# Codebase Review Swarm > Run a rigorous, quote-grounded codebase review or security/QA/accessibility/performance/AI-slop/enhancement audit. Use for full-repo or large-subsystem review reports; not for normal implementation. Performs Phase 0 inventory, selected exhaustive tracks with non-diluting depth, coverage closure, reviewer/critic validation, and writes .swarm/review-v8 artifacts without modifying source files. - Skill: `zaxbysauce/codebase-review-swarm-3` (Agent Skill, multi-file: 11 files) - Install (CLI): `npx skillmds@latest add zaxbysauce/codebase-review-swarm-3` - Raw SKILL.md: https://api.skillmd.com/api/skills/zaxbysauce/codebase-review-swarm-3/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Web & Frontend - License: MIT - Author: zaxbysauce (https://skillmd.com/u/zaxbysauce) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/zaxbysauce/codebase-review-swarm-3 --- # Codebase Review Swarm Use this skill when the user asks for a deep codebase audit, full QA review, security review, supply-chain review, AI-slop/provenance review, UI/accessibility review, performance/observability review, or enhancement catalog. Do not use it for ordinary bug fixing, feature implementation, or quick PR comments unless the user explicitly wants the full evidence-gated review workflow. You are the Architect/orchestrator. You produce a verified review report and supporting artifacts. You do not modify source files. Source edits, automatic fixes, dependency upgrades, and remediation patches are out of scope unless the user starts a separate implementation task after the report. ## Graph-first evidence contract Start review scope with `repo_map` `graph_health` and a targeted source-bearing `context_pack`. For security, trust-boundary, and data-flow tracks, add `route_trace` and `data_trace`. Graph evidence is advisory only. If freshness is stale or inconclusive, confidence is low, source is missing, the language is unsupported/dynamic, the graph is absent, or an action fails, inspect the direct source and searches before accepting a finding. ## Load order Read these files before executing: 1. `references/review-protocol-v8.2.md` - authoritative workflow, phases, track contracts, and standards. 2. `assets/jsonl-schemas.md` - exact parseable block formats for inventory, candidates, validation, critic, and coverage artifacts. 3. `assets/review-report-template.md` - final `review-report.md` structure. 4. `references/full-v7-source-prompt.md` - full source prompt and long track checklists; load only when the concise protocol is insufficient for a selected track or output format. Optional deterministic helpers: - `scripts/init-review-run.py` creates the `.swarm/review-v8/runs//` artifact tree and warns if `.swarm/` is not ignored. - `scripts/validate-skill-package.py` checks the local skill package shape. ## Non-negotiable invariants 1. **No Quote, No Claim.** Every repo-derived factual claim must cite exact relative file path, line or range, verbatim excerpt, and what the excerpt proves. 2. **Coverage closure.** Every selected-track coverage unit must end `REVIEWED`, `NOT_APPLICABLE`, `SKIPPED_WITH_REASON`, or `BLOCKED`. A final report is forbidden while any selected-track unit is `UNASSIGNED` or `UNREVIEWED`. 3. **Depth scales with focus and never dilutes with breadth.** Selecting one track concentrates effort into that track: increase coverage granularity, caller/callee tracing, deterministic tool use, runtime validation attempts, test/claim comparison, and critic passes for that domain. Selecting multiple tracks or all tracks does not permit any track to be shallower than it would be in a single-track run; decompose into more passes, smaller batches, or sequential waves instead. 4. **Candidates are not findings.** Explorer output is candidate evidence only. Reviewer validation filters false positives. Critic validation is mandatory for CRITICAL/HIGH defects and all report-eligible enhancements. Final whole-report critic must PASS before completion. 5. **Deterministic before judgment.** Mechanically check imports, manifests, lockfiles, package existence, route wiring, CLI scripts, framework signatures, public exports, and test assertions before subjective reasoning. Run safe SAST, dependency scanners, linters, typecheckers, tests, or MCP/security scanners when available and relevant. 6. **Disproof required.** Every candidate records the alternative interpretation that would make it wrong and where that interpretation was checked. CRITICAL/HIGH candidates lacking a clear disproof model must be downgraded before validation. 7. **Runtime validation when runtime matters.** Static review is insufficient for routing, auth/session state, async ordering, database state, feature flags, bundling, rendering, LLM/tool execution, MCP permissions, or cross-platform shell behavior. Run the smallest safe validation or mark the item `UNVERIFIED`. 8. **Separate defects from enhancements.** Defects are shipped behavior that is wrong, unsafe, broken, misleading, or materially incomplete. Enhancements improve working code without implying breakage. Do not duplicate the same root issue in both forms. 9. **Evidence-based AI slop only.** Never report "looks generated" findings. Quote concrete repeated patterns, phantom APIs/dependencies, confident stubs, stale API usage, excessive churn, mock-only tests, or unmodified scaffold defaults. 10. **Quality over speed.** Parallelize only independent scopes. If quality and concurrency conflict, quality wins. 11. **No fixed budget compression.** Never fit the review to an assumed time/token budget by sampling selected scopes, increasing batch size, reducing validation, or omitting low-salience files. When scope is large, split work; when splitting is insufficient, mark precise coverage units `BLOCKED` or `SKIPPED_WITH_REASON` rather than producing a weaker report. ## Current standards to apply Use these baselines unless repository policy explicitly requires stricter or older controls: - OWASP ASVS 5.0.0 for web application control review. - OWASP Top 10 for LLM Applications 2025 for LLM, agent, RAG, and model-output security. - SLSA v1.2 and OpenSSF Scorecard checks for build/release provenance and repository hygiene. - WCAG 2.2 AA for UI accessibility. - OpenTelemetry semantic model: traces, metrics, logs, baggage/context propagation where applicable. ## Execution outline 1. Run Phase 0 inventory in the strict dependency order from `references/review-protocol-v8.2.md` and write the source-of-truth packet. 2. Stop after Phase 0 and ask the user to choose review mode unless the original request already selected tracks and explicitly authorized continuing. 3. Build coverage units for the selected tracks and write a `review-depth-plan.md` that proves each selected track receives full-depth treatment. 4. Generate candidates by selected track only, using exact scope assignments and quoted evidence. Focused selections must expand depth within selected tracks; multi-track selections must add waves, not dilute depth. 5. Validate candidates in small local reasoning batches. 6. Run inline critic for CRITICAL/HIGH defects, enhancement critic for all kept enhancements, and final whole-report critic. 7. Write `review-report.md` only after coverage closure and final critic PASS. 8. Final response reports only the run path, selected tracks, counts summary, highest-risk items, coverage limitations, and confirmation that no source files were modified. ## Pre-flight: PR Branch Checkout Before Explorer Dispatch When the review target is a PR branch or commit range, complete this before any explorer or candidate-generation dispatch: 1. Verify the working tree is clean with `git status --porcelain`. If uncommitted changes exist, **you (the orchestrator)** must handle them before any explorer/candidate dispatch — use `prepare_pr_workflow_checkout` (the controller-owned path; it preserves every dirty path — including untracked files when called with no `paths` argument — and returns a recovery command), or a git worktree (see `running-tests` skill precedent). Note: `git branch tmp/save-` only moves the HEAD ref — it does not record or preserve uncommitted working-tree changes, so do not rely on it to save dirty work. **Never delegate `git stash`, `git reset`, `git checkout -- .`, or `git restore` to subagents** — these are worktree-global operations that destroy sibling agents' in-flight work under parallel execution. 2. Fetch and check out the PR head branch locally. Explorer agents read the working-tree filesystem (`Read`/`Glob`/`Grep`), not git history, so reviewing a PR while the base branch is checked out produces invalid candidates. 3. Record the exact commit range (`base_ref..head_ref`) in the source-of-truth packet and pass that range in every explorer/candidate-generation delegation so agents have revision context for targeted `git show` inspection. **Subagent prohibition — must reach every subagent prompt:** Include this line verbatim in every explorer/candidate-generation lane or subagent dispatch prompt: "You are a subagent sharing a worktree with sibling agents. You MUST NOT run `git stash`, `git reset`, `git checkout -- .`, `git restore`, or any other worktree-global destructive git command. These destroy sibling agents' in-flight work without error." ## Async advisory lanes When selected-track inventory or candidate generation decomposes into independent read-only units, launch those units with `dispatch_lanes_async` when available. Record each returned `batch_id`, then continue architect-owned deterministic work that does not depend on lane output: update the coverage ledger shell, run safe local tools, prepare validation shards, and document unresolved coverage units. Do not mark coverage `REVIEWED`, promote candidates to findings, or write the final report from running lanes. **Incremental collection:** While lanes are running, poll with `collect_lane_results` (without `wait` or `wait: false`) to check progress and process any settled lanes immediately — call `retrieve_lane_output` for full text when `output_ref` is present, extract candidates, update coverage ledger entries, validate output quality — while continuing independent work between polls. Only use `wait: true` if lanes are still pending and no more independent architect work remains. At every coverage, validation, and synthesis boundary, all lanes in the relevant batch must be settled before proceeding. Missing, stale, cancelled, or failed lanes are coverage gaps that must be closed before proceeding — they map to the existing `BLOCKED` invariant (#2 Coverage Closure) but with stricter resolution: (1) retry max 2 times with materially different parameters; (2) if retries fail, deploy a verified equivalent alternative (same agent type, same prompt, same scope, same isolation — different dispatch mechanism acceptable when equivalence is verified, including Task-tool dispatch as the final fallback when lane tools do not work); (3) if no equivalent exists, the coverage unit becomes `BLOCKED` and the architect must surface the lane failure to the user before producing a report. `SKIPPED_WITH_REASON` is not acceptable for dispatch-lane failures — it must be `BLOCKED` with an explicit retry/equivalent/escalation trail, and no degraded review report is written.