skill-orchestrator

name

skill-orchestrator

description

Orchestrate multi-skill pipelines through the CLI. Use when "run the creation pipeline", "execute the improvement workflow", "resume my pipeline", or when chaining multiple skills with decision points. Supports Creation, Improvement, and Library Management pipelines with state persistence and conditional branching. Do not use for single skill operations (use the skill directly) or exploratory tasks without defined workflow.

Purpose

Execute documented skill pipelines (Creation, Improvement, Library Management) automatically, with state persistence for resume capability and conditional branching based on evaluation results.

When to use

• Run a full Creation pipeline on a new skill
• Execute Improvement pipeline on existing skills
• Automate Library Management workflows
• Resume interrupted pipelines
• Chain multiple skills with decision points
• User says "run the pipeline", "execute workflow", "resume my pipeline"

When NOT to use

• Single skill operation needed (use that skill directly)
• Task is exploratory or undefined
• Real-time user input is required at every step

Procedure

1. Pipeline Selection

Identify the target pipeline. See references/pipeline-definitions.md for detailed specifications of all 3 pipelines.

Available Pipelines:

• Creation Pipeline (9 phases)
• Improvement Pipeline (4 phases)
• Library Management Pipeline (2 phases)

2. CLI Interface

Create a new pipeline:

./scripts/meta-skill-studio.py --mode cli --action create --brief "Create a skill for..."

Run a pipeline:

./scripts/meta-skill-studio.py --mode cli --action run --pipeline creation --target <skill-name>

Resume a halted pipeline:

./scripts/meta-skill-studio.py --mode cli --action resume --run-id <uuid>

3. Configuration Phase

Load or create pipeline configuration:

{
  "pipeline_id": "uuid-v4-string",
  "pipeline_type": "creation|improvement|library-management",
  "target_skill": "skill-name",
  "start_time": "2026-04-14T10:00:00Z",
  "phases": [
    {
      "phase_id": 1,
      "skill": "skill-creator",
      "status": "pending|running|completed|failed",
      "input": {},
      "output": {},
      "exit_code": null,
      "decision_branch": null
    }
  ],
  "current_phase": 0,
  "resume_from": null
}

4. Execution Phase

For each phase in the pipeline:

1. Load State: Check for existing state file at .meta-skill-studio/runs/{run_id}/state.json
2. Execute Skill: Run the target skill via CLI action
3. Capture Output: Store stdout, stderr, exit code, artifacts
4. Evaluate Results:
   • If exit code != 0 → Mark failed, optionally retry or halt
   • If exit code == 0 → Mark completed, check for decision branch
5. Conditional Branching: See Decision Rules below
6. Save State: Persist progress after each phase
7. Log Progress: Write to output panel

5. Decision Rules

Quality Score Branching:

If eval_score < 60:
  Insert skill-improver before next phase
  Set improvement_goal = "Address quality score {eval_score}"
  
If trigger_precision < 0.80:
  Insert skill-trigger-optimization
  Set focus = "description and trigger boundaries"

Failure Handling:

If phase fails (exit_code != 0):
  Option 1: Retry with same input (max 2 retries)
  Option 2: Halt pipeline, mark failed
  Option 3: Skip and continue (if non-critical)

6. State Persistence

State files stored at .meta-skill-studio/runs/{run_id}/:

• state.json - Current execution state
• log.md - Human-readable log
• final-report.json - Completion summary

7. Resume Capability

To resume a halted pipeline:

1. Read state file
2. Identify last completed phase
3. Set current_phase = last_completed + 1
4. Continue execution from that point

8. Completion

When all phases complete:

1. Generate final report with:
   • Total phases executed
   • Success/failure counts
   • Final quality scores
   • Generated artifacts list
2. Update skill lifecycle state if applicable
3. Clean up or archive state files (configurable)

Output Contract

Success Output

{
  "pipeline_id": "uuid",
  "status": "completed",
  "phases_executed": 9,
  "phases_successful": 9,
  "phases_failed": 0,
  "final_quality_score": 85,
  "artifacts": [
    "skills/skill-name/SKILL.md",
    "skills/skill-name/evals/trigger-positive.jsonl",
    ...
  ],
  "report_path": ".meta-skill-studio/runs/uuid/final-report.json"
}

Failure Output

{
  "pipeline_id": "uuid",
  "status": "failed",
  "failed_at_phase": 3,
  "failed_skill": "skill-evaluation",
  "error": "Exit code 1: validation failed",
  "resume_possible": true,
  "state_file": ".meta-skill-studio/runs/uuid/state.json"
}

Failure handling

Problem	Response
Pipeline config invalid	Report specific error, halt before execution
Phase fails with exit code != 0	Retry up to 2 times, then halt or skip based on config
State file corrupted	Report error, cannot resume — start fresh
Cannot resume (skill changed)	Warn user, suggest fresh pipeline run
Missing required output from previous phase	Halt, report missing artifact

References

• references/pipeline-definitions.md - Detailed pipeline specifications
• references/eval-artifact-schema.md - Eval artifact formats
• references/conditional-branching-rules.md - Decision tree documentation

Next steps

After pipeline completion:

• Review created skill → manual review of SKILL.md
• Install to local library → skill-installer
• Evaluate routing quality → skill-evaluation
• For failed phases → skill-improver or skill-trigger-optimization