ce-code-review

A structured code-review skill that evaluates code changes using dynamically selected reviewer personas. It spawns parallel sub-agents that return structured JSON, then merges and deduplicates their findings into a single report. It is meant to run before creating a pull request, after finishing a task during iterative implementation, or whenever feedback on a change is needed, and it can be invoked standalone or embedded inside a larger workflow.

Two operating modes share the same review pipeline. In default interactive mode it produces a Markdown report with pipe-delimited finding tables and an actionable-findings summary, and it applies safe, verified fixes and commits them when the pre-review working tree was clean. In mode:agent (with mode:headless as a deprecated alias) it is report-only: it returns one JSON object as a deterministic contract for programmatic and cross-harness callers and never mutates the working tree, leaving the caller to apply. A core operating principle is 'apply locally; never push' — it never pushes, opens PRs, or files tickets in any mode. It also avoids blocking prompts and never switches branches; passing a PR number, URL, or branch name selects review scope only, not permission to mutate the tree. Arguments are parsed for tokens such as base:<sha-or-ref>, plan:<path>, and grouping:auto/off/always, with explicit conflict rules that stop the run before dispatching reviewers.

Findings use a P0-P3 severity scale answering urgency, while autofix_class (gated_auto, manual, advisory) and owner describe the shape of follow-up rather than granting apply permission. Synthesis owns the final route, chooses the more conservative route on disagreement, and rejects unsafe classes such as safe_auto. A quick-review short-circuit can defer to the harness's built-in review for fast, light reviews, but mode:agent always runs the full multi-agent pipeline and returns JSON.