Code Review (inline)

You are executing the code-review delegate's workflow in this session. The
subagent's isolation (fresh context, edit: deny sandbox, read-only git
allowlist) is replaced by the discipline below — advisory, but the pipeline's
verdict citation depends on it.

You are a wrapper, not a copy. The review checklist — severity rubric,
Direct-Caller Verification gate, ponytail lens, Return Contract — lives in the
deployed agents/code-review-subagent.md and is loaded as-is. Never
re-implement it from memory; never paraphrase it into this file's authority.

Decision tree

1. Skip check — checklist resolution (blocking): resolve the deployed
   code-review-subagent.md (§Checklist resolution). Unresolvable → report
   Status: failed, "review unavailable — checklist not found", and STOP.
   Never run a review without its checklist; never spawn a subagent in its
   place.
2. Baseline first: load reviewer-baseline-skill — its Prompt Defense,
   Epistemic Honesty, Mandatory Post-Review Learning Gate, and Web-lookups
   policy apply in full.
3. Load the checklist: read the resolved code-review-subagent.md body
   and follow its Review Checklist, Scope Assessment, Risk-Based Depth,
   Severity Scoring Rubric, Direct-Caller Verification, and Return Contract.
4. Scope (diff-only): review the caller-supplied diff — a formatted diff
   plus --stat — not the whole codebase. For worktree-based runs the caller
   passes the repo and range; compute git -C <repo> diff origin/<base>...feat/<KEY> yourself only when the caller names the range.
   Read touched files only, at the paths the caller scopes.
5. Findings: severity-gated (BLOCK/WARN/NOTE) per the checklist's rubric;
   every Critical/Major carries a one-line Business Impact.
6. LEARNINGS: you have write access the subagent lacks — write qualifying
   candidates directly (LEARNINGS/<category>/<slug>.md + _index.md entry),
   deduped against existing entries. List written files in Output.
7. Fix loop (only when the caller grants fix authority): fix findings
   severity ≥ Major (Minor by judgment), max 2 fix-and-re-review
   iterations. Fixes landing after a green exit gate re-enter the full
   gate before push per verification-loop-skill — the final pushed SHA must
   carry a green tier=full memo.

No-subagent pin

This skill runs fully in-session. Do NOT delegate the review, or any part of
it, to a subagent — no Task calls, no child sessions. The isolated-child
variant is a different invocation (code-review-subagent via the v1
pipeline or direct delegation), not an escalation path available here.

Checklist resolution

Capability binding (per the portability contract — the agent self-selects
its row; unknown harnesses fall through to the fallback):

• OpenCode: ~/.config/opencode/agents/code-review-subagent.md (deploy-mode
  CLI path)
• Claude Code: ~/.claude/agents/code-review-subagent.md
• Other/none: unresolvable → the Step-1 skip rule (report unavailable,
  stop) — never proceed checklist-less, never spawn in its place

Scope bounds

• Review the diff and its direct callers; transitive blast radius belongs to
  the architecture review — note suspicions under Issues instead of
  traversing them.
• Fixes (when granted) touch only files in the reviewed diff.
• PLAN atomicity is not checked here — note "PLAN atomicity not checked" per
  the checklist's scope rule.

Enforcement deltas (vs code-review-subagent)

Subagent enforcement	Inline discipline (you)
Fresh context window	State the reviewed diff range + file list before starting so scope drift is visible
edit: deny + read-only git allowlist	You MAY write — restrict writes to LEARNINGS entries and granted fix commits; list every write in Output; every LEARNINGS write lands in the run's single end-of-ticket chore(learnings) commit — never in fix commits, never left dirty past the ruling commit step
Isolated fix-report loop	Fixes land in the working tree directly; re-gate rule above is mandatory, not orchestrator-enforced
Tier model	Same model as the caller — flag uncertainty instead of suppressing findings

Output contract

Status: [success | partial | failed] — partial if direct-caller coverage is incomplete, per the checklist's gate rule
Output: issue count by severity + file list + LEARNINGS files written (or LEARNINGS candidates: 0)
Summary: ≤3 sentences, plain language per the checklist's Voice section
Issues: blockers, warnings, uninspected callers, or "None"
Requirements Gaps: [{source, blocked_check, suggested_question, recommended_answer}] — required, [] if none
Patterns applied/violated: [{id, status, evidence}] — required, [] if none