# R-dev-review > Multi-domain code review (agents + Conventional Comments → findings + verdict). Triggers: "dev-review" | "code review" | "review changes" | "review PR #42" | "check my code" | "review my changes" | "review this PR" | "do a code review" | "review the diff" | "look at my code" | "/R-dev-review". Not the host natives /review or /doctor. - Skill: `roxabi/r-dev-review` (Agent Skill, multi-file: 14 files) - Install (CLI): `npx skillmds@latest add roxabi/r-dev-review` - Raw SKILL.md: https://api.skillmd.com/api/skills/roxabi/r-dev-review/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: Roxabi (https://skillmd.com/u/roxabi) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/roxabi/r-dev-review --- # Code Review ## Success I := F collected ∧ verdict posted (PR ∃) ∧ Phase 8 decision made V := `gh pr view {N} --comments | grep "## Code Review"` ∧ verdict ∈ {Approve, Request changes} Review branch/PR via fresh domain-specific agents → Conventional Comments → findings + verdict. **⚠ Flow: single continuous pipeline (Phases 1→4 + 6 + 8). ¬stop between phases. Decision response → immediately execute next phase. Stop only on: |Δ|=0, explicit Cancel, roster oracle review_halt, or Phase 8 completion.** ``` /R-dev-review → diff origin/${BASE}...HEAD (BASE = staging|main|master, first that exists) /R-dev-review #42 → gh pr diff 42 ``` Let: F := set of all findings | f ∈ F := single finding C(f) ∈ [0,100] ∩ ℤ — confidence | cat(f) ∈ {issue, suggestion, todo, nitpick, thought, question, praise} Δ := changed files | BASE := staging ∨ main τ := tier (S | F-lite | F-full) Q := present choice, wait for user reply **Stack:** Read `.dev/stack.yml` first — every `{field}` placeholder below resolves from it. ¬∃ → output: "`.dev/stack.yml` not found — run `/R-env-setup` to generate it." and stop. ## Pipeline | Phase | ID | Required | Verifies via | Notes | |-------|----|----------|---------------|-------| | 1 | gather-changes | ✓ | Δ listed | — | | 1.5 | secret-scan | ✓ | ∅ matches (or ACK) | — | | 2 | spec-compliance | — | criteria checked | spec ∃ | | 3 | multi-domain-review | ✓ | agents return | parallel · roster oracle | | 4 | merge-and-present | ✓ | F + verdict | dedup → classify → keep/drop → verdict | | 6 | post-to-pr | — | comment posted | PR ∃ | | 8 | next-step | ✓ | decision made | — | ## Pre-flight Success: F collected ∧ verdict posted ∧ Phase 8 decision made Evidence: `gh pr view {N} --comments | grep "## Code Review"` Steps: gather-changes → secret-scan → multi-domain-review → merge-and-present → post-to-pr → next-step ¬clear → STOP + ask: "Which branch/PR to review?" ## Phase 1 — Gather Changes 0. `BASE=$(. "${CLAUDE_SKILL_DIR}/../shared/lib.sh" && detect_base_branch)` 1. PR# → `gh pr diff <#>` | else → `git diff origin/${BASE}...HEAD` 2. Δ = `git diff --name-only origin/${BASE}...HEAD` (or `gh pr diff <#> --name-only`) 3. ∀ f ∈ Δ: read full (skip binaries, note) 4. |Δ| = 0 → halt 5. |Δ| > 50 → warn, suggest split ## Phase 1.5 — Secret Scan ```bash git diff origin/${BASE}...HEAD | grep -iE '(password|passwd|secret|api[_-]?key|auth[_-]?token|access[_-]?token|private[_-]?key)\s*[:=]\s*["\x27`][^"\x27`]{8,}' | head -20 ``` ∃ matches → WARN (redact to first 2 + last 2 chars): ``` ⚠️ Potential secrets found in diff — review before proceeding: : ``` → present choice **Review and proceed** | **Abort** ∅ → continue silently. ## Phase 2 — Spec Compliance **Resolve issue + spec (deterministic — #419):** 1. **issue_num** — priority: - `/R-dev-review #PR` → `gh pr view PR --json body,headRefName` → `(Fixes|Closes|Resolves) #(\d+)` in body; else first `N` from `feat/{N}-*` in `headRefName` - else current branch → `feat/{N}-*` match; else first `\d+` run (warn: legacy branch fallback) 2. **spec** ← lexicographically first `artifacts/specs/{issue_num}-*.md(x)` when issue_num set 3. **Approved σ only** — read frontmatter; `status: draft` → treat as spec ∄ for claim spawn (warn once); path-only roster 4. spec ∃ ∧ approved → ∀ criterion: met → ∅ | ¬met → `issue(blocking):` | ∀ met → `praise:` 4a. **Retain Σ (review-output display input, ¬a finding source):** Σ := [{criterion_text, verdict ∈ {met, missing}}] ∀ criterion — a **mirror of step 4**, same binary call, ¬a second judgement. `criterion_text` is the σ line already read in step 4 (verbatim, trimmed); `missing` ⟺ step 4 emitted `issue(blocking):` for that criterion. ¬`ac_id` (no AC-numbering scheme exists — `/R-spec` specifies binary criteria, ¬ids; a positional id would be fabricated), ¬`partial`, ¬scope-creep set: none has a producer in steps 1–4, and an unproduced row is an invented one. Σ carries ¬label, ¬C, ¬class: it adds no blocker and ¬enters F. 5. spec ∄ → skip (steps 4–5 unchanged when no spec) 6. SC→Test matrix (τ≠S): matrix ∃ in PR body → verify no silent gaps (every SC has a row), NO TEST reasons ∈ `{infra-not-wired, prompt-logic-only, ui-manual-only, out-of-scope}` enum. ¬matrix ∧ τ≠S → `issue(blocking):` missing SC→Test matrix. ## Phase 3 — Multi-Domain Review (Fresh Agents) Spawn fresh agents via Task (¬implementation context → ¬bias). ### Chunking (Slice 2 — O2) Before dispatching agents, partition Δ into chunks using the Python chunker (`${CLAUDE_SKILL_DIR}/chunker.py`). ```python # Pseudo-code — orchestrator executes this logic inline from chunker import parse_diff, chunk, compute_budget from digest import emit_all_digests, format_digest_for_agent raw_diff = ctx_window = files = parse_diff(raw_diff) budget = compute_budget(ctx_window) # 0.4 × ctx_window chunks = chunk(files, budget) # list[Chunk] digests = emit_all_digests(chunks) # list[BoundaryDigest] ``` - If `len(chunks) == 1` → single-chunk path (identical to pre-Slice-2 behaviour; all agents receive the full diff as before). - If `len(chunks) > 1` → per-chunk Lane A dispatch (see below). ### Roster oracle SOLE spawn decision for Phase 3. τ ← spec/plan frontmatter ∨ issue labels (default F-lite if unknown). `CHUNKS := |chunks|` from the chunker. **Global vs per-chunk.** Single-chunk: one `roster.sh --diff-list` (Δ = the chunk) — `computeRoster`. Multi-chunk: **one** allocate call (`--diff-list` full Δ + `--chunk-list` per chunk). `--chunk-list` count **defines** `chunks` (`--chunks` derived from `--chunk-list`; omit it; explicit disagreement → warning, file count wins). Scope check both directions: chunk path ∉ `--diff-list` ∨ in two chunks → warning; Δ path ∈ no chunk → **one aggregated warning** (that file is reviewed by nobody; ¬exit 1 — the chunker drops binaries that `--name-only` lists). Spawn exactly `chunk_agents[i]` on chunk i. `Spawned roster (this review): {agents}` uses the union. `R-adversarial` is the floor in every chunk. Removal order: candidates → COLLAPSE_ONCE → per-chunk `max_agents` → `max_agents_review` (`collapse → per-chunk max_agents` — collapse happens **before** the per-chunk cap). COLLAPSE_ONCE (`R-architect`, `R-devops`, `R-tester`, `R-axial-adr-review`) once per review (first chunk that gates them). `max_agents_review` default 0 = off; floors ¬capped (`R-adversarial` 1×/chunk, `stack:always`); forced > value → cap raised to forced count + warning. Off: COLLAPSE_ONCE bounds the four collapse roles, ¬`R-adversarial`. **Spawn exactly `chunk_agents[i]` (multi-chunk) or `agents[]` (single-chunk) from the JSON — the table below is documentation of the oracle's gates, ¬an independent decision surface.** `gates[]` carries the per-agent reason; `capped[]` = union of per-chunk capped sets (agents dropped by `max_agents`); `collapsed[]` / `capped_review[]` name review-level removals. `R-recall` and `R-finding-verifier` are ¬in `agents[]` (separate phases: `recall_eligible` / `verifier_enabled`). `R-product-lead` ¬∈ review roster — Phase 2 owns spec compliance. ```bash # spec_path from Phase 2; write Δ paths to a mktemp file (see tempfile-convention.md) REVIEW_TMP=$(mktemp -d -t "dev-core-review-delta-419-XXXXXX") trap 'rm -rf "$REVIEW_TMP"' EXIT printf '%s\n' "${DELTA_FILES[@]}" > "$REVIEW_TMP/delta.txt" # single-chunk: one computeRoster call (no --chunk-list) bash ${CLAUDE_PLUGIN_ROOT}/skills/dev-review/roster.sh \ --diff-list "$REVIEW_TMP/delta.txt" \ --tier "$TIER" \ --chunks "$CHUNKS" \ [--spec "$spec_path"] \ [--oracle-ok true|false] \ --json # multi-chunk: one allocate call — spawn chunk_agents[i] on chunk i; --chunks derived from --chunk-list printf '%s\n' "${CHUNK_I_FILES[@]}" > "$REVIEW_TMP/chunk_${i}.txt" bash ${CLAUDE_PLUGIN_ROOT}/skills/dev-review/roster.sh \ --diff-list "$REVIEW_TMP/delta.txt" \ --chunk-list "$REVIEW_TMP/chunk_0.txt" \ --chunk-list "$REVIEW_TMP/chunk_1.txt" \ --tier "$TIER" \ [--spec "$spec_path"] \ [--oracle-ok true|false] \ --json ``` After every invocation: `∀ w ∈ warnings[] → echo into the review output (¬silent)`; `review_halt: true → HALT` with the warning text. **R-tester gate (two-step):** first call ¬`--oracle-ok`. `delta_test_hit=true` in the JSON → run `bash ${CLAUDE_PLUGIN_ROOT}/skills/pr/run-falsify.sh --verify artifacts/reviews/{N}-falsify.json`, re-invoke `roster.sh` with `--oracle-ok true|false`. `delta_test_hit=false` → single call, R-tester ¬spawns. `delta_test_hit ∧ oracle_ok=missing → R-tester ¬spawns by design; the warning MUST appear in the output so the coverage gap is stated, ¬hidden`. Exit: `0` ok · `1` usage/IO error (incl. unreadable `--spec`, empty `--chunk-list`) · `2` σ priced-fence hygiene (σ has ≥1 priced fence ∧ ¬priced_claim_ok → spec-hygiene warning, emit `issue(blocking):` about the σ; ¬spawn R-security-auditor; JSON still printed on stdout). `claims` + `priced_claim_ok` are reported for the σ-hygiene finding only; they ¬gate any spawn. ### Agent dispatch | Agent | When | Focus | |-------|------|-------| | **R-adversarial** | **always** | red-team: bypass, fleet-regression, vacuous guards, assumption-kill + **OWASP lens** (secrets, injection, auth). R-security-auditor is independent when Δ ∩ auth/secrets/crypto (both may run) | | **R-security-auditor** | **`path_hit`** only — token match (path segments, ¬`\bauth\b`); covers oauth/session/jwt/login/password/rbac + `**/auth/**` | OWASP, secrets, injection, auth | | **R-tester** | `delta_test_hit ∧ oracle_ok=false` | coverage, AAA, edge cases, tautology | | **R-axial-adr-review** | ∃ axial ADR (`axial: true` ∈ `docs/architecture/adr/`) ∧ Δ ∩ {`infrastructure/`, `adapters/`, `domains/`, `stages/`} ≠ ∅ | Drift along non-primary axis (target × concern duplication) — read-only review agent (no Write/Edit/Bash tools) | | **R-frontend-dev** | `{frontend.path}` ∨ `{shared.ui}` non-empty → Δ ∩ those prefixes ≠ ∅; both empty → Δ ∩ {`.tsx`, `.jsx`, `.vue`, `.svelte`, `.css`, `.scss`} ≠ ∅ | FE patterns, components, hooks | | **R-backend-dev** | `{backend.path}` non-empty ∧ Δ ∩ that prefix ≠ ∅; empty → ¬spawn | BE patterns, API, errors | | **R-devops** | `τ=F-full ∧ Δ ∩ {scripts/, .github/, lefthook.yml, wrangler, deploy, Dockerfile} ≠ ∅` | config, deploy, infra — **the single infra agent** | | **R-architect** | `τ=F-full ∧ Δ ∩ infra = ∅` — mutually exclusive with R-devops | patterns, structure, circular deps | | **R-recall** | Phase 3b: `|chunks|>1 ∧ |Δ| > recall_min_delta` ∧ canonical class ∧ ≥3 callsites. Skip on single-chunk | class-join, uncited callsites — ¬Lane A | `max_agents` (default 4) is a **per-chunk** cap on `agents[]` (excl. R-recall / R-finding-verifier); truncated names land in `capped[]` (`capped[]` = union of per-chunk capped sets). Removal order: candidates → COLLAPSE_ONCE → per-chunk `max_agents` → `max_agents_review` (`collapse → per-chunk max_agents`). `max_agents_review` (default 0 = off) caps flattened Lane A after collapse; floors ¬capped (`R-adversarial` 1×/chunk, `stack:always`); forced > value → cap raised to forced count + warning; truncated names land in `capped_review[]`. COLLAPSE_ONCE kept in the first chunk that gates them; later drops land in `collapsed[]`. > **Note on R-axial-adr-review asymmetry (intentional):** The `/R-dev-review` condition is **structural** — it triggers when the diff touches `infrastructure/`, `adapters/`, `domains/`, or `stages/`. The spec phase (`/R-spec`) uses a **semantic/intent-based** condition (spec adds adapter/integration/target ∨ touches `infrastructure/`). The two are complementary: `/R-spec` catches intent-level N×M violations, `/R-dev-review` catches implementation-level ones. See `plugins/shared/references/axial-decomposition.md`. Skip: R-tester → ¬`delta_test_hit` ∨ `oracle_ok=true` ∨ `oracle_ok=missing` | R-frontend-dev → ¬FE Δ | R-backend-dev → `{backend.path}` empty ∨ Δ misses prefix | R-devops → τ≠F-full ∨ Δ ∩ infra = ∅ | R-architect → τ≠F-full ∨ Δ ∩ infra ≠ ∅ | R-axial-adr-review → ¬∃ axial ADR ∨ Δ misses AXIAL | R-security-auditor → **`¬spawn_security_auditor`** | R-recall → single-chunk ∨ |Δ| ≤ recall_min_delta ∨ ¬canonical class ∨ |callsites|<3 **Subdomain split (multi-chunk):** one allocate call (`--diff-list` full Δ + `--chunk-list` per chunk). `--chunks` derived from `--chunk-list` (omit `--chunks`; disagreement → warning, file count wins). Empty `--chunk-list` file → exit 1. Scope check both directions: chunk path ∉ `--diff-list` ∨ in two chunks → warning; Δ path ∈ no chunk → **one aggregated warning** (that file is reviewed by nobody; ¬exit 1 — the chunker drops binaries that `--name-only` lists). Spawn exactly `chunk_agents[i]` on `c_i.files`. `Spawned roster (this review): {agents}` = union. COLLAPSE_ONCE once per review. `max_agents` per-chunk; `max_agents_review` default 0 = off (floors ¬capped). R-recall is Phase 3b (not per-chunk Lane A). R-finding-verifier is Phase 4 (once per review). ### R-security-auditor scoping Only when R-security-auditor is actually spawned (`spawn_security_auditor` from the roster oracle — `path_hit`, ¬default): 1. ∀ f ∈ Δ: imports(f) = static `from '...'` ∪ dynamic `import('...')` 2. Resolve aliases: | Pattern | Resolution | |---------|-----------| | `./`, `../` | relative, try `.ts`, `/index.ts` | | `@repo/` | → `packages//src/index.ts` (skip vitest/playwright config) | | `@/*` | → `{frontend.path}/src/` + rest, try `.ts`, `.tsx`, `/index.{ts,tsx}` | | External | skip | 3. scope = Δ ∪ ⋃{resolve(imports(f)) | f ∈ Δ} ∪ `{backend.path}/src/auth/**` — deduplicate # SYNC REQUIRED: inline class list must match review-classes.yml slugs — see #149 # CROSS-SKILL CONSUMER: fix/SKILL.md Phase 0 reads this YAML via ${CLAUDE_PLUGIN_ROOT}/skills/dev-review/review-classes.yml — moving/renaming it breaks /R-fix (#286) ### Spawn template > **Note (orchestrator):** The `{format_digest_for_agent(d) for d in digests if d.chunk_index != i}` placeholder is a Python expression evaluated by the orchestrator (Claude main context) BEFORE the Task call — substitute its rendered value into the prompt string. It is NOT a runtime-resolved placeholder. All other `{...}` placeholders are simple value substitutions. **Single-chunk (|chunks| = 1):** agents receive full diff. **Use the same `Task()` template below** with `i=1, N=1` and Δ as the chunk — it is the only spawn carrier, so the `¬spawn` / `¬invoke` clause reaches single-chunk agents too (dual-use members — `R-tester`, `R-frontend-dev`, `R-backend-dev`, `R-devops`, `R-architect` — carry ¬spawn **only** via this prompt: their bodies serve `/R-dev-implement` where spawning is legitimate). At `N=1` drop the two chunk-relative sections: omit `---BOUNDARY DIGESTS---` (the set is empty) and replace `You are reviewing chunk 1 of 1. Review ONLY the files in this chunk.` with `You are reviewing the full diff.` R-adversarial prompt still includes the OWASP lens. ¬spawn R-recall (Phase 3b skipped). **Multi-chunk (|chunks| > 1) — Lane A per-chunk:** For each chunk `c_i`, spawn that chunk's `agents[]` (per-chunk oracle) in parallel: ``` Task( subagent_type: "dev-core:{agent}", description: "{agent} review — chunk {i}/{N} — {PR#|branch}", prompt: "Code review task. Focus: {focus}.\n\nSpawned roster (this review): {agents[]}. Sibling-drop rules key off THIS list — a concern whose owner is ¬in the list is YOURS: keep the finding. If you are R-adversarial: also apply an OWASP lens (secrets, injection, auth); the default panel is R-adversarial alone, so spec-scope, structure and coverage φ are yours unless the roster names R-architect/R-tester (R-product-lead is ¬in the roster at all — Phase 2 owns spec compliance). Output Conventional Comments findings only. ¬TaskCreate. ¬spawn agents (¬Task, ¬Skill). ¬invoke /R-dev-review. Review your assigned scope yourself.\n\nYou are reviewing chunk {i} of {N}. Review ONLY the files in this chunk.\n\nAdditionally audit each chunk against the systematic blind spots in `${CLAUDE_PLUGIN_ROOT}/skills/dev-review/review-blind-spots.md` — call out each applicable one explicitly (or note none apply).\n\nFormat per finding:\n