Skill Automatic Trigger and Composition Matrix¶
Status:
active. This page describes current package routing only. Earlier source versions and migration mappings do not define current behavior.
0. Minimum Sufficient Loading¶
PRIMARY_DOMAIN_SKILL_LIMIT = 1
DEFAULT_SUPPORTING_SKILL_LIMIT = 2
MAX_ACTIVE_SKILLS_WITHOUT_JUSTIFICATION = 4
- Select one primary domain Skill for each phase.
- Add at most two supporting Skills by default.
- Activate workflow Skills by phase; do not preload every delivery, review, documentation, and memory rule during initial analysis.
- More than four active Skills require a unique responsibility for each.
- Remove Skills from the active set when their phase ends.
- Explicit requester choices take priority but cannot expand permissions or bypass project rules.
Typical phases:
Analysis: domain or log Skill
-> Design: preimplementation review when needed
-> Implementation: domain + quality delivery
-> Stable diff: postimplementation review
-> Formal report: documentation Skill
-> Cross-session: long-running memory
1. Typical Single-Skill Triggers¶
| Skill | Should Trigger | Normally Should Not Trigger Alone |
|---|---|---|
backend-engineering |
“Analyze a Spring transaction, FastAPI async chain, Fastify API, or Gin service.” | A pure Vue styling or SQL-plan task |
frontend-engineering |
“Repair the request race after route navigation.” | Pure database-index analysis |
ai-engineering |
“Validate model schemas, RAG access, and agent tool calls.” | An ordinary backend bug with no model behavior |
data-middleware-infrastructure |
“Analyze a Redis hot key, database lock, or GPU resource issue.” | An ordinary server null dereference |
log-observability-analysis |
“Analyze these local logs and build a cross-service timeline.” | Refactoring without observability input |
engineering-quality-delivery |
“Modify, test, review, and commit locally.” | Explain one concept |
multi-agent-independent-review |
“Have several Reviewers comprehensively review the diff and reduce rework.” | Wording without behavior change |
technical-document-writing |
“Write a formal architecture design from the repository.” | Change one commit message |
long-running-task-memory |
“The task spans days; record each recoverable node.” | A one-session small repair |
2. Recommended Combinations¶
| Task | Recommended Combination |
|---|---|
| Backend bug repair in any language | General backend + quality delivery; add multi-agent review for medium/high risk |
| Backend plus Redis/MQ repair | General backend + data infrastructure + quality delivery; add review for high risk |
| AI, RAG, or GPU Worker incident | AI + logs; add backend or data infrastructure for the actual chain and delivery for changes |
| End-to-end frontend/backend SSE repair | Frontend + general backend + data infrastructure + quality delivery; add review for high risk |
| Architecture document from code | Documentation + actual technology Skill |
| Modify code and formal design | Technology + quality delivery + documentation |
| Large cross-session change | Technology + quality delivery + memory; add review after code stabilizes |
| Production deployment runbook | Documentation + data infrastructure + quality delivery |
| Comprehensive read-only review | Multi-agent review + technology Skill |
| Multi-day production observation | Logs + memory + domain; add quality delivery only when switching to writes |
3. Log-Analysis Trigger Tests¶
Should Trigger¶
- Analyze multiple application logs and compressed archives in a local directory.
- Build an incident timeline from Docker, Kubernetes pod, and Nginx logs.
- Read-only analysis of production Java service, connection-pool, and RabbitMQ logs.
- Correlate services by trace ID and distinguish root cause from symptoms.
- Cluster exceptions, count frequency, and propose validation steps.
Normally Should Not Trigger¶
- Pure code design with no observability data.
- README punctuation only.
- A new business feature without incident evidence.
- A theoretical concept unrelated to current logs.
Mode Selection¶
- Local static files: bounded decompression and temporary parsing, without overwriting originals.
- Local runtime: read-only by default; restart and modification separately authorized.
- Remote nonproduction: bounded command cost and no assumed write permission.
- Production read-only: bounded time, lines, and query cost; no cleaning, restart, deployment, or writes.
4. Multi-Agent Review Trigger Tests¶
Should Trigger¶
- The diff affects authorization, public APIs, databases, and historical data; run comprehensive multi-agent review.
- After repair, independently check function, regression, security, performance, data, and concurrency.
- Find issues once, attribute consistently, and repair centrally to reduce rework.
- The diff crosses backend, frontend, and messaging and benefits from parallel read-only review.
- A data migration is about to be committed and requires strict independent review.
Normally Should Not Trigger¶
- Markdown punctuation only.
- Explain an exception without changing code.
- Split existing commits without file changes.
- One local self-check explicitly requested for a low-risk nonbehavioral change.
Selection Rules¶
- Low risk: 0–1 Reviewer, normally
economyand Luna. - Medium risk: 1–2 Reviewers, normally
balanced. - High risk: 2–3 Reviewers, normally
deep; only a critical dimension may use Terra High. - Do not add duplicate responsibilities to meet a count.
- Do not modify before every first-round result returns.
- Defaults: parallel 3, total 6, two post rounds, two repair rounds, one Terra High; budget model, people, context, and rounds together.
5. Long-Running Memory Trigger Tests¶
Should Trigger¶
- Work will span sessions or days.
- The task requires updating task and progress documents after each small node.
- Multiple agents work in parallel and context compaction must remain recoverable.
- A production rollout requires continuous observation and recorded validation state.
- The conversation is long with multiple implementation phases remaining.
Small-Node Checkpoints¶
Update CURRENT_TASK.md and PROGRESS.md after:
- forming a call-chain conclusion;
- confirming or excluding a root cause;
- completing one functional-boundary change;
- completing a build, test, migration, or sample validation;
- dispatching, collecting, or consolidating one Reviewer round;
- completing centralized repair or targeted rereview;
- a blocker, scope change, commit, deployment, or pause.
One ls, grep, inconclusive read, or immediately reverted experiment does not need a checkpoint.
6. Documentation Skill Trigger Tests¶
Should Trigger Implicitly¶
- Write a system architecture design from the current repository.
- Restructure an existing Markdown technical proposal while preserving business definitions.
- Produce a database schema and index design.
- Write an incident analysis from logs and source.
- Prepare a formal management discussion report.
- Write deployment, staged-rollout, and rollback operations.
- Write API design and error-code documentation.
Normally Should Not Trigger Alone¶
- Add one comment to a Java method.
- Change a commit message to Chinese.
- Update one CHANGELOG entry.
- Explain Redis cache breakdown.
- Run
npm run buildand report the result.
7. Real Codex Routing Regression¶
Package resources:
tests/skill-routing-cases.json: required, optional, forbidden, and maximum active Skills;scripts/routing-eval.py: schema validation, observation templates, and scoring.
Usage:
python3 scripts/routing-eval.py validate
python3 scripts/routing-eval.py make-template --output routing-observations.json
Send each test prompt in Codex and record the actual displayed or reported Skills; never fill results from expectations. Then:
python3 scripts/routing-eval.py evaluate --results routing-observations.json
Package validation proves the cases and tool are valid, not that local Codex activated them. Rerun after changing a Skill name, description, global routing rule, or composition boundary.
Reviewer Isolation Scheduling¶
- TOML
read-onlyexpresses configuration intent only. - A
danger-full-accessorworkspace-writeparent normally permits onlylogical-readonlyclaims. - High-risk, production, and strict read-only tasks require an entirely read-only parent or valid system-isolation evidence.
- Self-review is not an independent Reviewer; logical read-only is not system isolation.
8. Four Primary Domains¶
| Scenario | Primary Skill | Optional support |
|---|---|---|
| Java, Python, Node.js, Go, .NET, Rust, or another server application | backend-engineering |
Quality delivery for changes; data infrastructure for data components |
| Browser, WebView, Renderer, and frontend frameworks | frontend-engineering |
Backend, AI, or data according to the actual call chain |
| Model calls, RAG, agents, AI evaluation, inference, and multimodal generation | ai-engineering |
Backend for SDK/Worker mechanics; data infrastructure for vector/GPU resources |
| Databases, caches, messaging, search/vector storage, files, GPU resources, containers, and networks | data-middleware-infrastructure |
Backend for application callers; AI for AI semantics |
Choose one primary domain per phase. Python does not imply AI, an AI project name does not imply the current task is AI work, and package.json does not imply frontend. A pure Node.js server routes to general backend.
Detailed frontend-domain routing¶
| Scenario | Primary Skill | Supporting Skill |
|---|---|---|
| Vue/Nuxt, React/Next/Remix, Preact, Angular, Svelte, Astro/Ember, or legacy/static pages | frontend-engineering |
Quality delivery when modifying |
| Server logic in SSR/full-stack frontend | frontend-engineering |
Relevant backend and data infrastructure |
| Microfrontend/monorepo | frontend-engineering |
Quality/documentation according to the change |
| Hybrid Web, WebView, or renderer | frontend-engineering |
Review native bridge, main process, and system capability separately |
| Electron/Tauri main process or native mobile | Do not use frontend-engineering |
Choose system/backend/security capabilities |
| Pure Node.js backend API/worker | Do not use frontend-engineering |
backend-engineering; add data infrastructure for databases or messaging |