compose

name	compose
description	Interactive workflow builder — helps create workflow.yaml files for the harness orchestrator. The ONE interactive entry point. Use when user wants to build a harness workflow, create a pipeline, or compose subsystems together.
user-invocable	true

Harness: Compose Workflow

Help the user build a workflow.yaml for the orchestrator to execute. This is the interactive counterpart to the headless /harness:run.

Step 1: Understand the goal

Ask the user what they want to accomplish. Common patterns:

• Build + Review: Implement features in parallel, then review
• Build + Optimize: Build features, then optimize the hot path
• Optimize only: Run a keep/revert optimization loop on an artifact
• Research: Search literature for a specific question
• Triage: Diagnose and fix an issue
• Full pipeline: Build → Review → Optimize → Research

Step 2: Pick subsystems

Every phase has plugin: harness, plus a type: and (for subagent phases) an agent: selecting which subsystem handles the work:

Subsystem	Phase types	Agents	Purpose
Build	agent-teams, subagent	lead, implementer, reviewer, code-reviewer	Parallel feature building with TDD + review
Optimize	loop	optimizer, advisor	Keep/revert optimization on any artifact + metric
Research	subagent	searcher, synthesizer, method-analyst, implementation-guide, architecture-evaluator	Literature search + synthesis
Triage	subagent	diagnostician, fixer, verifier	Issue diagnosis + auto-fix

Also check hardware targets for optimization phases:

cat "${HOME}/.harness/config.json" 2>/dev/null || echo "NO_CONFIG"

Step 3: Build phases interactively

For each phase the user wants, collect the required config:

Build phase (agent-teams)

• Branch name for the feature work
• Number of parallel teammates (1-5, default 3)
• Tasks: collect user stories inline or point to a tasks file
• Whether to run code review after

Optimize phase (loop)

• Which file to optimize (the artifact)
• What metric to measure
• Whether lower or higher is better
• The command to run
• Time budget per experiment
• Max experiments
• Target: local or server (if server, need ssh_host and remote cwd)

Research phase (subagent)

• The research question
• Which sources to search (arxiv, semantic-scholar, hf-papers, perplexity)
• Max papers to include

Triage phase (subagent)

• Issue URL or error log
• Whether to create a PR for the fix

Step 4: Wire dependencies and conditions

• Ask about phase ordering and dependencies
• Suggest conditions (e.g., "only research if optimization improved >10%")
• Add depends_on and condition fields

Step 5: Set defaults

Ask about quality commands:

defaults:
  quality:
    test: "npm test"        # or "pytest", "cargo test", etc.
    build: "npm run build"  # or "cargo build", etc.
    typecheck: "tsc --noEmit"  # if TypeScript

Step 6: Generate and write

Assemble the full workflow.yaml and write it:

mkdir -p .harness

Write the file to .harness/workflow.yaml.

Show the user the complete workflow and explain:

• How to run: /harness:run
• How to monitor: check .harness/ state files or the dashboard
• How to resume: just re-run /harness:run — it picks up where it left off

Example Output

name: build-and-optimize
description: "Build user management API, then optimize the hot path"

defaults:
  quality:
    test: "npm test"
    build: "npm run build"
    typecheck: "tsc --noEmit"

phases:
  - id: build
    plugin: harness
    type: agent-teams
    config:
      branch: "feat/user-api"
      max_teammates: 3
      code_review: true
      tasks:
        - id: "US-001"
          title: "Set up project structure"
          description: "As a developer, I want the Express project scaffolded with TypeScript"
          criteria:
            - "package.json has express dependency"
            - "src/ directory structure exists"
            - "TypeScript configured"

  - id: review
    plugin: harness
    type: subagent
    agent: reviewer
    depends_on: [build]

  - id: optimize
    plugin: harness
    type: loop
    depends_on: [review]
    config:
      artifact: "src/api/handler.ts"
      metric: "p99_latency_ms"
      direction: lower
      run_command: "npm run benchmark"
      time_budget: "2m"
      max_experiments: 20