performance-review

name

performance-review

description

[Debugging] Use when analyzing or optimizing performance bottlenecks: database queries, N+1 fan-out, indexing, API latency, memory/GC, concurrency and pool saturation, algorithmic complexity (O(n²)), network/protocol round trips, frontend rendering and Core Web Vitals, caching, and distributed/resilience paths. Calibration constants and domain laws (latency ladder, Little's Law, utilization knee, CWV thresholds, symptom→cause triage) live in references/performance-knowledge.md.

Codex compatibility note:

• Invoke repository skills with $skill-name in Codex; this mirrored copy rewrites legacy Claude /skill-name references.
• Task tracker mandate: BEFORE executing any workflow or skill step, create/update task tracking for all steps and keep it synchronized as progress changes.
• User-question prompts mean to ask the user directly in Codex.
• Ignore Claude-specific mode-switch instructions when they appear.
• Strict execution contract: when a user explicitly invokes a skill, execute that skill protocol as written.
• Subagent authorization: when a skill is user-invoked or AI-detected and its protocol requires subagents, that skill activation authorizes use of the required spawn_agent subagent(s) for that task.
• Do not skip, reorder, or merge protocol steps unless the user explicitly approves the deviation first.
• For workflow skills, execute each listed child-skill step explicitly and report step-by-step evidence.
• If a required step/tool cannot run in this environment, stop and ask the user before adapting.

Codex Project-Reference Loading (No Hooks)

Codex uses static project-reference loading instead of runtime-injected project docs. When coding, planning, debugging, testing, or reviewing, open project docs explicitly using this routing.

Always read:

• docs/project-config.json (project-specific paths, commands, modules, and workflow/test settings)
• docs/project-reference/docs-index-reference.md (routes to the full docs/project-reference/* catalog)
• docs/project-reference/lessons.md (always-on guardrails and anti-patterns)

Missing/stale context route: If docs/project-config.json, the docs index, lessons.md, CLAUDE.md, AGENTS.md, or any task-required reference doc is missing or stale, auto-run $project-init or the narrow setup route ($project-config, $docs-init, $scan-all, $scan --target=<key>, $claude-md-init) before ordinary project-specific work. If Codex mirrors or AGENTS.md are missing/stale, ask the user to run $sync-codex; do not auto-run it.

Situation-based docs:

• Project structure/architecture/tech-stack/deployment/setup (any layer — backend, frontend, or infra): project-structure-reference.md
• Backend/CQRS/API/domain/entity changes: backend-patterns-reference.md, domain-entities-reference.md
• Frontend/UI/styling/design-system: frontend-patterns-reference.md, scss-styling-guide.md, design-system/README.md
• Spec authoring, docs/specs/ pathing, or TC format: feature-spec-reference.md, spec-system-reference.md, spec-principles.md
• Behavior/public-contract changes or spec-test-code sync: workflow-spec-test-code-cycle-reference.md plus the spec docs above
• Derived spec indexes/ERDs/reimplementation guides: spec-system-reference.md and source Feature Specs under docs/specs/
• Integration test implementation/review: integration-test-reference.md
• E2E test implementation/review: e2e-test-reference.md
• Code review/audit work: code-review-rules.md plus domain docs above based on changed files

Do not read all docs blindly. Start from docs-index-reference.md, then open only relevant files for the task.

[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 — ALL 12: (1) query shape/over-fetching, (2) index/access path/data topology, (3) N+1 fan-out, (4) aggregation/join shape, (5) materialization/memory, (6) write path/locks/transactions, (7) caching, (8) API payload/frontend delivery/Core Web Vitals, (9) in-process compute/algorithmic complexity, (10) network/protocol round trips, (11) runtime/memory/GC pauses, (12) distributed resilience/load management (timeouts, retries, queue bounds). NEVER stop at 9 — 10-12 are the layers a code-only reading habitually never opens. [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. [IMPORTANT] MANDATORY MUST ATTENTION calibrate every number against a known anchor before assigning severity — latency ladder, utilization knee, Core Web Vitals thresholds, hit-ratio math (references/performance-knowledge.md); a breached anchor is a HYPOTHESIS to verify with local evidence, NEVER a finding on its own.

[PERFORMANCE-FIRST PRINCIPLES — three non-negotiable checks on every hot path, OOM first]

1. [MOST IMPORTANT] Hunt every OOM / out-of-memory bad practice. Unbounded read-all / SELECT * / no page bound, full materialization before paging/filtering, buffering a whole export/report instead of streaming/chunking, loading blobs / large JSON / tracked entities for list views, accidental multiple enumeration, unbounded caches / accumulators / queues / in-memory joins. Triage row COUNT before row SIZE, reduce rows AT THE SOURCE — a fast query pulling millions of rows still OOMs the process. Bound EVERY result set with a page/limit/cursor or proven business invariant.
2. Right data structure & algorithm for the stack. Match the structure to the access pattern via the runtime's efficient primitive — O(1) Set/Map/dict/hash lookup instead of a linear find/includes/contains/in list scan inside a loop; no O(n²) where O(n log n) / O(n) / O(1) exists; single-pass min/max/partition instead of redundant re-sort. Prove the complexity class at worst-case N, never by intuition.
3. Batch once, or parallelize — never serial fan-out. Collapse per-item query / API / cache calls into ONE batched call (IN / bulk / aggregate / prefetch dictionary); where independent calls remain, run bounded-parallel with a fresh safe resource per worker instead of sequential awaits — always preserving ordering, authorization, idempotency.

Performance Knowledge (calibration constants & domain laws) — the anchors severity depends on:

• Latency ladder 1 ns → 100 ns → 100 µs → 10 ms → 100 ms (L1 → RAM → SSD → disk seek → intercontinental), each rung ~100-1000×; ~1 ms RTT per 100 km of fiber is a hard floor no code fix beats.
• Utilization knee ~70-80% — queue wait ≈ service_time × ρ/(1−ρ): 80%→4×, 90%→9×, 95%→19×. Little's Law L = λ × W sizes every pool. Tail amplification — fan-out to 100 backends hits a p99 ~63% of the time, so a backend p99 becomes the user's median.
• Core Web Vitals LCP ≤2.5 s · INP ≤200 ms · CLS ≤0.1 · TTFB ≤800 ms, measured at p75 of real users (field), never a lab score alone.
• Cache hit-ratio math — 90%→99% cuts origin load 10×; percentiles are NEVER averageable.

MANDATORY MUST ATTENTION [BLOCKING at the severity/anchor moment] READ references/performance-knowledge.md — full ladder, universal laws, symptom→cause triage matrix, and deep tables for network/protocol, DB engine + isolation + sharding, caching, web/CWV, memory/GC, distributed resilience, measurement rigor. The read is REQUIRED — never optional — before you assign a severity or quote/compare any anchor constant; NEVER assign a severity or cite an anchor from memory or from the 4-bullet digest above. A scope-narrowed review that assigns no severity and quotes no constant may proceed on the digest alone. — why: the digest orders hypotheses but only the body carries the thresholds severity depends on, and quoting a constant without measuring THIS system is the guess-as-fact failure this skill exists to prevent.

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_progress when step starts, completed when 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.

Quick Summary

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, partition/shard skew), in-process compute (O(n²) scans, wrong data structures, ReDoS, serialize/clone churn), runtime cost (GC pauses, allocation pressure, blocked event loop), network round trips (handshake/keep-alive, chatty contracts, RTT floors), client delivery (Core Web Vitals, long tasks, payload/asset weight), and concurrency/resilience saturation (pool acquire-wait sized by Little's Law, timeouts, retries, unbounded queues) — every number calibrated against a known anchor, 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:

• Purpose & 8-phase pipeline (the main tasks): drive a target through Phase 0 Detect scope (+ symptom→cause triage) → Phase 1 Discover local context (grep 3+ patterns, read index/schema, map callers) → Phase 2 Baseline evidence + anchor calibration (or static risk + verify cmd) → Phase 3 twelve serial dimension passes → Phase 4 Findings + Severity → Phase 5 Optimize plan (behavior-preserving) → Phase 6 $why-review --validate-findings gate → Phase 7 validated-fix + full Phase-0 re-review — so every recommendation removes a real bottleneck, preserves behavior, is evidence-proven; an Architecture-Altitude lens applies the same gate at design time.
• Evidence is the gate, not intuition: capture a runtime baseline (query plan/explain, row counts, p95/p99 distributions, pool acquire-wait, GC pauses, call count × RTT, field CWV, microbench at worst-case N) or label the finding static risk with the exact verify command — never recommend below 60% confidence, never average percentiles, always name the load model.
• Calibrate against the anchors in references/performance-knowledge.md — latency ladder (1 ns → 100 ns → 100 µs → 10 ms → 100 ms), utilization knee ~70-80% (ρ/(1−ρ)), Little's Law, tail amplification, CWV thresholds, cache hit-ratio math — a breached anchor is a hypothesis to prove locally, NEVER a finding by itself.
• Walk dimensions ONE pass at a time — (1) query shape/data-minimization → (2) index/access-path/data-topology → (3) N+1/fan-out → (4) aggregation/join/pipeline → (5) materialization/memory → (6) write/locks/transactions → (7) cache/reuse → (8) API payload/frontend/CWV → (9) compute/algorithmic → (10) network/protocol → (11) runtime/memory/GC → (12) distributed resilience/load — never all at once; reduce rows at the source before trimming columns or caching, and size pools by Little's Law (replica count × per-instance pool) when a fast op shows high p99.
• No finding is fixable until $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:

1. Detect - Classify scope and bottleneck type; order hypotheses via the symptom→cause matrix.
2. Discover - Read local code, metrics, docs, query/index definitions, similar patterns.
3. Measure - Capture baseline against a known anchor, or mark static-only risk.
4. Analyze - Run 12 serial dimension passes with evidence.
5. Plan - Propose smallest fix preserving behavior.
6. Verify - Re-measure, run tests, and record evidence.
7. Validate Findings - Run $why-review --validate-findings <report-path> before any fix.
8. Fix + Full Re-Review - Fix only validated findings, then restart from Detect over the full target.

Key Rules:

• MANDATORY ALWAYS measure before/after; static review findings need explicit verification command.
• MANDATORY ALWAYS calibrate a number against a known anchor before assigning severity; an anchor breach alone is a hypothesis, never a finding.
• MANDATORY ALWAYS push row filters to data source before projection/caching; row-count reduction beats column trimming.
• MANDATORY ALWAYS verify index usability with query shape/order, not index existence alone.
• MANDATORY ALWAYS count call count × RTT on a remote path, and check the timeout/retry/queue-bound before optimizing inside a call.
• NEVER recommend caching until query shape, indexes, pagination, batching, and data volume are understood; NEVER call a cache done without its measured hit ratio and bound.
• NEVER average percentiles, and NEVER trust a throughput number whose load model (open vs closed) is unstated.
• Findings are not eligible for fix until $why-review --validate-findings confirms them; every validated fix restarts the full performance review from Phase 0.

$ARGUMENTS

---

Phase 0: Detect Scope

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
Network/protocol	chatty call count, per-request handshake, no keep-alive, large payload, cross-region hop	call count × RTT, connection reuse state, TLS/DNS timing, payload size, HTTP version
Runtime/GC	latency spikes uncorrelated with load, pauses, RSS growth, blocked event loop	GC log/pause histogram, allocation rate, RSS vs heap, thread states, event-loop lag
Resilience/load	retry storm, no timeout, unbounded queue, cold-start blip, one tenant degrades all	timeout/retry config, queue depth AND age, breaker state, per-tenant rate limits

Skip reason allowed only when target explicitly narrows scope and evidence proves dimension irrelevant.

Triage accelerator (symptom → usual cause). MUST ATTENTION use the symptom→cause matrix in references/performance-knowledge.md §3 to pick the FIRST evidence to pull — it maps signatures AI habitually misreads, e.g. p99 bad + p50 fine → GC pause / lock contention / fan-out tail / cold cache (NOT a slow query); latency scales with result size → N+1; sudden cliff at some load → utilization knee or pool exhaustion; degrades over days, fine after restart → leak/bloat/connection leak; slow for one tenant only → hot key/partition skew. NEVER let the matrix replace evidence — it orders the hypotheses, Phase 2 proves one.

---

Architecture-Altitude Performance Review

When to apply: design/architecture reviews (e.g. architect agent) — judge performance as a structural property of the design BEFORE it ships, not a tactical query fix after a bottleneck appears. Dimension passes stay the tactical tool; this section is the design-level lens.

Evaluate the layer model as a design concern, not a symptom site:

Performance as architecture
├── Database  — data access shape baked into the model (projection, paging, N+1 surface, index strategy, partition/shard key)
├── API       — serialization/processing cost, batched vs per-item queries, response-DTO contracts
├── Network   — payload size & call-count designed into the contract (batch endpoints vs chatty waterfalls), endpoint placement vs RTT budget
├── Frontend  — bundle/lazy-load topology, change-detection/list-keying/virtual-scroll as default architecture
├── Runtime   — allocation profile & collector choice, event-loop discipline, pool sizing, working-set target
└── Background jobs — bounded parallelism (local concurrency-limited primitive) + bulk write (local batch API) as the shape, not an afterthought

Architecture-altitude rules (decide at design time — cheapest to fix here):

• Bound every result set and project only needed columns/fields in the contract itself — never design an unbounded read-all or SELECT * endpoint; unbounded reads spike memory/latency under real data volume.
• Design out N+1 at the boundary — eager-load / batch-fetch is the default access pattern; per-item lookups are a design smell, not a tuning detail.
• Caching is a design decision, not a patch — choose request-scope memoization vs bounded shared cache up front, with key dimensions (tenant/user/auth/version), TTL/invalidation, size limits, privacy constraints specified; never cache to hide an unbounded query.
• Async I/O is structural — never design a path blocking threads with .Result; bounded parallelism for fan-out is part of the design, with a fresh safe scope/context per worker.
• Make the cost visible — design slow-operation + query logging in from the start so regressions are observable in production.
• Size pools and parallelism, never default them — derive connection/thread/permit pool size from Little's Law (in-use = arrival-rate × hold-time), state the assumptions; shrink hold-time (release the resource across non-DB / external-wait spans) before growing the pool; size a shared backend against fleet-aggregate demand (replica count × per-instance pool), not one instance — local per-instance tuning becomes a thundering herd on the shared dependency.
• Budget the round trips and the geography in the contract — count call count × RTT for every designed interaction and place the endpoint (edge/region/replica) against the latency budget; ~1 ms RTT per 100 km and a 2-RTT TCP+TLS handshake are floors no later optimization removes, so a chatty contract or a distant endpoint is a permanent design cost, not a tuning detail.
• Design the load-management controls in, not on — a decreasing timeout budget per hop, backoff + full jitter + retry budget + idempotency keys, breaker/bulkhead/shedding, and a BOUND on every queue belong in the design; leave them out and the system amplifies its own partial failures. Plan capacity below the ~70-80% utilization knee (wait ≈ service × ρ/(1−ρ)) and autoscale on a leading indicator (queue depth/concurrency), never lagging CPU.
• Choose the runtime cost profile deliberately — allocation rate and collector choice set the tail (GC pauses are correlated fleet-wide and invisible in the mean); an event-loop runtime must keep CPU work off the loop by design; state the working-set target so the RAM/page-cache cliff is a known bound, not a surprise.

DB index strategy at design time → dimension 2 below (composite key order, covering/partial indexes, write-cost analysis). The tactical evidence gate (measure baseline, prove with plan/explain) still applies to every recommendation at this altitude.

---

Phase 1: Discover Local Context

MANDATORY discovery before findings (MUST ATTENTION):

• ALWAYS search local standards: performance, index, query, pagination, projection, database, profiling, cache, timeout, retry, pool, contributing, style guide.
• search 3+ similar local query/API patterns before proposing a fix.
• read target code and index/migration/schema files controlling the queried data.
• map callers and frequency using available graph/call-trace/profiler tools; if none exist, use grep/import/call hierarchy. When .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.
• identify data shape: tenant/security-review filters, cardinality, expected max rows, selected columns/fields, sort, joins, aggregation/grouping, cache keys, partition/shard key, primary vs replica routing.
• ALWAYS discover the local SLA/budget (latency target, page-size cap, throughput/SLO) before judging any number — the local budget outranks every anchor in references/performance-knowledge.md.
• ALWAYS read the local resilience + delivery configuration the new dimensions rest on: HTTP client/keep-alive and pool settings, timeout/retry/breaker policy, queue and consumer bounds, rate limits, GC/runtime and container memory limits, CDN/asset caching headers, and whatever RUM/field-metrics source exists.
• NEVER hardcode project names, repository paths, ID formats, DB engines, ORMs, runtime/GC flags, HTTP clients, or framework defaults; derive every one from discovered files.

---

Phase 2: Baseline Evidence

Prefer runtime proof. If unavailable, label finding static risk and include exact command/query needed to verify.

MANDATORY baseline for DB findings:

• ALWAYS capture query source: file:line and generated SQL/query/ORM expression when available
• ALWAYS capture volume: input size, rows matched, rows returned, rows examined/scanned, page size/limit
• ALWAYS capture access path: query plan/explain, used index, sort/group strategy, join method when available
• ALWAYS capture timing: p50/p95/p99, elapsed query time, query count, allocation or response size
• ALWAYS capture context: endpoint/job/consumer frequency and worst-case fan-out

MANDATORY baseline for compute/CPU findings:

• ALWAYS capture input size N and the growth assumption (expected and worst-case N)
• ALWAYS capture operation count vs N (constant / linear / quadratic+) and the nested-loop or repeated-scan source file:line
• ALWAYS capture timing: microbench / console.time / profiler or flame-graph sample at representative AND worst-case N

MANDATORY baseline for saturation/pooling findings:

• ALWAYS capture offered concurrency and arrival rate (RPS / worker count / threads.max)
• ALWAYS capture resource hold-time vs total request time (a connection/lock/permit is held only for the fraction it is actually used, not the whole request)
• ALWAYS capture pool state: size, active/idle/pending, and acquire-wait time / queue depth at the pool entrance
• ALWAYS capture aggregate demand on shared dependencies: replica count × per-instance pool → total connections/cores the shared backend must serve

MANDATORY calibration + measurement rigor on EVERY baseline (references/performance-knowledge.md §1-2, §10):

• ALWAYS state which anchor the number violates (ladder rung, utilization knee, CWV threshold, hit-ratio target) — a raw number with no anchor cannot carry a severity.
• ALWAYS report distributions, never means: p50/p90/p99/p99.9 + max, segmented by endpoint/tenant/region. NEVER average percentiles across instances or windows — aggregate histograms instead.
• ALWAYS name the load model behind any throughput/latency number: open-model (arrival-rate) exposes queueing collapse, closed-model (fixed VUs) HIDES it; flag suspected coordinated omission when a tool reports an implausibly clean tail.
• ALWAYS state data volume and cache state of the measurement — a benchmark on toy data or a warm-only cache is fiction; soak/endurance is the only shape that surfaces leaks, fragmentation, and bloat.
• ALWAYS warm up (JIT + caches), measure steady state, repeat, and name the environment before comparing to a baseline; NEVER present a microbenchmark as system behavior.
• NEVER quote an anchor from the reference as a project requirement — local SLA/spec/config wins; the anchor calibrates, it does not govern.

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.

---

Phase 3: Serial Dimension Passes

MANDATORY apply one focused pass per dimension. NEVER scan all dimensions at once. 12 dimensions — 1-9 are the in-process/data-access core, 10-12 cover the layers a code-only reading habitually skips (network round trips, runtime/GC, resilience under load). references/performance-knowledge.md carries deep tables for network/protocol (§4), database (§5), caching (§6), web/CWV (§7), memory/GC (§8), and distributed resilience (§9); the remaining dimensions calibrate against the ladder, universal laws, and triage matrix (§1-3) instead of a dedicated table.

1. Query Shape And Data Minimization

Think: Which rows/columns load? Are filters, projection, sorting, and limits executed by data source before materialization?

MUST ATTENTION find:

• unbounded list/read-all APIs without page, limit, cursor, or bounded business invariant
• filter after materialization (ToList/array/load-all before Where/filter)
• projection after materialization; full entity/document loaded for list/summary view
• unused includes/joins/lookup data; large text/blob/json fields in list queries
• client-side sort/group/distinct; offset pagination on very deep pages where cursor/keyset fits better
• missing tenant/auth/status/date filters in hot-path queries

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.

2. Index, Access Path And Data Topology

Think: Can existing indexes satisfy equality/range filters, joins, sort, grouping, and projection in the actual query order? Sargability first: for EVERY filter/join predicate, is the indexed COLUMN left bare, or is it wrapped in a function/transformation that the DB must compute per row (killing the index)? Then: does the query reach the data through the right partition/shard/replica?

MUST ATTENTION — Non-sargable predicate spot-check (any ORM/SQL). Wrapping a column in a function/cast/transformation inside a query predicate translates to func(column) = $param — the DB CANNOT use an index on that column and full-scans. Scan every query expression for a transformation on the COLUMN side, not the parameter side: .ToLower()/.ToUpper()/.Trim()/.Substring() on a column, col1 + " " + col2 == x (concatenation), .Date/date-part extraction, Convert/cast/collation change, leading-wildcard LIKE '%x', or a computed expression compared to a value. Fix — keep the column bare and move the transformation to the in-memory PARAMETER (e.g. case-insensitive via a candidate list col == x || col == xLower), OR persist a normalized indexed column, OR add a functional/expression index. ALWAYS prove with EXPLAIN/query plan: Index Scan/Seek expected, Seq Scan = the smell confirmed.

Find:

• no index for high-cardinality filters, joins, foreign keys, sort columns, or frequent group keys
• composite index field order mismatched with equality -> range -> sort access pattern
• non-sargable predicate: an indexed column wrapped in a function/cast/concat/date-part/transformation (see spot-check above) — the single most common silent index-loss; also incompatible type/collation, leading wildcard, broad OR, negative predicate, or low selectivity
• sort spill/filesort because index order does not match filter + order by
• covering/partial/filtered index opportunity for hot narrow query
• index bloat from adding every field without write-cost analysis
• leftmost-prefix violation — a query filtering only on the SECOND column of a composite index gets no seek from it
• selectivity not established — "add an index" proposed without the selectivity number; above ~5-20% selectivity a sequential scan legitimately beats random index lookups
• stale statistics — plan/explain shows estimated rows far from actual rows; the plan is wrong for a reason no rewrite fixes (refresh stats/analyze first)
• partition pruning lost — partitioned table queried without the partition key, so every partition is scanned
• shard/partition key skew — monotonic (timestamp/auto-increment) or low-cardinality key creating a hot shard/partition; per-partition throughput ceilings hit by one key
• replica read correctness-vs-lag — read-your-writes broken by replication lag, or a lag-sensitive read pointed at a replica
• random-UUID primary key destroying index locality and inflating index size (time-ordered UUIDv7/ULID fits)

Prefer fixes: add/adjust smallest useful index, reorder composite keys to match query, rewrite predicate to be sargable, refresh statistics, carry the partition/shard key into the predicate, salt or re-key a hot partition, route lag-sensitive reads to primary (or a sticky/LSN-aware window), verify with plan/explain before/after, include write-cost risk. Escalate in order — tune query/index → cache → vertical → read replicas → partition → shard; NEVER propose sharding before the earlier rungs are proven exhausted (why: resharding and cross-shard joins are the most expensive reversal in the ladder).

3. N+1 And Fan-Out

Think: Does work scale with item count instead of request/job count?

Find:

• query/API/cache call inside loop, map, serializer, resolver, template/render loop, event handler loop
• per-item existence/count lookup; per-item lazy-loaded relation
• repeated same lookup with different IDs that could be one IN/batch/group query
• nested fan-out across services, queues, jobs, or retries
• sequential awaits where independent calls can batch or run bounded parallel with separate safe resources

Prefer fixes: batch IDs once, join/include only needed fields, prefetch dictionaries, aggregate counts in one query, use bounded concurrency, preserve ordering/authorization semantics.

4. Aggregation, Join, And Pipeline Shape

Think: Does the pipeline reduce data before expensive join/unwind/group/sort/window stages?

Find:

• join/unwind/group before selective filter
• cartesian joins or duplicate expansion not collapsed
• grouping/sorting without pre-filter or supporting index
• aggregation loads all related rows/documents when only existence/count/min/max needed
• repeated post-processing that database can compute safely

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.

5. Materialization And Memory

Think: What enters memory? Is it bounded, streamed, and tracking-free when read-only?

Find:

• large collection materialized before paging/filtering
• read-only queries tracking entities/objects unnecessarily
• blob/file/large JSON fields loaded for lightweight responses
• buffering entire export/report when streaming/chunking fits
• accidental multiple enumeration re-running query

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.

6. Write Path, Locks, And Transactions

Think: Does write work batch safely and keep locks/transactions small?

Find:

• per-row save/update/delete inside loop
• long transaction wrapping remote calls or heavy reads
• unnecessary unique checks per row instead of bulk validation
• lock escalation/hot-row contention/counter updates without batching
• parallel writes sharing unsafe session/context/unit-of-work
• long-running or idle-in-transaction connection — under MVCC it pins old row versions and drives bloat/vacuum pressure fleet-wide (a slow-motion outage, not a local slowdown)
• isolation level mismatched to the invariant — lost update at Read Committed, or write skew at Snapshot/Repeatable Read where Serializable (or an explicit lock/version column) is required; read-modify-write done in application code instead of one atomic UPDATE
• inconsistent lock acquisition ORDER across code paths (deadlock source), or no retry on the deadlock error
• schema/migration change taking a blocking lock proportional to table size instead of an online pattern (nullable add → batched backfill → NOT VALID constraint → validate; concurrent index build; expand/contract)
• durability setting silently traded for throughput without the trade named (fsync/commit-sync relaxation)

Prefer fixes: bulk write, chunk, shorten transaction, move remote calls outside transaction, use idempotent commands, create fresh safe scope/context per parallel worker, pick the isolation level the invariant needs (or an explicit FOR UPDATE/version column), make write conflicts atomic in one statement, order lock acquisition consistently and retry deadlocks, use the online migration pattern for large tables.

7. Cache And Reuse

Think: Is repeated expensive work stable, safe to reuse, and invalidated correctly?

Find:

• same lookup repeated within request/job
• hot reference data fetched every request
• cache key missing tenant/user/auth/filter/version dimensions
• cache hides unbounded query or stale security-sensitive data
• no hit-ratio evidence — a cache added without measuring the ratio; the ratio IS the value (90%→99% cuts origin load 10×, so a 60% hit ratio is barely a cache)
• stampede/thundering-herd exposure — hot key expiring sends every request to origin at once; no single-flight/request-coalescing, no per-key lease, no TTL jitter, or a whole key class expiring simultaneously
• cold-start blindness — post-deploy/failover empty cache indistinguishable from an origin outage; no warming and no LB slow-start
• unbounded cache (a memory leak with a friendly name): no size bound, no entry lifetime, no eviction policy matched to access skew — and cache thrash once the working set exceeds cache size (a cliff, not a slope)
• missing negative caching, so nonexistent keys generate repeated miss-storms
• schema/build version absent from the key, so a deploy can serve poisoned entries

Prefer fixes: request-scope memoization first, then bounded shared cache with explicit key, TTL/invalidation, size limits, privacy constraints, and hit/miss metrics. Add single-flight + TTL jitter for hot keys, stale-while-revalidate where staleness is acceptable, negative caching (or a Bloom filter) for absent keys, a version segment in the key, and an eviction policy matched to the access skew (LRU default, LFU/W-TinyLFU for skewed). NEVER treat "we added a cache" as a completed fix without the measured hit ratio and the bound.

8. API Payload, Frontend Delivery And Rendering

Think: Does the user-perceived time come from payload size, render/interaction work on the main thread, or asset delivery? Judge against the Core Web Vitals thresholds at p75 of real users, never a single lab run.

Find:

• endpoint returns more payload than the view needs; response DTO shaped by the table, not the screen
• CWV breach — LCP > 2.5 s, INP > 200 ms, CLS > 0.1, TTFB > 800 ms (references/performance-knowledge.md §7)
• long task > 50 ms blocking the main thread (destroys INP); CPU-bound work never yielded or moved to a Worker
• layout thrashing — interleaved DOM read/write forcing a synchronous reflow per iteration
• animation on layout-triggering properties (width/top/left) instead of compositor-only transform/opacity
• CLS source — image/ad/embed with no reserved space (width/height/aspect-ratio); FOIT from missing font-display
• render-blocking synchronous CSS/JS in <head>; critical CSS not inlined
• JS weight/parse cost — the most expensive byte class (parse + compile + execute, unlike an image); no code splitting, no route-level lazy load, no tree-shaking
• third-party scripts loaded eagerly (tag managers, chat, analytics) — habitually the #1 regression source
• repeated fetch, client-side request waterfall (N+1 over HTTP), missing list virtualization, unstable render keys/track-by
• hydration cost scaling with component count; rendering strategy (CSR/SSR/streaming/SSG/islands) never chosen as a performance decision
• HTTP caching wrong: assets not hashed+immutable, HTML not revalidated, Vary incorrect (cache poisoning), no Brotli/gzip on text
• missing resource hints where they pay (preconnect saves DNS+TCP+TLS, preload for late-discovered critical assets, fetchpriority)

Prefer fixes: shape the payload to the view, batch/aggregate server-side, break or yield long tasks, batch DOM reads then writes, animate compositor-only properties, reserve space for media, defer third-party and cold routes, virtualize long lists, stabilize keys, hash+immutable asset caching with correct Vary, Brotli text compression, AVIF/WebP + srcset + lazy below-fold. ALWAYS confirm with a browser profile/network waterfall AND field (RUM/CrUX) data — a lab score locates the cause, field data defines the truth.

9. Compute And Algorithmic Complexity

Think: Does in-process work grow super-linearly with input size, independent of any query or network call?

MUST ATTENTION find:

• nested iteration over the same/related collection (O(n²)+): loop-in-loop, map inside map, repeated full re-scan
• linear membership/lookup inside a loop — .find/.includes/.indexOf/in list/.contains where a Set/Map/dict gives O(1)
• wrong data structure for the access pattern: array used as a keyed store; repeated .filter().length for existence
• string built by concatenation in a loop; repeated JSON.parse/stringify/deep-clone/serialize per iteration
• catastrophic-backtracking regex on user- or attacker-sized input (ReDoS — cross-link $security-review)
• pure-CPU result recomputed every call when inputs are stable (memoization candidate, distinct from data cache)
• redundant sort/re-sort, or sorting when a single-pass min/max/partition suffices

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.

10. Network And Protocol Efficiency

Think: How many round trips does this path cost, and what is the RTT floor it can never beat? Count calls × RTT before optimizing anything inside a single call.

MUST ATTENTION find:

• per-request connection setup — no keep-alive/connection pooling/reused client, so every call pays TCP (1 RTT) + TLS (1-2 RTT) + possibly cold DNS; the single largest and most common network defect
• chatty contract — many small sequential remote calls where one batch endpoint or server-side aggregation fits; call count grows with items (network N+1, distinct from DB N+1)
• RTT floor ignored — latency budget already consumed by geography (~1 ms per 100 km) or cross-region hops, with a code-level fix proposed instead of an edge/replica/CDN move
• payload not compressed (no Brotli/gzip on text), or over-large for the consumer; critical response exceeding the ~14 KB initial congestion window when first-round-trip delivery matters
• protocol left on the table: HTTP/1.1 head-of-line blocking with 6-conn/origin limits, domain sharding retained under HTTP/2 (now an anti-pattern), lossy/mobile path that would benefit from HTTP/3/QUIC
• small-write RPC path suffering Nagle + delayed-ACK (~40 ms stalls) without TCP_NODELAY
• infra exhaustion limits unchecked — file descriptors, listen backlog, ephemeral ports (~28k default), TIME_WAIT accumulation, conntrack table, NAT/SNAT ports: these present as "random" latency or errors, never as a slow function
• load balancing weak: naive round-robin where least-connections/power-of-two-choices fits, no health check or outlier ejection, no slow-start for new instances (a cold node given full traffic times out)
• sticky sessions used where stateless + external session store fits, blocking rebalancing

Prefer fixes: reuse connections (keep-alive + pooled clients), collapse chatty calls into one batch/aggregate endpoint, move the endpoint closer (edge/CDN/regional replica) when RTT is the floor, compress and shrink payloads, enable the protocol version that matches the path, raise/verify the OS and infra limits, and configure LB algorithm + health checks + slow-start. ALWAYS quantify as call count × RTT before and after — why: a faster handler behind 12 avoidable round trips is not a fix.

11. Runtime, Memory And GC

Think: Does the runtime itself inject latency the code cannot see — collector pauses, allocation pressure, a blocked event loop, memory that never returns?

MUST ATTENTION find:

• GC pause as a tail-latency source — latency spikes uncorrelated with load, invisible in the mean and correlated across the fleet; allocation RATE (not heap size) driving collection frequency
• heap mis-sized: too small → continuous GC; too large → long pauses and swap risk; no headroom left for off-heap/native buffers/thread stacks
• managed-language leak shapes — unbounded caches, un-removed listeners/subscriptions, closures capturing large scopes, static collections, thread-locals on pooled threads
• RSS ≠ heap confusion — container/pod killed on RSS while heap looks healthy (fragmentation, native buffers, ~1 MB per thread stack)
• swap active on a latency-sensitive service (prefer fail-fast OOM over swap thrash); page cache double-buffered against an app cache
• working-set cliff — data outgrowing L3 → RAM → page cache, producing a step change rather than a gradual slope
• blocked event loop / blocking call in an async path — one CPU-bound task stalling every connection; .Result/.await-blocking on a thread-pool thread
• thread/worker pool mis-sized for the workload class (CPU-bound ≈ cores; I/O-bound ≈ cores × (1 + wait/compute))
• contention shaped wrong: one coarse global lock (the Amdahl serial section) where sharded/striped locks, lock-free counters, or immutable/copy-on-write data fit
• cache-line issues on genuinely hot paths: false sharing on adjacent hot counters, NUMA-remote allocation (~2× local), random access where sequential is available (10-100× on the same bytes)

Prefer fixes: cut allocation rate before tuning the collector, right-size the heap with headroom, bound every cache and unregister every listener, measure RSS not heap against the container limit, disable swap for latency-critical services, move CPU work off the event loop, size pools by workload class, reduce lock granularity, and restore sequential access order. Prove with GC-pause histogram, allocation profile, RSS trend, event-loop lag, or off-CPU flame graph — NEVER from code reading alone.

12. Distributed Resilience And Load Management

Think: Under load or partial failure, does this path degrade gracefully — or amplify the failure? Performance and resilience share the same queues, so a missing timeout is a latency defect.

MUST ATTENTION find:

• missing or non-decreasing timeout budget — no timeout anywhere, or a callee timeout ≥ the caller's remaining budget; a hung dependency then exhausts threads/pool and takes the caller down
• retry amplification — retries without exponential backoff + FULL JITTER, no cap, no retry budget (≤~10% of traffic), or retries on non-idempotent writes with no idempotency key → a partial outage becomes total
• no circuit breaker on a failing dependency; no bulkhead (shared pool lets one slow dependency consume every thread); no load shedding (slow timeouts served where a fast 429/503 is correct)
• unbounded queue/buffer — converts a throughput problem into unbounded latency then OOM; queue AGE not monitored (only depth); no DLQ or poison-message handling; no backpressure propagated to the producer
• per-item message publish — producer/consumer emits one message or event per item where one batched message or bulk event fits; broker round trips and consumer invocations then grow linearly with item count (the messaging form of N+1)
• fan-out tail amplification — scatter-gather over many backends where a backend p99 becomes the user's median (~63% hit rate across 100 calls); no hedged requests or per-shard timeout
• dual write to DB + broker instead of a transactional outbox/CDC; exactly-once assumed instead of at-least-once + idempotent consumer
• cross-service workflow with no saga/compensation, or 2PC on a latency-sensitive path; consensus/quorum round trips on the hot path
• wall-clock used for cross-machine ordering (NTP skew is ms-to-s) instead of monotonic/logical/hybrid clocks; leader action without a fencing token/lease (a GC-paused leader still believes it leads)
• autoscaling lag — scaling on lagging CPU rather than a leading indicator (queue depth, concurrency), boot+warmup exceeding the spike, no pre-scale for known events, no headroom below the utilization knee
• metastable failure risk — system stays broken after the trigger clears (retry storm + cold cache) with no explicit shedding path to recover
• no per-tenant quota/rate limit (token bucket / leaky bucket / sliding window) — one loud tenant becomes everyone's outage; correlated failure via a shared dependency or shared config push defeating nominal redundancy

Prefer fixes: set a decreasing timeout budget per hop, add backoff+jitter with a retry budget and idempotency keys, add breaker + bulkhead + load shedding, bound every queue and propagate backpressure, reduce fan-out or hedge it, replace dual writes with an outbox, order events by logical clock and fence leaders, autoscale on a leading indicator with headroom, and enforce per-tenant quotas. Prove with timeout/retry config file:line + queue depth AND age + breaker state + per-tenant limits — why: these defects are invisible at low load and only surface as the outage they cause.

---

Phase 4: Findings And Severity

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:

• Critical: outage/OOM/data corruption risk, unbounded hot path, lock storm, runaway fan-out.
• High: p95/p99 timeout risk, full scan on large/hot table/collection, N+1 on user-visible list, missing page bound.
• Medium: avoidable over-fetch, suboptimal index, repeated lookup, moderate memory waste.
• Low: cleanup with small measurable benefit or future-proofing.

NEVER inflate severity without production-like scale/frequency evidence.

---

Phase 5: Optimize Plan

Before code changes (MUST ATTENTION):

• present baseline, proposed change, behavior invariants, risks, verification commands, and rollback path.
• preserve functional behavior, authorization, ordering, pagination semantics, consistency, and idempotency.
• inspect affected tests/specs/docs when behavior, SLA, public contract, or limits change.
• NEVER change query semantics only to improve speed unless user approves changed behavior.
• NEVER add broad indexes/caches without write-cost, storage-cost, invalidation, and privacy analysis.

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.

---

Sub-Agent Routing

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/CWV/network waterfall	frontend or performance specialist
Runtime/GC pauses, allocation profile, pool + event-loop sizing	performance-optimizer (runtime evidence: GC log, allocation profile, RSS trend, off-CPU profile)
Timeouts/retries/breakers/queue bounds, resilience under load	performance-optimizer with the resilience config in scope; escalate design-level gaps to architect

Sub-agent prompt MUST include target, detected scope, local context evidence, required dimensions, the calibration anchors in play (references/performance-knowledge.md), report path, and "return summary only; write full report incrementally."

---

Phase 6: Why-Review Findings Validation Gate (MANDATORY when findings exist)

Purpose: Validate performance findings before optimization work. Performance reports overstate easily 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:

1. Read own finalized report from plans/reports/performance-{date}-{slug}.md or the exact report path written by the caller.
2. Invoke $why-review --validate-findings <performance-report-path>.
3. Read the validation verdict path returned by why-review, expected as plans/reports/why-review-validate-{date}.md.
4. If why-review demotes/removes any finding: update the performance report with revised severity, removed false positives, and a ## Why-Review Validation Notes section.
5. If why-review confirms all findings: append ## Why-Review Validation stating all findings were re-validated against measurement/static evidence.
6. If the report changed after validation: re-run this validation gate, maximum 2 validation passes, until the report's remaining findings are validated or zero findings remain.

Skip conditions (record explicit reason if skipping):

• Verdict is unconditional PASS with zero findings.
• Why-review skill itself is the active context.

---

Phase 7: Validated Fix + Full Performance Re-Review Loop (MANDATORY when validated findings remain)

Trigger: Phase 6 returns CLEAN/validated and the performance report still has one or more findings that must be fixed.

Protocol:

1. Create a fresh fix-cycle task list before editing. Do not reuse the review tasks.
2. Fix only findings that survived $why-review --validate-findings; if this skill is running inside a workflow, route implementation through the parent $plan + $feature-implement flow.
3. Re-measure or run the verification command named in the finding.
4. Restart the full $performance-review review from Phase 0 over the complete current target, not only the fixed files.
5. The restarted pass MUST create brand-new review tasks, re-detect scope, rediscover local context, rerun baseline/graph/profiler checks where applicable, and analyze all dimensions again from the beginning.
6. Repeat validate → fix → full performance re-review until a complete pass has zero findings.
7. If the same validated blocker repeats across 3 full invocations with no progress, stop and ask the user for a decision.

Non-negotiable rules:

• Never fix a performance finding before $why-review --validate-findings validates it.
• Never mark performance review clean after a targeted before/after check only; the clean verdict must come from a full Phase 0 restart.
• Never review only fixed files during the recursive pass.
• Never reuse old todo/task items for the recursive review pass.

---

Output

MANDATORY final report sections:

• Scope and detected bottleneck type
• Baseline evidence and unknowns — each number with the anchor it is calibrated against, the load model behind it, and the dimensions covered vs explicitly skipped (with reason)
• Findings ordered by severity
• Optimization plan and rejected alternatives
• Verification plan with before/after metrics
• Test/spec/doc impact or explicit skip reason
• Confidence and assumptions

If evidence insufficient, output: Insufficient evidence. Verified: [...]. Not verified: [...]. Next evidence needed: [...].

---

Graph-Assisted Investigation — MANDATORY when .code-graph/graph.db exists.

HARD-GATE: MUST ATTENTION run at least ONE graph command on key files before concluding any investigation.

Pattern: Grep finds files → trace --direction both reveals full system flow → Grep verifies details

Task	Minimum Graph Action
Investigation/Scout	trace --direction both on 2-3 entry files
Fix/Debug	callers_of on buggy function + tests_for
Feature/Enhancement	connections on files to be modified
Code Review	tests_for on changed functions
Blast Radius	trace --direction downstream

CLI: python .claude/scripts/code_graph {command} --json. Use --node-mode file first (10-30x less noise), then --node-mode function for 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 task tracking sub-task, work through it with file:line evidence, 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?

Scenario Stress & Resilience Evaluation — CONDITIONAL, evidence-gated, business-criticality-aware. The top-down companion to SYNC:scale-technique-gate: instead of "is technique X present?", put the system UNDER concrete failure/load scenarios and judge whether it SURVIVES, SELF-HEALS, and whether its BUSINESS needs it to. ADVICE-ONLY: emit the Scenario Stress Matrix as guidance; NEVER mutate any score, verdict band, or gate pass/fail.

1. Reuse the scale tier derived by SYNC:scale-technique-gate (or derive it identically from evidence); also derive business-criticality B0–B3 from specs/SLA/product docs + the domain, cite file:line + confidence. B0 best-effort · B1 important · B2 business-critical · B3 mission-critical/regulated. Unknown → state the assumption, do NOT default to B3/T3. Criticality-signal floor (both-directions safety): regulated / PII / financial / health data, money movement, auth/identity, or legal-compliance scope raises B to at least B2 even absent SLA/SLO docs; anti-over-engineering lowers hardening ONLY when NO such signal is present. B (blast if it fails) and T (scale of load/data) are independent — a low-traffic payroll run is low-T, high-B.
2. Select in-scope scenarios — only those the system's B/T combination warrants (a B0 internal PoC skips region-loss/DR entirely; a B3/T0 regulated service still needs backups + DR by BUSINESS, not scale).
3. Walk each in-scope scenario: simulate the stimulus → trace the break path → name the failure signature → answer the self-heal/recovery question (auto-recover? MTTR? manual runbook?) → name the trade-off it forces. Families: traffic spike · sustained growth · data-volume growth · write/ingest burst · dependency down/slow · instance/node loss · zone/region loss · data loss/corruption · poison-message/retry-storm · cascading failure/backpressure · cold-start/deploy-blip · clock-skew/duplicate-delivery.
4. Assign one verdict per scenario: WITHSTANDS · DEGRADES-GRACEFULLY · FAILS-HARD (→ advise only) · N/A-by-business (not warranted → skip, not a gap) · OVER-HARDENED (resilience beyond business need → advise AGAINST, cite carrying cost).
5. Anti-over-engineering guard (first-class): a lean system whose business does not need HA/DR is a PASS; OVER-HARDENED flags resilience the business does not warrant. This guard is symmetric with the criticality-signal floor above — never under-harden a B2+ system just because its traffic is low.
6. Output — Scenario Stress Matrix: scenario | in-scope (B/T)? | verdict | self-heal | trade-off | evidence (file:line/config/infra). Full catalog + Business×Scale in-scope baseline + verdict/tier tables → .claude/docs/scenario-stress-catalog.md. ADVISORY-ONLY: NEVER mutate any /20, /24, verdict band, or gate pass/fail. Drift-guard: scenarios/verdicts/business-tiers are AUTHORITATIVE in the catalog — update it FIRST, then re-run .claude/scripts/inject_scenario_stress_gate.py. Scale tier stays single-sourced in scale-technique-catalog.md.

BLOCKED until: - [ ] scale tier + business-criticality (with criticality-signal floor) derived from evidence - [ ] in-scope scenarios selected - [ ] matrix emitted - [ ] over-hardening guard applied - [ ] advisory-only (no score/verdict mutation) confirmed

Validated-Finding Fix + Full Re-Review Loop — Re-review is triggered by a validated finding fix cycle, not by a round number. Review purpose: review → validate findings → fix validated findings → full re-review until a complete review pass finds no issues. A clean review ENDS the loop — no further rounds required.

aka Self-Review Convergence Loop. The name is historical — there is NO 2-round cap; "double-round-trip" only means a validated-finding fix cycle forces at least one fresh re-review. It runs until a clean pass, bounded by the 5-round ceiling below.

Round cap — 5 rounds MAX (a ceiling, NEVER a target). A clean pass ENDS the loop immediately at ANY round — round 1 included; the cap never obliges you to keep spinning. Hitting round 5 with validated findings still open → STOP and escalate by asking the user directly with the still-open findings listed; NEVER emit a silent "good enough" PASS on cap exhaustion, and NEVER let the cap substitute for the clean-review requirement. The 3-repeated-no-progress blocker rule stays an EARLIER exit — escalate at whichever trips first.

Universal scope (any new output/judgment): any newly produced output or judgment gets ≥1 self-review; any new judgment gets ≥1 $why-review --validate-findings pass; anything flagged to re-check is re-checked ≥1 time — before that output is treated as final. This loop is the default convergence contract for ANY work-producing skill, not review skills only.

Routing invariant (author-facing): a skill that validates findings MUST route them through $why-review --validate-findings (the terminal validator) — NEVER fork an inline finding-validation. Routing through why-review is what makes the finding-survival bar and this loop apply; the verify-review-validate-coverage sensor enforces this exact route mechanically.

Round 1: Main-session review. Read target files, build understanding, note issues. Output findings + verdict (PASS / FAIL).

Decision after Round 1:

• No issues found (PASS, zero findings) → review ENDS. Do NOT spawn a fresh sub-agent for confirmation.
• Issues found (FAIL, or any non-zero findings) → run the active review skill's findings-validation gate first; for review skills the default gate is $why-review --validate-findings <report-path>. Fix only validated findings, then restart the full review protocol from the beginning with a fresh task breakdown.

Fresh full re-review after every fix cycle: Re-run the whole review protocol over the current full target. When sub-agents are part of that protocol, spawn NEW spawn_agent calls — never reuse prior agents. Reviewers re-read ALL files from scratch with ZERO memory of prior rounds. See SYNC:fresh-context-review for the spawn mechanism and SYNC:review-protocol-injection for the canonical Agent prompt template. Each fresh full review must catch:

• Cross-cutting concerns missed in the prior round
• Interaction bugs between changed files
• Convention drift (new code vs existing patterns)
• Missing pieces that should exist but don't
• Subtle edge cases the prior round rationalized away
• Regressions introduced by the fixes themselves

Loop termination: After each full re-review, repeat the same decision: clean → END; issues → validate findings → fix → restart from the first review phase. Continue until a complete review pass finds zero issues, capped at 5 rounds. Escalate by asking the user directly at whichever comes first: the same validated finding repeats for 3 full invocations with no progress · a fix requires product/owner input · round 5 completes with validated findings still open. NEVER loop past 5 rounds, and NEVER convert cap exhaustion into a PASS.

Rules:

• A clean Round 1 ENDS the review — no mandatory Round 2
• NEVER fix unvalidated findings; validate first using the caller's validation gate
• Every surviving finding must additionally clear the finding-survival bar defined in why-review's Findings Validation Routine (a deliberately higher bar than the generic act-gate — "keep this finding?" is a stricter question than "act on this evidence?"); a finding below the bar is demoted or dropped, not kept
• NEVER skip the full re-review after a fix cycle (every fix invalidates the prior verdict)
• NEVER reuse a sub-agent across rounds — every iteration that uses sub-agents spawns NEW Agent calls
• Main agent READS sub-agent reports but MUST NOT filter, reinterpret, or override findings
• The 5-round cap NEVER replaces the clean-review requirement — it bounds runaway looping, it does not authorize shipping an un-clean review; a clean pass ends the loop early at any round, and cap exhaustion escalates rather than passes
• Enforce the round cap of 5 alongside the 3 repeated-no-progress blocker rule; both are escalation triggers, neither is a completion criterion
• Track recursive invocation count and repeated blockers in conversation context (session-scoped)
• Final verdict must incorporate ALL rounds executed

Report must include ## Round N Findings (Fresh Sub-Agent) for every round N≥2 that was executed.

Goal Contract Satisfaction Loop — Persist the user goal in an external file, execute against it, and loop review/fix until every saved required criterion passes or a blocker escalates. Bounded closed loop — NEVER open-ended autonomous exploration.

1. Resolve the active goal (in order): active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create a new Goal Contract from the current user request (template: .claude/templates/goal-contract-template.md).
2. Required sections: Original Request, Purpose, Success Criteria (checkboxes; mark required vs optional), Constraints, Evidence Required, Iteration Log, Goal Satisfaction matrix.
3. Before work: read the active goal and map planned work to saved success criteria — execution serves the saved criteria, never chat memory alone.
4. After execution/verification: append an Iteration Log entry — result, evidence references (file:line, command output, report path), remaining gaps.
5. Review gate: emit a Goal Satisfaction matrix — | Success Criterion | Evidence | Status | with PASS/FAIL/BLOCKED. Overall PASS requires every required criterion PASS.
6. Loop rule (retry): required criterion FAIL → validate the gap is real → fix → re-review only the affected criteria. Stop cleanly when all required criteria PASS.
7. Escalation rule (stop): two consecutive iterations with no criterion progressing, or a blocker needing user input → mark the criterion BLOCKED with a user-facing reason and escalate. NEVER loop indefinitely.
8. Skip rule: tiny conversational tasks may skip the goal file ONLY with a recorded one-line reason. User-accepted gate skips are recorded in the goal file with reason and scope.
9. Security: NEVER store secrets, tokens, credentials, or private customer data in goal files — store evidence references and redact sensitive values.

Blocked until: active goal resolved (or skip reason recorded) · saved success criteria read before edits · iteration evidence appended after execution · Goal Satisfaction matrix emitted before any PASS verdict.

Trade-Off Interrogation Gate — ALWAYS ask these THREE questions before ANY verdict, score, finding, or recommendation — about the thing under review AND about every recommendation YOU make. — why: naming a benefit without its price is an endorsement, not a review; the costliest trade-offs are the ones nobody wrote down.

1. Is there any trade-off? Name what it SACRIFICES. "None" / "pure win" is an unfinished analysis, NOT an answer — to claim none, state which dimensions you checked and why each is unaffected: future change cost · complexity · performance/latency · memory/cost · coupling · reversibility · migration burden · operational load · blast radius · security posture · testability · team skill/ramp · delivery time · UX.
2. Is it worth it? Weigh gain against sacrifice EXPLICITLY — what is gained (with a metric) · what it costs · WHO pays · WHEN it comes due — then emit WORTH IT / NOT WORTH IT / UNCLEAR. "Better" with no metric and no cost FAILS this question. NOT WORTH IT → withdraw or replace the recommendation, never keep it as-is.
3. Is the trade-off material enough to CONFIRM WITH THE USER? A material trade-off is the user's call, never yours. MATERIAL when ANY holds: irreversible / one-way door (data migration, public contract, storage format, vendor lock-in) · cost shifted onto someone else (another team, ops/on-call, future maintainer, end user) · one quality attribute traded for another (correctness↔speed, security↔convenience, latency↔cost, simplicity↔flexibility) · a boundary crossed (client↔server tier, service contract, event contract, shared library) · a high-consequence path (auth, money, data integrity, breaking change, High/Medium residual risk) · the worth-it verdict is UNCLEAR.

MATERIAL → STOP and confirm by asking the user directly BEFORE the verdict stands — state the trade-off, both options, what each sacrifices, and your recommendation. NOT material → record it inline with a one-line justification and proceed.

Non-asking execution contexts — ESCALATE BY HANDOFF, never by silence. ask the user directly reaches only the main interactive agent: a sub-agent cannot ask the user, and a terminal/verdict-only mode asks nothing by design. When you are running in such a context, the obligation is redirected, never waived — do ALL of: (a) complete questions 1 and 2 normally; (b) decide materiality and record it in the Trade-Off Assessment row with confirmed? = NO — cannot ask from this context; (c) name the unconfirmed MATERIAL trade-off explicitly in your returned summary/verdict so the CALLER (or parent orchestrator) escalates it by asking the user directly on your behalf — a material trade-off mentioned only inside a report file on disk is NOT a handoff; (d) do not emit an unqualified PASS — mark the verdict as carrying an unconfirmed material trade-off, so the caller's gate stays closed until the user answers. The caller inherits the escalation duty the moment it reads your return.

This carve-out is about reachability, not convenience: it applies ONLY where the tool genuinely cannot reach the user (spawned sub-agent, terminal validate/verdict-only mode, non-interactive/headless run). It is NEVER a licence to skip the question, to self-approve a one-way door, or to downgrade materiality because asking is inconvenient — if you CAN ask, you MUST ask.

Emit a Trade-Off Assessment row per reviewed decision and per recommendation: | decision | sacrifices | gain (metric) | who pays, when | WORTH IT/NOT/UNCLEAR | material? | confirmed? |.

BLOCKED until: trade-off named (or dimensions-checked justification given) · worth-it verdict emitted · materiality decided · every MATERIAL trade-off either confirmed with the user OR — in a non-asking context — handed off in the returned verdict for the caller to confirm. A MATERIAL trade-off that is neither confirmed nor handed off can NEVER be PASS, and NEVER gets buried as a Low-severity note.

NEVER answer "no trade-off" without checking · decide a material trade-off silently on the user's behalf · let convergence/delivery pressure authorize walking through a one-way door · bundle several material trade-offs into one vague "proceed?".

• MANDATORY Large changeset → batch by size cap (≤8 files OR ≤2000 diff-lines), one parallel sub-agent per batch; never review many files one-by-one.
• MANDATORY > 6 categories OR > 40 files → add the hierarchical synthesis tier; each concern-synthesizer emits cross-concern interaction candidates and the orchestrator runs the cross-concern pass before concluding.