hermes - 💡(How to fix) Fix Post-tool-call nudge fires before prefill for structured-reasoning thinking models (e.g. qwen3-vl-8b-thinking via OpenRouter) [1 pull requests]

curl -s "https://openrouter.ai/api/v1/chat/completions" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-vl-8b-thinking",
    "provider": {"order": ["alibaba", "parasail"], "allow_fallbacks": false},
    "reasoning": {"enabled": true, "effort": "medium"},
    "stream": true,
    "messages": [
      {"role": "user", "content": "what is the weather in Casablanca?"},
      {"role": "assistant", "content": "", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "web_search", "arguments": "{\"query\":\"weather Casablanca\"}"}}]},
      {"role": "tool", "tool_call_id": "call_1", "content": "Casablanca 22 C sunny."}
    ],
    "tools": [{"type": "function", "function": {"name": "web_search", "description": "Search", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]}}}]
  }'

---

# Fires first
if (
    _prior_was_tool
    and not getattr(agent, "_post_tool_empty_retried", False)
    and not _has_inline_thinking  # <-- only guards against inline <think>, not structured reasoning
):
    # NUDGE — emits ⚠️ warning, adds synthetic (empty) assistant + nudge user, continues
    ...
    continue

# Unreachable for OpenRouter-style structured reasoning models
_has_structured = bool(
    getattr(assistant_message, "reasoning", None)
    or getattr(assistant_message, "reasoning_content", None)
    or getattr(assistant_message, "reasoning_details", None)
    or _has_inline_thinking
)
if _has_structured and agent._thinking_prefill_retries < 2:
    # PREFILL — never gets here when nudge fires first
    ...

---

_has_structured = bool(
    getattr(assistant_message, "reasoning", None)
    or getattr(assistant_message, "reasoning_content", None)
    or getattr(assistant_message, "reasoning_details", None)
    or _has_inline_thinking
)
if (
    _prior_was_tool
    and not getattr(agent, "_post_tool_empty_retried", False)
    and not _has_structured  # widened from _has_inline_thinking
):
    # NUDGE
    ...

---

12:32:45 INFO API call #2: ... in=16346 out=360 total=16706 latency=3.9s  (tool call)
12:32:46 INFO tool web_search completed (1.29s, 2164 chars)
12:32:48 INFO API call #3: ... in=17514 out=36  total=17550 latency=1.2s  (reasoning-only)
12:32:48 INFO Empty response after tool calls — nudging model to continue processing
12:32:52 INFO API call #4: ... in=17225 out=404 total=17629 latency=4.5s  (final answer after nudge)