← ClaudeAtlas

swe-review-looplisted

Run the Generate → Review → (Revise | Regenerate) → Verify loop on a software issue. Supports `review_guided`, `best_of_n`, and `hybrid` strategies. Tracks iterations, accumulates token usage, supports early stopping.
KevinZhangNothing/swe-review-plugin · ★ 1 · Code & Development · score 64
Install: claude install-skill KevinZhangNothing/swe-review-plugin
# swe-review-loop > Wraps `swe_review.LoopSkill` and `LoopSubAgent`. ## When to use Use when the user wants end-to-end closed-loop processing: a generator produces a candidate, the reviewer evaluates it, the reviser fixes defects, optionally a verifier runs tests. Iterate until approve, or budget exhaustion. ## Strategies - `review_guided` (default): single candidate + iterative revision via review defects. - `best_of_n`: generate N candidates, review each, pick the first `approve` or highest-confidence. - `hybrid`: `best_of_n(n=3)` first; if none approved, fall back to `review_guided`. ## Hard constraints 1. **NEVER** inject `golden_patch`, `gold_patch`, or `oracle` through the loop to a reviewer/reviser/generator. If you have a gold patch, gate it through `VerifySkill.execute(..., oracle=...)` after the loop. 2. The loop returns the chosen `final_pr_diff` and a structured `LoopResult` you can serialize to JSON. 3. Default `max_iterations=5`; respect early_stop unless explicitly disabled. ## How to invoke ```python import asyncio from swe_review import ( LoopSkill, ReviewSkill, ReviseSkill, GenerateSkill, VerifySkill, ClaudeCodeAdapter, ) tool = ClaudeCodeAdapter() review = ReviewSkill(tool_adapter=tool) revise = ReviseSkill(tool_adapter=tool) generate = GenerateSkill(tool_adapter=tool) verify = VerifySkill(repo_path="/path/to/repo") loop = LoopSkill( review_skill=review, revise_skill=revise, generator_skill=generate, verify_skill=verify,