fix: persist and emit agent response during verification stop loop

[yingliang-zhang]

Problem

When verify-on-stop or pre_verify triggers, the agent produces a complete response (e.g., a detailed summary with tables and analysis), but that response is suppressed from both the durable transcript (state.db) and the UI. The user only sees the terse post-verification reply — never the full answer.

This is the bug behind repeated user reports of "系统消息冲走了回复" (system messages flushed the response). PR #62657 attempted to fix this by removing the synthetic flag from final_msg, but only addressed DB persistence — the response was still not emitted to the UI.

Root Cause (Three Layers)

Layer 1: DB persistence (fixed by #62657)

final_msg["_verification_stop_synthetic"] = True → _is_ephemeral_scaffolding() returns True → _flush_messages_to_session_db() skips it → not persisted

Layer 2: UI emission (fixed by initial #62676)

No call to _emit_interim_assistant_message() → not displayed

Layer 3: Streaming buffer overwrite (fixed by this update)

Even with _emit_interim_assistant_message() called, the Desktop gateway streams the response via stream_delta_callback. _interim_content_was_streamed() detects already_streamed=True, so the callback only calls on_segment_break() (paragraph break) instead of on_commentary() (standalone message bubble). The streamed text lives in the streaming buffer and gets overwritten by subsequent verification messages.

Fix

run_agent.py — _emit_interim_assistant_message() gains force_display parameter:

• When True, already_streamed is forced to False
• The gateway callback calls on_commentary(text) → standalone permanent message bubble
• The full answer survives the verification loop

agent/conversation_loop.py — both verify-on-stop and pre_verify paths:

1. Remove synthetic flag from final_msg → persists to DB
2. Call _emit_interim_assistant_message(final_msg, force_display=True) → standalone UI bubble
3. Call _flush_messages_to_session_db() → persist before verification loop
4. Change nudge from role="user" to role="system" → internal instruction, not durable user message

Tests

File	Changes
test_verification_continuation_budget.py	Role sequence updated; final_msg no longer synthetic; new test asserts already_streamed=False (force_display)
test_verification_stop_caching.py	Assistant response now persists (not dropped); nudge still dropped

Results: 5/5 pass (test_verification_continuation_budget.py), 50/50 pass (test_verification_stop.py)

Supersedes #62657

PR #62657 only fixed DB persistence. This PR is a complete fix: DB persistence + UI emission + streaming buffer survival.

[teknium1]

Thanks for tracing both the persistence and UI paths; the suppression is real on the checked-out main (agent/conversation_loop.py:5153-5176, 5210-5227).

Problems

• The new role="system" nudges at agent/conversation_loop.py:5199 and its pre_verify counterpart are not repaired away before the next request. repair_message_sequence only repairs assistant/tool/user shapes (agent/agent_runtime_helpers.py:408-550), while the request builder forwards every live message and prepends the stable system prompt (agent/conversation_loop.py:787-851). This creates a mid-history synthetic system turn.
• Flushing the blocked answer before the synthetic nudge means the nudge is omitted by persistence (run_agent.py:1834-1848) but both assistant reports survive. On the next request, consecutive assistants are merged (agent/agent_runtime_helpers.py:429-467), so the attempted and verified reports do not remain distinct. The new callback test does not cover this persisted/resume path.

Suggested changes

• Keep a protocol-valid continuation mechanism rather than inserting a system turn mid-history.
• Separate UI preview from durable transcript state, or add an end-to-end persistence/resume test for both verification paths.

Automated hermes-sweeper review.