dsh-subagent-watchdog

In a DSH workflow, the most common frustration when delegating a multi-step task to a continuable subagent is that the child burns through its output-token budget mid-task and stops. The unfinished work sits there, and the delegating parent has to notice the failure, decide it is recoverable, and manually continue the child. dsh-subagent-watchdog turns that single recovery step into a deterministic automation: it watches exclusively for the max-tokens terminal signal, sends exactly one continuation through the official subagents.followup() seam, and lets the same durable session cold-resume from its existing state. No loops, no polling, zero extra dependencies.

The safety boundary is the design priority. The continue-once guarantee is enforced against the child's durable session id plus a continuation marker written into the child's own log, so plugin restarts or duplicate settlements can never trigger a second automatic continuation. If the recovered run fails again, whether with another max-tokens or a runtime error, Watchdog stops intervening immediately and delivers a single notice to the parent with the failure facts and official manual options. It never guesses and never auto-retries.

If you routinely delegate to continuable subagents inside DSH workflows, once this plugin is in place a subagent that hits its token ceiling simply picks up where it left off, and you no longer need to hand-write retry logic or babysit child-agent liveness.