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

Thanks for the detailed analysis. A few corrections and notes:

On the numbers: the repo ships 189 binding YAMLs, not 163. The 122 vs 141 difference is real but intentional — 122 are the runtime capabilities with deterministic Python baselines (no API key needed), 141 is the full registry count after the composability wave added 19 new capabilities. The changelog documents that explicitly, but I agree it should be clearer in the README itself. Will fix.

On the admission rubric: this already exists. SKILL_ADMISSION_POLICY.md defines a 7-item checklist, 4-channel promotion model (local → experimental → community → official), and a Canonical-First Rule. promotion_package.py implements it with different strictness per channel. Happy to get feedback on it.

On side-effect classification and idempotency declarations: fair point, and probably the most actionable gap you identified. Right now capabilities don’t declare side-effect class, retry semantics, or cacheability in their YAML contracts. That metadata exists conceptually in the architecture but isn’t enforced per-capability yet. Going to work on that.

On a hard end-to-end demo: agreed. The pieces are there — checkpoints, approval gates, fallback chains, webhook events — but there’s no single showcase workflow that exercises all of them together in a realistic scenario. On the list.

On the hybrid architecture: ORCA.md already describes this explicitly — the model is not “replace prompting” but “relocate it.” Planning above, structured execution in the middle, tool/model calls below. But I take the point that the paper should lean harder into that framing.

Appreciate the time. Some of the structural recommendations map well to what’s already in the roadmap.