spectre-code_review

code_review

Purpose

Try to prove the work wrong, unsafe, unreachable, or unable to meet requirements. Accept correctness only if evidence survives adversarial scrutiny; never defend the implementation.

Inputs

• $ARGUMENTS: feature/root or descendant artifact, focus, diff/base range, source-plan path, immutable BASE_SHA/HEAD_SHA/DIFF_SHA256 tuple, and --orchestrated.
• Scope: completed changes, changed files, direct dependencies/importers, tests, and requirements/ACs. Source requirements from matching tasks.json parents, else an explicit source-plan path before plan.md, then request+diff. execute.md only locates tasks.

Working Set

• Reuse a managed FEATURE_ROOT only when explicit/current-thread evidence ties it to this work (physical directory wins; never branch/recency/lifecycle/scans); distinct work ignores ambient roots. Otherwise, including on collision, standalone MUST first load and follow Skill(spectre-feature-root) through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.
• Repair stale feature/root metadata in artifacts this review touches.
• REVIEW_REPORT={FEATURE_ROOT}/reviews/comprehensive_code_review.md; replace it for the current candidate. Git history preserves prior decisions without accumulating sibling reports.
• Late-bind diff/base and tuple, changed files, requirement/AC paths and task ids, dependencies/importers, tests, and exclusions. Reviewers read diffs/task graphs.

Outputs + DONE

REVIEW_REPORT follows references/adversarial-review.md: bounded scope and tuple; verdict; evidence-backed findings; coverage; per-requirement delivery; scope/dead-path audits; actions; route metadata.

DONE when schema validation passes; scope/tuple are unchanged; every requirement/AC has one evidence-backed status; audit and route metadata exist; findings satisfy identity/evidence rules; and only the report changed.

Method / guardrails

1. Dispatch the fixed review prompt. Read references/adversarial-review.md and send it verbatim to one fresh reviewer, followed only by structured context: feature/root and REVIEW_REPORT; focus; immutable diff/base tuple; requirement/AC source paths and ids; changed files, direct dependencies/importers, and tests; exclusions; route metadata and fallback reason. Reviewers read the supplied diffs/task graphs; the primary does not paraphrase or augment the template.
2. Use one external-first review route from repo root; do not probe alternatives. Codex primary runs: claude -p --model opus --effort high --permission-mode dontAsk --allowedTools "Read,Grep,Glob,LS,Bash(mkdir -p *),Bash(git diff *),Bash(git show *),Bash(git status *),Bash(git rev-parse *),Write" --output-format text "$REVIEW_PROMPT" Claude primary runs: codex exec -C "$PWD" -m gpt-5.6-sol -c 'model_reasoning_effort="high"' -s workspace-write "$REVIEW_PROMPT" Record route metadata as Claude Code|opus|high|Codex -> Claude Code or Codex|gpt-5.6-sol|high|Claude Code -> Codex. Keep a 20-minute launcher-side poll limit; do not pass duration guidance to the reviewer. Quiet output is not failure. The reviewer writes only REVIEW_REPORT. Primary semantic self-review is prohibited.
3. Fallback once. Missing/non-zero opposing CLI, report-write failure, or unusable review permits one clean-context @spectre_reviewer with the same verbatim template and structured context. It returns the report in-thread for unchanged persistence. Record native-subagent|runtime-native|inherited|native-fallback plus reason. A usable review ends semantic review.
4. Verify deterministically. A candidate tuple requires all fields and canonical hash recomputation with git diff --binary --full-index --no-ext-diff --no-color --no-renames before dispatch and after report creation; mismatch makes the report stale. Validate the template's required sections, scope, exhaustive delivery coverage, and metadata. The primary may mechanically normalize report-only counts, paths, citations, metadata, sections/tables, and severity enums from existing reviewer semantics, but may not originate or materially reinterpret findings. An unusable native report remains incomplete.