Skip to main content

Python SDK

Cohort support was introduced in colabhive 0.7.0 as an additive, opt-in API. Since 0.7.1 it also exposes durable endpoint scaling and physical minimum-replica placement.

Install the published release:

python -m pip install colabhive==0.11.0

PyPI wheel SHA-256: 3fd7fc4542a3b83a12eca76c76d5556739f3538608280ff5d9734d4117874754.

The additive namespace is client.cohorts. Existing calls without execution retain their exact Classic request body.

  • capabilities(endpoint_id=None); endpoint-scoped preflight is recommended.
  • plan(...), run(...), get(task_id).
  • events(task_id, cursor=None, limit=100) and finite iter_events(...).
  • iter_events(..., follow=True, timeout=..., cancel_event=...) resumes from the last event ID, deduplicates overlap and retries transient 429/5xx responses.
  • explanation(task_id), cancel(task_id), and bounded/cancellable wait(task_id, timeout=..., cancel_event=...).
  • client.endpoints.infer(..., execution={...}) performs endpoint-scoped capability preflight so an old/off server cannot silently run Classic.
  • client.endpoints.set_scaling(...) persists replica bounds and optional required_node_ids; scale-to-zero remains the default.

Classic inference remains the default when execution is omitted.

The SDK exposes the production Cohort control surface only. It does not install or enable the query-blind textual-memory policy evaluated by R4-M. R4-CEM textual messages and R4-M textual memories belong to frozen research harnesses, not to an implicit SDK execution mode; neither study validates hidden-state or KV-cache efficacy.