ORCA: A Cognitive Runtime Layer for Agent Systems (paper + open source)

Good point on CapabilitySpec being dict-based — that’s the kind of observation that comes from actually reading the models module, not just the docs. I’m going to promote properties and safety to proper typed dataclasses in the next release.

On the 9-field proposal: I think the realistic scope is narrower. retry_policy and cache_hint make sense as new fields. compensation_strategy and cost_latency_class feel like distributed-systems concepts that don’t belong in a capability contract at this level. The rest (idempotent, side_effects, trust_level) already exist — they just need to graduate from dict keys to enforced types.

Curious about one thing — have you actually run any of the skills locally? The whole baseline-first design means you can pip install orca-agent-skills and execute the full suite without any API key. I’d be interested in your take on the actual execution ergonomics, not just the contract layer. Things like: does the scaffold wizard actually produce usable skills? Does the benchmark-lab output make the binding tradeoffs legible? That’s the kind of feedback that’s harder to get from reading the source.

If you want to try it, cloning the repo gives you the full test suite too — 1694 tests, all deterministic. Would appreciate a star if you find it worth exploring.