epic-harness

6 commands. Auto-trigger skills. Self-evolving.

01 / QUICKSTART

Quickstart

Install (Plugin)

Install directly from the Claude Code plugin marketplace.

/plugin marketplace add epicsagas/plugins
/plugin install epic@epicsagas

Install (Binary)

Install via Homebrew, cargo-binstall, or build from source.

# Homebrew (macOS)
brew install epicsagas/tap/epic-harness

# cargo-binstall
cargo binstall epic-harness

# From source
git clone https://github.com/epicsagas/epic-harness.git
cd epic-harness && cargo install --path .
epic install

First Session

Run /spec to define what you want to build, then /go to execute. The harness auto-detects your stack and applies cold-start presets.

/spec Build a REST API with authentication
# → Creates SPEC-*.md in .harness/specs/

/go
# → Auto-plans, delegates to subagents, verifies

02 / CORE CONCEPTS

Core Concepts

4-Ring Architecture

Ring 0 (Autopilot): Hooks run invisibly — session restore, auto-format, guard rails, observation. Ring 1 (Commands): 6 commands you call — /spec, /go, /check, /ship, /team, /evolve. Ring 2 (Auto Skills): Context-triggered — tdd, debug, secure, perf, simplify, document, verify. Ring 3 (Evolve): Observes tool usage, analyzes failures, auto-generates skills.

Unified Memory (SQLite)

All agents share a single knowledge graph in ~/.harness/memory.db (SQLite + FTS5). Uses composite scoring: recency(25%) + importance(35%) + access_frequency(15%) + FTS_match(25%). Node types: decision(0.9), resolution(0.8), concept(0.7), pattern(0.5), error(0.4), session(0.2).

Multi-Tool Support

Works with Claude Code and 6 additional AI coding tools: Codex CLI, Gemini CLI, Cursor, OpenCode, Cline, Aider. All share the same ~/.harness/ data directory.

03 / REFERENCE

Reference

6 Commands

The complete command reference.

/spec    Define what to build — clarify requirements, produce a spec
/go      Build it — auto-plan, TDD subagents, parallel execution
/check   Verify — parallel code review + security audit + performance
/ship    Ship — isolated pre-flight test, then PR, CI, merge
/team    Create and sync org-level agent teams across projects
/evolve  Manual evolution trigger / status / rollback

Memory CLI

Manage the knowledge graph directly from the terminal.

harness mem recall "auth refactor" --project my-project
harness mem add --title "JWT rotation" --type decision --body "..." 
harness mem query --type decision --project my-project
harness mem search "JWT"
harness mem context --project my-project
harness mem serve          # → http://localhost:7700

8 Auto Skills (Ring 2)

Skills trigger automatically based on context.

tdd        Test-Driven Development — write test first, then implement
debug      Systematic debugging — hypothesize → isolate → fix → verify
secure     Security review — OWASP Top 10 + injection + access control
perf       Performance review — detect N+1, leaks, unnecessary computation
simplify   Code simplification — extract, rename, reduce
document  Auto-documentation — generate JSDoc, docstrings, comments
verify     Pre-completion verification — build + test + lint must all pass
context    Context window management — summarize, compact, preserve state

04 / CONFIGURATION

Configuration

Project Data Directory

Each project gets its own data directory with organized subfolders.

~/.harness/projects/{slug}/
  memory/      # Session memory files
  sessions/    # Session logs
  obs/         # Observations (JSONL)
  evolved/     # Evolved skills
  team/        # Team configurations
  specs/       # Specification files

~/.harness/
  memory.db    # SQLite knowledge graph
  graph.json   # Graph edges

Cold-Start Presets

Auto-detects your stack on first session and applies preset skills for Node.js/TypeScript, Go, Python, and Rust.

05 / EXAMPLES

Examples

Typical Workflow

The /spec → /go → /check → /ship loop covers the full development cycle.

# 1. Define
/spec Add user authentication with JWT
# → Creates spec with requirements and acceptance criteria

# 2. Build
/go
# → Auto-plans tasks, delegates to subagents with TDD

# 3. Verify
/check
# → Parallel code review + security audit + perf analysis

# 4. Ship
/ship
# → Pre-flight test, PR creation, CI verification, merge