hermes - 💡(How to fix) Fix vision_analyze fails with kimi-coding provider: native vision fast path blocked [1 pull requests]

Bug Description

When using <code>kimi-coding</code> provider with model <code>kimi-for-coding</code> (Kimi K2.6), the <code>vision_analyze</code> tool fails with:

<pre>Error analyzing image: No LLM provider configured for task=vision provider=auto. Run: hermes setup</pre>

This happens even though Kimi K2.6 supports multimodal/vision natively.

Root Cause Analysis

Two issues chain together:

1. Provider not in native vision allow-list

<code>_supports_media_in_tool_results()</code> in <code>tools/vision_tools.py</code> does not include <code>kimi-coding</code>, so the native fast path is skipped.

2. Config model name not resolvable in models.dev

Config uses friendly name <code>kimi-for-coding</code>, but models.dev cache stores the canonical ID <code>k2p6</code>. <code>_find_model_entry()</code> cannot bridge this gap, so <code>get_model_capabilities()</code> returns <code>None</code>.

This causes <code>decide_image_input_mode()</code> to fall back to <code>'text'</code> mode, which triggers the legacy auxiliary-LLM path. Since no auxiliary vision provider is configured, it fails.

Reproduction Steps

1. Set config: <code>model.default: kimi-for-coding</code>, <code>model.provider: kimi-coding</code>
2. Send an image to the agent
3. Agent calls <code>vision_analyze</code>
4. Error: <code>No LLM provider configured for task=vision</code>

Expected Behavior

<code>vision_analyze</code> should use the native fast path, returning a multimodal tool-result envelope with the base64-encoded image, allowing the main model (Kimi K2.6) to see the pixels directly.

Environment

• Provider: kimi-coding
• Model: kimi-for-coding (Kimi K2.6)
• Hermes version: main @ e85592591

Proposed Fix

1. Add <code>kimi-coding</code>, <code>kimi-coding-cn</code>, <code>kimi</code>, <code>moonshot</code> to <code>_supports_media_in_tool_results()</code>
2. Add <code>_MODEL_ALIASES</code> mapping in <code>_find_model_entry()</code> to bridge friendly config names to canonical models.dev IDs

---

I have a branch ready with the fix.