Short definitions for the vocabulary used across the Argus docs.
- Reviewer
- A single LLM (via an API route or a CLI) that examines the diff and returns findings. The 15 reviewers live in the registry.
- Route
- How a reviewer is reached — e.g. a direct provider API (z.ai, MiniMax, DeepSeek), OpenRouter, or a paid CLI. Dual-route reviewers have more than one.
- Roster
- The concrete set of reviewers chosen for a run, after a profile and host rules are applied.
- Profile named roster
- A named, reusable roster in config.yaml: quick, standard, panel, security, deep, favorites, direct, leaderboard-top5. See Profiles.
- Route preference
- The
route_preferenceknob (openrouter|direct) that decides which provider a dual-route reviewer tries first. See Configuration. - Confidence threshold
- The minimum effective confidence (80) a finding must reach to survive the merge. Lower-confidence findings are dropped.
- Corroboration boost
- +15 confidence (capped at 100) applied when ≥2 reviewers flag the same file within ±3 lines — agreement is evidence.
- Fixture
- A labeled benchmark case: a
diff.patchplus aground-truth.jsonof the findings a good reviewer should catch. Includes a clean-baseline control. See Benchmarks. - F1
- The harmonic mean of precision and recall:
2·P·R / (P + R). The headline benchmark ranking axis. - Host CLI
- The CLI Argus is running inside (claude / codex / gemini / opencode / unknown). The matching reviewer is always skipped so a CLI never reviews its own invocation. See host-CLI awareness.
- aichat
- The universal LLM client Argus uses for API-routed reviewers. Its config is generated by
install_aichat.py; keys are read fromAICHAT_<CLIENT>_API_KEYat dispatch. - Overlay
- A prompt add-on layered on the base reviewer prompt to focus a run — e.g. the
securityanddeepoverlays used by the matching profiles.