kimi-effort Plugin

kimi-effort is a Kimi Code CLI plugin designed to automatically detect the reasoning and thinking effort capabilities of third-party provider models (such as google/gemini-3.8-flash, openai/gpt-5.6-sol, openai/gpt-6-astra, cpa-claude/claude-opus-4-8, claude-opus-4-8) and provide seamless adjustment via the /effort command.

Rather than relying purely on hardcoded rules, kimi-effort combines active live API probing with intelligent heuristics to inspect provider models, write detected capabilities into your ~/.kimi-code/config.toml, and allow real-time tuning of thinking budgets and effort levels.

---

Features

• Live Capability Probing:
  • Probes OpenAI-compatible endpoints with reasoning_effort (low, medium, high) or checks model metadata (e.g. verified on openai/gpt-5.6-sol and google/gemini-3.8-flash).
  • Probes Anthropic endpoints with thinking parameters (budget_tokens), mapping to standard levels (low, medium, high, max) (e.g. verified on cpa-claude/claude-opus-4-8).
  • Probes Google GenAI endpoints with thinkingConfig (thinking_budget), mapping to (low, high).
• Intelligent Heuristics Sniffing:
  • Automatically identifies model families (e.g. gpt-5.6-sol, gpt-6-astra, claude-opus-4-8, gemini-3.8-flash, deepseek-r1) when network probing is unavailable or inconclusive.
  • Correctly marks non-reasoning models (e.g. openrouter/minimax/minimax-m3:free) as non-thinking models.
• Seamless /effort Commands & Skill:
  • Run /effort to view current model (google/gemini-3.8-flash) thinking status and valid effort levels.
  • Run /effort <level> to quickly switch between low, medium, high, max.
  • Run /effort detect or /effort probe to run capability detection on your models.
  • Run /effort list to inspect all configured models at a glance.
• Preserves Configuration:
  • Intelligently updates ~/.kimi-code/config.toml while preserving comments, indentation, structure, and existing settings.
• SessionStart Lifecycle Hook:
  • Automatically checks if your active model has reasoning effort configuration when a session starts, sniffing and setting capabilities in the background without blocking your workflow.

---

Directory Structure

kimi-effort/
├── kimi.plugin.json           # Plugin manifest (name, skills, commands, hooks)
├── README.md                  # Plugin documentation
├── commands/
│   └── effort.md              # Slash command mapping /effort -> skill
├── hooks/
│   └── session-start.mjs      # Fail-open SessionStart lifecycle hook
├── scripts/
│   ├── config-manager.mjs     # Safe parser and editor for config.toml
│   ├── effort-detector.mjs    # Live API prober and heuristic detector
│   └── effort-cli.mjs         # CLI tool for status, detect, set, and list
└── skills/
    └── effort/
        └── SKILL.md           # Kimi Code agent skill for /effort

---

Installation

Method 1: Install from GitHub (Recommended)

Run the /plugins install command inside Kimi Code CLI:

/plugins install https://github.com/WENGENG-boop/kimi-effort-plugin

Method 2: Local Plugin Installation via CLI

Run /plugins install with your local directory path:

/plugins install C:/Users/weo/plugins/kimi-effort

Or copy the directory to the managed plugins path:

mkdir -p ~/.kimi-code/plugins/managed/effort
cp -r C:/Users/weo/plugins/kimi-effort/* ~/.kimi-code/plugins/managed/effort/

After installation, reload your session:

/reload

---

Usage Examples

1. View Current Effort Status

Type /effort or /effort status:

/effort

Outputs the active model, provider, thinking status, current effort level, and supported options. If the active model has not been inspected yet, detection is triggered automatically.

Example Output:

=== Kimi Code Reasoning Effort Status ===

   Config File:     C:\Users\weo\.kimi-code\config.toml
   Model Alias:     google/gemini-3.8-flash (default_model)
   Actual Model:    gemini-3.8-flash
   Provider:        google [openai]
   Thinking State:  Enabled (Capability: Supported)
   Current Effort:  medium (model default)
   Supported Levels:[ low, medium, high ]

2. Adjust Effort Level

Pass the desired effort level directly to /effort:

# Adjust effort for active model (e.g. google/gemini-3.8-flash)
/effort high
/effort low

# Or explicitly target a configured model:
/effort high openai/gpt-5.6-sol
/effort low openai/gpt-6-astra
/effort max cpa-claude/claude-opus-4-8

The plugin validates the requested level against the model's supported choices, updates config.toml, and reminds you if a /reload is recommended.

3. Detect / Re-probe Models

Force re-detection using live API probes and heuristics:

# Probe active model:
/effort detect

# Probe a specific model:
/effort detect openai/gpt-5.6-sol

# Probe all configured models at once:
/effort detect all

Real Detection Results on Configured Models:

Model Alias	Provider	Reasoning Model	Supported Efforts	Default Effort	Detection Method
google/gemini-3.8-flash	google	Yes	[low, medium, high]	medium	Live API Probe
openai/gpt-5.6-sol	openai	Yes	[low, medium, high]	medium	Live API Probe
openai/gpt-6-astra	openai	Yes	[low, medium, high]	medium	Heuristics
cpa-claude/claude-opus-4-8	cpa-claude	Yes	[low, medium, high, max]	high	Live API Probe
claude-opus-4-8	custom	Yes	[low, medium, high]	medium	Live API Probe
openrouter/minimax/minimax-m3:free	openrouter	No	[]	(none)	Heuristics

4. List All Configured Models

View all configured models and their reasoning capabilities:

/effort list

---

CLI Runner Direct Invocation

You can also execute the standalone CLI runner directly with Node.js:

# Check status of active or specific model
node scripts/effort-cli.mjs status
node scripts/effort-cli.mjs status openai/gpt-5.6-sol

# Probe all configured models and update config.toml
node scripts/effort-cli.mjs detect all

# Set thinking effort
node scripts/effort-cli.mjs set high
node scripts/effort-cli.mjs set max cpa-claude/claude-opus-4-8

# List all models
node scripts/effort-cli.mjs list

---

Supported Reasoning Levels by Model Family

Model Family / Example	Detected Effort Options	Default Effort	Detection Method
OpenAI GPT-5.6 Sol (gpt-5.6-sol)	low, medium, high	medium	Live API Probe / Heuristic
OpenAI GPT-6 Astra (gpt-6-astra)	low, medium, high	medium	Heuristic / Live API Probe
Anthropic Claude Opus 4.8 (claude-opus-4-8)	low, medium, high, max	high	Live API Probe / Heuristic
Google Gemini 3.8 Flash (gemini-3.8-flash)	low, medium, high	medium	Live API Probe / Heuristic
DeepSeek R1 (deepseek-r1)	default, high	high	Heuristic / Metadata
Non-Reasoning Models (minimax-m3:free)	(none / unsupported)	(none)	Catalog / Heuristic

---

Frequently Asked Questions (FAQ)

Q: Why doesn't /effort in the TUI show the new model's effort levels (e.g. max) after switching models?

A: This is due to how Kimi Code CLI caches configuration in memory:

1. When you launch Kimi Code, it reads config.toml into memory (availableModels).
2. When you switch models (via /model) or when the plugin detects/updates support_efforts in config.toml, the active TUI session retains the previously loaded in-memory model specifications until refreshed.
3. Solution: Simply type /reload in Kimi Code. This tells Kimi Code to re-read config.toml from disk, instantly updating the in-memory /effort picker with the newly active model's full set of effort options (such as max).
4. Model Alias Verification: Also check if you have multiple aliases configured for the same model. For example, if you configured both claude-opus-4-8 under an OpenAI proxy (which only supports [low, medium, high]) and cpa-claude/claude-opus-4-8 under native Anthropic (which supports [low, medium, high, max]), ensure you selected the Anthropic-backed alias (cpa-claude/claude-opus-4-8).

---

License

MIT