fix(server): task_status reports how long a running delegated task has been silent#23898

Problem

A delegated Codex child can sit in Working with no provider events at all. Neither the parent agent (task_status) nor the user can tell it apart from a turn that is doing work. This is the silent post-tool case in #17503. Below is the notification sequence @juliusmarminge asked for, from the last tool result through the end of each silence.

Versions: T3 Code Nightly 0.1.1-nightly.20261020.3411 (macOS), codex-cli 0.163.0 over codex app-server, model gpt-6.1-sol via a Responses-compatible CLIProxyAPI on loopback. Codex stream settings are defaults.

Silence 1: a hung request after a tool result (15m35s, never answered)

Silence 2: the same thread after the restart (5m13s before first output, then normal)

The proxy logged 499 | 15m35s for the first request when the client closed it, and 200 | 5m15s for the second. It logged no retries. These are slow or hung upstream requests that never return response headers, not a 429 retry sequence. Codex's stream idle timeout never fires, so it emits nothing and T3 has no signal to act on. A 5-minute silence can also be healthy, as the second one was, so this PR reports idleness and does not time anything out.

Change

• task_status (contract, server, tool description, docs) now returns three fields for running tasks:

  • lastActivityAt: the last recorded change to the child thread.
  • idleForMs: time since then.
  • hasRunningItem: whether an open item belongs to the run's active attempt. Items left open by an attempt that a restart superseded don't count.

  A long idleForMs with workState: working, hasRunningItem: false and no pending request points to a stalled provider. The parent can then restart or cancel the task. All three are null/false for queued and finished tasks.

• Web Working row: after 3 minutes without item activity, with nothing in progress and no pending request, the row adds a quiet "· No activity for N" hint. A pure helper in TimelineWorkGroupRows.logic.ts derives this from data the client already has.

• Not included:

  • Timeouts and auto-fail.
  • Adapter changes.
  • Mobile. Its working indicator is separate and would be a follow-up.
  • Streaming items with no output. An open text or reasoning item still counts as in progress, because reasoning can legitimately be silent. Codex's stream idle timeout covers a hang in the middle of a stream.

Scope and approval

This addresses #17503, specifically its "no no-progress signal on an active turn" gap. The reporter also asked for that state to be "visible … to parent agents through task_status for delegated tasks". The PR stays informational (no new timeouts or retries) so that it fits the problem the issue already establishes. It does not address the issue's systemError recovery or 429 backoff parts.

Before / after

MessagesTimeline rendered on its own with a running turn whose last activity was 15 minutes ago. Upstream ignores the idle input, and the two "before" captures are byte-identical.

Fixture-driven component captures, not a live stuck provider. The desktop and mobile shells were not exercised. Full sanitized notification sequence.

Verification

• OrchestratorMcpService.activity.test.ts (using TestClock) covers:
  • idleForMs growing as time advances;
  • open items from a superseded attempt vs. the active attempt;
  • an open item with no provider turn (workspace preparation);
  • completed items and items from inactive runs;
  • finished tasks returning nulls.
• TimelineWorkGroupRows.logic.test.ts covers:
  • the 3-minute threshold;
  • suppression while an item is in progress;
  • the restart baseline;
  • ignoring items from other runs;
  • missing attempt or start returning null;
  • formatting.
• The new tests fail on upstream, where the fields and helper don't exist.
• vp test run on the MCP service, activity and toolkit integration tests (44 passed) and the web logic tests (6 passed).
• vp run check:changed upstream/main passed: format, lint, knip, typecheck for 29 packages, and 191 tests.
• Three independent review rounds, none blocking. Their wording, query-cost, item-filter, render-flash and test-gap findings were fixed, including tests that fail when the lazy-loading guard is removed.
• Not checked: a live stalled provider end to end, Electron, mobile.

Done by Claude Opus 5.5 in Claude Code, running in T3 Code.

🤖 Generated with Claude Code