End Of Turn Review

End-of-turn review runtime

This client-agnostic hook utility is opt-in, not a shared agent skill. Wire it from the target client after reviewing the gateway, cost, and stop-hook behavior.

A second-opinion pass over work just produced. Designed to catch: bugs the implementer missed, edge cases the planner glossed over, regressions in unmodified behavior, and plans that look reasonable but have a load-bearing flaw.

The reviewer never edits. It produces a verdict and structured feedback. The primary decides what to do with it.

When to use

• End of any non-trivial coding turn (multi-file edit, new feature, refactor, bugfix beyond a typo)
• After a planner subagent produces a plan, before the worker starts executing it
• On explicit "review", "critique", "second opinion", "is this safe", "what would break"
• Pre-commit, pre-push, pre-PR-create

When NOT to use

• Trivial edits (1–3 lines, single-file typo fixes, comment-only changes)
• Read-only turns (research, log digests, no state change)
• Documentation-only edits (README, comments, doc strings)
• When the user explicitly says "skip review" / "just commit"

Tools

scripts/review.sh

Runs the review. Accepts the work-to-review on stdin (a diff, a plan, or freeform text) and prints a structured critique to stdout.

Usage:

# Review a code diff
git diff | scripts/review.sh --kind=diff

# Review a plan markdown
cat plan.md | scripts/review.sh --kind=plan

# Mixed: pass a turn transcript
scripts/review.sh --kind=turn < /tmp/turn-context.txt

Env vars:

• REVIEW_MODEL — model alias to call (default gpt-5.5). Must be configured in LiteLLM at LITELLM_BASE_URL.
• LITELLM_BASE_URL — defaults to https://litellm.drkmttr.dev/v1
• LITELLM_API_KEY — required; read from ~/.secrets/litellm-api-key if unset
• REVIEW_CONTEXT — optional path to a file with extra context (the user prompt, prior plan, etc.) the reviewer should consider

Output format — first line is one of:

LGTM
LGTM with notes
BLOCK — <one-line reason>

followed by a numbered list of specific issues (each with file:line if applicable) and an optional "what I'd change" section.

scripts/diff-review-hook.sh

Client-agnostic wrapper around review.sh. Designed to be called from any agent runtime's end-of-turn / session-stop event, or from a plain git pre-commit hook. It checks for a non-trivial uncommitted diff and pipes it through review.sh, printing the critique on stdout.

The hook makes no assumptions about the caller beyond (a) cwd is inside a git repo and (b) stdout is captured and surfaced somewhere — to the next agent turn, to the developer's terminal, or wherever the calling system routes it.

Not auto-installed. See reference/hook-setup.md for per-client wiring (Claude Code, opencode, Cursor, Aider, plain git).

Reference

• reference/system-prompt.md — the reviewer's full system prompt (loaded by review.sh)
• reference/hook-setup.md — how to wire the skill into Claude Code Stop hooks and opencode session-end hooks
• reference/output-schema.md — exact format the reviewer is constrained to produce