Review the session

Suggest changes that prevent recurring friction in future runs. Default to the current session; when the user names another, read its available logs, artifacts, commits, and verification results. Distinguish observed events from inference and missing evidence.

Load writing-for-agents, plain-english, and deslop-written through the Skill tool when available; otherwise read their SKILL.md files. Apply them to the report.

Find improvements

Trace each candidate to session evidence and inspect repository configuration before proposing a fix:

• Navigation: time spent finding files, hidden dependencies, and missing context pointers. Prefer a precise pointer to a duplicate document.
• Automated checks: a mistake that a deterministic check can catch, or an unwired or broken check. Read build scripts, CI, and hooks first. An absent hook warrants a proposal when the session shows a failure it would prevent or the repository requires one.
• Review guidance: judgment calls the review missed, such as cross-file consistency or domain meaning. Put mechanical rules in the repository's checker; reserve review prose for decisions a checker cannot make.
• Steering files: duplicated instructions, instructions that do not change behavior, or large always-loaded files that could point to focused references.
• Tool use: repeated expensive reads, avoidable output, unavailable capabilities, or scripts whose input or output causes recurring mistakes.
• Information access: missing logs or read access that blocked a decision. Name the specific information and the smallest access change that would expose it.

Keep implementation facts, review facts, and environment causes separate. A reviewer still needs the spec, standards, and relevant context; a diff alone does not establish intended behavior.