dsh-resume-turn

Slow or unstable model endpoints tend to drop a streaming reply midway through generation—time out, lose the transport, hit a rate-limit, or return a server error. The default dsh-llm-retry treats the failure as a brand-new call: everything that already streamed (visible text, the entire reasoning chain) is discarded, and a multi-minute generation can be repeated several times before it finally lands, burning tokens and patience along the way.

dsh-resume-turn intervenes at the exact point of failure. It pulls the already-streamed partial output—visible text plus reasoning chunks—from the session event log, crafts a steering message that quotes that partial content and instructs the model to pick up where it left off, then takes over recovery on the retry chain. The next request starts from the interruption point, not from zero. Up to three auto-resumes are allowed per turn before the default retry strategy takes back over; backoff starts at two seconds and doubles each attempt up to a thirty-second cap, with optional knobs for eligible failure codes and the maximum characters quoted into the resume prompt.

It is built for anyone running long conversations or extended thinking chains against slow, flaky inferencing endpoints—AMD DeepSeek included—where an occasional disconnect is the norm rather than the exception, and restarting from scratch each time simply does not scale.