Description

[Issue Summary]
Streamed assistant response cuts off mid-sentence with no period, no error event, no retry indicator. Symptomatically matches NVBug 6096024 (closed-verified 2026-04-24) —
suspected regression.

[Environment]

• macOS, NemoClaw latest (verified 2026-04-28)
• Model: moonshotai/kimi-k2.5 (NVIDIA Endpoints)
• OpenClaw Web UI at 127.0.0.1:18789/chat

[Steps to Reproduce]

1. nemoclaw onboard, open dashboard in browser.
2. Send this prompt in a fresh session:

Run these 3 commands in the sandbox shell, one at a time, and after each explain what its output means in 2 sentences: hostname / date / uptime. Then summarize all three
  in a single paragraph.

[Actual]
After date exec succeeded, the agent streamed:

 "2. date — The date command shows that the current system time is Tuesday, April 28, 2026 at 12:44:35 UTC. ... approximately 15 minutes ahead of the timestamp in your
  ▎ message which showed 12:29 UTC when you sent the initial"     

Token meter: Tool turn ↑14.1k ↓40 then Assistant turn ↑14.5k ↓127 — a new turn opened while the previous was visibly incomplete.

Reproduced across session=ns3retry1 and session=agent:main:ns3retry2. - Could check attachment

[Expected]
Sentence completes with proper boundary, OR UI shows a clear "response may be incomplete" indicator if the run was cut short.

[Impact]

• Multi-step responses become unreliable; users see partial text with no indication anything went wrong.
• Truncation appears to corrupt agent state — next turn forgets the in-progress task (see related bug 6122540).

[Suggested Fix Direction]
- Verify whether response.output_text.delta events are being silently dropped at certain context sizes / timeouts.

• Add a UI guard: assistant turns closed without a sentence-ending chunk should be flagged as incomplete.

Bug Details

[NVB#6122543]