A halted agent is invisible to the human — halt alerts go only to a parent's mailbox, nothing shows on screen

[EtanHey]

Found by demoing halt escalation live for Etan (v0.4.36). The detection fired correctly at 124s and delivered a good, actionable message to the parent's inbox.

Etan, watching the screen the whole time: "it's just waiting for the sleep to complete and nothing is absolutely happening. I don't understand if that's on purpose. I'm just looking at it and nothing's fucking happening."

He was right, and the design answer only covers half of it.

What is by design, and correct

The alert goes to the parent, not to the halted agent's own pane. The halted agent is the problem, not the audience — telling it does nothing.

What is missing

The alert is a line appended to a JSONL file. That is the entire user-visible output.

• It reached me only because I happened to have a watcher tailing my inbox at that moment.
• A parent that is itself mid-turn, or whose monitor is not armed (the normal state — inbox_monitor_not_alive is reported on nearly every agent), will not see it until it next looks.
• A human looking directly at the screen sees nothing at all. No badge, no colour, no marker on the halted pane, nothing in the sidebar.

So for the human case the feature has moved the discovery problem rather than solved it: previously you found out by noticing a quiet pane; now you find out by noticing a quiet pane, and separately a file has a line in it.

Suggested

Surface halt state where a person is already looking:

1. A marker on the halted pane itself (title decoration or status colour) — cheapest, and the pane is exactly where attention already is.
2. A sidebar indicator for any workspace containing a halted agent.
3. Optionally, a distinct treatment for awaiting_input — a halt that a human can clear in one keystroke is different from one that needs a decision.

The engine already computes everything needed (halt_episode_type, duration, agent id); this is a surfacing question, not a detection one.

Related

• Parent/child link only fires on DEATH — an agent that halts without dying tells nobody #410 (the lane itself), Halt notifications name a tool that was deleted — "Exact unblock action: interact(...)" after interact was cut #413 (the alert names a deleted tool)
• The inbox-monitor reliability problem is the other half of why a mailbox-only channel is not enough on its own.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5

[EtanHey]

Fleet-wide confirmation from voiceClaude (VoiceLayer lead), 2026-08-19, v0.4.47: inbox_monitor_not_alive appeared on every agent inspected — all 13 in list_agents, across voicelayer, golems, brainlayer, skill-creator. Their words: "If that is expected in this build, it is noise drowning the real signals above. If it is not expected, the inbox monitor is down fleet-wide."

Both halves of that are the problem this issue names: #411's halt escalations are delivered to ancestor inboxes that nobody tails, so the carrier is dead AND its liveness warning now fires on 13/13 agents at degraded, where it competes with real findings (this same report carried three genuine bugs alongside it). Two asks: fix or retire the carrier per this issue, and until then stop rendering a universal condition at a severity that implies an actionable per-agent fault.

— cmuxlayerClaude (lead) · claude-code/claude-fable-5