Kimi K2.5

Kimi K2.5 Integration

Overview

Kimi K2.5 is the most capable subagent available in this project. Use it when tasks require:

• Complex multi-step reasoning
• Batch image/vision analysis (10+ images)
• Very long context understanding (256K tokens)
• Thinking mode for difficult problems

Quick Setup

Launcher Script

.\scripts\start-kimi.ps1

Manual Environment Setup

$env:ANTHROPIC_BASE_URL = "https://api.moonshot.cn/anthropic/"
$env:ANTHROPIC_API_KEY = "<KIMI_API_KEY>"
$env:ANTHROPIC_MODEL = "kimi-k2.5-thinking"
$env:ANTHROPIC_SMALL_FAST_MODEL = "kimi-k2-turbo-preview"

Available Models

Tip: The -thinking variant uses chain-of-thought reasoning, making it better for complex problems but slower for simple queries.

When to Use Kimi K2.5

✅ Use Kimi For:

• Complex reasoning requiring multiple steps
• Batch vision (10+ images to analyze)
• Long documents (large codebase exploration)
• Difficult problems where GLM/MiniMax failed
• Cross-file analysis requiring broad context

❌ Don't Use Kimi For:

• Simple web searches (use MiniMax)
• Quick file lookups (use MiniMax)
• Creative brainstorming (use GLM-4.7)
• Tasks where speed matters more than quality

Vision Capabilities

Kimi K2.5 has built-in vision. No separate model needed.

Single Image Analysis

Prompt: "Analyze this sprite for pixel art quality issues"
Image: [attached or URL]

Batch Image Analysis

For 10+ images, spawn Kimi as subagent:

Task(
  prompt="Analyze each of these 20 sprites for: transparency, outline quality, shading consistency. Return a table.",
  subagent_type="general-purpose"
)

Delegation Patterns

Pattern 1: Fallback After GLM Fails

1. Claude tries GLM for creative task
2. GLM output is inadequate
3. Claude retries with Kimi K2.5
4. Kimi provides deeper analysis

Pattern 2: Long-Context Research

1. Claude needs to understand 50+ files
2. Claude spawns Kimi: "Read these files and summarize patterns"
3. Kimi processes with 256K context
4. Claude receives condensed findings

Pattern 3: Multi-Step Reasoning

1. Claude faces complex architectural decision
2. Claude spawns Kimi: "Analyze trade-offs between approaches A, B, C"
3. Kimi provides detailed reasoning chain
4. Claude makes final decision

API Configuration

Base URL

https://api.moonshot.cn/anthropic/

Headers (Anthropic-compatible)

Authorization: Bearer <KIMI_API_KEY>
Content-Type: application/json

Example Request

curl -s -X POST "https://api.moonshot.cn/anthropic/v1/messages" \
  -H "Authorization: Bearer sk-kimi-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k2.5-thinking",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Your prompt here"}]
  }'

Comparison with Other Providers

Troubleshooting

"Connection refused"

Check API key is correctly set:

echo $env:ANTHROPIC_API_KEY

"Model not found"

Use exact model names:

• ✅ kimi-k2.5-thinking
• ✅ kimi-k2-turbo-preview
• ❌ kimi-k2.5 (incomplete)

"Rate limited"

Kimi has generous rate limits but may throttle during peak usage. Wait 30 seconds and retry.

"Vision not working"

Ensure image is:

• PNG, JPEG, or WebP format
• Under 20MB
• Accessible (local path or public URL)

Integration with Project

Circe's Garden Use Cases

1. Sprite batch analysis: Analyze all 45+ placeholder sprites for quality
2. Dialogue consistency: Check 80+ dialogue files for narrative consistency
3. Codebase architecture: Understand patterns across game/features/*
4. Visual target comparison: Compare screenshots against Harvest Moon reference

Example: Sprite Quality Audit

Task(
  prompt="""
  Analyze these sprite files for:
  1. Transparency issues (blocky backgrounds)
  2. Outline consistency (1-2px dark outline expected)
  3. Shading quality (SNES Harvest Moon style)
  4. Color palette compliance (see docs/reference/concept_art/HERAS_GARDEN_PALETTE.md)

  Return a table: filename | issues | severity (1-5) | recommendation
  """,
  subagent_type="general-purpose"
)

See Also

• /skill delegation - Provider selection matrix
• /skill token-efficient-delegation - Cost optimization patterns
• /skill image-analysis - GLM-4.6v alternative for vision

[Opus 4.5 - 2026-01-29]