Configuration reference
auto-agents reads auto-agents.toml from repository root on each launch. Legacy dev.toml is still supported as fallback.
All fields are optional. Start with defaults and add only required overrides.
Frequently used top-level fields
Agent selection and model defaults
Use --agent per invocation, or persist model preferences in config.
--model <id> and AUTO_AGENTS_MODEL can override model choice for one run.
Skill path resolution
When [skills] entries are omitted, auto-agents resolves skill files in this order:
.auto-agents/skills/.../SKILL.md(recommended override layout in downstream repos)assets/skills/.../SKILL.md(bundled source layout)Materialized bundled copies under OS app-data directory
If both .auto-agents/skills and assets/skills exist, .auto-agents/skills wins.
Telemetry opt-out
Disable telemetry with any of these methods:
DO_NOT_TRACK=1AUTO_AGENTS_NO_TELEMETRY=1Config file:
Bot account setup for code review
code-review submits PR reviews through GitHub APIs and requires a bot identity that is separate from the PR author.
Recommended mode: GitHub App credentials:
DEV_BOT_APP_IDDEV_BOT_INSTALLATION_IDDEV_BOT_PRIVATE_KEY(path to PEM file)
Alternative mode: second-user PAT:
DEV_BOT_TOKENorDEV_BOT_TOKEN_PATH
Non-secret values can be saved to auto-agents.toml. Secrets are persisted through OS credential storage, not plaintext config.
Workspace feature
Use --workspace <name> to load context overrides from .auto-agents/workspaces/<name>/.
--workspace nonedisables workspace selection explicitly.When omitted and workspaces exist, auto-agents can prompt interactively.
Recommended configuration hygiene
Commit only non-secret defaults to VCS.
Keep API keys/tokens in environment variables or keychain storage.
Prefer explicit
workflow_presetin shared repositories.