brindle 0.3.5

Released Oct 12, 2026 · All releases

For everyone

• Four-tier model routing: light, medium, heavy, and advanced. Tasks can now be sized into four tiers (with "reasoning" and "complex" accepted as aliases for "advanced").
  • light: small, well-specified, mechanical fixes and docs (developer-light baseline).
  • medium: standard single-area features and fixes (developer baseline).
  • heavy: architecture design, broad cross-cutting refactors, or subtle bugs (developer-heavy on Claude Opus with high effort baseline).
  • advanced: complex, long-running system-design and implementation tasks (developer-fable on Claude Fable with high effort baseline).
• New built-in profiles: developer-fable and developer-gemini.
  • developer-fable runs on Claude Fable at high effort for Tier 4 advanced work.
  • developer-gemini runs on Google Antigravity (Gemini) at high effort as a Tier 3 heavy candidate and Tier 4 candidate.
• Provider capacity in cost tracking. The sidebar and collapsed costs header now report real-time provider headroom percentages (e.g. claude 85%) across active providers with quota tracking.
• Antigravity usage and spend persistence. Token usage and cost metrics are now captured and tracked for Antigravity workers alongside Claude and Codex.
• Autopilot and cost estimation priors. Autopilot ceiling resolution (max_weight), task priority queues, and cost estimation priors are updated for the 4-tier routing model.

Changes to note

• developer-heavy's baseline model is Claude Opus at high effort; Claude Fable is now dedicated to developer-fable in the advanced tier.
• Running Brindle sessions keep their loaded version until restarted. Reconnect Brindle (/mcp in Claude Code) or start a new session to load the new profiles and routing defaults.

curl -fsSL pawdelta.com/brindle/install | sh