Choosing a harness

A harness is the LLM driver — one way of calling a model. Every harness serves every runbook (the output schema is a parameter), so picking one is purely about how you want the model to run: which CLI/SDK, which auth, cloud or local. Select with --harness, env HARNESS, or config harness:; default claudecode.

Quick answer

You want…	Pick	Why
Zero-setup local reviews (default)	claudecode	Drives your claude CLI — subscription auth, git tools, explores beyond the diff
Predictable CI cost	pydantic	One structured API call, no agentic wandering
The OpenAI/Codex stack	codex	codex exec in a sandbox, OAuth or API key
Local / offline models	pi	Ollama / LM Studio / vLLM via ~/.pi/agent/models.json — no API key, nothing leaves the machine
To experiment with orchestration	deepagents	LangChain orchestrator — unstable, may skip the submit tool

Comparison

Harness	Install	Runs via	Auth	Status
claudecode	(core)	single claude -p subprocess	subscription or ANTHROPIC_API_KEY	stable
codex	junior[codex]	single codex exec in sandbox	OAuth or OPENAI_API_KEY	stable
pydantic	junior[pydantic]	single structured pydantic-ai call	API key required	stable
deepagents	junior[deepagents]	LangChain orchestrator + subagents	API key required	unstable
pi	(core)	single pi --mode json subprocess	provider key, auth.json, or none (local models)	stable

junior config list harnesses shows what’s installed and ready on your machine; junior config env --harness X lists the exact env vars one needs.

How each harness works

Every harness does a single structured call and returns a validated instance of the runbook’s output schema. They differ in how the schema is enforced and whether the model can read repository files itself (file_access — see Glossary):

	pydantic	claudecode	codex	deepagents	pi
Diff in user message	Yes	No — reads files via tools	No — reads files via sandbox	Yes	Small diffs yes; tools beyond
file_access	False	True	True	False	True
File tools	read_file, list_dir, grep (Python)	Read, Grep, Glob, Bash(git…) (built-in)	Sandbox filesystem access	read_file, ls, grep, glob (via deepagents)	read, grep, find, ls (read-only)
Output contract	Returns the output schema instance directly	Output schema via --json-schema	Output schema via --output-schema	submit_review tool whose schema is the output model	Schema embedded in system prompt + validated reply
Token usage	pydantic-ai usage	CLI JSON output	CLI stderr	langchain callback	JSON event stream (per-turn usage)

Small diffs (≤ 50k chars) are inlined for every harness — the diff is the review’s primary evidence; file tools serve for context beyond it.

[!NOTE] --harness / HARNESS / config harness: is canonical. The old --backend / BACKEND / backend: is kept as a deprecated alias for one version.

Going deeper

Per-harness internals (subprocess flags, event parsing, error handling): Claude Code · Codex · Pydantic AI · DeepAgents · Pi. Writing your own is one file — Adding runbooks & harnesses.