Burrow's MCP tools, tuned against real agent transcripts — one analyze call maps disk hotspots, killed runs say so instead of failing silently, and slow scans warn you up front.
improved
- One
burrow_analyzecall now maps disk hotspots. The engine reports one directory level per run, which forced agents into a call per directory (a real session made 15 in a row).analyzegainsdepth(descend into the largest subdirectories),limit, andmin_size, emits compact JSON instead of 35–60 KB pretty-printed blobs, and always reports what it pruned (entries_omitted/omitted_bytes,partial: truewhen the descent hits its time budget). (#303) - The slow tools now say they're slow.
analyze,clean,purge, andinstallerdescriptions warn that big scans take minutes, so agents scope to a specific folder instead of hanging past their client's patience and falling back to shell commands. The agent docs and theburrow-system-toolsskill now lead with the low-disk emergency playbook — the pattern that actually fires in practice. (#303)
fixed
- Killed runs no longer fail silently. A
burrow_cleanthat hit its time limit rendered as{"exit_code": 9, "output": ""}— nothing an agent could act on. Timed-out actions (and analyze) now returntimed_out: trueplus a hint. (#303) burrow_cleanup_historyexplains itself. When engine history is unavailable, the error now points atburrow_infoto check whether Burrow is recording at all. (#303)