We're starting to offer experimental support for Kobold and Mistral models. Since testing has been insufficient, I can't say they're working perfectly yet. (It's difficult for me to test them since I don't use either service.) If you use these providers and run into any issues, I would greatly appreciate it if you could report them.

Issue Resolved: 'Service Provider Selection' Button Missing in Mod Settings
I apologize for the oversight. Due to a rollback mistake, the button was inadvertently removed. I've now fixed it.
In addition, I've corrected some minor log errors. There are no major changes to the mod.

I've added a slider to control how talkative each colonist is. You can adjust it in the Character -> Persona Settings menu. There are levels from 1 to 5, and setting it to 0 will make the character completely silent.

I've strengthened the prompt combination logic, improving processes like JSON parsing and the automatic removal of special characters.

1. Refactored the God Class
The God Class, which previously contained a jumble of unrelated functionalities, has been refactored according to its specific concerns. The new, separated systems include:

- Conversation Topic Selection System
- Response Generation System
- Combat Topic Generation System

This work was done for development convenience and doesn't affect the final compilation.

2. Improved Topic Generation
I refined the prompts and adjusted the probability of generating topics for object observation and combat.

3. Adjusted Global Prompt
I made some adjustments to the global prompt to further reduce the rate of returned errors. This is a challenging issue that's difficult to resolve in a single pass. When it occurs, it temporarily causes all Gemini models to return an error, so I am approaching its improvement carefully and over the long term. While the error rate has significantly decreased compared to the initial version, inexplicable quota exceeded errors still occur periodically. (This appears to be caused by unexpected behavior in the Gemini API rather than an actual quota issue.)

I've added support for Ollama-based local LLM models. My tests show that they don't perform very well, so while I won't stop you from using them, I don't think they're suitable for the real-time conversation this mod aims for. I'm not planning any further patches for this feature.

You can now add various service providers, including OpenAI and DeepSeek. I've only thoroughly tested OpenAI and Gemini, so other providers may not work correctly. This mod's core logic is built around Gemini's key cycling, and I highly recommend using that method.

I've refined the topics for combat situations. You'll now see settlers react to various scenarios like attacking, being hit, getting injured, and witnessing combat. Please don't test this by killing your teammates.

I've also made several UI improvements and added some debugging logs. These logs will be visible until the mod is stable, and I'll remove them later.

1. You can now register multiple API keys and use them sequentially to distribute your quota more efficiently. To do so, check the "Use multiple API keys" option in the mode settings and register your keys. You can assign an alias to each key for easy identification.

2. I believe I've finally found the culprit for the "quota exceeded" error that was causing an endless loop. After moving the position of the Return statement in the API request prompt from the beginning to the end, the strange phenomenon where an identical prompt would switch between an error and a normal response was resolved. This issue has been fixed.

Fixed an issue where a speech bubble would not appear for a character speaking simultaneously with the first speaker.

The criteria for the dialogue overlap prevention logic has been relaxed. (I was experimenting with a feature that would cancel a planned statement and respond to another character's words if the need to reply was high, even if they had something to say.)

A dedicated topic is now generated for combat actions (e.g., attacking an enemy, attacking a structure, hunting).

Corrected the incorrect key usage mechanism. I've fixed an issue where the game couldn't find the pawn's identity, which was preventing dialogue from being generated.

1.
This mod has a higher average API quota consumption and uses more tokens on average than the original mod. As a result, you may occasionally exceed your quota and find the conversation stops after playing the game for a while.

This was happening more often than intended, and during debugging, I discovered a snowball effect caused by the design.

Originally, when generating dialogue, the mod would pass the top 3 memories by importance from all of a character's memory categories into the prompt. As the game progresses and pawns accumulate more events and conversations, all of their memory categories quickly fill up. When a pawn attempts to speak in this state, a very large amount of text is used as a prompt all at once.

However, the increased token consumption from passing on 'more memories' seems to have a minimal, intuitive effect on generating 'more unique responses.' Therefore, I've changed the mod to select and include just one core memory per utterance, based on a mix of 'importance, freshness, and randomness.' This is a patch aimed at a better trade-off.


2.
I've patched the API request logic to use a more modern method.

I've updated it to use the recommended new UnityWebRequest constructor with a dedicated UploadHandler. This also uses TaskCompletionSource to properly handle asynchronous operations, resulting in a cleaner and more efficient process.

This change improves performance and resource management, especially during periods of heavy API usage.


3.
I've optimized each API request prompt to further reduce the number of tokens requested. (e.g., plain English to JSON format)

Bug Fixes

Fixed a dialogue queue bug.
We found a bug where if you exited to the main menu and then re-entered the game without a full restart, the conversation queue would remain, preventing dialogue from proceeding. We've confirmed that deleting this cache at the start of each game session resolves the issue.

Fixed a dialogue momentum bug.
Another bug was discovered where a character's desire to speak would not properly reset after their turn, causing their willingness to speak (which builds up based on their experiences) to not accumulate correctly. This has also been fixed.