Environment

• OpenClaw version: 2026.3.23-1
• OS: macOS (Mac mini M4, Apple Silicon)
• Model: kimi/k2p5 (Moonshot Kimi K2.5)
• Channel: webchat (Control UI) + TUI
• Thinking: high then medium

Bug Summary

After tool calls return successfully, the agent produces no user-visible output. The agent run ends silently. User must send a follow-up message to wake the agent, which then responds immediately with new content (not delayed previous output).

Related Issues

• Agent silent failure: tool errors cause no response to user #12595 - Agent silent failure (closed as not planned)
• Agent ends silently after message tool without executing next planned tool #5336 - Agent ends silently after message tool
• [Bug]: Agent hangs indefinitely on failed tool calls — no timeout, no recovery, no fallback #8288 - Agent hangs on failed tool calls

Reproduction Steps

1. Send request requiring multiple parallel tool calls
2. Agent spawns 3-6 parallel tool calls
3. All tool results return successfully
4. Agent enters thinking phase
5. Agent run ENDS - no user-visible text generated
6. User sends follow-up message
7. Agent immediately responds with complete analysis

Key Observations

1. Frequency increases with context size (above 50% utilization)
2. Correlated with parallel tool calls (3-6 concurrent results)
3. Thinking mode amplifies the issue (high worse than medium)
4. Also occurs without tool calls (pure text replies)
5. TUI vs Web UI difference - TUI shows thinking status, Web UI disconnects

Timeline (2026-03-24 UTC+8)

• 11:18-11:24: 6 parallel searches -> silence -> user ping -> immediate response
• 11:37-11:39: 3 kimi_search -> silence -> user ping -> immediate response
• 11:52: Simple question -> silence -> user ping -> immediate response
• 12:42-12:43: No tool calls, text reply -> silence (twice in 1 minute)

Workarounds

1. reply-guard plugin (before_prompt_build hook injecting prependContext reminder)
2. Reduced thinkingDefault from high to medium
3. System prompt size reduction by 26%

Hypothesis

Agent loop state machine transitions to terminal state after thinking phase completes, without verifying user-visible output was generated. Related to #5336 root cause.