performance-review Skill

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]

[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.

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.

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 OR flagging one as a defect. Before changing or reporting a constant, limit, flag, cutoff, wording, or pattern, read nearby context and history, the CALLER's ordering, and 2+ sibling call sites of the same convention. A doc stating WHAT without WHY is missing rationale, not proof of a missing guard. Surface ambiguity before acting — don't pick silently. Multiple valid interpretations require an explicit question or stated assumption with risk. Assert the outcome your system owns, not the intermediate state your infrastructure owns. When verifying async work, assert the final business state — never the delivery/retry bookkeeping held in shared infrastructure that any co-running process can write. Such a check passes when run alone and flakes the moment anything else shares that infrastructure. 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:

Detect - Classify scope and bottleneck type; order hypotheses via the symptom→cause matrix.
Discover - Read local code, metrics, docs, query/index definitions, similar patterns.
Measure - Capture baseline against a known anchor, or mark static-only risk.
Analyze - Run 12 serial dimension passes with evidence.
Plan - Propose smallest fix preserving behavior.
Verify - Re-measure, run tests, and record evidence.
Validate Findings - Run $why-review --validate-findings <report-path> before any fix.
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.

<target>$ARGUMENTS</target>

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:

Read own finalized report from plans/reports/performance-{date}-{slug}.md or the exact report path written by the caller.
Invoke $why-review --validate-findings <performance-report-path>.
Read the validation verdict path returned by why-review, expected as plans/reports/why-review-validate-{date}.md.
If why-review demotes/removes any finding: update the performance report with revised severity, removed false positives, and a ## Why-Review Validation Notes section.
If why-review confirms all findings: append ## Why-Review Validation stating all findings were re-validated against measurement/static evidence.
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:

Create a fresh fix-cycle task list before editing. Do not reuse the review tasks.
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.
Re-measure or run the verification command named in the finding.
Restart the full $performance-review review from Phase 0 over the complete current target, not only the fixed files.
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.
Repeat validate → fix → full performance re-review until a complete pass has zero findings.
If the same validated blocker repeats across 2 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.

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.

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).

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.

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).

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.

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 clears the round's exit bar (see Severity floor below). 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 3-round ceiling below.

Round cap — 3 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 3 with blocking findings still open (severity floor applied) → 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 2-repeated-no-progress blocker rule stays an EARLIER exit — escalate at whichever trips first.

Severity floor — from round 3, LOW stops blocking. The exit bar tightens by round, so the loop converges on consequence instead of spinning on polish:

Define one predicate everywhere: blocking_findings(round, findings) returns all validated findings in rounds 1–2 and only validated CRITICAL/HIGH/MEDIUM findings in round 3+. A binary gate (test-green, security must-fix, required artifact) is exempt only when its owning invariant explicitly says so.

| Round | Exit bar — loop ENDS when the fresh full review has… | Must be fixed to continue | | ----- | ------------------------------------------------------------------------- | ------------------------------ | | 1-2 | zero validated findings at ANY severity | CRITICAL · HIGH · MEDIUM · LOW | | 3+ | zero validated CRITICAL / HIGH / MEDIUM findings — LOW-only is a PASS | CRITICAL · HIGH · MEDIUM only |

From round 3 onward LOW findings are NOT required to be fixed: a round whose validated findings are ALL LOW ENDS the loop immediately — do not open another round for them. Severity tiers are SYNC:severity-rubric (CRITICAL block-merge · HIGH must-fix · MEDIUM should-fix · LOW nice-to-fix); rounds 1-2 are unchanged, so an easy LOW still gets fixed early when it is cheap.

Severity-floor rules:

Never silently drop a deferred LOW. Every unfixed LOW is listed in the final report under ## Deferred LOW Findings (severity floor, round ≥3) with file, line, and description, so the owner can schedule it. Dropping it from the report is a protocol violation, not a clean pass.

Never re-tier a finding to trigger the exit. Downgrading a real CRITICAL/HIGH/MEDIUM to LOW so the loop can end is a FALSE PASS. Severity is set by consequence per SYNC:severity-rubric before the round bar is applied — never after, and never with the exit in view. — why: a floor that can be reached by relabeling is not a floor.

The floor bounds the loop, not the standard. It ends iteration; it never authorizes shipping a known CRITICAL/HIGH/MEDIUM, and it never lowers the finding-survival bar that admits a finding in the first place.

The floor never applies to a hard gate. Test-green gates (a suite must actually pass), security must-fix gates, and any gate whose criterion is binary rather than severity-rated are unaffected — a failing test is a failure, not a LOW finding.

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.

blocking_findings(round, findings) is non-empty → 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 against that round's exit bar: bar cleared → END; blocking findings remain → validate findings → fix → restart from the first review phase. Rounds 1-2 clear on zero findings at any severity; from round 3 the bar is zero CRITICAL/HIGH/MEDIUM, so a LOW-only round ENDS the loop (deferred LOWs go in the report). Capped at 3 rounds. Escalate by asking the user directly at whichever comes first: the same validated finding repeats for 2 full invocations with no progress · a fix requires product/owner input · round 3 completes with CRITICAL/HIGH/MEDIUM still open. NEVER loop past 3 rounds, and NEVER convert cap exhaustion into a PASS.

Rules:

A clean Round 1 ENDS the review — no mandatory Round 2

From round 3 on, a round whose validated findings are ALL LOW ENDS the loop — never open round N+1 to fix LOW alone; list those LOWs as deferred instead

NEVER re-tier a CRITICAL/HIGH/MEDIUM down to LOW to reach the round-3 exit — severity is assigned by consequence before the bar is applied

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 3-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 3 alongside the 2 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, plus ## Deferred LOW Findings (severity floor, round ≥3) whenever the loop ended on the round-3+ bar with LOWs still open.

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.

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).

Required sections: Original Request, Purpose, Success Criteria (checkboxes; mark required vs optional), Constraints, Evidence Required, Iteration Log, Goal Satisfaction matrix.

Before work: read the active goal and map planned work to saved success criteria — execution serves the saved criteria, never chat memory alone.

After execution/verification: append an Iteration Log entry — result, evidence references (file:line, command output, report path), remaining gaps.

Review gate: emit a Goal Satisfaction matrix — | Success Criterion | Evidence | Status | with PASS/FAIL/BLOCKED. Overall PASS requires every required criterion PASS.

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.

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.

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.

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.

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.

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.

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.

MANDATORY Classify findings Critical/High/Medium/Low by consequence; Critical/High block PASS until fixed or owner-accepted.
MANDATORY Score-based skills (sre 0-2, perf two-axis) map onto the same four tiers — no parallel severity vocabulary.

MANDATORY Derive review categories from file language + directory semantics + change nature; create a sub-task per category.
MANDATORY Derive each category's concerns from first principles with file:line evidence — never a fixed checklist.

IMPORTANT MUST ATTENTION scenario-stress gate: reuse the scale tier T0–T3 AND derive business-criticality B0–B3 from evidence first — apply the criticality-signal floor (regulated/PII/financial/health data · money movement · auth/identity · legal-compliance → at least B2 even absent SLA docs; do NOT default to B3). Select only the scenarios the B/T combination warrants, then walk each (simulate → trace → failure signature → self-heal/MTTR → trade-off) and assign WITHSTANDS/DEGRADES-GRACEFULLY/FAILS-HARD/N/A-by-business/OVER-HARDENED. Anti-over-engineering is first-class (a lean system that needs no HA/DR is a PASS) AND symmetric (never under-harden a B2+ system for low traffic). ADVICE-ONLY — emit the Scenario Stress Matrix as guidance; NEVER mutate any score, verdict band, or gate pass/fail. Full catalog → .claude/docs/scenario-stress-catalog.md (authoritative for scenarios/verdicts/business-tiers — on any change update the catalog FIRST, then re-run inject_scenario_stress_gate.py; scale tier stays single-sourced in scale-technique-catalog.md).

MANDATORY IMPORTANT MUST ATTENTION execute the review loop (aka Self-Review Convergence Loop): review → validate findings → fix validated findings → full re-review. A complete review pass with zero findings ENDS the review. Any newly produced output/judgment gets ≥1 self-review; any new judgment gets ≥1 $why-review --validate-findings pass before it is treated as final.
MANDATORY apply the severity floor: rounds 1-2 exit on zero findings at any severity; from round 3 the bar is zero CRITICAL/HIGH/MEDIUM — LOW findings are no longer required to be fixed, so a LOW-only round ENDS the loop. List every deferred LOW in the report; NEVER re-tier a real CRITICAL/HIGH/MEDIUM down to LOW to reach the exit, and NEVER apply the floor to a binary gate (test-green, security must-fix).
MANDATORY enforce the round cap of 3 — a ceiling, NEVER a target: a clean pass ends the loop immediately at any round (round 1 included), and round 3 completing with CRITICAL/HIGH/MEDIUM still open → STOP & escalate by asking the user directly, never a silent PASS. The 2-repeated-no-progress blocker rule is an earlier exit — escalate at whichever trips first. NEVER loop open-ended.

MANDATORY Resolve the active Goal Contract BEFORE work (active plan goal.md → plans/goals/{YYMMDD-HHmm}-{slug}/goal.md → create from current request) and read saved success criteria before editing.
MANDATORY Append iteration evidence after execution; emit a Goal Satisfaction matrix (PASS/FAIL/BLOCKED) before reporting PASS; loop on validated FAIL; escalate repeated no-progress or blockers. NEVER store secrets in goal files.

MANDATORY MUST ATTENTION ALWAYS ASK THE 3 TRADE-OFF QUESTIONS — on the thing under review AND on every recommendation you make: (1) is there any trade-off? name what it SACRIFICES (change cost · complexity · perf · coupling · reversibility · migration · ops load · blast radius · security · testability · delivery time · UX) — "none"/"pure win" is an unfinished analysis, so state the dimensions checked; (2) is it worth it? gain (with a metric) vs cost, WHO pays, WHEN → emit WORTH IT / NOT WORTH IT / UNCLEAR; NOT WORTH IT → withdraw or replace it; (3) is it material enough to confirm with the user? irreversible/one-way door · cost shifted onto another team/ops/maintainer/user · one quality attribute traded for another · a tier/service/event/library boundary crossed · auth/money/data-integrity/breaking-change/High-or-Medium-risk path · verdict UNCLEAR → STOP and confirm by asking the user directly BEFORE the verdict.
MANDATORY A MATERIAL trade-off with no user confirmation can NEVER be PASS; NEVER bury one as a Low-severity note, NEVER decide it silently, and NEVER let delivery or convergence pressure authorize a one-way door. — why: an un-walked-back one-way door is the user's call to make, not the reviewer's.
MANDATORY — non-asking 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. There the duty is REDIRECTED, not waived — still name the trade-off, still decide materiality, record confirmed? = NO — cannot ask from this context, state the unconfirmed MATERIAL trade-off in your RETURNED verdict/summary so the CALLER escalates it (a note only in an on-disk report is not a handoff), and never emit an unqualified PASS. Applies ONLY where the user is genuinely unreachable (spawned sub-agent, terminal validate mode, headless run) — if you CAN ask, you MUST ask.

Parallel Sub-Agent Dispatch — Plan parallelism the moment a task breakdown exists, BEFORE executing it — running provably independent tasks sequentially wastes wall-clock. Applies to every multi-step job: workflow steps, planning, batch updates, investigation, research, scans, reviews, doc sync. Plan execution is metadata-gated, NEVER default-parallel — fan-out follows ONLY what the plan declares (PAR/SEQ tags + per-phase write set); an untagged plan runs sequentially — why: a derived write set cannot see cascade or generated writes.

Tag every task PAR or SEQ. PAR = inputs exclude every pending task's output AND write set disjoint from every other PAR. Else SEQ — MUST ATTENTION name the dependency forcing it.

Group PAR into waves. No edge between members. Two writers of one file NEVER share a wave. Read-only work (search, investigation, review, research) parallelizes freely.

Declare before dispatch: Parallel plan: wave 1 = [...] · wave 2 = [...] · SEQ = [...] (reason).

Spawn each wave in ONE message — every spawn_agent call in one response, NEVER dripped per turn. Route each task to its specialist (.claude/skills/shared/sub-agent-selection-guide.md); NEVER code-reviewer as catch-all.

Brief each sub-agent self-contained: goal · scope + owned files · reference docs · return contract (summary + Full report: path, per SYNC:subagent-return-contract) · incremental persistence to plans/reports/ (per SYNC:incremental-persistence).

Barrier per wave. Advance ONLY after EVERY member returns (a skipped conditional counts as returned). Merge, mark each task completed/skipped, THEN dispatch the next wave. Mutating steps wait for the barrier.

One level deep. A dispatched sub-agent executes its own brief; further fan-out stays the orchestrator's job unless that agent's .claude/agents/*.md definition authorizes it.

NEVER parallelize: tasks sharing a write target · a task consuming a pending task's output · trivial single-file work (dispatch overhead > gain) · an order a skill or workflow explicitly fixes · gates awaiting user approval.

Blocked until: MUST ATTENTION every task tagged PAR/SEQ with a named reason per SEQ · waves declared + write-set disjointness checked · each wave spawned in ONE message · barrier honored before the next wave.

MANDATORY After planning tasks, tag each PAR/SEQ and spawn every PAR wave as parallel sub-agents in ONE message — default parallel for workflows, batch updates, investigation, research, reviews; plan execution fans out ONLY on what the plan declares.
MANDATORY Disjoint write sets per wave · all-return barrier before the next wave · specialist routing · sub-agents NEVER fan out further unless their own agent definition authorizes it.

Project Protocol Overlay — Before executing this skill, resolve any PROJECT overlay rules layered onto it: match this skill's name against the Target column of the project's skill-protocol index (docs/project-reference/skill-protocols-reference.md by default; a referenceDocs entry in docs/project-config.json overrides the path), taking the most specific matching tier ONLY — exact name > glob > *. That precedence orders overlays against EACH OTHER, never against this skill. Read ONLY the matched bodies, resolved as <protocols-dir>/<Name>.md; a row's Body link is display text, never a read path. A matched body that is missing or malformed is REPORTED and skipped — never reconstructed from the index Description. No index, or no match -> proceed with no overlay, silently. Full contract: .claude/skills/project-skill-protocol/references/registry.md.

Overlays are ADDITIVE ONLY: they ADD rules on top of this skill's own protocol and NEVER replace, override, disable, or reinterpret a rule it already states — removing every overlay must return this skill to exactly its documented behavior. An overlay is a BRIEF, not an authority escalation: it can NEVER waive a workflow gate, git discipline, a review gate, or a user-confirmation gate. A genuine overlay-vs-skill conflict, or two equally-specific overlays that directly contradict -> surface both to the user; NEVER resolve silently.

MUST ATTENTION resolve project protocol overlays for this skill BEFORE executing — most specific matching tier only (exact > glob > *, which ranks overlays against each other, NEVER against this skill), read only matched bodies at <protocols-dir>/<Name>.md; a missing or malformed body is reported, never reconstructed. Overlays are ADDITIVE ONLY (they never replace this skill's own rules) and are a brief, NEVER an authority escalation; an equal-specificity contradiction goes to the user.

Closing Reminders

IMPORTANT MUST ATTENTION Goal: Ensure every shipped performance fix removes a measured (or static-risk-labeled) bottleneck — across data access, in-process compute, runtime/GC, network round trips, client delivery, and concurrency/resilience saturation — with 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.

Protocols in force (concise digest of the SYNC/shared blocks this skill carries):

Critical Thinking: Traced file:line proof per claim; NEVER present a guess as fact.
AI Mistake Prevention: verify generated content against evidence, trace downstream references, verify all affected outputs, re-read after context loss, surface ambiguity.
Graph-Assisted Investigation: ALWAYS run a graph trace on key files when graph.db exists.
Severity Rubric: Classify by consequence; Critical/High block PASS until resolved.
Category Review Thinking: Derive per-category concerns from first principles, NEVER a fixed checklist.
Systematic Batching: Large changeset → size-capped parallel batches, then reduce.
Performance Knowledge (references/performance-knowledge.md): latency ladder · universal laws (Little, utilization knee, Amdahl, USL, tail amplification) · symptom→cause triage · network/DB/cache/web/memory-GC/distributed deep tables · measurement rigor. Calibrates severity; NEVER governs over local SLA/spec.
Parallel Sub-Agent Dispatch: Tag tasks PAR/SEQ, group PAR into disjoint-write-set waves, spawn each wave in ONE message, barrier before advancing.

IMPORTANT MUST ATTENTION run ALL 8 phases in order — Detect scope (+ symptom→cause triage) → Discover local context → Baseline evidence + anchor calibration → 12 serial dimension passes → Findings+Severity → Optimize plan → Why-Review validation gate → Validated-fix + full Phase-0 re-review; NEVER skip a phase or jump to a fix — why: AI forgets its own steps and ships unmeasured, unvalidated changes. IMPORTANT MUST ATTENTION cover ALL 12 dimensions one pass each — (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 — why: a single combined scan silently drops a dimension, and 10-12 are the layers a code-only reading habitually never opens. IMPORTANT MUST ATTENTION calibrate every number against a known anchor before assigning severity — latency ladder (1 ns → 100 ns → 100 µs → 10 ms → 100 ms, ~1 ms RTT per 100 km as a hard floor), utilization knee ~70-80% (wait ≈ service × ρ/(1−ρ); 90%→9×), Little's Law, tail amplification (fan-out to 100 backends hits a p99 ~63% of the time), CWV (LCP 2.5 s/INP 200 ms/CLS 0.1/TTFB 800 ms at field p75), cache hit-ratio math (90%→99% = 10× less origin load) — and treat a breached anchor as a HYPOTHESIS needing local proof, never a finding — why: an uncalibrated number cannot carry a severity, and a quoted constant with no local measurement is guess-as-fact. IMPORTANT MUST ATTENTION on any remote path count call count × RTT FIRST and verify connection reuse (keep-alive/pooled client) — why: per-request TCP+TLS handshakes and chatty contracts dominate paths where every individual handler is already fast. IMPORTANT MUST ATTENTION check the resilience controls as performance defects — decreasing timeout budget per hop, backoff + FULL JITTER + retry budget + idempotency keys, breaker/bulkhead/load-shedding, and a BOUND on every queue (watch age, not only depth) — why: these are invisible at low load and surface only as the outage they cause; an unbounded queue turns a throughput problem into unbounded latency and then OOM. IMPORTANT MUST ATTENTION report distributions not means (p50/p90/p99/p99.9 + max, segmented), NEVER average percentiles, always name the load model (open-model exposes queueing collapse, closed-model hides it), and flag suspected coordinated omission — why: the aggregate mean hides exactly the tail users complain about. IMPORTANT MUST ATTENTION apply the Performance-First Principles on every hot path — (1) [MOST IMPORTANT] hunt every OOM bad practice: bound every result set, reduce rows at the source, stream instead of buffer, triage row-count before row-size; (2) pick the data structure/algorithm that matches the access pattern (O(1) Set/Map over linear scan-in-loop, no needless O(n²)) and prove the complexity class at worst-case N; (3) batch per-item calls into one, else run bounded-parallel — never serial fan-out — why: unbounded memory OOMs the process, the wrong structure melts at scale, and serial fan-out multiplies latency. 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 — 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 — NEVER stop at 9 — why: split attention misses violations, and 10-12 are the layers a code-only reading habitually never opens. 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. | | "Every handler is fast, so the path is fast" | Count call count × RTT and check connection reuse. 12 avoidable round trips beat any handler micro-optimization. | | "Latency is high, optimize the code" | Check the RTT/geography floor first (~1 ms per 100 km) — physics and handshakes are not fixable in code; only moving the endpoint is. | | "Spikes are random / just noise" | Correlate against GC pauses, cold cache, deploys, and pool wait before calling anything random. Uncorrelated-with-load spikes are usually the runtime. | | "Added a cache, that's the fix" | Show the measured hit ratio AND the size/TTL bound. 60% hit ratio is barely a cache; unbounded is a leak. | | "Resilience is not a performance concern" | A missing timeout, jitterless retry, or unbounded queue IS a latency defect — same queues, and it converts partial failure into total. | | "Lighthouse score is green" | CWV verdicts come from field p75 (RUM/CrUX). Lab locates causes; field defines truth. | | "Only 9 dimensions matter, 10-12 are infra" | 10-12 (network, runtime/GC, resilience) are where code-only reviews are blindest. NEVER drop a dimension without an evidence-backed skip reason. | | "Anchor says it's slow, that's the finding" | An anchor breach is a hypothesis. Promote it with file:line + measurement or an explicit static risk label and verify command. | | "Digest is enough, skip the references body" | The digest orders hypotheses; only the body carries the thresholds. NEVER assign a severity or quote an anchor constant from memory or the digest — read it. |

[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); calibrate the number against a known anchor, and treat an anchor breach as a hypothesis, never a finding. IMPORTANT MUST ATTENTION walk ALL 12 dimensions one pass each — dimensions 10-12 (network/protocol, runtime/GC, distributed resilience) are the ones a code-only reading skips. 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-purpose

Each batch sub-agent receives: its full file list; SYNC:category-review-thinking as 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 to plans/reports/ (per SYNC: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.

Hookless Prompt Protocol Mirror (Auto-Synced)

Source: .claude/.ck.json + .claude/skills/shared/sync-inline-versions.md (:full blocks) + .claude/scripts/lib/hookless-prompt-protocol.cjs

[WORKFLOW-EXECUTION-PROTOCOL] [BLOCKING] Workflow Execution Protocol — MANDATORY IMPORTANT MUST CRITICAL. Do not skip for any reason.

Generic portability boundary: Reusable skills and protocol text stay project-neutral; project-specific conventions are discovered from docs/project-config.json and docs/project-reference/. Apply shared AI-SDD from shared/sdd-artifact-contract.md. Read docs/project-config.json and docs/project-reference/docs-index-reference.md, then open the project reference docs named there. For spec, test-case, behavior-change, public-contract, or docs/specs/ work, route through the local spec docs named by the docs index: feature-spec-reference.md, spec-system-reference.md, spec-principles.md, and workflow-spec-test-code-cycle-reference.md when specs/tests/code must stay synchronized. If either file or a required reference doc is missing or stale, auto-run $project-init (or the narrow lower-level route such as $project-config, $docs-init, $scan-all, or $scan --target=<key>) before ordinary project-specific work. Any supported AI tool may execute when this shared context and local docs are available.

DETECT: If the prompt starts with an explicit slash skill/workflow command, execute it directly. Otherwise match the prompt against the workflow catalog and skill list.
ANALYZE: Choose the best option: execute directly, invoke a skill, activate a standard workflow, or compose a custom step combination.
AUTO-SELECT: Pick the best option yourself. Do not ask the user to choose between direct execution, skill, standard workflow, or custom workflow.
ACTIVATE: For a selected workflow, call $start-workflow <workflowId>; for a selected skill, invoke that skill; for a custom workflow, sequence custom steps directly; for direct execution, proceed with the task.
CREATE TASKS: task tracking for ALL workflow/skill/custom steps before execution when the selected path has multiple steps.
PARALLELIZE: Before executing the task list, tag each task PAR (independent inputs + write set disjoint from every other PAR task) or SEQ (name the blocking dependency), group PAR tasks into waves, declare the wave plan, and spawn each wave's sub-agents in ONE message — all-return barrier per wave, fan-out one level deep unless a sub-agent's own definition authorizes further fan-out. Sequential-by-default is a defect when tasks are independent; do not parallelize shared write targets, output-consuming tasks, trivial single-file work, ordering a skill or workflow explicitly fixes, or user-approval gates.
EXECUTE: Advance per the Workflow Step Advancement & Parallel Phases rule in your context instructions — model-driven; a sub-agent completion advances a step identically to an inline call; a parallel-phase group is an all-return barrier (advance only after ALL members return, never serialize it)

Shared AI-SDD Protocol Markers

Source: .claude/skills/shared/sync-inline-versions.md

SYNC:ai-sdd-artifact-contract

AI-SDD Artifact Contract — Shared spec-driven development rules stay portable and source-owned.

Keep reusable AI-SDD principles in .claude; put repository-specific paths, commands, owners, products, and formats in project config/reference docs.

Preserve cycle: spec -> plan -> tasks -> implement -> verify -> update spec/docs.

Trace every requirement or invariant through decision, task, TC/test, source evidence, and docs/spec update.

Treat code-to-spec extraction as reference-only until accepted by the canonical spec owner.

Any supported AI tool may plan, implement, review, or verify with synced context; using multiple tools is optional.

Update .claude source first, then sync generated mirrors; do not manually edit .agents, .codex, or AGENTS.md. — why: mirrors are generated artifacts; hand-edits are overwritten on the next sync

If docs/project-config.json, root instruction files, or a required project-reference doc is missing or stale, auto-run $project-init or the narrow lower-level route before ordinary project-specific work.

Active reference: shared/sdd-artifact-contract.md in the active skills root.

SYNC:ai-sdd-artifact-contract:reminder

MANDATORY Apply shared/sdd-artifact-contract.md; keep reusable AI-SDD in .claude and local rules in project docs.
MANDATORY Code-to-spec extraction is reference-only until canonical acceptance; any supported AI tool may execute with synced context.
MANDATORY Update .claude source before syncing generated mirrors; do not manually edit .agents, .codex, or AGENTS.md.
MANDATORY Missing or stale project config, root instruction files, or required reference docs route project-specific work through $project-init or the narrow setup route automatically. [TASK-PLANNING] [MANDATORY] BEFORE executing any workflow or skill step, create/update task tracking for all planned steps, then keep it synchronized as each step starts/completes.

[LESSON-LEARNED-REMINDER] [BLOCKING] Task Planning & Continuous Improvement — MANDATORY. Do not skip.

Break work into small tasks (task tracking) before starting. Add final task: "Analyze AI mistakes & lessons learned".

Extract lessons — ROOT CAUSE ONLY, not symptom fixes:

Name the FAILURE MODE (reasoning/assumption failure), not symptom — "assumed API existed without reading source" not "used wrong enum value".
Generality test: does this failure mode apply to ≥3 contexts/codebases? If not, abstract one level up.
Write as a universal rule — strip project-specific names/paths/classes. Useful on any codebase.
Consolidate: multiple mistakes sharing one failure mode → ONE lesson.
Recurrence gate: "Would this recur in future session WITHOUT this reminder?" — No → skip $learn.
Auto-fix gate: "Could $code-review/$code-simplifier/$security-review/$lint catch this?" — Yes → improve review skill instead.
BOTH gates pass → ask user to run $learn. [CRITICAL-THINKING-MINDSET] Apply critical thinking, sequential thinking. Every claim needs traced proof, confidence >80% to act. Anti-hallucination principle: 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 Attention principle (Primacy-Recency): Put the 3 most critical rules at both top and bottom of long prompts/protocols so instruction adherence survives long context windows. Goal-driven execution: Define success criteria first, loop until verified, and stop only when observable checks pass. Tests verify intent: Tests must protect business rules/invariants and fail when the protected intent breaks, not only mirror current behavior.

Common AI Mistake Prevention (System Lessons)

Re-read files after context compaction. Edit requires prior Read in same context; compaction wipes read state. Re-read before editing.
Grep for old terms after bulk replacements. AI over-trusts find/replace completeness. Grep full repo after bulk edits for missed refs in docs/configs/catalogs.
Check downstream references before deleting. Deletions cascade doc/code staleness. Map referencing files before removal.
After memory loss, check existing state before creating new. Compaction wipes prior-work memory. Query current state to resume — never blindly duplicate.
Verify AI-generated content against actual code. AI hallucinates APIs, class names, method signatures. Grep to confirm existence before documenting/referencing.
Trace full dependency chain after edits. Changing a definition misses downstream consumers. Trace the full chain.
When renaming, grep ALL consumer file types. Some file types silently ignore missing refs (no compile error). Search code, templates, configs, generated files.
Trace ALL code paths when verifying correctness. Code existing ≠ code executing. Trace early exits, error branches, conditional skips — not just happy path.
Update docs that embed canonical data when source changes. Docs inlining derived data (workflows, schemas, configs) go stale silently. Update all embedding docs alongside source.
Verify sub-agent results after context recovery. Background agents may finish while parent compacted — grep-verify output, don't trust assumed completion.
Cross-check full target list against sub-agent assignments. Parallel sub-agents by category miss boundary items. Reconcile union of assignments against target list before proceeding.
Sub-agents inherit knowledge only from their agent .md definition — use custom agent types, not built-in Explore. Tool adoption = permission + knowledge + enforcement (numbered workflow step).
Persist sub-agent findings incrementally, not as a final batch. Long sub-agents hit cutoffs before final write — findings lost. Instruct append-per-section to report file.
When debugging, ask "whose responsibility?" before fixing. Trace caller (wrong data) vs callee (wrong handling). Fix at responsible layer — never patch symptom site.
Test failure → record a provisional verdict before trace/edit, then investigate. Use the full five-way taxonomy: SOURCE-WRONG (production violates intent), TEST-WRONG (assertion/setup is stale), TEST-NOT-OPTIMAL (valid but fragile or low-signal test), ENVIRONMENT-BLOCKED (external state prevents a verdict), or AMBIGUOUS (intent/evidence cannot choose safely). Then trace root cause and triangulate against the governing spec (docs/specs/** if one exists) AND source. NEVER weaken an assertion, add a skip, relax a timeout, or change source merely to force green.
Grep ALL removed names after extraction/refactoring. Primary file "done" ≠ secondary files clean. Grep entire scope for every removed symbol before declaring complete.
Assume existing values are intentional — ask WHY before changing OR flagging one as a defect. Pattern-matching as "wrong" skips context. Before changing or reporting any constant/limit/flag/cutoff: read comments, git blame, the CALLER's ordering (the guarantee that makes the value correct usually lives in code running immediately BEFORE the cited line), and 2+ sibling call sites of the same convention. A doc stating WHAT without WHY is missing rationale, not proof of a missing guard — and in a validation pass, an accurate file:line citation proves the transcription, never the defect.
Verify ALL affected outputs, not just the first. One build green ≠ all green. Multi-stack changes (backend/frontend/tests/docs) require verifying EVERY output.
Evaluate fit before copying a nearby pattern. Closest example ≠ matching preconditions — verify the new context shares the same constraints, base classes, scope, lifetime.
Holistic-first debugging — resist nearest-attention trap. Don't dive into first plausible cause. List EVERY precondition (config, env vars, paths, DB, endpoints, creds, versions, DI, data). Verify each against evidence (grep/query — not reasoning). Ask "what would falsify this?" — if nothing, it's not a hypothesis. Most expensive failure: going deeper in "obvious" layer while bug sits in layer never questioned.
Surgical changes — apply the diff test (context-aware). Two modes: (1) Bug fix → every line traces to the bug; no restyling; orphan cleanup only for imports YOUR changes made unused. (2) Review/enhancement → implement improvements AND announce as "Enhancement beyond main request: [what]". Never silently scope-creep. Diff test: "Would this line exist if I wasn't asked to do X?" — if no, delete or announce.
Surface ambiguity before coding — don't pick silently. Multiple valid interpretations → present each with effort: "[Request] could mean (1) [N h], (2) [N h]. Which matters?" List scope/format/volume/constraints assumptions first. If simpler path exists, say so. Never silently pick.
[MANDATORY FIRST ACTION] ALWAYS activate a suitable skill or workflow BEFORE responding. Match task against workflow catalog + skill list; invoke via skill invocation or $start-workflow <workflowId>. NEVER answer or write code before checking. Skip = protocol violation.
Why-Review adversarial mindset — apply when reviewing any plan, decision, or design. Default SKEPTIC not VALIDATOR: steel-man a rejected alternative, invert each stated reason ("what does it sacrifice?"), stress-test top 2-3 assumptions, run pre-mortem ("ships, fails in 3 months — what breaks?"), surface 1-2 alternatives author missed. Section presence ≠ quality; quality = causal reasoning + concrete mitigations + evidence, not "it's better" or "monitor closely".
Front-load report-write in sub-agent prompts for large reviews. Many-file sub-agents hit budget before final write — findings lost. Design prompts so: (1) report-write is first explicit deliverable, (2) append per-file/section (not batched), (3) scope bounded so reads don't exhaust budget. Truncated mid-sentence with no report file → spawn narrower scope, don't retry same prompt.
After context compaction, re-verify all prior phase outcomes before continuing. Summaries describe intent, not environment state (git index, filesystem, processes). On resume, FIRST audit: git status, re-read modified files, verify filesystem. Every "completed" claim is an untested hypothesis until evidence confirms.
OOM/memory: check row count before row size. Triage: (1) Unbounded query — no DB filter for trigger? Push filter to DB; eliminates OOM. (2) Large rows? Projection reduces proportionally. Row reduction > projection in ROI.
Assert the outcome your system OWNS, never the intermediate state your INFRASTRUCTURE owns. When testing anything asynchronous (queue/broker delivery, retries, background jobs, caches, replication), assert the final business/entity state. NEVER assert the delivery bookkeeping — consume/send status, attempt counts, last-error, row existence or counts in a broker, scheduler, or outbox/inbox table. That bookkeeping lives in shared infrastructure that ANY co-running process (a peer worker, a second replica, a leftover local container) can write, usually under a deterministic shared key, so the assertion silently tests the developer's environment instead of the system: green when run alone, flaky the instant anything else shares that broker + database. Gate question for every assertion: "would this hold no matter WHICH process did the work?" — if no, assert the converged data state instead. Corollary: process-local fault injection and in-process telemetry cannot gate work any process may perform — use them as stress amplifiers (arm → bounded window → disarm → assert convergence), never as preconditions.
Keep domain concepts out of generic/shared/infrastructure layers. Reusable layer (shared library, framework, infra module) must reference NO consumer-specific domain concept — tenant/customer/product IDs, business entities, feature rules. Leak compiles + runs → passes review silently while coupling the "reusable" layer to one consumer. Keep shared type domain-free; push domain fields/logic down into the consumer via subclass/composition. — why: a layer coupled to one consumer's domain is no longer reusable.

Agent Skills: performance-review

Install this agent skill to your local

Skill Files

Codex Project-Reference Loading (No Hooks)

Quick Summary

Phase 0: Detect Scope

Architecture-Altitude Performance Review

Phase 1: Discover Local Context

Phase 2: Baseline Evidence

Phase 3: Serial Dimension Passes

1. Query Shape And Data Minimization

2. Index, Access Path And Data Topology

3. N+1 And Fan-Out

4. Aggregation, Join, And Pipeline Shape

5. Materialization And Memory

6. Write Path, Locks, And Transactions

7. Cache And Reuse

8. API Payload, Frontend Delivery And Rendering

9. Compute And Algorithmic Complexity

10. Network And Protocol Efficiency

11. Runtime, Memory And GC

12. Distributed Resilience And Load Management

Phase 4: Findings And Severity

Phase 5: Optimize Plan

Sub-Agent Routing

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

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

Output

Closing Reminders

Hookless Prompt Protocol Mirror (Auto-Synced)

[WORKFLOW-EXECUTION-PROTOCOL] [BLOCKING] Workflow Execution Protocol — MANDATORY IMPORTANT MUST CRITICAL. Do not skip for any reason.

Shared AI-SDD Protocol Markers

SYNC:ai-sdd-artifact-contract

SYNC:ai-sdd-artifact-contract:reminder

[LESSON-LEARNED-REMINDER] [BLOCKING] Task Planning & Continuous Improvement — MANDATORY. Do not skip.

Common AI Mistake Prevention (System Lessons)