Delegate-review

Review a PR or diff with eight finder angles, but run the finders on cheap external models instead of same-model Claude subagents. The orchestrator (you, ideally Fable/Opus - see delegate's orchestrator check) gathers the diff, writes each finder's cold-start prompt, dispatches in parallel, verifies every candidate itself, and writes the report. If the delegate skill's couriers or routing table aren't already in context, load it first: invoke the delegate skill via the Skill tool, or read ../delegate/SKILL.md relative to this skill's base directory (the two ship side by side). Its rules (self-contained prompts, size-bounded structured output, verify before integrating) all apply here.

Phase 0 - Gather the diff

For a PR: gh pr view <n> --json title,body,author,baseRefName,headRefName,state,additions,deletions,changedFiles and gh pr diff <n>. For a ref range or the working tree: git diff <range> / git diff HEAD. The diff is the review scope. Note the repo's absolute path and the head commit - finder prompts must reference PR-version files via git show <commit>:<path> and git grep <pattern> <commit>, never the local working tree (it may not match the PR).

Phase 1 - Dispatch finders (parallel, one block)

Eight angles, each returning up to 6 candidates as JSON {file, line, summary, failure_scenario}:

Correctness (route to gpt via codex-runner, read-only mode - an independent model family catching what Claude-family models miss is the point):