| name | harness-generator |
| description | Use when you need to create a new multi-agent harness for a domain not yet covered in this repository, or when asked to 'build a harness', 'design an agent team', or 'create agents for [domain]'. Analyzes the domain, selects one of 6 architectural patterns (Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, Hierarchical), then generates all agent .md files, SKILL.md files, and HARNESS.md following Copilot CLI file-bus conventions. Does NOT port existing Claude Code harnesses — use PORTING-NOTES.md and guides/porting-from-claude-code.md for that. Also triggers on: re-run harness design, update agent team, revise harness architecture, supplement harness with additional agents. |
| metadata | {"category":"meta-skill","type":"harness-generator"} |
Harness Generator
Generates a complete multi-agent harness from a domain description: agent definitions,
orchestrator skill, and HARNESS.md — all following this repository's file-bus conventions.
Execution Mode
Sequential Generation — Six phases produce all files needed for a working harness.
Read skills/harness-generator/references/copilot-adaptation-rules.md before starting.
Pre-Flight Checks
Phase 1: Domain Analysis
task(agent_type="general-purpose",
description="Analyze the requested domain and answer these questions:
1. What is the primary workflow? (linear / parallel / conditional branching)
2. What are the 4-6 specialist roles needed? Name each with a clear domain boundary.
3. What is the natural information flow between roles? Draw an ASCII DAG.
4. Which of the 6 patterns best fits the workflow:
Pipeline | Fan-out/Fan-in | Expert Pool | Producer-Reviewer | Supervisor | Hierarchical
Read skills/harness-generator/references/agent-design-patterns.md for definitions.
5. What domain knowledge frameworks are relevant? (e.g. OWASP, DORA, SOLID, TOGAF)
6. What are 3 concrete trigger phrases a user would say?
Create _workspace/ and _workspace/messages/ directories.
Write analysis to _workspace/00_domain_analysis.md with sections:
WORKFLOW_TYPE, SPECIALIST_ROLES, INFORMATION_FLOW_DAG, PATTERN_SELECTION (with rationale),
DOMAIN_FRAMEWORKS, TRIGGER_PHRASES.")
Phase 2: Team Architecture Design
task(agent_type="general-purpose",
description="Based on _workspace/00_domain_analysis.md, design the agent team:
- Harness number: check harnesses/ directory for next available nn- prefix
- Team name: {nn}-{domain-name} (kebab-case)
- Agent list: 4-6 agents with role, expertise, and output file
- Information flow: which agent outputs feed into which (with _workspace/ file names)
- Workspace file layout: 00_input.md through {n}_{final}.md
- Message handoff points: {from}-to-{to}.md for each transition
Read templates/orchestrator-skill-template.md and
skills/harness-generator/references/orchestrator-template.md for structure guidance.
Write to _workspace/01_team_architecture.md with sections:
HARNESS_ID, AGENT_TABLE (name | role | output_file | receives_from),
WORKSPACE_LAYOUT (full _workspace/ tree), PATTERN_IMPLEMENTATION.")
Phase 3: Agent File Generation
task(agent_type="general-purpose",
description="Generate all agent .md files based on _workspace/01_team_architecture.md.
For each agent, follow templates/agent-template.md EXACTLY:
1. YAML frontmatter: name, description (trigger-first: 'Use when...')
2. Identity: Role, Expertise, Output format
3. Core Responsibilities: 4-5 numbered items
4. Working Principles: Why-First format (no ALWAYS/NEVER — explain the reason instead)
5. Input Contract: list all _workspace/ files this agent reads
6. Output Contract: exact output file path and format
7. Message Protocol: exact _workspace/messages/{from}-to-{to}.md path and STATUS format
8. Domain Knowledge: embed relevant frameworks/standards/checklists
9. Quality Gates: self-validation checklist
Read skills/harness-generator/references/skill-writing-guide.md for best practices.
Read skills/harness-generator/references/copilot-adaptation-rules.md for Copilot CLI rules.
Write each agent file to harnesses/{nn}-{name}/agents/{agent-name}.md
Confirm no SendMessage() references — use file-bus only.")
Phase 4: Skill File Generation
task(agent_type="general-purpose",
description="Generate the orchestrator SKILL.md and 1-2 domain skill files.
For orchestrator SKILL.md (harnesses/{nn}-{name}/skills/{name}/SKILL.md):
Follow templates/orchestrator-skill-template.md EXACTLY:
1. trigger-first description with Does NOT cover + follow-up keywords
2. Execution Mode: File-Bus Team
3. Agent Composition table
4. Pre-Flight Checks
5. Phase-by-phase task() calls (each referencing the agent .md file)
6. Scale Modes table (Full / Reduced / Minimal)
7. Error Handling table
8. Test Scenarios (3 types)
9. Attribution section
For domain skills (1-2 supporting skills), embed relevant frameworks/checklists.
These go in harnesses/{nn}-{name}/skills/{skill-name}/SKILL.md
Validate all SKILL.md descriptions start with 'Use when'.
Read skills/harness-generator/references/qa-agent-guide.md for validation checklist.")
Phase 5: HARNESS.md Generation
task(agent_type="general-purpose",
description="Generate HARNESS.md for harnesses/{nn}-{name}/ with ALL of:
- H1 title: '{nn} — {Harness Name}'
- Overview paragraph (what it does, when to use)
- Structure section: directory tree with file descriptions
- Usage section: trigger phrases (3 examples)
- Workspace Artifacts: full _workspace/ layout with file descriptions
- Agent Team table: agent | role | output
- Scale Modes: Full Pipeline / Reduced / Minimal
- Quick Start: cp command + trigger phrase
- Attribution:
'Generated harness — not ported from harness-100.
Pattern: {Pattern Name} (see skills/harness-generator/references/agent-design-patterns.md)'
OR if adapted from harness-100:
'Adapted from [revfactory/harness-100](...) under Apache 2.0 License.'
Write to harnesses/{nn}-{name}/HARNESS.md")
Phase 6: Validation
task(agent_type="general-purpose",
description="Validate the generated harness against all quality gates:
1. File existence:
- harnesses/{nn}-{name}/HARNESS.md ✓
- All agent .md files listed in Agent Composition table ✓
- All skill SKILL.md files ✓
2. Content checks:
- Each agent has Input Contract + Output Contract + Message Protocol sections ✓
- Orchestrator SKILL.md description starts with 'Use when' ✓
- Orchestrator SKILL.md contains 'Does NOT cover' ✓
- Orchestrator SKILL.md contains task(agent_type ✓
- No SendMessage() in any file ✓
- HARNESS.md contains Attribution section ✓
3. Run: npm test
Report any test failures with file and line context.
If issues found: fix them directly. If ambiguous: report to user with options.")
Error Handling
| Error Type | Strategy |
|---|
| Ambiguous domain | Ask user for the 3 most important agent roles before continuing |
| Harness number conflict | Increment to next available number |
| Missing domain knowledge | Embed best available general knowledge; note limitation in HARNESS.md |
| Test failure | Fix the specific file causing the failure; re-run npm test |
