "Testing and Code Review" to "Testing and Validation"

`plugins/code/scripts/run-loop.sh`

Add run_post_loop_review() function (~130 lines) invoked after both completion-promise and max-iterations exit paths
Detects changes via working tree + staged + committed diffs against START_SHA
Runs code-review:start --base $START_SHA, reads verdict.json from the session directory
If verdict is not approve, invokes code-review:fix with --max-cycles 2
Write CLOSEDLOOP_START_SHA to config.env in create_state_file()

`plugins/code-review/skills/fix/SKILL.md` (new)

6-step pipeline: parse findings → verify with sonnet agents (parallel) → fix serially → build-validator → conditional re-review → summary
Only acts on BLOCKING/HIGH severity findings; MEDIUM/LOW are logged
Auto-discovers most recent cr-* session directory when not provided
Capped at configurable --max-cycles (default 2) to prevent infinite loops

`plugins/code-review/commands/fix.md` (new)

Thin command wrapper for manual /code-review:fix invocation

Version bumps

code: 1.6.0 → 1.7.0
code-review: 1.4.0 → 1.5.0

Test plan

Run run-loop.sh on a project with an intentional bug — verify post-loop review catches it and fix skill addresses it
Run on clean code — verify review returns approve and no fix cycle triggers
Run /code-review:fix manually on a prior review session directory
Verify CLOSEDLOOP_START_SHA is written to config.env
Verify Ctrl+C during post-loop review is handled gracefully
Verify Phase 5 no longer launches code:code-reviewer

🤖 Generated with Claude Code

Add comprehensive code review as the final step in the run-loop pipeline, replacing the lightweight in-loop code reviewer with the full code-review:start pipeline after implementation completes. - Remove Phase 5 Step 2 (single-agent code reviewer) from prompt.md, rename phase to "Testing and Validation" - Add run_post_loop_review() to run-loop.sh that invokes code-review:start after loop completion, reads verdict.json, and triggers code-review:fix if issues are found - Write CLOSEDLOOP_START_SHA to config.env for diff scoping - Create code-review:fix skill (verify findings with sonnet agents, fix serially, run build-validator, re-review up to max-cycles) - Add thin command wrapper at commands/fix.md for manual invocation - Bump code plugin to 1.7.0, code-review plugin to 1.5.0 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

shafty023

finished review

…eview