Highlights
- OpenRouter adapter — runs OpenSwarm's native agentic tool loop against any OpenRouter model, with OAuth PKCE (or
OPENROUTER_API), ZDR (data_collection: deny) for non-OpenAI models, automatic Anthropic prompt caching, and reasoning-off for mechanical roles. (#63) - LM Studio adapter — dedicated OpenAI-compatible endpoint with auto model selection. (#60)
- Repo knowledge loop — workers learn each repository across tasks: outcomes are stored as per-repo memories and recalled into future worker prompts. (#63)
- L0–L6 benchmark suite — synthetic L0–L5 with deterministic grading + L6 = real SWE-bench Lite instances graded by the official harness. Hybrid mode (frontier read-only diagnosis + lightweight implementer) resolved 3/3 attempted instances. See benchmarks/RUBRIC.md. (#63)
Harness hardening (from SWE-bench findings)
Working-directory injection, bash exit-code surfacing, raised compaction thresholds (fixes an infinite re-read loop), final-answer turn, no-edit guard, verification-file protection (protectedFiles), configurable bash timeout (bashTimeoutMs).
Other changes
- Worker success judged primarily by git diff (no structured JSON block required).
- Benchmark-driven default model routing: lightweight worker + frontier escalation; frontier planner/reviewer.
- Standalone mode: runs without Discord/Linear credentials.
- All repository docs are now English. (#65)
Full changelog: CHANGELOG.md