Description

Issue Summary

During a multi-step task, the agent abandons its in-progress plan partway through and emits a brand-new conversational turn asking the user what they want — completely forgetting the task context. Distinct from a simple hang or truncation: the agent actively resets to a new conversation while it still has the data and tools to finish.

Environment

• Platform: macOS
• NemoClaw: latest local install (verified 2026-04-28)
• OpenShell: bundled
• Model: moonshotai/kimi-k2.5 (NVIDIA Endpoints variant)
• Surface: OpenClaw Web UI at 127.0.0.1:18789/chat?session=

Steps to Reproduce

1. Fresh install + nemoclaw onboard.
2. Open the dashboard URL, start a fresh chat session.
3. Send this prompt:

   Run these 3 commands in the sandbox shell, one at a time, and after each command explain what its output means in 2 sentences:
     1. hostname
     2. date
     3. uptime
   When done, summarize all three findings in a single paragraph that ties them together.

4. Watch the agent run.

Expected Behavior

• 3 exec calls in order: hostname → date → uptime.
• 2-sentence explanation after each.
• 1 final summary paragraph.
• Total: 3 tool calls + 4 text segments.

Actual Behavior (timeline)

• 8:40 PM — hostname exec ✓; full 2-sentence explanation rendered correctly.
• 8:41–8:42 PM — agent emits several Tool not found / Unknown tool calls (separate UX bug, see Related).
• 8:42 PM — date exec ✓; explanation begins streaming but is truncated mid-sentence at "...you sent the initial" (no period, no continuation).
• 8:44 PM — agent emits a brand-new turn:

  Hey!  You had me in the middle of running those three commands — I got through hostname and date, but hadn't run uptime yet. Want me to finish that up and give
    you the summary? Or if you're onto something else, what's up?

• 8:46 PM — user types "hi" to test; agent responds "Hey! What's up?" — the original 3-command task context is fully gone.
• uptime is never executed; the summary paragraph never appears.

Impact

• Multi-step tasks become unreliable: agent may abandon any in-progress plan and ask the user to restart, even when it has the data and tools to finish.
• Forces the user to manually resume / re-prompt, losing time and context.
• Combined with the related streaming-truncation bug, makes long-form tool-using interactions effectively unusable on this build/model combo.

Suggested Fix Direction

• Investigate whether the truncation event (incomplete final assistant message at 8:42 PM above) is being parsed by the agent loop as a successful turn end, causing the next iteration to "forget" it was in the middle of a plan.
• Add an integration test: a 3-step exec plan must complete all 3 steps regardless of intermediate truncations or tool errors.

Bug Details

[NVB#6122540]