dsh-continue

Long-running inference tasks often break not on complex reasoning but on a single transient timeout or a routine confirmation that freezes the pipeline. dsh-continue is built for exactly this: it deterministically recovers from brief network interruptions, narrowly automates unique recommendation selection, binary continue prompts, and valid Plan Mode exits, while delegating every ambiguous, risky, or unverified case back to the official UI.

Its core capabilities center on safe unattended continuation: adaptive backoff recovery for TIMEOUT, TRANSPORT, and SERVER codes (RATE_LIMIT is deliberately left to dsh-camel); decision memory that only records verified manual selections; and continuation candidates that remain suggest-only and never bypass the risk boundary. Every executable capability is off by default; installation alone changes no retry, learning, or approval behavior.

Best suited for projects running long inference pipelines where low-risk prompts should not block progress, yet no privilege escalation, credential access, payment, or irreversible action should ever be automated. If your primary need is generic model-error retry or web-UI message rewriting rather than task-level continuation, the comparison table in the README will point you to a more fitting tool.