## TL;DR
A frozen console gives no error to handle: the session simply stops answering, and an agent without supervision waits forever. Add the supervision the console lacks: heartbeat checks on the session, a hard inactivity limit after which the worker is declared dead, safe restart that requeues the in-flight sample, and a cap on restarts so a poison sample that freezes every console is quarantined for manual handling instead of looping the fleet.

```text
the triage agent hung when the sandbox console froze mid-detonation and returned an error
```

## Use this when
- The agent shows a detonation in progress for far longer than any normal analysis.
- The sandbox console session is visibly frozen or unresponsive.

## Not for this skill when
- The sample is genuinely long-running within platform limits. Check platform timing before restarting workers.
- The whole sandbox platform is down. Console restarts will not help; use the outage path.

## Steps
1. Confirm the freeze: check the console session responsiveness and the sample elapsed time against normal detonation durations. Verify: freeze versus slow is decided with timings.
2. Terminate the stuck worker and requeue its sample once, recording the freeze against that sample hash. Verify: the worker restarts clean and the sample is tracked.
3. Add inactivity supervision: if a console session shows no progress event within the limit, the supervisor restarts it automatically. Verify: a simulated freeze is recovered without human action.
4. Set a per-sample freeze limit: after repeated freezes, route the sample to manual analysis and flag it as a possible sandbox-evasion attempt. Verify: the poison sample path exists.
5. Run the agent through a normal campaign batch. Verify: detonations complete, and any freeze is recovered and logged.

## Variant phrasings
### sandbox console frozen mid detonation agent hang
The frozen console variant.
### detonation worker supervision restart poison sample
The supervision variant.

Compatibility: Self-managed and vendor sandbox consoles with agent browser or API supervision.

## Why it happens
Consoles are built for a human who notices a frozen screen. An agent sees only the absence of the next event, and without an explicit inactivity model, absence has no deadline, so the hang lasts until a human notices the queue.

## Edge cases / pitfalls
- Freeze patterns are intelligence: samples engineered to crash analysis are telling you something about the campaign.
- Restart storms can cascade across a worker fleet sharing one bad sample. The per-sample cap prevents it.
- Keep console screenshots at freeze time; they often show the evasion behaviour mid-act.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_RLP2ia9r6iLyQi_cZxgKGA