Timmy_Foundation/hermes-agent
15
0
Fork 0
Code Issues 64 Pull Requests 4 Actions 1 Packages Projects Releases Wiki Activity

Implement reasoning extraction and enhance assistant message handling

Browse Source 
- Added a new method `_extract_reasoning` to extract reasoning content from assistant messages, accommodating multiple formats from various providers.
- Updated message handling to ensure all assistant messages include reasoning content for API compatibility, preserving multi-turn reasoning context.
- Enhanced logging to capture reasoning details for debugging and analysis.
- Modified the TODO.md to reflect changes in planning and task management, emphasizing the need for structured task decomposition and progress tracking.
This commit is contained in:
teknium1
2026-02-01 22:48:18 -08:00
parent 9b4d9452ba
commit e114f09f70
2 changed files with 128 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files

94
TODO.md
View File

@@ -177,56 +177,48 @@ These items need to be addressed ASAP:
---
## 2. Context Management (complements Subagents)
## 2. Planning & Task Management 📋
**Problem:** Context grows unbounded during long conversations. Trajectory compression exists for training data post-hoc, but live conversations lack intelligent context management.
**Problem:** Agent handles tasks reactively without explicit planning. Complex multi-step tasks lack structure, progress tracking, and the ability to decompose work into manageable chunks.
**Ideas:**
- [ ] **Incremental summarization** - Compress old tool outputs on-the-fly during conversations
- Trigger when context exceeds threshold (e.g., 80% of max tokens)
- Preserve recent turns fully, summarize older tool responses
- Could reuse logic from `trajectory_compressor.py`
- [ ] **Task decomposition tool** - Break complex requests into subtasks:
```
User: "Set up a new Python project with FastAPI, tests, and Docker"
- [ ] **Semantic memory retrieval** - Vector store for long conversation recall
- Embed important facts/findings as conversation progresses
- Retrieve relevant memories when needed instead of keeping everything in context
- Consider lightweight solutions: ChromaDB, FAISS, or even a simple embedding cache
Agent creates plan:
├── 1. Create project structure and requirements.txt
├── 2. Implement FastAPI app skeleton
├── 3. Add pytest configuration and initial tests
├── 4. Create Dockerfile and docker-compose.yml
└── 5. Verify everything works together
```
- Each subtask becomes a trackable unit
- Agent can report progress: "Completed 3/5 tasks"
- [ ] **Working vs. episodic memory** distinction
- Working memory: Current task state, recent tool results (always in context)
- Episodic memory: Past findings, tried approaches (retrieved on demand)
- Clear eviction policies for each
- [ ] **Progress checkpoints** - Periodic self-assessment:
- After N tool calls or time elapsed, pause to evaluate
- "What have I accomplished? What remains? Am I on track?"
- Detect if stuck in loops or making no progress
- Could trigger replanning if approach isn't working
- [ ] **Explicit plan storage** - Persist plan in conversation:
- Store as structured data (not just in context)
- Update status as tasks complete
- User can ask "What's the plan?" or "What's left?"
- Survives context compression (plans are protected)
**Files to modify:** `run_agent.py` (add memory manager), possibly new `tools/memory_tool.py`
- [ ] **Failure recovery with replanning** - When things go wrong:
- Record what failed and why
- Revise plan to work around the issue
- "Step 3 failed because X, adjusting approach to Y"
- Prevents repeating failed strategies
**Files to modify:** `run_agent.py` (add planning hooks), new `tools/planning_tool.py`
---
## 3. Self-Reflection & Course Correction 🔄
**Problem:** Current retry logic handles malformed outputs but not semantic failures. Agent doesn't reason about *why* something failed.
**Ideas:**
- [ ] **Meta-reasoning after failures** - When a tool returns an error or unexpected result:
```
Tool failed → Reflect: "Why did this fail? What assumptions were wrong?"
→ Adjust approach → Retry with new strategy
```
- Could be a lightweight LLM call or structured self-prompt
- [ ] **Planning/replanning module** - For complex multi-step tasks:
- Generate plan before execution
- After each step, evaluate: "Am I on track? Should I revise the plan?"
- Store plan in working memory, update as needed
- [ ] **Approach memory** - Remember what didn't work:
- "I tried X for this type of problem and it failed because Y"
- Prevents repeating failed strategies in the same conversation
**Files to modify:** `run_agent.py` (add reflection hooks in tool loop), new `tools/reflection_tool.py`
---
## 4. Tool Composition & Learning 🔧
## 3. Tool Composition & Learning 🔧
**Problem:** Tools are atomic. Complex tasks require repeated manual orchestration of the same tool sequences.
@@ -257,7 +249,7 @@ These items need to be addressed ASAP:
---
## 5. Dynamic Skills Expansion 📚
## 4. Dynamic Skills Expansion 📚
**Problem:** Skills system is elegant but static. Skills must be manually created and added.
@@ -286,7 +278,7 @@ These items need to be addressed ASAP:
---
## 6. Task Continuation Hints 🎯
## 5. Task Continuation Hints 🎯
**Problem:** Could be more helpful by suggesting logical next steps.
@@ -336,7 +328,7 @@ These items need to be addressed ASAP:
---
## 8. Resource Awareness & Efficiency 💰
## 6. Resource Awareness & Efficiency 💰
**Problem:** No awareness of costs, time, or resource usage. Could be smarter about efficiency.
@@ -373,7 +365,7 @@ These items need to be addressed ASAP:
---
## 10. Project-Local Context 💾
## 7. Project-Local Context 💾
**Problem:** Valuable context lost between sessions.
@@ -393,7 +385,7 @@ These items need to be addressed ASAP:
---
## 11. Graceful Degradation & Robustness 🛡️
## 8. Graceful Degradation & Robustness 🛡️
**Problem:** When things go wrong, recovery is limited. Should fail gracefully.
@@ -414,7 +406,7 @@ These items need to be addressed ASAP:
---
## 12. Tools & Skills Wishlist 🧰
## 9. Tools & Skills Wishlist 🧰
*Things that would need new tool implementations (can't do well with current tools):*
@@ -481,7 +473,7 @@ These items need to be addressed ASAP:
---
## 13. Messaging Platform Integrations 💬
## 10. Messaging Platform Integrations 💬
**Problem:** Agent currently only works via `cli.py` which requires direct terminal access. Users may want to interact via messaging apps from their phone or other devices.
@@ -525,7 +517,7 @@ These items need to be addressed ASAP:
---
## 14. Scheduled Tasks / Cron Jobs ⏰
## 11. Scheduled Tasks / Cron Jobs ⏰
**Problem:** Agent only runs on-demand. Some tasks benefit from scheduled execution (daily summaries, monitoring, reminders).
@@ -570,7 +562,7 @@ These items need to be addressed ASAP:
---
## 15. Text-to-Speech (TTS) 🔊
## 12. Text-to-Speech (TTS) 🔊
**Problem:** Agent can only respond with text. Some users prefer audio responses (accessibility, hands-free use, podcasts).
@@ -601,7 +593,7 @@ These items need to be addressed ASAP:
---
## 16. Speech-to-Text / Audio Transcription 🎤
## 13. Speech-to-Text / Audio Transcription 🎤
**Problem:** Users may want to send voice memos instead of typing. Agent is blind to audio content.

107
run_agent.py
View File

@@ -980,6 +980,49 @@ class AIAgent:
# Check if there's any non-whitespace content remaining
return bool(cleaned.strip())
def _extract_reasoning(self, assistant_message) -> Optional[str]:
"""
Extract reasoning/thinking content from an assistant message.
OpenRouter and various providers can return reasoning in multiple formats:
1. message.reasoning - Direct reasoning field (DeepSeek, Qwen, etc.)
2. message.reasoning_content - Alternative field (Moonshot AI, Novita, etc.)
3. message.reasoning_details - Array of {type, summary, ...} objects (OpenRouter unified)
Args:
assistant_message: The assistant message object from the API response
Returns:
Combined reasoning text, or None if no reasoning found
"""
reasoning_parts = []
# Check direct reasoning field
if hasattr(assistant_message, 'reasoning') and assistant_message.reasoning:
reasoning_parts.append(assistant_message.reasoning)
# Check reasoning_content field (alternative name used by some providers)
if hasattr(assistant_message, 'reasoning_content') and assistant_message.reasoning_content:
# Don't duplicate if same as reasoning
if assistant_message.reasoning_content not in reasoning_parts:
reasoning_parts.append(assistant_message.reasoning_content)
# Check reasoning_details array (OpenRouter unified format)
# Format: [{"type": "reasoning.summary", "summary": "...", ...}, ...]
if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details:
for detail in assistant_message.reasoning_details:
if isinstance(detail, dict):
# Extract summary from reasoning detail object
summary = detail.get('summary') or detail.get('content') or detail.get('text')
if summary and summary not in reasoning_parts:
reasoning_parts.append(summary)
# Combine all reasoning parts
if reasoning_parts:
return "\n\n".join(reasoning_parts)
return None
def _get_messages_up_to_last_assistant(self, messages: List[Dict]) -> List[Dict]:
"""
Get messages up to (but not including) the last assistant turn.
@@ -1318,22 +1361,20 @@ class AIAgent:
for msg in messages:
api_msg = msg.copy()
# For assistant messages with tool_calls, providers require 'reasoning_content' field
# Extract reasoning from our stored 'reasoning' field and add it as 'reasoning_content'
if msg.get("role") == "assistant" and msg.get("tool_calls"):
# For ALL assistant messages, pass reasoning back to the API
# This ensures multi-turn reasoning context is preserved
if msg.get("role") == "assistant":
reasoning_text = msg.get("reasoning")
if reasoning_text:
# Add reasoning_content for API compatibility (Moonshot AI, Novita, etc.)
# Add reasoning_content for API compatibility (Moonshot AI, Novita, OpenRouter)
api_msg["reasoning_content"] = reasoning_text
# Remove 'reasoning' field - it's for trajectory storage only
# The reasoning is already in the content via <think> tags AND
# we've added reasoning_content for API compatibility above
# We've copied it to 'reasoning_content' for the API above
if "reasoning" in api_msg:
api_msg.pop("reasoning")
# Remove 'reasoning_details' if present - we use reasoning_content instead
if "reasoning_details" in api_msg:
api_msg.pop("reasoning_details")
# Keep 'reasoning_details' - OpenRouter uses this for multi-turn reasoning context
# The signature field helps maintain reasoning continuity
api_messages.append(api_msg)
if active_system_prompt:
@@ -1694,14 +1735,16 @@ class AIAgent:
# Reset retry counter on successful JSON validation
self._invalid_json_retries = 0
# Extract reasoning from response if available (for reasoning models like minimax, kimi, etc.)
# Extract reasoning from response for storage
# The reasoning_content field will be added when preparing API messages
reasoning_text = None
if hasattr(assistant_message, 'reasoning') and assistant_message.reasoning:
reasoning_text = assistant_message.reasoning
elif hasattr(assistant_message, 'reasoning_content') and assistant_message.reasoning_content:
reasoning_text = assistant_message.reasoning_content
# Extract reasoning from response if available
# OpenRouter can return reasoning in multiple formats:
# 1. message.reasoning - direct reasoning field
# 2. message.reasoning_content - alternative field (some providers)
# 3. message.reasoning_details - array with {summary: "..."} objects
reasoning_text = self._extract_reasoning(assistant_message)
if reasoning_text and self.verbose_logging:
preview = reasoning_text[:100] + "..." if len(reasoning_text) > 100 else reasoning_text
logging.debug(f"Captured reasoning ({len(reasoning_text)} chars): {preview}")
# Build assistant message with tool calls
# Content stays as-is; reasoning is stored separately and will be passed
@@ -1723,6 +1766,14 @@ class AIAgent:
]
}
# Store reasoning_details for multi-turn reasoning context (OpenRouter)
if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details:
assistant_msg["reasoning_details"] = [
{"type": d.get("type"), "text": d.get("text"), "signature": d.get("signature")}
for d in assistant_message.reasoning_details
if isinstance(d, dict)
]
messages.append(assistant_msg)
# Execute each tool call
@@ -1810,6 +1861,10 @@ class AIAgent:
current_tokens=self.context_compressor.last_prompt_tokens
)
# Save session log incrementally (so progress is visible even if interrupted)
self._session_messages = messages
self._save_session_log(messages)
# Continue loop for next response
continue
@@ -1865,11 +1920,11 @@ class AIAgent:
self._empty_content_retries = 0
# Extract reasoning from response if available
reasoning_text = None
if hasattr(assistant_message, 'reasoning') and assistant_message.reasoning:
reasoning_text = assistant_message.reasoning
elif hasattr(assistant_message, 'reasoning_content') and assistant_message.reasoning_content:
reasoning_text = assistant_message.reasoning_content
reasoning_text = self._extract_reasoning(assistant_message)
if reasoning_text and self.verbose_logging:
preview = reasoning_text[:100] + "..." if len(reasoning_text) > 100 else reasoning_text
logging.debug(f"Captured final reasoning ({len(reasoning_text)} chars): {preview}")
# Build final assistant message
# Content stays as-is; reasoning stored separately for trajectory extraction
@@ -1879,6 +1934,14 @@ class AIAgent:
"reasoning": reasoning_text # Stored for trajectory extraction
}
# Store reasoning_details for multi-turn reasoning context (OpenRouter)
if hasattr(assistant_message, 'reasoning_details') and assistant_message.reasoning_details:
final_msg["reasoning_details"] = [
{"type": d.get("type"), "text": d.get("text"), "signature": d.get("signature")}
for d in assistant_message.reasoning_details
if isinstance(d, dict)
]
messages.append(final_msg)
if not self.quiet_mode: