Codex Desktop repeatedly routes to wait(noop) without a running exec cell, causing tool-call loops and stalled subagents

Environment

• Product: Codex Desktop
• OS: macOS (Darwin arm64)
• Date observed: 2026-07-18
• Workspace: FAIRY
• GitHub account used to file: Rinai-R

Summary

Codex Desktop repeatedly routed tool calls to wait.wait with cell_id: "noop" even though no exec cell was running. This created a tool-call loop and prevented the main task from making progress. In the same session, subagent waiting/interruption state also appeared unreliable: waits timed out, interrupted agents could remain listed as running/interrupted, and follow-up routing to completed/interrupted subagents was not dependable.

Symptoms

• The assistant repeatedly attempted a wait call like:

{ "cell_id": "noop", "yield_time_ms": 1000 }

• The runtime returned:

Script error: exec cell noop not found

• There was no known running exec cell corresponding to noop.
• After the error, later attempts still routed back to wait(noop) instead of recovering to normal execution.
• Subagent-related calls also showed stalled or inconsistent state, including repeated:

Wait timed out

• Interrupting a subagent returned state indicating it had been running, e.g.:

"previous_status":"running"

• After timeout/interrupt flows, child tasks could remain running or interrupted, and follow-up messages to completed/interrupted tasks did not reliably resume or complete the original work.

Expected behavior

• wait should only be called with a real active exec cell id previously returned by exec.
• If a stale or invalid cell id is somehow selected, Codex should recover and continue normal tool routing instead of repeatedly calling wait(noop).
• Subagent wait/interrupt/follow-up state should converge: a timed-out or interrupted child should not leave the parent in a stalled loop, and follow-up routing should have predictable behavior.

Actual behavior

• Codex Desktop repeatedly attempted wait(noop).
• The runtime rejected it with exec cell noop not found.
• The parent task was effectively stalled by repeated invalid wait routing.
• Subagent wait/interruption/follow-up handling appeared to compound the stall.

Related issues

These look related but do not appear to cover the exact wait(noop) / nonexistent exec-cell route:

• #24951
• #31036
• #32640
• #33719
• #32997
• #32203
• #32756

Notes

I searched for exact strings such as "exec cell noop", "wait(noop)", and "cell_id" "noop" in openai/codex before filing and did not find an exact duplicate.