kimi-code-model-management

name	kimi-code-model-management
description	Switch kimi-code CLI models — add new model definitions, change default_model, test, and verify. Covers the config.toml schema, provider binding, model capability flags, and the Kimi Code managed OAuth provider. Load when Arif says "tukar model kimi-code", "switch kimi model", "pakai K3 untuk kimi", "update kimicode", or any task touching /root/.kimi-code/config.toml model config.

Kimi Code Model Management

Config Location

/root/.kimi-code/config.toml

Kimi Code binary at /root/.kimi-code/bin/kimi.

Model Docs

Official model reference: https://www.kimi.com/code/docs/en/kimi-code/models

Provider Architecture

Two providers in the config:

Provider	Type	Base URL	Auth	Models
managed:kimi-code	kimi	https://api.kimi.com/coding/v1	OAuth (file)	K3, K2.7 Code, K2.7 HighSpeed
minimax-coding-plan	anthropic	https://api.minimax.io/anthropic	API key	MiniMax M2/M2.1/M2.5/M2.7/M3

Known Model IDs

Model ID (kimi-code)	Actual Model	Context	Capabilities	Notes
k3	Kimi K3 (2.8T)	256k–1M (plan-tiered)	thinking, always_thinking, image_in, video_in, tool_use	Latest flagship. Reasoning: max only. Released Jul 16 2026. Context: Moderato=256k, Allegretto+=1M. Massive quota burn on cheap plans.
kimi-for-coding	Kimi K2.7 Code	256k	thinking, always_thinking, image_in, video_in, tool_use	Mature/stable coding model
kimi-for-coding-highspeed	Kimi K2.7 Code (6×)	256k	thinking, always_thinking, image_in, video_in, tool_use	6× speed, 3× quota usage

Adding a New Model

1. Add a [models."kimi-code/<model-id>"] block after existing model definitions
2. Required fields: provider, model, max_context_size, capabilities, display_name
3. For managed:kimi-code provider models, use provider = "managed:kimi-code"

Template

[models."kimi-code/<model-id>"]
provider = "managed:kimi-code"
model = "<model-id>"
max_context_size = <262144 for Moderato / 1048576 for Allegretto+>
support_efforts = [ "low", "high", "max" ]
default_effort = "high"
capabilities = [ "thinking", "always_thinking", "image_in", "video_in", "tool_use" ]
display_name = "<Human Name>"

CRITICAL for K3: always_thinking is mandatory. Without it, K3 silently routes to K2.6. The docs say: "K3 / K2.7 without Thinking routes to K2.6."

Switching Default Model

• --output-format text gives plain text output (default is streaming JSON)
• Expect: response identifying the new model
• Exit code 0 = working
• If OAuth error: run kimi login first (see OAuth pitfall below)

Pitfalls

Planning / Configuration

• 🚨 TEST BEFORE DOCUMENTING: The #1 pitfall — Arif WILL say "context is wrong" if you skip this. Do NOT change default_model and then immediately update AGENTS.md claiming the new model is active. Always run a test prompt first: kimi -m <model> -p "what model are you?" --output-format text and verify the model identity in the response. If the test fails (OAuth: auth.login_required, 404: Model not exist, 429: rate limit, or wrong model in response), REVERT the config change AND the AGENTS.md change immediately. Only update AGENTS.md after a successful test. 2026-07-18 case study: Changed default_model to kimi-code/k3 without verifying OAuth. Empty token = auth.login_required. Arif: "context is wrong u setup." Reverted both config.toml and AGENTS.md. Lesson: OAuth check FIRST.
• AGENTS.md OVERRIDES config.toml + DOUBLE-STALE DRIFT: Kimi-code injects /root/.arifos/agents/kimi/AGENTS.md as system prompt. The **Model:** line there becomes the model identity the LLM sees, overriding default_model in config.toml. When switching models, update BOTH files. When testing, run from /tmp (no project AGENTS.md) to isolate config.toml behavior. AGENTS.md is the single highest-drift file — in practice it's often TWO upgrades behind (e.g. deepseek-v4-pro while config is MiniMax-M3, or MiniMax-M3 while config is kimi-code/k3). Always read AGENTS.md even when you think you only need to touch config.
• Ask plan tier before setting context: K3 max_context_size is plan-tiered. Moderato=262144, Allegretto+=1000000, Andante=not supported. Never assume 1M. Ask: "What Kimi plan are you on?"
• K3 always_thinking is MANDATORY: Without it in capabilities, K3 silently routes to K2.6. Docs say: "K3 / K2.7 without Thinking routes to K2.6."
• Config path resolution: /root/.kimi/config.toml resolves to /root/.arifos/agents/kimi/config.toml (separate file, not a symlink). The checklist references /root/.kimi/config.toml which works, but the canonical source is under the agent home. grep both paths to be safe; both files must agree on default_model.
• Model must exist in [models] before switching default_model to it.
• 🚨 PATCH DRIFT — default_model CAN SILENTLY REVERT: External processes (other agent sessions, kimi-code config refresh, v0.27.0 auto-refresh of model list) can overwrite default_model back to a prior value. After ANY config.toml patch, always verify: head -1 /root/.kimi-code/config.toml — confirm it says the intended model before moving on. 2026-07-20 case study: Patched default_model to kimi-code/k3. Later verification showed it had reverted to kimi-code/kimi-for-coding mid-session. File was externally modified. Patch → verify → patch again if drift.

OAuth & Auth

• managed:kimi-code uses OAuth (not API key). The api_key = "" in provider block is correct.
• OAuth login required: Before using managed:kimi-code provider, run kimi login. Without login, default_model = "kimi-code/k3" will silently fall back to minimax-coding-plan — no error in logs, no warning. The response says "MiniMax-M3" and you won't know K3 was attempted. Always verify with --model kimi-code/k3 first.
• default_model may not take effect for -p mode: Due to silent fallback. Always verify model identity in the response by asking "what model are you?"

Quota & Capacity

• K3 quota burn: Massive on cheap plans. A Moderato user may exhaust quota in one coding session. Warn before switching.
• Quota exhaustion: If Arif says "quota finish", revert immediately. Don't argue or suggest upgrading — just switch back to MiniMax-M3.
• K3 capacity: Upstream limited after release. May return 429 errors. If flaky, fall back.
• K3 reasoning_effort: Only max supported (Jul 16 2026). Other levels promised, not available.
• --no-sandbox flag: Does NOT exist in v0.26.0. Causes error: unknown option.
• -y (yolo) + -p (prompt) mutual exclusion (v0.27.0): Cannot combine --yolo/-y with --prompt/-p. Returns error: Cannot combine --prompt with --yolo. These are separate invocation modes — use -y alone for autonomous mode (no prompt arg, Kimi drives itself) or -p "..." for explicit prompting. They are not flags to stack.

K3 Plan Tier / Alternative Paths

K3 availability varies by path (2026-07-18 tested):

Path	Status	Context	Quota	Notes
kimi-code/k3 (managed:kimi-code)	⚠️ OAuth required	256K (Moderato) / 1M (Allegretto+)	Plan-limited	kimi login device-code flow needed
openrouter/moonshotai/kimi-k3	✅ WORKS	1M full	OpenRouter billing	Wired in OpenClaw fallback chain
bailian-token-plan/kimi-k3	❌ 404	N/A	N/A	Model not yet available on Alibaba Bailian

• K3 via OpenRouter in OpenClaw: openrouter/moonshotai/kimi-k3 gives full 1M context regardless of Kimi Code plan tier. Use this for K3 coding when direct OAuth isn't working or plan limits context. Wired in /root/.openclaw/openclaw.json as fallback and auto-discovered model.
• K3 via Bailian Token Plan: Does NOT work (404 Model not exist as of 2026-07-18). Don't add kimi-k3 to the bailian-token-plan provider — it will fail at runtime.
• Cheap plan (Moderato): K3 works via managed:kimi-code but at 256K only. Stick with MiniMax-M3 as default for daily use. K3 via --model kimi-code/k3 for heavy tasks. K3 via OpenClaw/OpenRouter for coding — bypasses quota.

AAA TOOLBENCH ALIGNMENT (Model-Card Sync)

When you change Kimi Code's model, 5+ files across the AAA toolbench must be updated or the registry, Warga card, A2A cards, and agent identity will all disagree. Arif catches this fast.

Files to Audit After Any Model Change

Model changes ripple across TWO layers: identity cards (what the agent claims to be) and federation registries (what the federation believes). Missing either layer = Arif catches drift immediately.

Layer 1 — Identity Cards (7 files)

#	File	What to update
1	/root/.kimi/config.toml	default_model
2	/root/.arifos/agents/kimi/AGENTS.md	**Model:** line
3	/root/AAA/agents/_external/kimi-code/WARGAAA_CARD.md	model field in identity table + version/date
4	/root/AAA/agents/_external/kimi-code/agent-card.json	model field
5	/root/AAA/a2a-server/agent-cards/harnesses/kimi-code.json	model field
6	/root/AAA/a2a-server/agent-cards/forge/fi-008-kimi-code.json	model + description + binary
7	/root/AAA/agent-cards/harnesses/fi-003-kimi-code/agent-card.json	model + description + binary + fix FI number if wrong

Layer 2 — Federation Registries (4 files — MOST COMMONLY MISSED)

#	File	What to update
8	/root/AAA/registries/AAA_AGENTS_REGISTRY.json	model + ensure entry EXISTS (was missing until 2026-07-18)
9	/root/AAA/registries/forge_instruments.yaml	model + binary + version + note
10	/root/AAA/ROOT_AGENT_CONFIG.yaml	model + version (often forgotten — was stale at v0.18.0/kimi-for-coding until 2026-07-18)
11	/root/AAA/a2a/registry/agents.yaml	Entry must exist with correct model in notes

CRITICAL: Layer 2 is the most commonly missed. The agent can work fine with stale identity cards, but the federation dashboard, A2A discovery, and forge instrument registry will all show wrong model info. After model changes, always audit both layers. 2026-07-18 case study: ROOT_AGENT_CONFIG.yaml still said model: kimi-for-coding, version: 0.18.0 while actual was MiniMax-M3 v0.26.0. This file is especially prone to drift because it's not in the card directories.

Common Staleness Patterns

• Model field says kimi-k2 / kimi-for-coding but config says MiniMax-M3
• FI number mismatch: CIV-33 uses FI-003, AGENTS.md says FI-008
• Binary path: says /root/.local/bin/kimi, actual is /root/.kimi-code/bin/kimi
• Description says "K2.7 Code model" but running MiniMax
• Agent missing from AAA_AGENTS_REGISTRY.json: FI-008 was izin GRANTED since 2026-06-10 but never added to the primary agent registry. Check: grep kimi-code /root/AAA/registries/AAA_AGENTS_REGISTRY.json — empty = gap.
• Agent missing from a2a/registry/agents.yaml: A2A peer discovery won't find the agent. Check: grep kimi-code /root/AAA/a2a/registry/agents.yaml
• After any model change, run: grep -r "kimi-for-coding\|kimi-k2" /root/AAA/agent-cards/ /root/AAA/a2a-server/agent-cards/ /root/AAA/agents/_external/kimi-code/ /root/AAA/registries/ — any hits = stale

Verification After Fix

# Layer 1 — identity cards (7 files)
for f in /root/AAA/agents/_external/kimi-code/WARGAAA_CARD.md \
  /root/AAA/agents/_external/kimi-code/agent-card.json \
  /root/AAA/a2a-server/agent-cards/harnesses/kimi-code.json \
  /root/AAA/a2a-server/agent-cards/forge/fi-008-kimi-code.json \
  /root/AAA/agent-cards/harnesses/fi-003-kimi-code/agent-card.json \
  /root/.kimi/config.toml /root/.arifos/agents/kimi/AGENTS.md; do
  grep -i "MiniMax\|kimi-for-coding\|kimi-k2" "$f" 2>/dev/null | head -1
done
echo "==="
# Layer 2 — federation registries (4 files)
for f in /root/AAA/registries/AAA_AGENTS_REGISTRY.json \
  /root/AAA/registries/forge_instruments.yaml \
  /root/AAA/ROOT_AGENT_CONFIG.yaml \
  /root/AAA/a2a/registry/agents.yaml; do
  echo "--- $(basename $f) ---"
  grep -i "kimi\|MiniMax" "$f" 2>/dev/null | head -2
done

All 11 files must agree on the current model. Zero references to stale K2.7/kimi-for-coding. Layer 2 files must all contain a kimi-code entry (gaps = registry drift). ROOT_AGENT_CONFIG.yaml is the most common drift file — verify model and version match live state.

K3 Direct API Key (for OpenCode / OpenClaw / Hermes)

K3 can also be called directly via API key (sk-ki...) at https://api.kimi.com/coding/v1 — this is the Chat Completions endpoint, not the OAuth-managed kimi-code provider. Use this to wire K3 into OpenCode, OpenClaw, or Hermes as a regular OpenAI-compatible provider.

Property	Value
Base URL	https://api.kimi.com/coding/v1
Auth	Bearer sk-ki... (API key from Kimi platform console)
Transport	openai_chat / openai-completions
Models	k3 (2.8T, 1M ctx), kimi-for-coding (256K), kimi-for-coding-highspeed (256K)
api.moonshot.cn	❌ Does NOT work with these keys — returns 401

Verify: curl -s https://api.kimi.com/coding/v1/models -H "Authorization: Bearer $KIMI_API_KEY" — returns k3, kimi-for-coding, kimi-for-coding-highspeed.

Pitfall — K3 thinking consumes tokens before content: K3 with always_thinking may return empty content on short max_tokens because all tokens go to reasoning. Use max_tokens ≥ 200 for simple prompts, ≥ 4096 for coding tasks.

Reference Files

• references/config-anatomy.md — detailed config.toml field reference
• references/kimi-code-audit-checklist.md — 10-step systematic audit for kimi-code configuration
• references/openclaw-k3-openrouter.md — wire K3 into OpenClaw via OpenRouter (bypasses Kimi quota)
• references/kimi-k3-specs.md — Kimi K3 specifications: architecture, benchmarks, plan tiers, availability paths, quota warnings