AgentSkillsCN

Orchestrator

编排者

SKILL.md

Feature Orchestrator Skill

When to use

Use this skill to execute a complete feature through the workflow, automatically coordinating agents and subagents. The Orchestrator reduces manual overhead by managing the entire feature lifecycle.

How to invoke

  • Slash command: /orchestrate or /run-feature or /orchestrate-feature
  • Example: /orchestrate feature F-010
  • Example with parallel: /orchestrate feature F-010 --parallel
  • Or mention: "run feature F-010" or "orchestrate feature F-010"

Context required

  • Feature ID from FEATURES.md
  • Feature must be defined with clear scope and acceptance criteria
  • Architecture decisions (ADRs) must be available if needed

What this skill does

1. Analysis Phase (main agent)

  • Reads FEATURES.md to locate the feature
  • Determines feature complexity
  • Identifies parallelization opportunities
  • Creates execution plan
  • Determines if feature has independent components

2. Implementation Phase

Simple Feature (Sequential):

  • Invokes /coder skill
  • Waits for implementation completion

Complex Feature (Parallel Subagents):

  • Spawns subagents for independent components
  • Each subagent uses /coder skill
  • Coordinates implementation completion
  • Ensures all implementation reports are filed

3. Review Phase (always sequential)

  • Invokes /reviewer skill
  • Waits for approval before proceeding
  • If changes required, returns to implementation phase

4. Test Phase (always sequential)

  • Invokes /tester skill
  • Validates feature completion
  • If tests fail, returns to implementation phase

5. Documentation Phase (conditional)

  • If documentation needed, spawns /docs subagent
  • Can run in parallel with next feature if appropriate
  • Otherwise runs sequentially after testing

Execution modes

Mode 1: Sequential (default)

code
Main Agent → /coder → /reviewer → /tester → /docs (if needed)

Mode 2: Parallel (with --parallel flag)

For features with independent components:

code
Main Agent
  ├─> Subagent 1: /coder (Component A)
  ├─> Subagent 2: /coder (Component B)
  └─> Subagent 3: /docs (Documentation)
  
Then sequential:
  → /reviewer (integration review)
  → /tester (end-to-end validation)

State management

  • All state updates go to FEATURES.md
  • All reports go to /reports/
  • Orchestrator tracks progress but doesn't modify feature status directly
  • Each agent skill updates status according to WORKFLOW.md
  • Artefact-driven state is maintained throughout

Constraints

  • Never skip quality gates (review/test)
  • Always maintain artefact-driven state
  • Subagents must produce reports, not just code
  • Quality gates (review/test) are always sequential
  • Implementation can be parallelized if components are independent

Outputs

  • Complete feature execution through workflow
  • All required reports filed in /reports/
  • Feature status updated in FEATURES.md
  • Execution summary with outcomes

Related skills

  • Individual agent skills can be used for manual control
  • /planner - If feature needs definition first
  • /architect - If architectural decisions are needed

Boot sequence (automatic)

When invoked, this skill automatically:

  1. Reads FEATURES.md to locate feature
  2. Reads WORKFLOW.md to understand workflow rules
  3. Analyzes feature complexity
  4. Creates execution plan
  5. Begins execution

Definition of Done (Orchestrator Perspective)

Orchestration is complete when:

  • Feature status is done in FEATURES.md
  • All required reports are filed
  • All quality gates have passed
  • Documentation is updated if needed

Error handling

  • If any phase fails, orchestrator reports the issue
  • Returns control to user with clear next steps
  • Maintains state in FEATURES.md
  • Does not attempt to fix issues automatically