Make a hand-written agent loop durable

Goal: end this skill with the user's own agent loop — their prompt, their tools, their model call — running as a Dapr Workflow, where every model call and every tool call is its own activity. And with proof: a worker killed in the middle of a tool call resumes at that tool, and the model is not asked again for a turn it already answered. A loop that only survives when nothing crashes has gained nothing.

Nothing here depends on where the Dapr sidecar runs. The same code runs against a local sidecar, on Kubernetes or on Catalyst; section 13 is the only part specific to Catalyst.

1. This skill, or an adapter?

This skill is for a loop the user wrote: code that calls a model API directly — Anthropic, OpenAI, Bedrock, a local model — reads the tool calls out of the reply, runs them, appends the results and calls again.