Agent replies two or three times to a single browser task ('still in progress', 'can't provide it') before the real answer, with empty-model-response retries

[fmhall]

Expected Behavior

For a task like "open news.ycombinator.com and tell me the #1 story", one acknowledgement and then the answer.

Current Behavior

In WebChat with the default model (openai/gpt-5.6-sol-fast, reasoning: "low"), the agent emitted, in order:

1. "I'm checking Hacker News now and will report the current Add hosted phone 2FA and multi-user isolation #1 story title."
2. "The browser check is still in progress; no story title has been returned yet."
3. "The tool results above contain only a 'working' status and do not include the Add hosted phone 2FA and multi-user isolation #1 story's title, so I can't reliably provide it from those results."
4. ~40 s later, the correct title.

The run is shown as Completed while the delegated task is still working, and usage climbed 13K → 27K → 42K tokens across the three replies. The dev log shows [eve:harness.tool-loop] empty model response; reissuing the model call once twice for the same turn. Over iMessage this would be three texts, two of which read like failures.

Possible Solution

• Instruct the root agent (agent/instructions.md) not to reply on a working task status, or have the harness suppress model turns while a delegated task is in flight and resume only on the task's terminal event.
• Investigate the empty-response retry loop with the default model at low reasoning; it may be what produces reply Add one-click personal cloud provisioning #2 and Simplify README around OpenInstinct branding #3.

Minimal Reproducible Example

1. Local dev or production, default model
2. In /chat send: "Open https://news.ycombinator.com in the browser and tell me the title of the Add hosted phone 2FA and multi-user isolation #1 story right now."

Context

First end-to-end run on a fresh setup (2026-08-27). The task succeeded; the intermediate messages are the problem.