Review All

Full Review

Orchestrate all domain-specific review skills as parallel subagents, then consolidate into a unified report.

Workflow

1. Scope and explore

• Confirm scope with the user: full codebase, specific packages/directories, changed files only (PR or branch diff), or specific concern.
• Resolve scope to a file/package list. Based on what the user requested:
  • Changed files (PR or branch): Run git diff --name-only --diff-filter=d <base>...HEAD to get changed files (default <base> is main). If the user references a PR number, use gh pr diff <number> --name-only instead. Derive affected Go packages from the file paths (unique parent directories containing .go files).
  • Explicit paths/packages: The user may specify directories (e.g. internal/auth/), Go package patterns (e.g. ./internal/auth/...), or individual files. When given a directory or package pattern, include all files under it. Derive Go package paths for static analysis tool invocations.
  • Full codebase: No filtering. Explore everything (default).
• Pass the resolved scope (file list, derived package paths, and file-type flags below) to each review subagent in step 3 so they skip their own scope confirmation and use the provided scope directly.
• Resolve pr_url for deep-linking (display-only, used in the final consolidated report). Run gh pr view --json url -q .url 2>/dev/null to capture the PR URL for the current branch (or for the PR number the user supplied via gh pr view <num> --json url -q .url). Empty string if no PR exists. Pass pr_url to each subagent in step 3 and to the summarization subagent in step 4. Subagents wrap finding path:line references using ~/.claude/scripts/pr-deeplink.sh "$pr_url" <path> <line>. See Finding link wrapping below.
• Classify the resolved files to determine which reviews to launch:
  • has_code: any source files (.go, .rs, .ts, .tsx, .js, .jsx, .swift, .kt, .kts, .py, .rb)
  • has_go: any .go files
  • has_proto: any .proto files
  • has_sql: any .sql files
  • has_iac: any Dockerfiles/Containerfiles, k8s manifests, Terraform (.tf/.tofu), Helm charts (Chart.yaml), service mesh / gateway CRDs (Linkerd/Istio/Gateway API/Ingress/Envoy bootstrap)
  • has_ci: any GitHub Actions workflows (.github/workflows/*.yml), composite actions (action.yml), Dependabot/Renovate configs, or CI configs (.circleci/config.yml, .buildkite/pipeline.yml, .gitlab-ci.yml, Jenkinsfile, azure-pipelines.yml, cloudbuild.yaml, bitbucket-pipelines.yml)
  • has_infra: shorthand for has_iac || has_ci (kept for backwards compatibility with existing review-* subagents)
  • has_api_specs: any .proto, OpenAPI/Swagger specs (openapi.{yml,yaml,json}, swagger.{yml,yaml,json}), or GraphQL schemas (*.graphql, *.gql)
  • has_docs: any .md files or OpenAPI/Swagger specs
  • has_manifest: any dependency manifest (go.mod, package.json, requirements.txt, pyproject.toml, Cargo.toml, Gemfile, mix.exs, composer.json)
  • has_changes: true when scope is "changed files (PR or branch)", or when scope is "explicit paths" and those paths have a diff against the base ref (run git diff --name-only --diff-filter=d <base>...HEAD -- <paths> to check; default base is main). False for full-codebase reviews with no diff baseline.
• Detect opt-in flags from the user's request phrasing:
  • include_performance: true when the user explicitly asks for performance, perf, benchmark, profiling, pprof, hot-path, or latency review alongside the comprehensive request. Default false. Never auto-enable from file types.
  • include_deep_security: true when the user explicitly asks for a deep / thorough security review, "CodeQL", "capability audit", "supply-chain audit", or otherwise signals willingness to pay the multi-minute cost of CodeQL database builds and capslock VTA. Default false. Standard review-security still runs without this flag; the flag only enables the deep static-analysis tier. Never auto-enable from file types.
• Determine which review types are applicable using the flags above:
  • review-security: applicable if has_code or has_infra
  • review-reliability: applicable if has_code or has_iac
  • review-code: applicable if has_code or has_proto
  • review-database: applicable if has_sql, or database-interacting code exists (check imports for DB drivers like pgx, pq, database/sql, sqlx, diesel, sqlalchemy, etc.)
  • review-coverage: applicable if has_go and has_changes
  • review-documentation: always applicable
  • review-infrastructure: applicable if has_iac
  • review-ci: applicable if has_ci
  • review-observability: applicable if has_code (observability gaps are code-level; configs alone aren't enough)
  • review-api-compat: applicable if has_api_specs AND has_changes (it's a diff-aware review; no diff baseline = nothing to compare)
  • review-performance: applicable ONLY if include_performance is true. Never auto-launched.
  • evaluate-dependency: applicable if has_manifest AND has_changes. Diff the manifest to find added or version-bumped dependencies (git diff <base>...HEAD -- <manifest>); run it in review mode once per added/bumped dependency. Skip when the only manifest change is a removal or a checksum-only lockfile churn.

1a. Detect conformance mode

If the user's request includes phrases like "full conformance", "pattern discovery", "check patterns", or "discover patterns", set conformance_mode to full. Otherwise default to lightweight. This flag is passed to review-code in step 3.

1b. Load REVIEW.md (if present)

Check for a REVIEW.md file at the repository root. If it exists, read it and extract:

• Always check rules: these become mandatory check items passed to all subagents (flagged at HIGH severity)
• Skip rules: filter these paths/patterns out of scope before passing to subagents (apply alongside the file-type classification above)
• Domain-specific sections (Security, Reliability, Database, Protobuf & API, Go conventions, Documentation): route each section to the corresponding review subagent as additional context

If no REVIEW.md exists, proceed without it. All review skills have their own reference checklists.

REVIEW.md schema extensions used by this skill (all subsections optional):

## Open context

- Skip: true                    # disable open-work-context lookup entirely
- Include PRs: false            # disable GH PR source
- Include issues: false         # disable GH issue source
- Jira project: AUTH            # enable Jira source with this project key
- Recency days: 60              # override the 30-day window

## Policy gate

- Skip: true                    # suppress the gate prompt; always proceed with original scope
- Policy files: [glob, ...]     # additional globs appended to the default policy-file list

Open context controls step 1d (below). Policy gate controls step 1c (above).

1c. Policy-change detection (gate)

If diff scope is in use, check whether any changed file is a policy file — a rule/config artifact that affects review of the whole repo, not just itself:

• REVIEW.md, CLAUDE.md, CLAUDE.local.md, AGENTS.md (at any path depth; typically repo root)
• Any file under .claude/rules/ or .cursor/rules/ (recursive)
• Any glob declared in REVIEW.md under Policy gate: Policy files: [glob, ...] — appended to (not replacing) the default list

Detection:

POLICY_HITS=$(printf '%s\n' "$CHANGED_FILES" \
  | grep -E '(^|/)(REVIEW|CLAUDE|CLAUDE\.local|AGENTS)\.md$|^\.claude/rules/|^\.cursor/rules/' \
  || true)
# Append matches for any REVIEW.md-declared globs to POLICY_HITS.

When POLICY_HITS is empty, skip the rest of step 1c and continue to 1d. When non-empty, apply the gate below.

Opt-out: if REVIEW.md declares Policy gate: Skip: true, suppress the prompt and continue with the original diff scope silently. Record Scope: diff (kept; policy gate skipped via REVIEW.md) in the run metadata.

Gate behavior when POLICY_HITS is non-empty and not opted out:

1. Surface to the user a brief block listing the policy files in POLICY_HITS, plus the one-line explanation: "Policy/rule changes affect the whole codebase; diff review only audits the policy itself."
2. Ask: "Switch to full-repo review against this policy? [y/N]"
3. y → rescope: clear pr_url, set scope to full codebase (run the file-type classification on the entire tree the way a no-PR / no-base-ref invocation would), set conformance_mode=full. Append to run metadata: Scope: full-repo (escalated from diff due to policy change in <files>).
4. N → keep diff scope unchanged. Append to run metadata: Scope: diff (kept despite policy change in <files>).

Continue to 1d once the gate decision is recorded.

1d. Resolve open-work context

Surface in-flight work that may overlap with the (post-gate) scope so subagents can flag conflicts and duplicates. Skip the entire step if REVIEW.md declares Open context: Skip: true.

Keyword extraction from the changed file list (output of step 1):

1. Unique top-level dirs from changed paths.
2. Unique immediate parent dir names.
3. For .go files: filepath.Base(pkg_dir).
4. For .proto files: the declared package line.

De-dup, lowercase, drop the stop-list (internal|pkg|cmd|test|tests|vendor|gen|api|proto|src|lib). If the resulting keyword set is empty (e.g. all changed files sit inside stop-listed dirs), skip step 1d entirely — no useful filter is possible.

The keyword set drives the three source queries below (PRs, issues, Jira).

Source: open GitHub PRs

Skip if REVIEW.md declares Open context: Skip: true or Open context: Include PRs: false.

gh pr list --state open --limit 50 \
  --search "updated:>$(date -v-30d +%Y-%m-%d 2>/dev/null || date -d '30 days ago' +%Y-%m-%d)" \
  --json number,title,headRefName,labels,updatedAt,url 2>/dev/null

Post-filter the JSON result: keep entries where title || headRefName || labels[*].name contains ≥1 keyword (case-insensitive substring). Sort by updatedAt desc, cap at 10.

Fail-soft: if gh errors (no auth, no remote, command not found), record one line PR lookup unavailable: <reason> in the run metadata and continue with the remaining sources.

Source: open GitHub issues

Skip if REVIEW.md declares Open context: Skip: true or Open context: Include issues: false.

gh issue list --state open --limit 100 \
  --search "updated:>$(date -v-30d +%Y-%m-%d 2>/dev/null || date -d '30 days ago' +%Y-%m-%d)" \
  --json number,title,labels,updatedAt,url 2>/dev/null

Post-filter: keep entries where title || labels[*].name contains ≥1 keyword. Sort by updatedAt desc, cap at 10.

Label boost: when a hit carries any of bug, regression, flaky, security, prefix the rendered row with ★ (visual emphasis only — ordering is unchanged).

Fail-soft: same as the PR source — record Issue lookup unavailable: <reason> and continue.

Source: open Jira tickets (opt-in)

Run this substep only when REVIEW.md declares Open context: Jira project: <KEY>. No default; absence means no Jira lookup.

Use the MCP tool claude_ai_Atlassian_Rovo:searchJiraIssuesUsingJql with this JQL template (substitute <KEY> and <RECENCY>; <RECENCY> defaults to 30, override via Open context: Recency days:):

project = <KEY> AND statusCategory != Done AND updated >= -<RECENCY>d

Post-filter: keep entries where summary || description contains ≥1 keyword (case-insensitive substring). Sort by updated desc, cap at 10. Render each as <KEY>-<NUM>: <summary> with status.

Fail-soft: if the MCP is unavailable or returns an auth error, record Jira lookup unavailable: <reason> and continue.

Rendered block — Open work context:

## Open work context

Filter: changed-path keywords `<keyword-list>`; updated within last 30 days.

**Open PRs (n)**
| # | Title | Branch | Updated |
| --- | --- | --- | --- |
| [#412](url) | <title> | <branch> | YYYY-MM-DD |

**Open issues (n)**
| # | Title | Labels | Updated |
| --- | --- | --- | --- |
| ★[#523](url) | <title> | bug | YYYY-MM-DD |

**Open Jira (n)**
| Key | Summary | Status | Updated |
| --- | --- | --- | --- |
| [AUTH-2583](url) | <summary> | In Progress | YYYY-MM-DD |

Omit each sub-block when count is 0. Omit the whole section when all three counts are 0 (do not render an empty ## Open work context heading).

2. System overview

Produce a brief architecture summary covering:

• Services, ports, and transport (gRPC, HTTP, etc.)
• Data stores and external dependencies
• Authentication and authorization mechanisms
• Deployment model (if discernible)

Map the critical hot paths:

Client → Transport
  → step 1 (local / I/O annotation)
  → step 2 (DB round-trip #1)
  → step 3 (external call, round-trip #2)
  → response

Annotate each step: local vs. I/O, serial vs. parallel, cached vs. uncached.

This system overview is shared context for all review subagents.

2b. Run pattern discovery (if full conformance mode)

If conformance_mode is full, resolve the review output directory first. .reviews/ is gitignored on first use; review outputs are working artifacts, not source:

REVIEW_DATE=$(date +%Y-%m-%d)
REVIEW_DIR=".reviews/${REVIEW_DATE}"
if [ -d "$REVIEW_DIR" ]; then REVIEW_DIR=".reviews/${REVIEW_DATE}-$(date +%H%M)"; fi
~/.claude/scripts/ensure-gitignore.sh '.reviews/'
mkdir -p "$REVIEW_DIR"

Then launch a /discover-patterns subagent (subagent_type="generalPurpose", model: opus per subagent-model-routing — architecture-level pattern discovery) with the resolved scope and REVIEW_DIR, instructing it to write to ${REVIEW_DIR}/PATTERNS.md. Pass REVIEW_DIR to review-code's prompt so its Conformance Check subagent reads ${REVIEW_DIR}/PATTERNS.md. Other review subagents (security, reliability, database, documentation) can launch in parallel with this step since they don't depend on it; only review-code must wait for it to complete.

3. Launch review subagents in parallel

Launch applicable review skills concurrently using the Task tool (max 4 at a time; if more than 4, launch the first 4 and the remaining after one completes). Each subagent is subagent_type="generalPurpose", model: sonnet (per subagent-model-routing — structured analysis with code-level findings).

For each subagent, include in its prompt:

• The system overview and flow mapping from step 2
• The resolved file list and package paths from step 1 (the subagent should use this scope directly and skip its own scope confirmation)
• The has_changes flag, base ref, and changed file list from step 1 (so change-aware subagents like review-code's Regression History can use them)
• The pr_url from step 1 (used to wrap path:line finding references via ~/.claude/scripts/pr-deeplink.sh; empty string disables wrapping)
• The conformance_mode flag from step 1a (for review-code only)
• The include_deep_security flag from step 1 (for review-security only)
• If REVIEW.md was loaded in step 1b: the "Always check" rules (for all subagents) and the relevant domain-specific section for that subagent (e.g. Security section → review-security, Database section → review-database). Instruct the subagent to treat "Always check" rules as HIGH severity and domain-specific rules as MEDIUM severity, in addition to its own reference checklist.
• Instructions to follow the corresponding skill's workflow (read the SKILL.md for reference on what each skill does)
• Request that it return the full findings output (including tracking annotations and tool availability sections)
• If the Open work context block from step 1d is non-empty, include it in the subagent prompt under a dedicated heading:

  # Open work context that may overlap
  <paste the rendered block from step 1d here>
  
  Flag in your findings if any item below conflicts with, duplicates, or would be invalidated by your recommendations. Do not treat the existence of an open PR as license to skip a finding.
  

  This is the sole injection point for the open-context block. The system overview from step 2 is not modified to carry it — keep step 2 focused on architecture, step 3's per-subagent prompt focused on reviewer-facing context.

Review subagents to launch:

Each subagent should NOT write its own output file; it returns findings to this orchestrator. (evaluate-dependency is launched once per added/bumped dependency, batched under the concurrency cap; each returns a GO/CAUTION/NO-GO row.)

Concurrency cap. Launch up to 4 subagents at a time. With all skills enabled the dispatch can exceed 4; queue the rest and launch them as earlier ones complete.

4. Launch summarization subagent

After all review subagents complete, launch a single summarization subagent (subagent_type="generalPurpose", model: opus per subagent-model-routing — cross-cutting dedup and prioritization across all review domains) with the full findings from each review subagent. Pass pr_url so it can preserve and apply the Finding link wrapping convention when rewriting tables.

Prompt it to:

1. Deduplicate overlapping findings across all reviews. Common overlaps to watch for:
   • security ↔ reliability (e.g. unbounded result sets)
   • security ↔ infrastructure (e.g. inline secrets in TF / k8s)
   • security ↔ ci (e.g. PR-target script injection)
   • security ↔ dependency-eval (e.g. a new dep's vuln history overlaps govulncheck output)
   • reliability ↔ observability (e.g. missing error spans on hot paths)
   • reliability ↔ infrastructure (e.g. k8s probes vs. shutdown contract — review-infrastructure covers probe presence, review-reliability covers shutdown semantics)
   • code ↔ api-compat (e.g. a proto change flagged for design AND for wire compat)
2. Cross-reference each deduplicated finding to its source review and IDs.
3. Compute unified tracking for each deduplicated finding. Populate tracking: {status, sources, possibly_overlaps} per reference-tracking.md. Combines in-repo signals (TODO / FIXME etc., always evaluated) with open-work signals (tier-1 path mention / tier-2 symbol mention / tier-3 keyword match against the step-1d set). Tier-1 and tier-2 promote to tracked; tier-3 stays untracked with a possibly_overlaps annotation. Terminal-state items (closed / merged PRs, closed issues, statusCategory = Done Jira) never produce a source entry.
4. Prioritize and lay out per reference-tracking.md § Findings layout in SUMMARY.md: two flat tables (untracked above the fold; tracked collapsed in a <details> block), severity desc then category alpha, tracking rendered inline per reference-tracking.md § Badge rendering.
5. Recommend fix order, considering dependencies between findings and effort estimates. Tracked findings (tier-1 or tier-2 source) may be deprioritized when the source indicates an active plan. A possibly_overlaps annotation (tier-3 only) does not affect ordering.
6. Tool Availability summary. Consolidate from all reviews into a summary listing which automated tools ran successfully, which were skipped, and why.

5. Present results

If REVIEW_DIR was resolved in step 2b, reuse it. Otherwise, resolve it now:

REVIEW_DATE=$(date +%Y-%m-%d)
REVIEW_DIR=".reviews/${REVIEW_DATE}"
if [ -d "$REVIEW_DIR" ]; then REVIEW_DIR=".reviews/${REVIEW_DATE}-$(date +%H%M)"; fi
~/.claude/scripts/ensure-gitignore.sh '.reviews/'
mkdir -p "$REVIEW_DIR"

Capture run metadata for the output header (see Run metadata header below). When scope is diff-based, also capture BASE_REF and BASE_COMMIT=$(git rev-parse "$BASE_REF").

Write the summarization output to ${REVIEW_DIR}/SUMMARY.md, structured as:

1. Run metadata header
2. Tool availability summary
3. System overview (from step 2)
4. Open work context (from step 1d; omit this section entirely if the block was empty)
5. Findings — untracked (one flat table; see reference-tracking.md § Findings layout in SUMMARY.md)
6. Findings — tracked (collapsed <details> block, one flat table; omit when empty)
7. Recommended fix order (untracked first; tracked items only if they're high-severity blockers despite an owner)

Render both findings sections per reference-tracking.md § Findings layout in SUMMARY.md — flat tables (categories are a column, not a header), severity desc then category alpha, empty-group rules defined there. Per-domain Output Templates below describe what each subagent emits; the summarization subagent compresses them into the unified shape. Present the report to the user.

Run metadata header

The metadata-capture snippet and header template live in reference-metadata.md. Capture once near REVIEW_DIR resolution and prepend to every output document this skill writes (subagents that write their own files do the same).

Finding link wrapping

When pr_url is non-empty (resolved in step 1), every path:line reference inside finding cells in the consolidated tables below is wrapped as a Markdown link to the PR's "Files changed" tab, anchored at the line. The display text stays path:line; only the link target carries the URL, so table widths don't blow up.

Use the helper to build each link:

~/.claude/scripts/pr-deeplink.sh "$pr_url" <path> <line>
# → [path:line](https://github.com/owner/repo/pull/N/files#diff-<hash>R<line>)

~/.claude/scripts/pr-deeplink.sh "$pr_url" <path>
# → [path](https://github.com/owner/repo/pull/N/files#diff-<hash>)   (file-level)

~/.claude/scripts/pr-deeplink.sh "" <path> <line>
# → path:line   (no PR scope; plain text)

Notes:

• Right-side anchor (R<line>) is the default and almost always correct; findings call out added/modified code.
• Use L as the fourth argument only when a finding is specifically about removed code on the left side of the diff.
• The diff anchor format (#diff-<sha256(path).first32>) is GitHub's stable but undocumented convention. If GitHub ever changes it, only pr-deeplink.sh needs updating.
• For file-level findings (no specific line), call the helper without <line> to emit a file-anchor link.
• Tracked column entries that include path:line (e.g. TODO in foo.go:42) follow the same wrapping rule.
• Findings themselves follow terse-comments: concrete fix, optional bug:/risk:/nit:/unsure: prefix, no praise or restating the diff.

This applies to the consolidated tables below and to per-category finding tables emitted by each subagent (reproduced into the consolidated report).

The Tracked column in those tables renders the badge defined in reference-tracking.md § Badge rendering.

Output Templates

The Tracked column in every table below renders the unified tracking badge defined in reference-tracking.md § Badge rendering. The example cell values (—, TODO in file:line) are placeholders — real rows render [tracked: …] or [→ possibly overlaps: …] per that section's rules.

Consolidated security findings

| Severity | ID | Finding | STRIDE | OWASP | Tracked |
|----------|----|---------|--------|-------|---------|
| CRITICAL | 1 | Description with code references | S1, E1 | A01, A07 | — |
| HIGH | 2 | Description with code references | T2 | A04 | TODO in file:line |

Consolidated reliability findings

| Priority | Finding | Impact | Effort | Tracked |
|----------|---------|--------|--------|---------|
| P0 | Description with code references | Impact on availability/latency | trivial / small / moderate / large | — |

Consolidated code findings

| Severity | ID | Finding | Source | Tracked |
|----------|----|---------|--------|---------|
| HIGH | 1 | Description with code references | ARCH1, DEP2 | — |
| MEDIUM | 2 | Description with code references | GO1, SA3 | TODO in file:line |
| MEDIUM | 3 | Description with code references | PB2, PBL1 | — |
| HIGH | 4 | Description with code references | REG1 | — |
| MEDIUM | 5 | Description with code references | CONF1, CONF2 | — |

Consolidated documentation findings

| Severity | ID | Finding | Source | Tracked |
|----------|----|---------|--------|---------|
| HIGH | 1 | Description with code references | DOC1, DOC4 | — |
| MEDIUM | 2 | Description with code references | DOC2 | TODO in file:line |

Consolidated infrastructure findings

| Priority | Surface | Finding | Impact | Effort | Tracked |
|----------|---------|---------|--------|--------|---------|
| P0 | k8s | Description with code references | Impact | trivial / small / moderate / large | — |
| P1 | terraform | Description with code references | Impact | Effort | FIXME in file:line |

Consolidated CI findings

| Priority | Workflow | Finding | Impact | Effort | Tracked |
|----------|----------|---------|--------|--------|---------|
| P0 | release.yml | Description with code references | Supply chain / security | trivial / small / moderate / large | — |

Consolidated observability findings

| Priority | Signal | Finding | Impact | Effort | Tracked |
|----------|--------|---------|--------|--------|---------|
| P0 | tracing | Description with code references | MTTR / debuggability | trivial / small / moderate / large | — |

Consolidated API compatibility findings

| Priority | Surface | Change | Class | Recommendation | Tracked |
|----------|---------|--------|-------|----------------|---------|
| P0 | proto | `pkg.Service.Method` removed at file:line | wire-breaking | Deprecate first; remove in next major version | — |

Consolidated performance findings

Only emitted when include_performance is true.

| Priority | Category | Finding | Impact | Effort | Evidence | Tracked |
|----------|----------|---------|--------|--------|----------|---------|
| P1 | allocation | Description with code references | Allocations on hot path | small | profile needed | — |

Consolidated coverage findings

Per-package coverage (omit Delta column when no prior baseline exists):

| Package | Coverage | Delta | Affected Functions |
|---------|----------|-------|--------------------|
| internal/auth | 78.4% | +2.1% | 3 changed, 1 uncovered |
| internal/store | 64.2% | — | 5 changed, 4 uncovered |

Uncovered functions (grouped by package, sorted by severity):

| Severity | ID | Package | Function | File:Line | Tracked |
|----------|----|---------|----------|-----------|---------|
| HIGH | COV1 | internal/auth | `verifyToken` | auth/verify.go:42 | — |
| MEDIUM | COV2 | internal/render | `(*Page).Render` | render/page.go:104 | — |
| LOW | COV3 | internal/store | `formatRowKey` | store/key.go:18 | TODO in store/key.go:15 |

Re-evaluation table (for follow-up reviews)

| Finding | Status | What Changed |
|---------|--------|--------------|
| ~~1. Description~~ | FIXED | Brief explanation of the fix |
| 2. Description | Still applicable | No changes |

Guidelines

• Search the organization's codebase (Sourcegraph, GitHub) for existing patterns before recommending new dependencies or approaches.
• Cross-reference findings between reviews to avoid duplicate entries in the consolidated tables.
• Include effort estimates to help prioritize implementation.
• When the user asks for a follow-up review, find the most recent review directory containing SUMMARY.md. Primary lookup: ls -d .reviews/*/ 2>/dev/null | sort | tail -1. Legacy fallback if empty: ls -d reviews/*/ 2>/dev/null | sort | tail -1. Re-evaluate all prior findings against the current code state, and update with the re-evaluation table appended.
• Findings must cite probed evidence (path:line, grep output, command result), not pattern-matched suspicion. Per ~/.claude/rules/probe-not-assume.md.