Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
117 lines
4.8 KiB
Markdown
117 lines
4.8 KiB
Markdown
# System-Hint Agent Changelog
|
|
|
|
## 2026-07-18 - kimi-k3 (reasoning model) + offline status-bar preview
|
|
|
|
### Changes Made
|
|
|
|
- **Default model is now `kimi-k3`** (Moonshot's current flagship reasoning model),
|
|
replacing the retired `kimi-k2-*-preview` line. `provider="kimi"`/`"moonshot"`
|
|
both resolve to `kimi-k3` unless `--model` overrides it.
|
|
- **Reasoning-model handling**: `kimi-k3` returns a separate `reasoning_content`
|
|
field alongside the final `content`. The agent reads the answer from `content`
|
|
(FINAL ANSWER detection unchanged), calls with `max_tokens=8192` (well above the
|
|
reasoning minimum of 2048), and forces `temperature=1` for reasoning models via
|
|
`_reasoning_safe_temperature()` (kimi-k3 / gpt-5). Assistant turns are replayed
|
|
with `model_dump()` (which includes `reasoning_content`); the Moonshot API accepts
|
|
this on follow-up calls, so multi-turn tool loops work unchanged.
|
|
- **Offline status-bar preview** (`python main.py --mode preview`): renders all five
|
|
status-bar techniques as before/after comparisons with **no API key and no LLM
|
|
call**. Honors `--no-timestamps` / `--no-counter` / `--no-todo` / `--no-errors`
|
|
/ `--no-state` to isolate individual techniques.
|
|
|
|
## 2025-09-30 - Trajectory Logging Enhancement
|
|
|
|
### Changes Made
|
|
|
|
#### 1. Full LLM Messages in Trajectory
|
|
Added tracking of the complete messages list sent to the LLM, including system hints:
|
|
|
|
- **Added field**: `last_llm_messages` to `SystemHintAgent` class
|
|
- Stores the full messages array sent to the LLM, including the system hint appended as a user message
|
|
- This differs from `conversation_history` which only stores the base conversation without the dynamic system hints
|
|
|
|
- **Modified methods**:
|
|
- `__init__`: Initialize `last_llm_messages = None`
|
|
- `execute_task`: Capture `messages_to_send` before LLM call and store as `self.last_llm_messages`
|
|
- `_save_trajectory`: Include `last_llm_messages` in the trajectory JSON output
|
|
- `reset`: Reset `last_llm_messages` to `None`
|
|
|
|
#### 2. Real System Time (No Mock Time)
|
|
Verified and ensured real system time is used throughout:
|
|
|
|
- **Default configuration**: `simulate_time_delay = False` (line 68 in agent.py)
|
|
- When `False`, uses `datetime.now()` for all timestamps
|
|
- When `True` (only for demos), uses simulated time
|
|
|
|
- **Timestamp sources**:
|
|
- `_get_timestamp()`: Uses `datetime.now()` when `simulate_time_delay=False`
|
|
- `trajectory_data['timestamp']`: Always uses `datetime.now().isoformat()`
|
|
- Tool call timestamps: Always use `datetime.now().isoformat()`
|
|
- TODO item timestamps: Always use `datetime.now().isoformat()`
|
|
|
|
### Benefits
|
|
|
|
1. **Complete LLM Context**: The `last_llm_messages` field in trajectory.json now shows exactly what was sent to the LLM, including dynamic system hints about current state, TODO list, timestamps, etc.
|
|
|
|
2. **Debugging**: Easier to debug agent behavior by seeing the complete context the LLM received, not just the conversation history
|
|
|
|
3. **Accurate Timestamps**: All timestamps reflect real system time for accurate trajectory analysis and debugging
|
|
|
|
### Example Trajectory Structure
|
|
|
|
```json
|
|
{
|
|
"timestamp": "2025-09-30T20:26:32.057323",
|
|
"iteration": 1,
|
|
"provider": "kimi",
|
|
"model": "kimi-k3",
|
|
"conversation_history": [
|
|
{"role": "system", "content": "..."},
|
|
{"role": "user", "content": "[2025-09-30 20:26:00] Task..."},
|
|
{"role": "assistant", "content": "..."}
|
|
],
|
|
"last_llm_messages": [
|
|
{"role": "system", "content": "..."},
|
|
{"role": "user", "content": "[2025-09-30 20:26:00] Task..."},
|
|
{"role": "assistant", "content": "..."},
|
|
{"role": "user", "content": "=== SYSTEM STATE ===\nCurrent Time: 2025-09-30 20:26:32\n..."}
|
|
],
|
|
"tool_calls": [...],
|
|
"todo_list": [...],
|
|
"current_directory": "/path/to/dir",
|
|
"final_answer": null,
|
|
"config": {
|
|
"enable_timestamps": true,
|
|
"enable_tool_counter": true,
|
|
"enable_todo_list": true,
|
|
"enable_detailed_errors": true,
|
|
"enable_system_state": true,
|
|
"timestamp_format": "%Y-%m-%d %H:%M:%S",
|
|
"simulate_time_delay": false
|
|
}
|
|
}
|
|
```
|
|
|
|
### Differences: conversation_history vs last_llm_messages
|
|
|
|
- **conversation_history**: Permanent record of the conversation between user and assistant
|
|
- System prompt
|
|
- User messages (with timestamps if enabled)
|
|
- Assistant responses
|
|
- Tool call messages and results
|
|
|
|
- **last_llm_messages**: Complete snapshot of what was sent to LLM in the last call
|
|
- Everything from conversation_history
|
|
- PLUS: Dynamic system hint appended as final user message
|
|
- Shows current system state, TODO list, directory, time
|
|
- This is what the LLM actually sees when making decisions
|
|
|
|
### Testing
|
|
|
|
All changes have been tested and verified:
|
|
- ✅ `last_llm_messages` correctly captured and saved
|
|
- ✅ Real system timestamps used (not simulated time)
|
|
- ✅ Trajectory JSON format validated
|
|
- ✅ No linter errors
|
|
- ✅ Backward compatible with existing code
|