claude-harness-forge

Claude Harness Forge

A Claude Code plugin that builds software the way a well-run engineering team would -- from requirements to production, with independent verification at every step.

v3.0 (May 2026) is the current spec. It retrofits the v2.0 pipeline with: initializer/coding-agent split, feature_list.json as the project-completion contract, mandatory browser-automation E2E gate, Ralph Loop exit interception, per-workflow LLM routing, Plan Mode subagent, Extended ReAct, 5-stage adaptive compaction, spec-gap backpropagation, monotonic-improvement guards, instinct extraction, tree-structured sessions, and YAML recipes. Full spec: brd/v3.0.md. Operational plan: brd/v3.0-implementation-plan.md. Live punch list: feature_list.json. 19 agents · 27 hooks · 47 skills · 23 commands.

You describe what you want to build. The forge runs 19 specialized agents through a 9-phase pipeline (v3.0 §7): gathering requirements through Socratic interview, challenging your architecture decisions, decomposing work into stories, generating code with parallel agent teams, and verifying everything by actually running the application. Not by reading the code and saying "looks good."

One command starts it. Human approval gates the creative decisions (BRD, architecture, design). Everything after that -- implementation, testing, verification, self-healing -- runs autonomously, bounded by the feature_list.json contract.

# Load as plugin and scaffold
claude --plugin-dir ~/claude-harness-forge
> /scaffold

# Run the full pipeline
> /build

---

What the Forge Does

Builds and Verifies Autonomously

The forge doesn't just generate code -- it runs your app, hits your API endpoints, drives a browser through Playwright, and checks for console errors. A 200 response with "Failed to connect" in the body is a failure. An empty list when data should exist is a failure. If something breaks, it diagnoses the issue, fixes it, and re-verifies -- up to 3 attempts per gate before escalating.

Catches What Tests Miss

Tests pass. The app crashes. This is the most common failure mode in AI-generated code, and the forge addresses it structurally:

• Three-level verification -- liveness (does it respond?), behavior (does it work correctly?), integration (do features work together?)
• Smoke launch with real data -- every build group must start the app with actual production data, not test fixtures. This gate cannot be disabled.
• Spec gaming detection -- catches agents deleting tests to make suites pass, writing tautological assertions (expect(true).toBe(true)), inflating coverage with dead code. Also cannot be disabled.
• Mutation testing -- injects small bugs and verifies your tests actually catch them. The mutation score ratchets: once it reaches 72%, it can never drop below 72%.

Learns and Improves

The forge has multiple feedback loops that compound over time:

• Self-healing loop -- when a gate fails, the system diagnoses the failure (14 categories), spawns a targeted fix with the structured failure context and prior attempt history (so it doesn't retry the same fix), re-runs only the failed gate, and extracts a learned rule on success.
• Cross-project learning -- agents read prior stack decisions, failure patterns, and integration notes before making recommendations. Mistakes from project A inform project B.
• Findings reporter -- opt-in, anonymized feedback to the forge itself. A passive hook collects build findings; you review everything before submitting. The forge improves from real-world usage.
• Change management -- mid-build requirement changes don't get lost. /change logs them with version tracking, runs impact analysis, and cascades updates through only the affected stories, design, and code.

Adapts to Your Project

The architect analyzes your requirements and activates only what's relevant:

Project Type	What Activates
CRUD	Standard architecture review, gates 1-8
ML	+ ML pipeline design, compliance gate, model cards, bias/fairness audits
Agentic	+ Agentic architecture round, OWASP Agentic Top 10, agentic UX patterns
RAG	+ RAG scaffolding, vector DB selection, chunking/embedding guidance

Projects can match multiple types. The forge also supports 4 execution modes -- from Full (all 12 gates, production-grade) to Solo (3 gates, weekend projects) -- so you control cost and rigor.

Scales with Agent Teams

For large story groups, the generator spawns parallel sub-agents that each own a slice of work. A dependency handshake identifies shared files before work begins, preventing merge conflicts. Each teammate gets the full context: learned rules, architecture constraints, and test requirements.

---

The Pipeline