Summary

When using provider: copilot-acp with model: claude-sonnet-4.6, many queries that cause the Copilot CLI to attempt a tool call (e.g. shell commands to read config files) result in empty responses delivered to the user as (empty).

Root Cause

The Copilot CLI, when running in --acp --stdio mode, is a full agent that:

1. Generates agent_thought_chunk events (internal reasoning) while thinking
2. Sends a session/update with sessionUpdate: "tool_call" to announce a tool it wants to run
3. Sends session/request_permission to request execution approval

Hermes currently always denies session/request_permission with outcome: cancelled. When the CLI receives a cancellation, it terminates the turn with stopReason: end_turn — emitting zero agent_message_chunk events (no visible text output).

Hermes then receives:

• content = "" (empty)
• reasoning = "<thoughts>" (non-empty)

This triggers run_agent.py's thinking-only response detection, which prefill-retries twice, exhausts its budget, and delivers (empty) to the user.

Observed session log

[0] role=user   content='Help me configure matrix access, based on openclaw matrix config'
[1] role=assistant  finish=incomplete  content=''   ← prefill retry 1
[2] role=assistant  finish=incomplete  content=''   ← prefill retry 2  
[3] role=assistant  finish=stop        content='(empty)'  ← exhausted

And in the UI:

↻ Thinking-only response — prefilling to continue (1/2)
↻ Thinking-only response — prefilling to continue (2/2)
⚠️ Empty response from model — retrying (1/3)

Reproducer

# config.yaml
model:
  default: claude-sonnet-4.6
  provider: copilot-acp

Send any message that causes the CLI to try a shell/file-system tool, e.g.:

"Help me configure matrix access based on openclaw config at ~/.openclaw"

Fix

After a session/prompt turn that produces only thought chunks (tool denied), send a follow-up session/prompt in the same session asking the model to answer from existing knowledge without running commands. If the follow-up also produces no text, surface the original reasoning as the reply text. This ensures run_agent.py always receives a non-empty response and the thinking-only loop never fires.