メニュー
[Debugging] Use when analyzing or optimizing performance bottlenecks: database queries, N+1 fan-out, indexing, API latency, memory, concurrency, algorithmic complexity (O(n²)), frontend rendering, caching, and distributed paths.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/duc01226/EasyPlatform --skill performance-reviewコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
[Architecture] Use when designing solution architecture across backend, frontend, deployment, monitoring, testing, and code quality.
[Utilities] Use when you need to answer technical and architectural questions.
[Content] Use when you need to brainstorm as a PO/BA — structured ideation for problem-solving, new product creation, or feature enhancement.
SOC 職業分類に基づく
| name | performance-review |
| version | 3.3.0 |
| description | [Debugging] Use when analyzing or optimizing performance bottlenecks: database queries, N+1 fan-out, indexing, API latency, memory, concurrency, algorithmic complexity (O(n²)), frontend rendering, caching, and distributed paths. |
[IMPORTANT] MANDATORY MUST ATTENTION stay project-generic: discover local stack, conventions, query APIs, index definitions, metrics, and report paths before judging. [IMPORTANT] MANDATORY MUST ATTENTION prove every performance claim with measurement or static evidence:
file:line, query text/shape, row counts, query plan/explain output, trace, profile, or logs. [IMPORTANT] MANDATORY MUST ATTENTION review performance one dimension at a time: over-fetching, filters, indexes, N+1 fan-out, batching, aggregation/join shape, materialization, writes, caching, in-process compute/algorithmic complexity, concurrency/pool saturation. [IMPORTANT] MANDATORY MUST ATTENTION include in-process compute, not just I/O: flag O(n²)+ nested scans, linear membership lookups inside loops, ReDoS-prone regex, and per-iteration serialize/clone — CPU bottlenecks need the same evidence rigor as queries. [IMPORTANT] MANDATORY MUST ATTENTION when an operation is fast but p95/p99 is high, suspect saturation not the query: measure pool/thread acquire-wait and queue depth, and size pools by Little's Law (in-use = arrival-rate × hold-time) × replica count.
Critical Thinking Mindset — Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination: Never present guess as fact — cite sources for every claim, admit uncertainty freely, self-check output for errors, cross-reference independently, stay skeptical of own confidence — certainty without evidence root of all hallucination.
AI Mistake Prevention — Failure modes to avoid on every task:
Re-read files after context changes. Context compaction, resume, or long-running work can make memory stale; verify current files before acting. Verify generated content against source evidence. AI hallucinates APIs, names, claims, and document facts. Check the relevant source before documenting or referencing. Check downstream references before deleting or renaming. Removing an artifact can stale docs, generated mirrors, configs, and callers; map references first. Trace the full impact chain after edits. Changing a definition can miss derived outputs and consumers. Follow the affected chain before declaring done. Verify ALL affected outputs, not just the first. One green check is not all green checks; validate every output surface the change can affect. Assume existing values are intentional — ask WHY before changing. Before changing a constant, limit, flag, wording, or pattern, read nearby context and history. Surface ambiguity before acting — don't pick silently. Multiple valid interpretations require an explicit question or stated assumption with risk. Keep shared guidance role-relevant. Universal guidance must help every receiving skill or agent; code-specific obligations belong only in code-specific protocols.
[BLOCKING] Execute skill steps in declared order. NEVER skip, reorder, or merge steps without explicit user approval. [BLOCKING] Before each step/sub-skill call, update task tracking: set
in_progresswhen step starts,completedwhen step ends. [BLOCKING] Every completed/skipped step MUST include brief evidence or explicit skip reason. [BLOCKING] If task tools unavailable, maintain equivalent step-by-step tracker with synchronized statuses.
Goal: Ensure every shipped performance fix removes a measured (or static-risk-labeled) real bottleneck — across database waste (rows/columns, missing/unused indexes, query-in-loop fan-out, unbounded materialization, slow joins/aggregations, write amplification), in-process compute (O(n²) scans, wrong data structures, ReDoS, serialize/clone churn), and concurrency saturation (pool/queue acquire-wait sized by Little's Law) — while preserving behavior, authorization, and semantics, proven by before/after evidence, validated via /why-review before any fix, and confirmed by a clean full Phase-0 re-review — never a guess-driven change that hides waste or breaks correctness.
Summary:
static risk with the exact verify command — never recommend below 60% confidence./why-review --validate-findings confirms it (Phase 6); each validated fix then restarts the FULL review from Phase 0 over the whole target (Phase 7) — a targeted before/after check alone never earns a PASS.Renamed: formerly
/performance— that name no longer resolves as a slash command; use/performance-review.
Workflow:
/why-review --validate-findings <report-path> before any fix.Key Rules:
/why-review --validate-findings confirms them; every validated fix restarts the full performance review from Phase 0.$ARGUMENTS
Classify before analysis. Detection drives dimensions, evidence, sub-agent choice.
| Scope | Signals | Primary evidence |
|---|---|---|
| DB read | slow query, full scan, sort spill, high rows examined | query text/ORM expression, row count, plan/explain, indexes |
| DB write | slow save, lock waits, per-row updates, transaction bloat | write loop, batch size, lock/deadlock logs, transaction scope |
| N+1/fan-out | loop with query/API call, lazy loading, per-item lookup | caller trace, query count, loop source |
| API latency | high p95/p99, timeout, slow endpoint/job | trace/profile/logs, call chain |
| Saturation/Queueing | high p99 while the operation itself is fast, pool exhausted/timeout, threads blocked on acquire | pool active/idle/pending, acquire-wait time, threads/workers vs pool size, replica count × pool |
| Memory/OOM | large materialization, blobs, no paging, buffering | allocation profile, result size, collection loads |
| Frontend | slow render, huge bundle, repeated fetch, DOM churn | browser profile, network waterfall, component/render trace |
| Distributed | message lag, cross-service waterfall, retry storm | trace spans, queue metrics, consumer/producer chain |
| Compute/CPU | hot loop, nested iteration, quadratic scaling, regex stall, heavy serialize/clone | input N, operation count vs N, profiler/flame-graph sample, microbench |
Skip reason allowed only when target explicitly narrows scope and evidence proves dimension irrelevant.
When to apply: design/architecture reviews (e.g. via the
architectagent) — judging performance as a structural property of the design before it ships, not a tactical query fix after a bottleneck is observed. The dimension passes below stay the tool for tactical work; this section is the design-level lens layered on top.
Evaluate the bottleneck layer model as a design concern, not just a symptom site:
Performance as architecture
├── Database — data access shape baked into the model (projection, paging, N+1 surface, index strategy)
├── API — serialization/processing cost, parallel tuple queries, response-DTO contracts
├── Network — payload size & call-count designed into the contract (batch endpoints vs chatty waterfalls)
├── Frontend — bundle/lazy-load topology, OnPush/track-by/virtual-scroll as default architecture
└── Background jobs — bounded parallelism (`ParallelAsync` with `maxConcurrent`) + batch (`UpdateManyAsync`) as the shape, not an afterthought
Architecture-altitude rules (decide at design time — cheapest to fix here):
SELECT * endpoint; unbounded reads spike memory and latency under real data volume..Result; bounded parallelism for fan-out work is part of the design, with a fresh safe scope/context per worker.For database index strategy at design time, see the database-optimization skill (composite key order, covering/partial indexes, write-cost analysis). The tactical evidence gate (measure baseline, prove with plan/explain) in the phases below still applies to every recommendation made at this altitude.
MANDATORY discovery before findings (MUST ATTENTION):
performance, index, query, pagination, projection, database, profiling, contributing, style guide..code-graph/graph.db exists, run a graph blast-radius pass (trace --direction downstream on the hot path) to size the fan-out before proposing a fix — see the Graph-Assisted Investigation gate below.Prefer runtime proof. If unavailable, label finding static risk and include exact command/query needed to verify.
MANDATORY baseline for DB findings:
file:line and generated SQL/query/ORM expression when availableMANDATORY baseline for compute/CPU findings:
file:lineconsole.time / profiler or flame-graph sample at representative AND worst-case NMANDATORY baseline for saturation/pooling findings:
Confidence:
| Confidence | Action |
|---|---|
| 95%+ | Recommend fix freely. |
| 80-94% | Recommend with caveats and verification command. |
| 60-79% | List unknowns first; gather more evidence before fix. |
| <60% | STOP. Do not recommend. |
MANDATORY apply one focused pass per dimension. NEVER scan all dimensions at once.
Think: Which rows/columns load? Are filters, projection, sorting, and limits executed by data source before materialization?
MUST ATTENTION find:
ToList/array/load-all before Where/filter)Prefer fixes: push predicates to data source, select only needed fields, bound result set, use cursor/keyset for deep sequential access, keep reusable predicates near domain/query-owner layer discovered locally.
Think: Can existing indexes satisfy equality/range filters, joins, sort, grouping, and projection in the actual query order?
Find:
OR, negative predicate, or low selectivityPrefer fixes: add/adjust smallest useful index, reorder composite keys to match query, rewrite predicate to be sargable, verify with plan/explain before/after, include write-cost risk.
Think: Does work scale with item count instead of request/job count?
Find:
IN/batch/group queryPrefer fixes: batch IDs once, join/include only needed fields, prefetch dictionaries, aggregate counts in one query, use bounded concurrency, preserve ordering/authorization semantics.
Think: Does the pipeline reduce data before expensive join/unwind/group/sort/window stages?
Find:
Prefer fixes: filter early, project early, aggregate at source, reduce join cardinality, use existence/count queries, repeat necessary post-expansion filters when array/child semantics require it.
Think: What enters memory? Is it bounded, streamed, and tracking-free when read-only?
Find:
Prefer fixes: page/chunk/stream, use no-tracking/read-only mode when local stack supports it, project lightweight DTOs, move filter before load, memoize intentionally.
Think: Does write work batch safely and keep locks/transactions small?
Find:
Prefer fixes: bulk write, chunk, shorten transaction, move remote calls outside transaction, use idempotent commands, create fresh safe scope/context per parallel worker.
Think: Is repeated expensive work stable, safe to reuse, and invalidated correctly?
Find:
Prefer fixes: request-scope memoization first, then bounded shared cache with explicit key, TTL/invalidation, size limits, privacy constraints, and hit/miss metrics.
Think: Is slow work caused by network waterfall, payload size, render churn, or background fan-out?
Find:
Prefer fixes: batch API, reduce payload, add backpressure, virtualize large lists, stabilize render keys, lazy-load cold assets/routes, measure browser/network trace.
Think: Does in-process work grow super-linearly with input size, independent of any query or network call?
MUST ATTENTION find:
map inside map, repeated full re-scan.find/.includes/.indexOf/in list/.contains where a Set/Map/dict gives O(1).filter().length for existenceJSON.parse/stringify/deep-clone/serialize per iteration/security-review)Prefer fixes: build a Set/Map/dict index once and look up in O(1); hoist invariant work out of the loop; accumulate into an array + single join instead of +=; precompute/memoize stable pure results; anchor/bound regex and cap input length; pick the data structure that matches the access pattern. Prove with a microbench/profiler sample at representative AND worst-case N — never reasoning alone.
Finding format:
- [Severity] [file:line] [dimension] Problem. Evidence: metric/plan/query count. Impact: user/system effect. Fix: smallest behavior-preserving change. Verify: command/query/metric.
Severity:
NEVER inflate severity without production-like scale/frequency evidence.
Before code changes (MUST ATTENTION):
Spec-Loop Discipline (Dual-Feedback half — tailored). Performance is orthogonal to functional correctness, so the property/metamorphic generation and the MUTATION-SCORE assertion gate are scoped to functional core-logic and do NOT apply here — N/A. Apply only the dual-feedback half: when a finding establishes or moves a behavior-defining boundary — an SLA/latency budget (p95/p99 target), a result-set bound, a max-rows/page-size limit, a pool-size assumption — feed it BOTH (a) the spec — record the SLA/limit as a §5 invariant / documented constraint so the budget is intended contract, not an undocumented tuning value — AND (b) a guarding test — a benchmark/assertion that fails when the budget or bound regresses. A fix that improves the number but leaves the boundary undocumented OR unguarded is INCOMPLETE, never a code-only change.
Use specialized help when available:
| Detected focus | Sub-agent |
|---|---|
| DB/query/N+1/memory/backend hot path | performance-optimizer |
| Auth, PII, tenant isolation, sensitive cache keys | security-auditor first, then performance-optimizer |
| Cross-service architecture, caching policy, capacity/SLO trade-off | architecture/performance specialist |
| Frontend render/bundle/network waterfall | frontend or performance specialist |
Sub-agent prompt MUST include target, detected scope, local context evidence, required dimensions, report path, and "return summary only; write full report incrementally."
Purpose: Validate performance findings before optimization work. Performance reports are easy to overstate when evidence is static-only, a plan lacks production-like scale, or a proposed index/cache changes write cost or data freshness risk.
Trigger: Any performance finding or optimization recommendation (Critical, High, Medium, Low, WARN, or static risk). Skip ONLY when the report's verdict is unconditional PASS with literally zero findings.
Protocol:
plans/reports/performance-{date}-{slug}.md or the exact report path written by the caller./why-review --validate-findings <performance-report-path>.plans/reports/why-review-validate-{date}.md.## Why-Review Validation Notes section.## Why-Review Validation stating all findings were re-validated against measurement/static evidence.Skip conditions (record explicit reason if skipping):
Trigger: Phase 6 returns CLEAN/validated and the performance report still has one or more findings that must be fixed.
Protocol:
/why-review --validate-findings; if this skill is running inside a workflow, route implementation through the parent /plan + /feature-implement flow./performance-review review from Phase 0 over the complete current target, not only the fixed files.Non-negotiable rules:
/why-review --validate-findings validates it.MANDATORY final report sections:
If evidence insufficient, output: Insufficient evidence. Verified: [...]. Not verified: [...]. Next evidence needed: [...].
Graph-Assisted Investigation — MANDATORY when
.code-graph/graph.dbexists.HARD-GATE: MUST ATTENTION run at least ONE graph command on key files before concluding any investigation.
Pattern: Grep finds files →
trace --direction bothreveals full system flow → Grep verifies details
Task Minimum Graph Action Investigation/Scout trace --direction bothon 2-3 entry filesFix/Debug callers_ofon buggy function +tests_forFeature/Enhancement connectionson files to be modifiedCode Review tests_foron changed functionsBlast Radius trace --direction downstreamCLI:
python .claude/scripts/code_graph {command} --json. Use--node-mode filefirst (10-30x less noise), then--node-mode functionfor detail.
Severity Rubric — Classify every finding by consequence, not by how easy it is to fix. One scale across all reviews so a "High" means the same thing everywhere.
Severity Action Definition CRITICAL Block merge Silent runtime failure, data corruption, validation bypass, security hole HIGH Must fix Incorrect behavior, invariant gap, architectural violation MEDIUM Should fix Design debt, maintainability, likely future bug LOW Nice to fix Convention, documentation, minor clarity Score-based skills map their numeric scale onto these tiers — do not invent a parallel vocabulary:
- 0-2 criterion scoring (e.g. production-readiness-review):
0= CRITICAL/HIGH (criterion unmet, blocks production readiness),1= MEDIUM (partial, should fix),2= pass (no finding).- Two-axis scoring (e.g. performance-review, impact × likelihood): map the resulting cell to the nearest tier — high-impact + high-likelihood → CRITICAL/HIGH; low-impact OR low-likelihood → MEDIUM/LOW.
A finding's tier drives the gate: CRITICAL/HIGH must be resolved or explicitly accepted by the owner before PASS; MEDIUM/LOW may ship with a tracked follow-up.
Category Review Thinking — A thinking framework for reviewing any category of changed files. NOT a fixed checklist — derive concerns from domain knowledge; the examples are starting points only. Your knowledge of the category exceeds any list here — trust it.
Step 1 — Understand the category's role. What is this category responsible for in the overall system? What invariants must it uphold? What are its consumer contracts (who depends on it, what do they expect)?
Step 2 — Read project conventions for this category. Search for reference docs, style guides, ADRs, or READMEs specific to this area. Grep 3+ existing similar files — extract naming conventions, structural patterns, shared base classes. If no docs exist, derive conventions empirically from existing code.
Step 3 — Derive concerns from first principles. Apply all that are relevant; expand beyond this list based on the actual category:
- Correctness: Does the logic match the intent? Trace happy path AND error path.
- Boundary contracts: Are interfaces/APIs/events/protocols honored? No implicit coupling introduced?
- Project conventions: Does new code follow the patterns found in Step 2? Evidence-confirmed, not assumed.
- Security: Auth enforced at every entry point? Input validated at boundaries? No secrets in the diff?
- Performance: Unbounded operations? N+1 patterns? Blocking calls in async context? Unindexed queries?
- Maintainability: DRY? Single responsibility? Complexity within reason? Names reveal intent?
- Test coverage: Are the changed paths covered by tests? Are existing tests still valid after the change?
- Documentation: Do related docs, specs, or READMEs reflect the changes?
Step 4 — Create sub-tasks and execute. For each identified concern: create a
TaskCreatesub-task, work through it withfile:lineevidence, mark done. No findings without proof.Illustrative concern examples by category type (not exhaustive — trust your knowledge beyond this):
- Server-side logic: handler/service structure conventions, validation layer placement, side-effect isolation, cross-service boundary enforcement, data-access layer separation, error propagation strategy
- Client-side logic: component lifecycle management, resource cleanup (subscriptions, listeners, timers), state management patterns, API integration layer separation, reactive stream composition
- Data/Schema: migration reversibility (rollback script), lock impact on table volume, backfill idempotency, index coverage for query patterns, deployment ordering
- Configuration: present in ALL environments? No secrets in diff? App fails fast if config missing (not silently null)? Documented in setup guide?
- Infrastructure: dev/prod parity? No hardcoded dev values (localhost, debug flags)? Pinned image/dependency versions? CI/CD secret requirements documented?
- Styles/Assets: follows project naming conventions? Uses design variables/tokens (no hardcoded magic values)? Correct scope (no global side effects from component styles)?
- Documentation: accurate? Links valid? Examples still match current code/behavior? Covers new scenarios?
- Tests: assertions verify specific outcomes (not just "no exception")? Idempotent (repeatable N times)? Covers edge cases, not just happy path?
- Security artifacts: all code paths reach the gate? Negative tests exist (unauthorized denied)? Both enforcement AND display control updated?
- Build/Tooling: rule changes apply consistently? No exceptions that silently swallow violations? Impact on CI runtime documented?
file:line evidence — never a fixed checklist.IMPORTANT MUST ATTENTION Goal: Ensure every shipped performance fix removes a measured (or static-risk-labeled) bottleneck while preserving behavior, authorization, and semantics — proven by before/after evidence, validated via /why-review before any fix, and confirmed by a clean full Phase-0 re-review — never a guess-driven change that hides waste or breaks correctness.
Protocols in force (concise digest of the SYNC/shared blocks this skill carries):
file:line proof per claim; NEVER present a guess as fact.graph.db exists.IMPORTANT MUST ATTENTION prove every performance claim with measurement or static evidence — file:line, query text/shape, row counts, query plan/explain, trace, profile, or logs; confidence >80% to act, 60-79% gather more, <60% STOP — why: a number without a measured baseline is a guess that ships unverified waste.
IMPORTANT MUST ATTENTION review performance one dimension at a time — over-fetching, filters, indexes, N+1 fan-out, batching, aggregation/join shape, materialization, writes, caching, in-process compute/algorithmic complexity, concurrency/pool saturation — why: split attention misses violations.
MANDATORY search 3+ similar local query/API patterns before proposing a fix, and read the index/migration/schema files controlling the data — why: local conventions override generic framework defaults; the closest example must match preconditions (base class, scope, cardinality) before you copy it.
MANDATORY ALWAYS measure before/after; static review findings need an explicit verification command attached.
MANDATORY ALWAYS verify index usability with actual query shape/order and plan/explain — index existence alone is not proof.
IMPORTANT MANDATORY MUST ATTENTION ALWAYS push row filters to the data source before projection/caching; row-count reduction beats column trimming — why: fewer columns from too many rows still scans the rows.
MANDATORY size pools/parallelism by Little's Law (in-use = arrival-rate × hold-time) × replica count, and shrink hold-time before growing the pool — why: a fast op with high p99 is saturation at the pool entrance, not a slow query.
MANDATORY Break work into small tracked tasks before starting; one in_progress at a time; mark each completed immediately after its evidence lands — why: compaction wipes memory and untracked review scope silently goes uncovered.
MANDATORY when a finding moves a behavior-defining boundary (SLA/p95 budget, result-set bound, page-size limit, pool-size assumption), feed it BOTH the spec (record as a §5 invariant) AND a guarding test/benchmark — why: a faster number left undocumented OR unguarded regresses silently.
MANDATORY add a final review task checking doc/test/spec staleness.
Anti-Rationalization:
| Evasion | Rebuttal |
|---|---|
| "Bottleneck obvious, skip baseline" | No measurement = guess. Capture metric or label static risk with the verify command. |
| "Index exists, so query fine" | Show plan/explain and access path. Existing unused index proves nothing. |
| "Projection enough" | First reduce rows. Loading fewer columns from too many rows still wastes work. |
| "Just cache it" | Fix query shape/index/bounds first. Cache can hide stale, unsafe, unbounded work. |
| "Only one query in code" | Trace loops, serializers, resolvers, consumers, and retries. Fan-out often hides upstream. |
| "Loop is fine, the list is small" | Show N and worst-case N. O(n²) that's fine at 10 melts at 10k. Bench at real scale. |
| "Query is fast, so the endpoint is fast" | Measure pool acquire-wait and queue depth. A 2ms query behind a saturated pool still yields a 200ms p99 — the wait is at the pool entrance, not in the query. |
| "Found one similar pattern, good enough" | Grep 3+ and verify preconditions match. One nearby example ≠ a fit; cite file:line. |
| "Fix it where it errors/spikes" | Trace caller (wrong data) vs callee (wrong handling); fix at the layer owning the invariant, not the symptom site. |
| "Validated nothing, just fix the obvious one" | No fix until /why-review --validate-findings confirms it; then restart the FULL review from Phase 0. |
[TASK-PLANNING] Break work into small tracked tasks before starting; update each status immediately.
IMPORTANT MUST ATTENTION prove every claim with measurement/static evidence + file:line (confidence >80% to act, <60% STOP).
IMPORTANT MUST ATTENTION push row filters to the data source before projection/caching; verify index usability via plan/explain, never existence alone.
IMPORTANT MUST ATTENTION no fix before /why-review --validate-findings; after every validated fix restart the full review from Phase 0 before claiming PASS.
Systematic Review Batching (map-reduce) — When a changeset is large, do NOT review files one-by-one. Partition into size-capped batches, fire one specialized sub-agent per batch in parallel, then reduce. This bounds EVERY context — each batch agent AND the orchestrator — so coverage stays complete as file count grows.
Trigger ladder (one ordered escalation — not competing thresholds):
- < 10 changed files → sequential per-file review (default; no batching).
- ≥ 10 changed files → switch to systematic parallel mode. Announce:
"Detected {N} changed files. Switching to systematic parallel review protocol."Then: categorize → size-capped batches → flat consolidation.- categories > 6 OR files > 40 → additionally insert the hierarchical synthesis tier (below). Everything from rung 2 still applies.
Step 1 — Categorize. Group changed files into logical categories derived from the project's actual structure (not forced). Category is the concern axis; orient with these examples, derive what fits the repository:
Category Type Example Groupings Agent/Tooling AI scripts, hooks, skill definitions, workflow configs, linting rules Root config/docs Root README, project config, CI/CD pipeline configs Reference docs Architecture docs, patterns references, setup guides Feature/domain docs Business feature documentation, spec files, ADRs Backend logic Service/handler/controller source (infer from project structure) Frontend logic UI component/state/API source (infer from project structure) Data/Schema Migrations, schema files, seed data Tests Unit, integration, E2E test files Infrastructure Docker, k8s, CI/CD, cloud manifests Step 2 — Size-capped batches. One sub-agent per batch of ≤8 files OR ≤2000 diff-lines, whichever hits first. Category stays the concern axis, but any category exceeding a cap splits into multiple size-capped batches (30 backend files → 4 batches). Size caps — not category caps — make "many files" safe: a category cap alone lets one giant category blow a single agent's context.
Step 2a — Sub-agent type per batch (match the batch's dominant concern):
- Code logic (any stack) →
code-reviewer- Security-sensitive changes →
security-auditor- Performance-critical paths →
performance-optimizer- Docs, plans, specs, configs, infra →
general-purposeEach batch sub-agent receives: its full file list;
SYNC:category-review-thinkingas its primary thinking model — derive each category's concerns from first principles, NOT a fixed checklist (if the consuming skill does not carry that block, apply category-first thinking directly); project reference docs relevant to its concern (discover via*patterns*,*conventions*,*style-guide*); cross-reference verification instructions (counts, tables, links). All batch agents run in parallel and write findings toplans/reports/(perSYNC:task-tracking-external-report); reducers read from disk, never from memory.Step 3 — Reduce.
- Flat reduction (rung 2, ≤6 categories AND ≤40 files): the orchestrator collects each batch report, cross-references counts/tables/contracts ACROSS batches, detects gaps visible only across categories (feature in code but missing from docs; new API endpoint with no client call), and consolidates into one categorized holistic report.
- Hierarchical reduction (rung 3, > 6 categories OR > 40 files): insert a mid-tier — each concern gets ONE synthesizer agent that reads only its own batch reports and emits a single concern-synthesis. The orchestrator reads the concern-syntheses (~5), never the raw batch reports — keeping the reducer's context O(#concerns), not O(#files).
- Cross-concern interaction pass (mandatory at rung 3 — closes the synthesis-tier blind spot): concern-siloed synthesis can drop an interaction spanning two concerns AND two batches (tainted source in data-layer/batch 7 → sink in api/batch 3). So: (a) each concern-synthesizer MUST emit an explicit "cross-concern interaction candidates" list — entities/symbols/contracts it touched that plausibly bind to another concern (shared DTOs, event names, table/collection names, exported symbols); (b) the orchestrator MUST run the Step-3 cross-reference/gap step over those candidate lists across all concern-syntheses, not only within a batch, before concluding. Without this pass the tier trades completeness for context-bounding on exactly the large diffs it targets.
Step 4 — Holistic assessment. With all findings combined, judge: overall coherence as a unified intent; cross-category sync (docs match code? contracts match callers?); risk areas where categories interact; missing doc/spec updates for changed artifacts.
No silent truncation. If any cap forces sampling or a batch is dropped for budget, ANNOUNCE the dropped/sampled scope explicitly — bounded coverage must never read as complete coverage.
MUST ATTENTION apply critical + sequential thinking — every claim needs appropriate traced evidence (file:line for repo/code claims; source URL or artifact section for research, product, content, and docs claims); confidence >80% to act, <60% DO NOT recommend. Anti-hallucination: never present guess as fact, admit uncertainty freely, cross-reference independently, stay skeptical of own confidence.
MUST ATTENTION apply AI mistake prevention — verify generated content against evidence, trace downstream references before deleting or renaming, verify all affected outputs, re-read files after context loss, and surface ambiguity before acting.