Skip to content

The desktop app shows a team session whose turn the harness stopped as failed, and says why in words

Decision 760 · accepted · 2026-10-03 · issue #354

Issue #354, defect D53's first bullet, following 745 and 750. - A plane's row is read for its reason. rowFromPlane maps failed_reason to FleetRow.failed as {reason, detail: null}, the field a daemon's rows fill (745), so the list's row, the launcher's and the review queue's status say Failed for a team session as they do for a local one, and nothing that reads failed has a new shape to learn. detail is null because a plane holds no content (750). A plane from before the column, or a row without one, says nothing failed. - Each reason has its words. failedTitle says "The agent kept crashing and the session stopped" for agent_failed, as before, and "A tool kept failing and the harness stopped the turn" for tool_failures: the failure guard counts one tool's failures in a row, so it is a tool and not the agent. A reason the app does not know reads "The turn failed ()" rather than nothing. - A local session still cannot say tool_failures. The daemon's failed is agent_failed only (745), so those words reach a local row only once a daemon lists it, and the transcript still shows such a turn as the failure guard's question answered stop, as 745 left it. A notification that reads a row's failed follows, though a plane's rows count nothing unseen and so raise none. - Proof: the client's fleet test (a plane row with each reason, one with none, one from before the column) and the desktop app's team-failed test (a fake plane lists a session stopped on each reason beside one that finished; the launcher's recent rows and the list say Failed, with the reason's words on hover), both failing on the tip; and the web build against pnpm fake, which now seeds a run stopped on tool_failures.

Governs: clients/gui