feat: support for Moonshot LLMs (Kimi 2.5 etc)

[stebbins]

Summary

• add support for Moonshot LLMs
• add moonshot provider declaration
• support reasoning_content when these models use thinking mode + tool call
• source for context etc https://platform.moonshot.ai/docs/introduction#kimi-k2-model

Type of Change

• Feature
• Bug fix
• Refactor / Code quality
• Performance improvement
• Documentation
• Tests
• Security fix
• Build / Release
• Other (specify below)

AI Assistance

• This PR was created or reviewed with AI assistance

Testing

• CI tests
• manual build and testing with Moonshot providers
• NOTE: I do not have OpenAI access, so cannot directly manually "regression" test OpenAI

Related Issues

Closes #6902

Screenshots/Demos (for UX changes)

Before:

After:

[Tyler-Hardin]

Have you tested a longer session? I rebased your PR onto main and it seems to fail with the same error after multiple turns.

$ ./target/release/goose
starting session | provider: custom_kimi model: kimi-k2.5
    session id: 20260210_7
    working directory: /x/qscratch/tyler/goose

goose is running! Enter your instructions, or try asking what goose can do.

Context: ○○○○○○○○○○ 0% (0/131072 tokens)
( O)> Write me an install.sh script that runs cargo build --release and, if it succeeds, copies ./target/release/goose to /export/mana/bin, but first creates a hardlink of the existing copy to a file in the same dir called goose.bk.$(date %Y%m%d.%H%M%S). not sure if my directions will actually achieve what i want -- i want to backup the original without modifying the timestamps. i kinda prefer this hardlink method because it seems more atomic. if we mv the original file first, for a split second there will be no goose installed. but if hardlink causes the backup to accidentally be overwritten, then use mv instead.

you attempted this previously. look at what's currently there in install.sh and tell me if you think it's good to use

─── text_editor | developer ──────────────────────────
path: /x/qscratch/tyler/goose/install.sh
command: view

Looking at the current `install.sh`, here's my assessment:

## What's Good

The script is **well-designed and safe to use**. It actually implements a **better** approach than your original description:

| Your Request | What the Script Does | Assessment |
|-------------|---------------------|------------|
| Hardlink existing, then copy new | Copy new to temp → hardlink existing → atomic mv | ✅ **Better** - The `mv` is truly atomic; your approach would have a window where the file is being copied |
| `goose.bk.$(date %Y%m%d.%H%M%S)` | `goose.bk.$(date +%Y%m%d.%H%M%S)` | ✅ Correct (the `+` is required for `date` command) |
| Fallback to mv if hardlink fails | Catches `ln` failure with `2>/dev/null` and falls back | ✅ Implemented |
| Atomic operation | Uses temp file + atomic `mv` | ✅ Yes - at no point is there no `goose` binary |

## Potential Issue

The backup could be overwritten if you run multiple installs within the same second (same timestamp). To fix this:
─── text_editor | developer ──────────────────────────
path: /x/qscratch/tyler/goose/install.sh
command: str_replace
old_str:     BACKUP="$DEST_DIR/goose.bk.$(date +%Y%m%d.%H%M%S)"
new_str:     BACKUP="$DEST_DIR/goose.bk.$(date +%Y%m%d.%H%M%S).$$"

Ran into this error: Request failed: Bad request (400): thinking is enabled but reasoning_content is missing in assistant tool call message at index 5.

Please retry if you think this is a transient or recoverable error.

⏱️  Elapsed time: 34.95s
Context: ●○○○○○○○○○ 5% (7171/131072 tokens)

[Tyler-Hardin]

My guess is that moonshot might not be echoing back the full reasoning chain. They may only send you back the most recent one. Would be hypocritical of them, but would also explain the error. Idk tho, didn't look deeply.

[stebbins]

@Tyler-Hardin thank you for testing, i will take a look

[Tyler-Hardin]

Thanks for all your work on this. It was already working well for me before your last commit. (Not to suggest the last commit was unneeded, just that the last edge case you fixed wasn't something I had encountered yet.)

[Tyler-Hardin]

By the way, fun fact, there's a bug introduced by 2bc3689 that causes text_editor view to return nothing. I thought it might be a bug in your patch but it's actually in main, predating your changes.

[stebbins]

@Tyler-Hardin cheers and thanks for the heads up. I've been running into more "edge cases" where the reasoning context is lost so trying a few different approaches

[stebbins]

@Tyler-Hardin I think we're good!

[DOsinga]

I think this looks good, but would be good to make sure it still works with like deepseek

[DOsinga]

you are removing the existing comment here, which makes me slightly worried - is this now compatible with both Kimi & Deepseek?

[DOsinga]

ok, I merged the other one, I think we can close this now. let me know if not. and thanks for the contribution @stebbins