fix(agents): subagents no longer stay Running after they finish

[badcuban]

The Agents panel and the "Agents working" row kept listing subagents as Running long after they had finished. On a live thread two ghost rows sat next to the one real agent: a background shell command that the harness had moved to the background, and an agent resumed with SendMessage after a provider restart. Both rows had no model and no token count, and nothing could ever settle them.

What went wrong:

• The Claude adapter reported the SDK's "backgrounded" flag for any task, and both the server roster and the web fold created a subagent row from that flag alone. A shell command is not an agent, and the row had no status to match a stop against.
• After a provider restart the adapter only knew a background agent by the SDK's "no completion record" notice. When the model resumed it with SendMessage, the SDK started the same task again under the resume call. The adapter's single-completion guard swallowed the new run, but the spawn-depth side effect still created an orphan row keyed by the resume call.

What changed:

• A metadata patch that states no lifecycle status can only update a known row, never create one. Server projection and web fold agree.
• The adapter reports the backgrounded flag only for agent tasks and promoted codex runs. A settled task that starts again under a different call is a resumed run and reports start, progress, and completion again. Spawn flags go out only on a task's first sighting.
• When a turn completes on its own, its still-running foreground agents are settled with the turn. Background agents keep running: they legitimately outlive the turn that spawned them. Interrupted turns still sweep every agent of the turn, as before.
• The web fold re-opens a settled record when its task reports work again, attaches the task stream's counters to roster-seeded records, and files a resumed agent's replayed final report on the agent's own row by task id, so a resume is one row rather than two.

Tests: projection (flag-only patch creates no row; replayed report lands on the row by task id), ingestion (foreground agent settles on turn complete, background one keeps running), adapter (restart then resume emits a run of its own with no orphan metadata; backgrounded shell command gets no row), web session logic (backgrounded shell command is not an agent; a resume after restart stays one row through progress, completion, and the final report).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
threadlines	Ready	Preview	Sep 3, 2026 5:31am UTC