Inspect a run
task_status describes execution, not independently verified business success. A completed reflection checkpoint may still produce no proposal or an error. Read the nested outcome and reason.
Wait for the right phase
Await these methods. Export completion does not prove that asynchronous platform persistence or evaluation has finished. A settlement wait timeout can leave underlying work running; it is not the same as shutting down the runtime.
run() coordinates task/evaluation/reflection, but validation settles separately. Release task resources before waiting for replay, and settle required validation before closing Reflect. Close your SDK client separately when the application shuts down.
Tracing and cost
Task, reflection, and validation traces carrypandaprobe.reflect metadata with their role, execution ID, and original task session. Internal work uses separate session attribution and is excluded from recursive task detection. Its model/tool usage still belongs in total operational cost.
Reflect uses the SDK’s native HTTP resources for platform operations. Primary evaluation, optional diagnosis, reflection model calls, and paired replay all add work. Sampling and the configuration budgets bound different parts of that work; no one setting is a complete monetary cap.
The reusable API is also available independently as pandaprobe.api.PlatformClient in Python and PlatformClient from pandaprobe/api in TypeScript. It has a separate HTTP lifecycle from trace export and reuses your client’s resolved configuration.
Common outcomes
A successful task result is preserved if later reflection fails. An exception from the task itself propagates; Reflect does not replace it with a reflection answer or automatically continue the failed invocation.
Recovery and uncertain evaluation creation
Known evaluation run IDs are persisted and polled. Useresume_evaluations() / resumeEvaluations() after restart, or settle a particular pending execution. Safe reads have bounded retries; evaluation creation is sent once.
If a create response is lost, the SDK cannot know whether the server created a billable run. It records creation_unknown and refuses an automatic replacement. Inspect the platform and local record before deciding how to recover. The current API has no idempotency-key contract; deleting the claim and retrying can create a duplicate.
Late feedback needs an explicit continuation. An interrupted claimed reflection checkpoint is not silently re-entered across restart because arbitrary host code may have side effects.

