aiwf-orchestrator
Implicit routing control for the AIWF skill pack. Use at the start of non-trivial local AI, coding, security, startup growth, funding, platform, model, dataset, robotics, embedded, web, research, mult
SKILL.mdskills/aiwf-orchestrator/SKILL.md
GitHubAIWF Orchestrator
Control Policy
AIWF_MAX_DOWNSTREAM_SKILLS: 4
AIWF_MAX_AGENT_SPAWN: 3
AIWF_MAX_LOOPS_WITHOUT_PROGRESS: 2
AIWF_AI_AVOIDANCE_LEVEL: 1.0
AIWF_DEEP_RESEARCH_EXTRA_URLS: ""
AIWF_TORCHIE_DEFAULT: false
These are agent-read policy values. They guide this skill when it is loaded; they do not reconfigure the Codex host, model, context window, or tool limits. Provider-specific settings remain controlled by the host.
Core Rule
Route first, then load only the skills that materially change the work. This orchestrator is the pack's only implicitly invoked skill; downstream skills are explicit routes so they do not compete for every prompt.
Workflow
- Honor skills explicitly named by the user.
- Read project guidance and use
references/routing-map.mdfor inferred routes. - For broad prompts, run the advisory classifier:
python <this-skill>\scripts\route_prompt.py --prompt "<user request>"
- Select no more than
AIWF_MAX_DOWNSTREAM_SKILLS. Prefer the skill that owns the first risky or irreversible decision, then add only real cross-boundary owners. - Announce the route and reason in one line.
- Read every selected downstream
SKILL.mdcompletely before editing, running services, starting expensive work, or publishing. - Stop repeated attempts after
AIWF_MAX_LOOPS_WITHOUT_PROGRESSloops with no new evidence, useful diff, or passing check. Change the route or report the exact blocker.
Route Shape
Use focused skills instead of the retired broad wrappers:
- Repository preflight and diff integrity:
aiwf-repo-sentinel. - C, C++, Python 3.10, Python 3.12, CUDA, FastAPI, Gradio, React, Vue/VitePress, CSS, JavaScript, TypeScript, or Android/Kotlin: the matching focused coding skill.
- Cross-layer API/UI wiring:
aiwf-ui-electricianafter the relevant backend and frontend skills. - Model loading, training, evals, serving, pipelines, GPU failures, RAG, or data stores: the matching runtime skill.
- Robotics, physics, networking, embedded systems, field pilots, or Windows local runtime work: the matching systems skill.
- Security work:
aiwf-security-guardrailsplus at most two focused security owners, leaving room for the repo or domain skill. - Startup marketing, funding, AEO/GEO, Meta, Google Ads, or YouTube/AdSense: the matching growth or platform skill.
- Shared Codex, Claude, Grok, context, handoff, or resource coordination:
aiwf-multi-agent-workspace. - Repeated QA, bug passes, test-fix loops, or severity-based convergence:
aiwf-qa-convergencewith the repository and domain owners. - External research:
aiwf-deep-research; durable plan verification:aiwf-agent-mokonly when needed. - Design, generated-image QA, technical SEO, release hygiene, service intake, or Torchie voice: the matching output skill.
If a task spans more than four owners, route in phases instead of loading the whole pack.
Gates
- No large model or dataset downloads unless asked.
- No training, GPU-heavy generation, or long benchmarks unless asked.
- No public network exposure, firewall changes, or credential changes unless asked.
- No destructive data, filesystem, device, database, or git operations unless asked.
- No major dependency upgrades, package-manager switches, broad rewrites, weakened tests, or security-boundary reductions as hidden side effects.
- Keep Torchie off unless Shawn asks or
AIWF_TORCHIE_DEFAULTis deliberately enabled. - Pass
AIWF_DEEP_RESEARCH_EXTRA_URLSinto a deep-research source plan when it is non-empty.
Output
Use:
Route: aiwf-orchestrator -> <skill names>; reason: <short reason>
For routing-only questions, return the route. For implementation work, continue through the selected skills and report actual validation.