harness-orchestrator 2.2.6

Contract-driven multi-agent development orchestrator for Cursor & Codex

中文

harness-orchestrator

Contract-driven multi-agent development framework — run a full plan-build-review-ship pipeline inside Cursor with one command.

AI coding tools excel at single-shot tasks. Continuous development needs more: goal tracking, quality gates, adversarial review, and audit trails. Harness organizes these into a contract-driven engineering loop that runs inside your Cursor IDE — no separate orchestrator process, no complex setup. For CI/CD and headless automation, an optional orchestrator mode drives Cursor and Codex agents via external CLI.

Quick Start (Cursor-native, 3 minutes)

1. Install harness

pip install harness-orchestrator
harness --version   # verify (also works: python3 -m harness --version)

Alternative: install from source (for contributors)

git clone https://github.com/arthaszeng/harness-orchestrator.git
cd harness-orchestrator
pip install -e ".[dev]"

2. Initialize your project

cd /path/to/your/project
harness init

The wizard walks you through setup. When asked for Workflow Mode, choose cursor-native:

Step 5/9  Workflow Mode
  Choose how harness drives development:
  1. orchestrator -- External CLI process drives cursor-agent (default)
  2. cursor-native -- Skills + subagents inside Cursor IDE (no external process)
  Choose [2]: 2
  → cursor-native mode: will generate skills, subagents, and rules

This generates skills, subagents, and rules directly into your .cursor/ directory.

3. Use it in Cursor

Open your project in Cursor. You now have four skills available:

Skill	What it does
/harness-plan	Analyze a requirement, produce a spec and contract with adversarial review
/harness-build	Implement the contract, run CI, triage failures, write a structured build log
/harness-eval	Three-pass adversarial code review (Claude + Claude adversarial + GPT cross-model)
/harness-ship	Full pipeline in one command: plan → build → review → fix → commit → push → PR

Try it now — open Cursor chat and type:

/harness-ship add input validation to the user registration endpoint

Harness will plan the work, implement it, run a three-pass adversarial review, auto-fix trivial issues, create bisectable commits, and open a PR — all without leaving your IDE.

Updating

harness update          # upgrade to latest, reinstall agents, check config
harness update --check  # just check if a new version is available

---

What happens under the hood

You type /harness-ship "add feature X"
  → Rebase onto main, run tests
  → Three-pass adversarial review:
      Pass 1: Claude structured review (4 dimensions)
      Pass 2: Claude adversarial subagent (attack surface)
      Pass 3: GPT cross-model review (independent perspective)
  → Fix-First: auto-fix trivial issues, ask about important ones
  → Bisectable commits + push + PR

Three-pass adversarial review

Every code change goes through three independent reviewers:

1. Structured review — Claude scores on completeness, quality, regression, and design
2. Claude adversarial — A fresh Claude subagent hunts for security holes, race conditions, edge cases, and resource leaks
3. GPT cross-model — A GPT-based reviewer (default: gpt-4.1) provides perspective from a different model family

Passes 2 and 3 are dispatched in parallel for speed. Findings from 2+ passes are flagged as high confidence. The adversarial model is configurable in .agents/config.toml.

Fix-First auto-remediation

Review findings are classified before presenting:

• AUTO-FIX — High certainty, small blast radius, reversible. Fixed immediately and committed.
• ASK — Security findings, behavior changes, or low confidence. Presented to you for decision.

Trivial issues never block shipping. Important decisions always get human judgment.

Graceful degradation

Pass 1 (Structured)	Pass 2 (Claude)	Pass 3 (GPT)	Behavior
OK	OK	OK	Full three-pass synthesis
OK	OK	Failed	Two-pass, tagged [claude-only]
OK	Failed	OK	Two-pass without Claude subagent
OK	Failed	Failed	Single-reviewer mode
Failed	—	—	Fatal — cannot evaluate

---

Generated artifacts

When you choose cursor-native mode, harness init generates:

Artifact	Path	Purpose
/harness-plan	.cursor/skills/harness/harness-plan/SKILL.md	Plan and decompose a task with adversarial spec review
/harness-build	.cursor/skills/harness/harness-build/SKILL.md	Build: implement contract, run CI, triage failures
/harness-eval	.cursor/skills/harness/harness-eval/SKILL.md	Three-pass review with Fix-First auto-remediation
/harness-ship	.cursor/skills/harness/harness-ship/SKILL.md	Full automated pipeline: test → review → fix → commit → PR
Adversarial reviewer	.cursor/agents/harness-adversarial-reviewer.md	Cross-model code reviewer (configurable model, readonly: true)
Evaluator	.cursor/agents/harness-evaluator.md	Structured evaluator with JSON output (readonly: true)
Trust boundary	.cursor/rules/harness-trust-boundary.mdc	Always-on: Builder output is untrusted
Fix-First	.cursor/rules/harness-fix-first.mdc	Always-on: classify findings before presenting
Workflow conventions	.cursor/rules/harness-workflow.mdc	Commit format, branch naming, task state

To regenerate after config changes:

harness install --force

---

Configuration

Project settings live in .agents/config.toml:

Key	Default	Description
workflow.mode	"orchestrator"	orchestrator or cursor-native
workflow.profile	"standard"	lite / standard / autonomous
workflow.max_iterations	3	Max iterations per task
workflow.pass_threshold	3.5	Evaluator pass threshold (out of 5)
workflow.auto_merge	true	Auto-merge branch after pass
workflow.dual_evaluation	false	Add alignment review after quality review
workflow.branch_prefix	"agent"	Task branch prefix
native.adversarial_model	"gpt-4.1"	Cross-model reviewer model
native.adversarial_mechanism	"auto"	Adversarial dispatch: subagent / cli / auto
native.review_gate	"eng"	Which review layers are hard gates
autonomous.max_tasks_per_session	10	Max tasks per autonomous session
autonomous.consecutive_block_limit	2	Stop after this many consecutive blocks

Models (optional)

Per-role model selection under [models]. Harness only passes --model when the resolved value is non-empty.

Resolution order: role_overrides.<role> → driver_defaults.<driver> → models.default → empty.

[models]
default = ""

[models.driver_defaults]
# codex = "o3"
# cursor = "claude-4-opus"

[models.role_overrides]
# planner = "o3-pro"
# builder = ""  # explicit: always use IDE default

Workflow profiles

Profile	Flow	When to use
lite	planner → builder → eval (no spec/contract split; threshold cap 3.0; max 2 rounds)	Small changes, quick fixes
standard	planner → spec + contract → builder → eval (full review)	Day-to-day development (default)
autonomous	strategist → standard loop → reflector	Vision-driven autonomous mode

---

Task artifacts

All artifacts live under .agents/ at the project root:

.agents/
├── config.toml            # Project config
├── vision.md              # Project vision
├── state.json             # Runtime state
├── .stop                  # Stop signal
├── runs/
│   └── <session-id>/
│       └── events.jsonl   # Structured events
├── tasks/
│   └── task-001/
│       ├── spec-r1.md     # Spec: analysis and technical plan
│       ├── contract-r1.md # Contract (Markdown)
│       ├── contract-r1.json # Contract (JSON sidecar)
│       ├── evaluation-r1.md # Review (Markdown)
│       ├── evaluation-r1.json # Review (JSON sidecar)
│       ├── alignment-r1.md # Alignment review (if dual_evaluation)
│       ├── build-r1.log   # Builder log
│       └── ...
└── archive/               # Archived sessions

Every step is traceable. JSON sidecars suit automation and UIs without regex-parsing Markdown.

Local-first: All state stays on disk; no cloud dependency. The .agents/ tree is usually gitignored. To share config.toml or vision.md with your team, use git add -f .agents/config.toml.

---

Command reference

Command	Description
harness install [--force] [--lang]	Install agent definitions to local IDE
harness init [--name] [--ci] [--lang] [-y]	Initialize project configuration (interactive wizard)
harness vision	Create or update project vision
harness run <req> [--resume] [--verbose]	Run a single development task
harness auto [--resume] [--verbose]	Start the autonomous development loop
harness status	Show current progress
harness stop	Gracefully stop the current task
harness --version	Show version

---

Advanced: Cross-Client Orchestrator Mode

Cursor-native mode covers most interactive development workflows. For CI/CD pipelines, headless automation, or multi-IDE setups (Cursor + Codex), use orchestrator mode.

Prerequisites

Dependency	Requirement	Notes
Python	>= 3.9	Runs the Harness CLI
Cursor CLI and/or Codex CLI	At least one	Provides agent capability
Git	Any version	Project must be a Git repo

IDE CLI setup:

• Cursor: Command Palette → Install 'cursor' command
• Codex: npm install -g @openai/codex or from GitHub

Orchestrator vs Cursor-native

	Orchestrator	Cursor-native
How it runs	External harness CLI spawns agent processes	Skills + subagents inside Cursor IDE
Entry point	harness run / harness auto	/harness-plan, /harness-build, /harness-eval, /harness-ship
Cross-model review	Configurable per role	Adversarial subagent with a different model
When to use	CI/CD, headless, multi-IDE	Interactive development, Cursor-only

Role architecture

Role	Responsibility	Default backend (auto mode)
Planner	Analyze requirements; produce spec and contract	Codex
Builder	Implement against the contract; commit changes	Cursor
Evaluator	Independent review; four-dimensional scoring	Codex
Alignment Evaluator	Requirement alignment and intent drift detection	Codex
Strategist	Pick the next task from vision (autonomous mode)	Codex
Reflector	Distill lessons into long-term memory	Codex/Cursor

Each role's backend is configurable under [drivers.roles]. See docs/compatibility.md for CLI version requirements.

Orchestrator setup

# 1. Install agent definitions to IDE directories
harness install

# 2. Initialize (choose "orchestrator" mode)
cd /path/to/your/project
harness init

# 3. Create project vision
harness vision

# 4. Run
harness run "add user authentication"   # single task
harness auto                            # autonomous loop

# 5. Monitor
harness status
harness stop

Single-task flow (harness run)

Requirement
  → Planner: spec + iterative contract
  → Builder: implement and commit
  → Evaluator: four-dimensional score
      → Pass (≥ 3.5) → done
      → Fail → feedback to Builder, iterate
  → Max iterations (3) → blocked

Autonomous loop (harness auto)

Vision
  → Strategist: pick next task
  → Single-task flow
  → Reflector: distill lessons
  → Loop until: all done / stop signal / block limit / task limit

Dual Evaluator

With workflow.dual_evaluation = true, quality review is followed by alignment review:

• Quality — Code quality + regression (four-dimensional scoring)
• Alignment — Requirement coverage + contract fit + intent drift

If alignment returns MISALIGNED, the task iterates back to Builder. If CONTRACT_ISSUE, feedback goes to Planner to revise the contract instead.

[workflow]
dual_evaluation = true

---

Troubleshooting

Resuming interrupted work

harness run "original requirement" --resume
harness auto --resume

--resume reloads from state.json and continues from the interrupted phase.

Stop behavior

harness stop writes .agents/.stop. The task finishes its current phase and exits cleanly. For immediate abort, use Ctrl+C — Harness saves a checkpoint before exit.

IDE CLI not found

If you see Neither Cursor nor Codex CLI detected:

• Cursor: Command Palette → Install 'cursor' command
• Codex: npm install -g @openai/codex

Ensure the binary is on PATH. For cursor-native mode, Cursor CLI is optional — harness generates files that work directly in the IDE.

Reinstalling

If harness install fails or produces a broken setup:

harness install --force

This overwrites existing files, retries CLI installations, and regenerates native artifacts.

---

Observability

Each session writes structured events to .agents/runs/<session-id>/events.jsonl:

{"ts": "2026-03-31T10:00:00.000Z", "event": "agent_end", "role": "planner", "driver": "codex", "exit_code": 0, "elapsed_ms": 12340}

Event types: agent_start/agent_end, ci_result, state_transition, task_start/task_end.

---

Repository layout

harness-orchestrator/
├── src/harness/
│   ├── cli.py              # CLI entry (Typer)
│   ├── commands/            # Subcommand implementations
│   ├── orchestrator/        # Workflow core
│   ├── drivers/             # IDE agent invocation abstraction
│   ├── core/                # State, config, UI, events
│   ├── methodology/         # Evaluation, scoring, contracts
│   ├── native/              # Cursor-native mode generator
│   ├── agents/              # Role definitions (Cursor / Codex)
│   ├── templates/           # Prompt templates (orchestrator + native)
│   └── integrations/        # Git, Memverse
├── tests/                   # Test suite
├── docs/                    # State machine, compatibility
└── pyproject.toml

---

When it fits — and when it doesn't

Good fit:

• You use Cursor and want quality gates on agent output, not blind trust
• You want traceability across multi-step work
• You want adversarial review to catch what a single pass misses

Poor fit:

• Expecting a one-click "build the whole product" autopilot
• Enterprise approval workflows unrelated to coding
• Environments where you cannot install Python or any supported agent CLI (Cursor/Codex)

---

Internationalization

harness init --lang zh    # Chinese
harness init --lang en    # English (default)

Affects CLI messages, agent prompts, generated files, and installed agent definitions. Stored in .agents/config.toml under [project] lang.

---

Development

pip install -e ".[dev]"
pytest
ruff check src/ tests/
ruff format src/ tests/

Ruff targets Python 3.9 with line length 100. See docs/releasing.md for the release process.

---

Further reading

Doc	Description
docs/state-machine.md	Task state machine
docs/compatibility.md	CLI version requirements
docs/releasing.md	Release process and PyPI publishing
examples/todo-api-benchmark/	Benchmark: five tasks, three modes

---

License

MIT