| name | crow-pet-orchestrator |
| description | Route crow.pet tasks into records, policy, adapters, subagents, evaluation, and memory. |
crow.pet Orchestrator
Use this skill as the main crow.pet control plane.
Workflow
- Read the incoming request.
- Create a task record from
templates/task-record.md.
- Classify risk with
policies/risk-matrix.md.
- Apply approval rules from
policies/approval-policy.md.
- Route sensitive or state-changing actions through
crow-pet-policy-gate.
- Route ambiguous build requests through
crow-pet-spec-writer.
- Route approved spec-first builder work or workflow run instances through
crow-pet-work-queue.
- Route generated-agent, multi-harness, cross-harness critique, or persistent role decisions through
crow-pet-agent-assembly.
- Route new, changed, composed, compiled, or repeated workflows through
crow-pet-skill-architect.
- Select the specialist skill and adapter.
- Use the Capability Broker when the task asks for semantic abilities before a concrete adapter is known.
- Attach a bounded context capsule when memory is needed.
- Launch the subagent with
workflows/subagent-launch.md.
- Collect the handoff report.
- Evaluate the result.
- Record useful events in the Ledger when coordination should be searchable.
- Route learning through the Loop before Memory Garden promotion.
- Send a concise status reply.
Progressive Disclosure
Load templates/task-record.md when creating a task record.
Load policies/risk-matrix.md and policies/approval-policy.md before policy decisions.
Load ../crow-pet-policy-gate/SKILL.md when a task can read sensitive context, write files, call tools, publish, spend money, change external systems, or mutate crow.pet behavior.
Load ../crow-pet-spec-writer/SKILL.md when the user wants something built and the output contract, constraints, acceptance criteria, or builder surface need to be made explicit.
Load ../crow-pet-work-queue/SKILL.md when an approved SPEC.MD or workflow definition should become builder-ready queued work.
Load ../crow-pet-agent-assembly/SKILL.md when the task needs multiple agents, generated agents, cross-harness critique, or a persistent agent-profile decision.
Load ../crow-pet-skill-architect/SKILL.md when the task creates, changes, composes, compiles, or hardens a skill, adapter-backed workflow, repeated procedure, or multi-stage agent pattern.
Load workflows/subagent-launch.md only when delegation is needed.
Load ../../adapters/patterns/capability-broker.md only when a skill or role brief declares semantic capabilities such as web_research, browser_control, deploy_inspect, repo_read, repo_search, file_discovery, json_query, visual_understanding, or external_write.
Load ../../adapters/patterns/tool-capability-profile.md and ../../templates/tool-capability-profile.md only when a requested tool capability is repeated, ambiguous, permission-sensitive, or shared across runtimes.
Load ../../adapters/patterns/bus.md, ../../adapters/memory/ledger.md, and ../../adapters/patterns/loop.md only when agent communication, event search, or curated learning affects the task.
Load templates/context-capsule.md only when memory affects execution.
Load contract.md and rubric.md only for acceptance or evaluation.
Dependency Edges
routes_to: crow-pet-spec-writer, crow-pet-work-queue, crow-pet-agent-assembly, crow-pet-skill-architect, crow-pet-codex-delegate, crow-pet-claude-desktop-delegate, crow-pet-cicd, crow-pet-marketing, crow-pet-image, crow-pet-self-improvement, crow-pet-crow-voice, crow-pet-policy-gate.uses_adapter: channel, memory, pattern, harness, and tool adapters selected by task, including Capability Broker, Tool Capability Profiles, Bus, Ledger, and Loop when needed.loads_context: context capsule only when task-relevant memory is needed.evaluates_with: contract.md, rubric.md.
Rules
- Keep skills runtime-neutral.
- Use adapters for runtime-specific behavior.
- Use the Capability Broker for semantic capability requests before choosing browser, MCP, CLI, API, local app, tool, harness, or LLM capability surfaces.
- Use Tool Capability Profiles for reusable or sensitive tool mappings; profiles inform binding but never grant permission.
- Treat Unix-style commands as adapter bindings for local inspection, not as canonical skill requirements.
- Preserve parallel read, sequenced write for workflow orchestration.
- Preserve Crow-mediated agent interfaces: brief in, handoff out.
- Use Bus messages and Ledger events as evidence, not memory.
- Treat generated agents as disposable unless Self-Improvement promotes an agent profile.
- Use memory adapters for selective recall.
- Ask before high-risk actions.
- Preserve audit trails.
- Stop when policy blocks progress.
