agent-harness（Agent 线束）

Agent Harness

Subcommand Routing

When invoked as /agent-harness <subcommand>, route to the appropriate workflow:

Usage examples:

• /agent-harness init "My Project" - Initialize new project
• /agent-harness sprint "Add authentication" - Plan sprint
• /agent-harness code - Start coding session
• /agent-harness 996 - Execute parallel coding with subagent orchestration
• /agent-harness status - Show status

CRITICAL: Always operate on files in the current project directory, NOT in the skill directory. At the start of every workflow, run python3 scripts/resolve_project_dir.py to get the correct absolute path. Use this path for ALL reads/writes of features.json and progress.md.

Framework for executing complex, multi-session projects using Sprint-Coding agent cycles.

Overview

Agent Harness enables Claude to manage long-running projects by:

• Sprint Agent: Plans features and breaks down requirements
• Coding Agent: Implements features one at a time
• 996 Agent: Orchestrates parallel feature implementation with subagents
• Progress Tracking: Maintains context across sessions
• Quality Gates: Ensures working code at each step

Commands

Agent Harness provides slash commands with tab completion. Type /agent-harness to see all commands.

/agent-harness

Show available commands and usage.

Output:

Agent Harness Commands:
  /agent-harness init <name>           - Initialize new project tracking files
  /agent-harness sprint [req]          - Create or update sprint with feature breakdown
  /agent-harness code                  - Start coding session for next feature
  /agent-harness 996                   - Execute parallel coding with subagent orchestration
  /agent-harness status                - Show current project status
  /agent-harness archive               - Archive completed sprints
  /agent-harness force-archive         - Force archive ALL sprints (including incomplete)

/agent-harness init <name>

Initialize new project tracking files.

Example: /agent-harness init My App

Actions:

• Create features.json and progress.md
• Run python3 scripts/init_project.py "My App"
• Commit the initialized files: git add features.json progress.md && git commit -m "chore: initialize project tracking"

/agent-harness sprint [requirements]

Create or update sprint with feature breakdown.

Examples:

• /agent-harness sprint - Review current sprint and plan next
• /agent-harness sprint Add user authentication with email and social login
• /agent-harness sprint Build dashboard with charts and filters

Actions:

1. Archive completed sprints (if any)
2. Analyze requirements
3. Break into atomic features with acceptance criteria
4. Update only features.json and progress.md — do NOT create extra files like summary documents
5. Display sprint planning summary in terminal
6. Ask user to confirm the sprint plan before committing
7. After user confirms, commit: git add features.json progress.md && git commit -m "chore: sprint planning - <sprint-name>"
8. Read sprint-agent.md for detailed workflow

/agent-harness code

Start coding session for next feature.

Actions:

1. Review progress.md and features.json
2. Select next pending feature
3. Implement following session protocol
4. Commit code after implementation: git add -A && git commit -m "feat(<scope>): <description>"
5. Read coding-agent.md for protocol

/agent-harness 996

Execute parallel coding tasks with subagent orchestration.

Examples:

• /agent-harness 996 - Run parallel orchestration with default settings
• /agent-harness 996 --max-parallel=3 - Limit to 3 concurrent subagents

Actions:

1. Check for uncompleted sprint
2. Analyze dependencies and file conflicts
3. Dispatch subagents for parallel execution
4. Verify results and update tracking files
5. Commit: git add -A && git commit -m "chore: 996 orchestration complete - N features"
6. Read 996-agent.md for protocol

Use when: You want to accelerate sprint completion by running multiple coding tasks in parallel.

/agent-harness status

Show current project status.

Output:

• Current sprint and goal
• Feature completion statistics
• Next recommended feature
• Run python3 scripts/status.py

/agent-harness archive

Archive completed sprints.

Actions:

• List completed sprints
• Move to .agent-harness/archived/
• Run python3 scripts/archive_sprint.py
• Commit: git add -A && git commit -m "chore: archive completed sprint"

/agent-harness force-archive

Force archive ALL sprints (including incomplete ones).

Warning: This will archive all sprints regardless of their status.

Actions:

• List all sprints to be archived
• Confirm with user
• Move all to .agent-harness/archived/
• Run python3 scripts/archive_sprint.py --force
• Commit: git add -A && git commit -m "chore: force archive all sprints"

Quick Start

python3 scripts/init_project.py "Project Name" -d "Description"
python3 scripts/status.py
python3 scripts/validate_structure.py

Workflow

1. Sprint Planning (Sprint Agent)

When: New project or iteration

1. Analyze user requirements
2. Break into atomic features with acceptance criteria
3. Order by dependencies
4. Update features.json and progress.md

→ Read references/sprint-agent.md for workflow, schemas, and examples

2. Feature Implementation (Coding Agent)

When: Each development session

Session Protocol:

→ Read references/coding-agent.md for protocol, testing, and schemas

3. Iterate

Loop: Sprint Agent → Coding Agent → Coding Agent →... → Sprint Agent

Core Files

→ See references/sprint-agent.md or references/coding-agent.md for schemas

Scripts

All scripts support --project-dir / -p to specify the project directory. When using slash commands, --project-dir "$(pwd)" is passed automatically.

python3 scripts/resolve_project_dir.py [-s dir]                # Resolve project directory
python3 scripts/init_project.py <name> [-d desc] [-p dir]     # Initialize tracking
python3 scripts/status.py [-p dir]                             # Show status
python3 scripts/validate_structure.py [-p dir]                 # Validate features.json
python3 scripts/archive_sprint.py [-p dir] [--list|--dry-run]  # Archive completed sprints

CRITICAL: Always prefix with $env:PYTHONIOENCODING="utf-8"; when running scripts (required for Windows GBK compatibility).

Critical Rules

1. One Feature Per Session - Don't implement multiple
2. Always Leave Working Code - Never break the build
3. Test End-to-End - Verify as user would experience
4. Commit Frequently - Small commits enable rollback
5. Never Delete Features - Only change status
6. Use Progress Log - Record every session

References

By Agent Role:

• sprint-agent.md - Sprint Agent: workflow, feature breakdown, schemas
• coding-agent.md - Coding Agent: session protocol, testing, schemas
• 996-agent.md - 996 Agent: parallel orchestration, subagent dispatch, conflict detection

Complete Examples:

• examples.md - Realistic examples of features.json, progress.md, and workflows