daarion-admin/microdao-daarion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: DSML fallback — 3rd LLM call for clean synthesis + think tag stripping

Browse Source 
Router (main.py):
- When DSML detected in 2nd LLM response after tool execution,
  make a 3rd LLM call with explicit synthesis prompt instead of
  returning raw tool results to the user
- Falls back to format_tool_calls_for_response only if 3rd call fails

Router (tool_manager.py):
- Added _strip_think_tags() helper for <think>...</think> removal
  from DeepSeek reasoning artifacts

Gateway (http_api.py):
- Strip <think>...</think> tags before sending to Telegram
- Strip DSML/XML-like markup (function_calls, invoke, parameter tags)
- Ensure empty text after stripping gets "..." fallback

Deployed to NODE1 and verified services running.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Apple
2026-02-09 10:30:37 -08:00
parent 990e594a1d
commit 7887f7cbe9
3 changed files with 53 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
services/router/main.py
View File

@@ -1055,8 +1055,41 @@ async def agent_infer(agent_id: str, request: InferRequest):
# CRITICAL: Check for DSML in second response too!
if response_text and "DSML" in response_text:
logger.warning(f"🧹 DSML detected in second LLM response, clearing ({len(response_text)} chars)")
response_text = format_tool_calls_for_response(tool_results, fallback_mode="dsml_detected")
logger.warning(f"🧹 DSML detected in 2nd LLM response, trying 3rd call ({len(response_text)} chars)")
# Third LLM call: explicitly ask to synthesize tool results
tool_summary_parts = []
for tr in tool_results:
if tr.get("success") and tr.get("result"):
res_text = str(tr["result"])[:500]
tool_summary_parts.append(f"Tool '{tr['name']}' returned: {res_text}")
if tool_summary_parts:
synthesis_prompt = "Based on the following tool results, provide a helpful response to the user in their language. Do NOT use any markup or XML. Just respond naturally.\n\n" + "\n".join(tool_summary_parts)
try:
synth_resp = await http_client.post(
f"{cloud['base_url']}/v1/chat/completions",
headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
json={"model": cloud["model"], "messages": [
{"role": "system", "content": system_prompt or "You are a helpful assistant. Respond naturally."},
{"role": "user", "content": synthesis_prompt}
], "max_tokens": max_tokens, "temperature": 0.3, "stream": False},
timeout=cloud["timeout"]
)
if synth_resp.status_code == 200:
synth_data = synth_resp.json()
synth_text = synth_data.get("choices", [{}])[0].get("message", {}).get("content", "")
if synth_text and "DSML" not in synth_text and "invoke" not in synth_text:
response_text = synth_text
tokens_used += synth_data.get("usage", {}).get("total_tokens", 0)
logger.info("\u2705 3rd LLM call synthesized clean response from tool results")
else:
response_text = format_tool_calls_for_response(tool_results, fallback_mode="dsml_detected")
else:
response_text = format_tool_calls_for_response(tool_results, fallback_mode="dsml_detected")
except Exception as synth_err:
logger.warning(f"3rd LLM call failed: {synth_err}")
response_text = format_tool_calls_for_response(tool_results, fallback_mode="dsml_detected")
else:
response_text = format_tool_calls_for_response(tool_results, fallback_mode="dsml_detected")
if not response_text:
logger.warning(f"⚠️ {cloud['name'].upper()} returned empty response after tool call")