alchemishty/agent-harness

Agent Harness

A complete scaffolding system for autonomous agent-driven development. Humans steer (decisions, plans, merge). Agents execute (code, tests, review, cleanup, learning).

When to Activate

Activate these skills when:

• Setting up a new project or installing the harness into an existing codebase
• Planning and implementing features with TDD
• Creating pull requests with quality gates
• Cleaning up code and managing documentation
• Running post-implementation or post-review retrospectives
• Updating an existing harness installation to the latest version

Skill Bundles

Bootstrap

Skills for installing and configuring the harness in any project.

• install-harness — Entry point: detects greenfield vs existing, copies skills, delegates to bootstrap
• bootstrap-greenfield — Interactive setup for new projects: questions, scaffold, docs, enforcement
• bootstrap-existing — Deep-scan existing codebases and adapt without restructuring
• migrate-harness — Migrate project scaffolding to match latest harness version

Workflow

End-to-end development automation from plan to merge-ready PR.

• implement-feature — Full orchestrator: plan, TDD implementation, verification, review, completion
• create-plan — Structured implementation plans from feature descriptions
• create-tests — Tier-aware test generation (unit/integration/e2e)
• create-pr — PR creation with blocking standards gate
• create-qa — QA checklist generation from diff analysis
• retrospective — Post-implementation and post-review learning capture to memory

Maintenance

Codebase health and cleanup.

• deslop — Code cleanup: comments, naming, dead code, simplification
• garden — Entropy scan: file size, doc freshness, test gaps, scratch cleanup, memory freshness
• doc-split — Auto-split documents exceeding size threshold
• project-structure-validator — Architecture compliance validation agent

Key Concepts

Configuration-Driven

harness.yaml is the single source of truth. Every command, threshold, and rule lives here. Skills read from it at runtime — nothing is hardcoded.

Context Engineering

The harness applies research-backed principles: cross-session memory (memory/), observation masking (scratch/), plan re-reading, compression triggers, attention-favored placement, phase isolation, and retrospective learning.

Self-Healing Verification

Every commit passes a pre-commit gate (lint + test + enforcement). Failures are automatically diagnosed and fixed by the agent, up to a configurable retry limit.

References

• Context Management Guide
• Architecture Reference
• Conventions Reference
• Testing Reference
• Doc Structure Reference
• AGENTS.md Reference