Docs
Run forensics
When one agent run does something alarming, this is where you find out what it actually did. Pick a run and see the ordered timeline (tool calls, secret reads, file touches, egress markers), the blast-radius rollup of everything it reached (which files, which secrets, which repos, which external endpoints), and a step-through replay of the stored transcript. If the run is still live, the kill switch stops it before you finish reading.
Where to find it
- Localhost:
/run-forensics.html - API:
GET /api/run-forensics/runs(the picker),GET /api/run/:id/forensics,POST /api/run/:id/kill - Sidebar: Guard, then Run Forensics
- Hosted: team-scoped parity at
/team/run-forensics
What it does for you
Built vs. planned
The timeline, blast radius, replay, the picker and the kill switch all ship today, LLM-free and with no new capture. An unknown run id returns a 404 rather than a fabricated timeline. The token-revoke leg uses the optional secret broker: when the broker is not armed, that leg reports honestly that the token was not revoked while the process-signal and branch-block legs still run. On the hosted side the store is metadata-only by design, so the hosted blast radius shows what is streamed and states plainly that the file, secret and endpoint detail is reconstructed on the desktop. This is dynamic run forensics, which is a different thing from the static code-graph blast radius tab and from the quality-drift agent traces tab.
Last updated