Investigating Runs

Everything on this page is read-only — the agent only reads, so there is nothing it can change while answering.

Why a run failed, or is stuck

Ask things like “why did the nightly sync fail?”, “why is this task stuck?”, or “what did the last run return?”. The agent resolves fuzzy names to templates and runs, walks the execution tree — including referenced sub-tasks — to the node that actually caused the problem, and answers with the root cause plus clickable deep links into the UI for every hop of the path.

Where a run’s time went

Ask “why was last night’s import so slow?” and the agent calls task_report, which answers for a finished run in two parts. The first is the run’s table, always covering the whole run: one row per node that ran, with its run count (and how many failed or were cut short by the task’s end), and its work, wait, defer and re-dispatch numbers, each as a total plus the worst single run. The second is the waterfall: one entry per individual run, in order, with its wait / work / defer stretches, its error, the service name and version and the replica that ran it, and the sub-task a sub-task node started — which the agent can follow into a report of its own.

Pass a trace id and only the waterfall narrows to that one message path; the table stays whole-run, so it does not move under you. For a task that has not finished the tool refuses and says so: ask “what is it waiting on?” instead, which is a different question with different tools behind it.

Example prompts

• “Why did the nightly import fail?”
• “Where did the time go in the last run of the nightly import?”
• “Is anything stuck right now? What is it waiting on?”