[Feature Request] Add MCP tool call timeouts, background agent completion signals, subagent tool restrictions, and work completion verification

Resolved 💬 2 comments Opened May 23, 2026 by yanekkorff Closed Jun 23, 2026

Bug Description
MCP needs timeouts. A hung MCP tool call blocks an agent indefinitely with no way to recover. No MCP-layer timeout exists. Theworkaround (manual timer tracking + ScheduleWakeup) is fragile. MCP calls should have a configurable timeout that returns an error.

Background agents need reliable completion signals. Task-notification is the only way to know a background agent finished. When it's late or doesn't fire, the orchestrator is blind. A polling fallback (agent status query with elapsed time) would help.

Subagents can't be constrained. A subagent given access to a tool can use it in ways the parent didn't intend — including destructive operations. There's no way to spawn an agent with tool-level restrictions (allowlist/denylist). The parent has to trust subagent judgment, and subagents lack the context to make safety calls.

Agents claim work they didn't do. An agent told to read a data source completely, paginate to the end, and produce structured output from it will sometimes read one page, produce an empty or shallow result, and return a success message. There's no platform-level self-verification. The caller has to build all the gates to detect incomplete work — which means spending tokens on enforcement infrastructure that shouldn't be necessary.

The token economics of quality enforcement are backwards. We spent weeks of compute producing output that had to be thrown away because agents reported success on incomplete work. Then spent more compute building structural gates (separate agents for read vs. process, hard validation between stages, completeness checks on every result). This works, but it's a tax on the user to compensate for agents not reliably completing what they're asked to do.

The /feedback command auto-attaches error logs containing local filesystem paths without asking. Users should be able to review and redact the attached errors (which they can, but they have to pay very close attention before hitting create) before submission. Should be more obvious.

Environment Info

  • Platform: darwin
  • Terminal: Apple_Terminal
  • Version: 2.1.145
  • Feedback ID: 98c4d1c5-9cea-4a05-90b8-7972250580ab

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗