LLM providers

The split, reword, and squash verbs use a language model to propose hunk groupings, commit messages, and squash summaries. This page covers how to point regit at a provider, where credentials come from, and what the model is — and is not — allowed to do.

LLM features are opt-in

regit works fully without any LLM configured: redate, retitle, reauthor, drop-bodies, linearize, purge, redact, and undo never touch a model. Running an LLM-backed verb with nothing configured fails fast with a clear message telling you to pass --llm=provider/model or --agent=<cmd> — nothing else is affected.

Flag surface

Flag	Purpose
--llm=<provider>/<model>	Select an HTTP provider and model, e.g. --llm=ollama/qwen2.5-coder.
--llm-base-url=<url>	Override the provider's default endpoint (self-hosted, vLLM, proxies).
--agent=<cmd>	Drive a local agent CLI (e.g. claude) as the backend instead of HTTP.
--agent-args=<argv>	Advanced: replace the default agent argv; a {schema} placeholder is substituted with the JSON schema at call time.
--prompt=<hint>	Free-text guidance appended to the model's instructions.
--no-cache	Ignore the proposal cache and recompute.
--clear-cache	Wipe the feature's cached proposals and exit — nothing is proposed or applied.

Provider/model parsing

--llm is parsed on the first slash only, so models whose names contain slashes work naturally: --llm=openrouter/anthropic/claude-x means provider openrouter, model anthropic/claude-x.

Providers

Provider	Default base URL	API key env var
ollama	http://localhost:11434/v1	none (local)
lmstudio	http://localhost:1234/v1	none (local)
openai	https://api.openai.com/v1	OPENAI_API_KEY
openrouter	https://openrouter.ai/api/v1	OPENROUTER_API_KEY

All providers speak the OpenAI-compatible /v1/chat/completions API, so any compatible server works via --llm-base-url=.

API keys

API keys come from environment variables only — never from flags, the config file, or logs. Resolution order:

1. REGIT_LLM_API_KEY — tool-wide override, beats the provider-standard vars
2. Provider-standard var: OPENAI_API_KEY, OPENROUTER_API_KEY
3. None — local providers (ollama, lmstudio) need no key

export OPENAI_API_KEY=sk-...
regit reword HEAD --llm=openai/gpt-4o

Setting a default provider

Set llm.provider and llm.model in the config file, or export REGIT_LLM_PROVIDER / REGIT_LLM_MODEL, and the --llm flag can be omitted entirely:

llm:
  provider: ollama
  model: qwen2.5-coder

regit split HEAD          # uses ollama/qwen2.5-coder from config

Agent backend

--agent=<cmd> drives a local agent CLI as a subprocess instead of calling an HTTP API, reusing the host's existing agent session for auth — no API key needed. The default command drives claude headless:

claude -p --output-format json --bare --json-schema {schema}

The prompt is fed on stdin. --bare skips the host project's CLAUDE.md, hooks, plugins, and MCP servers so output stays deterministic. Calls are bounded by a 5-minute timeout. --agent-args= replaces this argv when you need a different agent binary or flags; keep the {schema} placeholder so regit can inject the JSON schema.

Trust model

The model only ever proposes structured data — a hunk grouping, a set of message variants, a squash summary. regit validates the proposal against a schema and its own invariants, then performs the git operations itself through the same guarded apply path every verb uses (backup ref, content verification, undo). The LLM never mutates the repository.

Agent scan is defense-in-depth, not a sandbox

reword --agent-scan (requires --agent) lets the agent browse repository files for extra context. Access is deliberately limited to the read-only file tools Read, Grep, and Glob — no shell of any kind: even innocent-looking read-only commands can escape (git show accepts --output=<file> to write arbitrary files; find allows -exec to run arbitrary commands), so no Bash(...) allowance is safe. regit also scrubs secret-bearing environment variables (names containing API_KEY, TOKEN, SECRET, PASSWORD, AWS_, GH_, GITHUB_, OPENAI, OPENROUTER, or ANTHROPIC) before spawning the agent or a --test-cmd, and denies the agent access to SSH/AWS/GPG key stores, /etc, /proc, and .env files. Treat all of this as defense-in-depth for a mostly-trusted local agent — not as a sandbox for running against untrusted repositories.

Caching

Proposals are cached repo-scoped under .git/regit/cache/, keyed by a fingerprint of every input that affects the output (commit content, feature parameters, prompt hint) plus the model name — changing any of them naturally misses the cache. --no-cache recomputes without reading the cache; --clear-cache wipes the feature's cache and exits without running.

Related pages

• Configuration — the llm, split, and reword config sections
• split — split one commit into semantic commits
• reword — regenerate commit messages, including --agent-scan
• squash — collapse commit runs with LLM summaries