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.md

GitHub

AIWF 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

  1. Honor skills explicitly named by the user.
  2. Read project guidance and use references/routing-map.md for inferred routes.
  3. For broad prompts, run the advisory classifier:
python <this-skill>\scripts\route_prompt.py --prompt "<user request>"
  1. 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.
  2. Announce the route and reason in one line.
  3. Read every selected downstream SKILL.md completely before editing, running services, starting expensive work, or publishing.
  4. Stop repeated attempts after AIWF_MAX_LOOPS_WITHOUT_PROGRESS loops 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-electrician after 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-guardrails plus 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-convergence with the repository and domain owners.
  • External research: aiwf-deep-research; durable plan verification: aiwf-agent-mok only 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_DEFAULT is deliberately enabled.
  • Pass AIWF_DEEP_RESEARCH_EXTRA_URLS into 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.