KevinZhangNothing

swe-review-loop

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.

swe-review-review

Agentic code review of a candidate PR — explore the repository, trace the call chain, and emit a structured JSON review report. Default mode is senior ENGINEERING review (8-dimension 100-point scoring + P0–P4 evidence-bound findings + 4-level decision); legacy bug-fix-centric modes available. Use when given an issue + a candidate PR (diff).

swe-review-analyze

Static analysis of a candidate PR diff — counts additions/deletions/hunks, surfaces public-API changes, flags TODO/FIXME and new exception paths, computes a complexity score. Useful as a cheap signal before/alongside an LLM review.

swe-review-explore

Collect repository context for a code-review task. Returns modified files, related files, file contents (truncated), call chain hints, keyword extraction, and a `root_hint` list to focus attention. Use when starting a review or revision that needs repository-grounded evidence.

swe-review-generate

Generate a candidate PR (unified diff) for a given software issue, with optional exploration context. Used by `best_of_n` and `hybrid` strategies in the loop. The output must be a `git apply`-compatible unified diff.

swe-review-revise

Code revision guided by structured review feedback — given a candidate PR and a list of defects (severity/description/location/suggestion), produce a new git-apply-compatible unified diff that addresses the defects with minimum changes.

swe-review-verify

Verify a candidate PR by applying the patch (in a sandbox tempdir if enabled) and optionally running `fail_to_pass` and `pass_to_pass` tests. Returns a structured pass/fail result. Sandbox by default to avoid polluting the user's working tree.