dsh-progress-reporter

Long-running tasks in DeepSeek Harness suffer from a quiet but real pain point: the model keeps calling tools in the background while the user sees only scattered log lines, with no clear signal about what has been completed, where things are stuck, or what comes next. dsh-progress-reporter fills exactly that gap by giving Harness a Codex-style progress communication layer, so the model reports at meaningful work stages rather than counting tool invocations mechanically.

Under the hood, the plugin injects a set of progress-communication rules into the model context at session start, agreeing that the model should briefly announce itself when it finishes a meaningful stage, hits a blocker, needs a user decision, or is about to begin a lengthy step. If roughly 60 seconds pass without visible text output, a short collapsed notice is inserted at the next official step boundary, prompting the model to state in the current session language what is done, what is in progress, and what is next or blocked. Work resumes immediately afterward without interrupting the Goal mechanism. The trigger is time-based rather than call-count-based, routine operations are never narrated one by one, and by default only the main agent is asked to report, preventing a flood of simultaneous sub-agent updates. Everything runs locally with no network requests, no credentials, and no persistent storage.

Pick this up if you frequently hand multi-step, long-horizon tasks to Harness and want the reassurance of a reliable colleague checking in, without being buried under verbose tool-call logs. The plugin does not modify Harness core code; it lightly inserts checkpoints at official extension-point boundaries, making upgrades and rollback straightforward.