development-harness

Compiles a ROADMAP.md into phased, validator-backed autonomous execution units.

Updated Mar 6, 2026
One-click install
npx skills add https://github.com/timi-ty/agent-forge --skill development-harness-timi-ty
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: development-harness
Source: https://github.com/timi-ty/agent-forge/tree/main/skills/development-harness
Command: npx skills add https://github.com/timi-ty/agent-forge --skill development-harness-timi-ty

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve? Turning a product roadmap into consistent, verifiable agent-driven development work is hard: agents lose track of progress between sessions, skip validation, and drift from the plan. This Skill installs a project-local control plane that compiles a ROADMAP.md into dependency-ordered phases and units, each backed by a validator, and drives execution one bounded batch at a time. ## Core Features & Use Cases - Roadmap compilation: Parses ROADMAP.md headings into a phase-graph.json skeleton, then guides the agent to refine phases, units, dependencies, and validation gates. - Seven workspace commands: create, invoke, update, state, sync, clear, and inject-issues cover the full harness lifecycle in both Cursor and Claude Code. - Validator-backed execution: Every unit requires acceptance criteria and a validation method; a stop hook keeps the agent iterating until goals are verified or blockers appear. - Optional parallel execution: Units marked parallel_safe with disjoint touches_paths can fan out into git worktrees with scope-checked merges. - Use Case: Point the agent at a new repo, say "create development harness", answer the setup questions, and the workspace gains a persistent execution plan you can resume any time with "/invoke-development-harness". ## Quick Start Ask the agent to create a development harness for this repository and then invoke it to start executing the first unit of work.

Frequently Asked Questions about development-harness

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I set up an autonomous development harness for my repo?

Run the create command by saying "create development harness". The skill detects your tool (Cursor or Claude Code), inspects the repo, asks targeted setup questions, compiles ROADMAP.md into phases, and generates all .harness/ artifacts plus workspace commands.

How does the harness decide what to work on next?

The select_next_unit.py script is the authoritative selector. It reads phase-graph.json for canonical phase and unit status, and the stop hook only continues execution when the selector output and the human-readable checkpoint agree.

Does the development harness work with both Cursor and Claude Code?

Yes. The create flow detects the host tool and generates tool-appropriate artifacts: .cursor or .claude directories, matching hook configs, and rule file formats. On Claude Code, autonomous runs use the native /loop skill because its Stop hook cannot force continuation.

Can the harness run multiple units in parallel?

Yes, when parallelism is enabled in config.json. Units marked parallel_safe with disjoint touches_paths are packed into batches, each dispatched to its own git worktree, and merged back with scope checks against declared paths.

What happens when the harness schema version changes?

By default the harness refuses to operate on a schema mismatch and requires re-running /create-development-harness. ROADMAP.md and PHASES documents are preserved and the harness is regenerated around them; no in-place migration tooling is provided.

How do I safely remove all harness artifacts from a project?

Use the clear command. It reads manifest.json, runs clear_harness.py in dry-run mode first, shows exactly what will be deleted or preserved, and only executes after explicit confirmation. Product-owned files are never touched.