- I implemented universal logging for all API providers in src/models/, ensuring calls to logger.js for text and vision logs.
- I added transformation of <thinking>...</thinking> tags to <think>...</think> in all provider responses before logging, for correct categorization by logger.js.
- I standardized the input to logger.js's log() function to be a JSON string of the message history (system prompt + turns).
- I removed unnecessary comments from most API provider files, settings.js, and prompter.js to improve readability.
Note: I encountered some issues that prevented final comment cleanup for qwen.js, vllm.js, and logger.js. Their core logging functionality and tag transformations (for qwen.js and vllm.js) are in place from previous steps.
I implemented comprehensive logging across all API providers in src/models/ using logger.js.
This includes:
- Adding log() and logVision() calls to each provider (Claude, DeepSeek, Gemini, GLHF, GPT, Grok, Groq, HuggingFace, Hyperbolic, Local, Mistral, Novita, Qwen, Replicate, VLLM).
- Ensuring logging respects 'log_normal_data', 'log_reasoning_data', and 'log_vision_data' flags in settings.js, which I added.
- I deprecated 'log_all_prompts' in settings.js and updated prompter.js accordingly.
I refactored openrouter.js and prompter.js:
- I removed the experimental reasoning prompt functionality ($REASONING) from openrouter.js.
- I removed a previously implemented (and then reverted) personality injection feature ($PERSONALITY) from prompter.js, openrouter.js, and profile files.
I had to work around some issues:
- I replaced the full file content for glhf.js and hyperbolic.js due to persistent errors with applying changes.
Something I still need to do:
- Based on your latest feedback, model responses containing <thinking>...</thinking> tags need to be transformed to <think>...</think> tags before being passed to logger.js to ensure they are categorized into reasoning_logs.csv. This change is not included in this update.
This update finalizes the implementation of three distinct vision modes:
- "off": This disables all my vision capabilities.
- "prompted": (Formerly "on") This allows me to use vision via explicit commands from you (e.g., !lookAtPlayer), and I will then summarize the image.
- "always": (Formerly "always_active") I will automatically take a screenshot every time you send a prompt and send it with your prompt to a multimodal LLM. If you use a look command in this mode, I will only update my view and take a screenshot for the *next* interaction if relevant, without immediate summarization.
Here are the key changes and improvements:
1. **Bug Fix (Image Path ENOENT)**:
* I've corrected `Camera.capture()` so it returns filenames with the `.jpg` extension.
* I've updated `VisionInterpreter.analyzeImage()` to handle full filenames.
* This resolves the `ENOENT` error that was previously happening in `Prompter.js`.
2. **Vision Mode Renaming**:
* I've renamed the modes in `settings.js` and throughout the codebase: "on" is now "prompted", and "always_active" is now "always".
3. **Core Framework (from previous work, now integrated)**:
* I've added `vision_mode` to `settings.js`.
* `Agent.js` now manages `latestScreenshotPath` and initializes `VisionInterpreter` with `vision_mode`.
* `VisionInterpreter.js` handles different behaviors for each mode.
* My vision commands (`!lookAt...`) respect the `off` mode.
* `History.js` stores `imagePath` with turns, and `Agent.js` manages this path's lifecycle.
* `Prompter.js` reads image files when I'm in "always" mode and passes `imageData` to model wrappers.
4. **Extended Multimodal API Support**:
* `gemini.js`, `gpt.js`, `claude.js`, `local.js` (Ollama), `qwen.js`, and `deepseek.js` have been updated to accept `imageData` in their `sendRequest` method and format it for their respective multimodal APIs. They now include `supportsRawImageInput = true`.
* Other model wrappers (`mistral.js`, `glhf.js`, `grok.js`, etc.) now safely handle the `imageData` parameter in `sendRequest` (by ignoring it and logging a warning) and have `supportsRawImageInput = false` for that method, ensuring consistent behavior.
5. **Testing**: I have a comprehensive plan to verify all modes and functionalities.
This set of changes provides a robust and flexible vision system for me, catering to different operational needs and supporting various multimodal LLMs.
Add Deepseek-R1 Support.
Deepseek-R1 uses "<think> [Thoughts] </think>" and the response can be cut short if it reasons through a command and it's syntax, so we cut it short before returning the response.
