📎 Source:openclaw-sumopod — view on GitHub & star ⭐

⚠️ JANGAN Update ke OpenClaw 2026.3.7 - 2026.3.10 — Kimi 2.5 Tool Calling BROKEN!

Status: 🟢 RESOLVED in 2026.3.11 — Update ke versi terbaru sekarang! ✅
Fixed Version: 2026.3.11+ ✅
Affected Versions: 2026.3.7, 2026.3.8, 2026.3.9, 2026.3.10
Last Known Good: 2026.3.2 ✅
Release Notes:v2026.3.11
Related Issues:#39907, #41297

🎉 GOOD NEWS: Kimi 2.5 Fixed di 2026.3.11!

┌─────────────────────────────────────────────────────────┐
│  ✅ UPDATE NOW: npm i -g openclaw@latest                │
│                                                         │
│  🎉 Tool calling KIMI 2.5 berfungsi kembali            │
│  🎉 Infinite loop issue sudah resolved                 │
│  🎉 Kimi K2.5 bisa dipakai lagi seperti biasa          │
│                                                         │
│  📖 Release Notes: github.com/openclaw/openclaw/       │
│                    releases/tag/v2026.3.11             │
└─────────────────────────────────────────────────────────┘

Update command:

pkill -f openclaw && npm install -g openclaw@latest && openclaw gateway restart

📝 Catatan Historis (Bug di 2026.3.7 - 2026.3.10)

🚨 Apa yang Terjadi?

Sebelum (2026.3.2 ✅)

User: "execute ls -la"

Kimi 2.5: [TOOL_CALL] exec({"command": "ls -la"})
         ↓
     [✅ EXECUTED] — Real tool card muncul!

Sesudah (2026.3.7+ ❌)

User: "execute ls -la"

Kimi 2.5: "exec({"command": "ls -la"})" 
         ↓
     [❌ FAILED] — Cuma text literal, gak jalan!
         ↓
     Atau looping: "Executing... ⏳" "Running now... ⏳" tanpa henti

🔍 3 Pola Failure yang Terjadi

Pattern	Deskripsi	Visual
🔄 Looping	Model bilang "Executing..." "Running now..." tapi gak pernah jalan	`⏳ ⏳ ⏳` infinite loop
📝 Literal Text	Model output `exec({"command": "..."})` sebagai chat text	`exec({...})` — cuma text, bukan tool call
🎭 Fake Success	Model bilang "✅ Done! Command executed" padahal gak jalan sama sekali	Tipu-tipu 🎪

🧠 Kenapa Bisa Begini?

Root Cause Analysis

┌──────────────────────────────────────────────────────────┐
│  REQUEST SIDE ✅ (Works fine)                           │
│  OpenClaw → convert tool schema → Kimi API format       │
│  ✓ Payload benar, Kimi nerima dengan baik               │
├──────────────────────────────────────────────────────────┤
│  RESPONSE SIDE ❌ (BROKEN in 2026.3.7+)                 │
│  Kimi API → return tool_use blocks                      │
│  OpenClaw parser → gak recognize format Kimi            │
│  Result → treat as plain text ❌                        │
└──────────────────────────────────────────────────────────┘

Bukti: Kimi API ITU BISA tool calling!

# Direct test ke Kimi API — TOOL CALLING WORKS! ✅
curl -s https://api.kimi.com/coding/v1/messages \
  -H "x-api-key: $KIMI_API_KEY" \
  -H "content-type: application/json" \
  -d '{
    "model": "k2p5",
    "tools": [{"name": "bash", "input_schema": {...}}],
    "messages": [{"role": "user", "content": "Run ls -la"}]
  }'

# Response (correct):
{
  "content": [{
    "type": "tool_use",           ← ✅ Proper tool_use block!
    "name": "bash",
    "input": {"command": "ls -la"}
  }]
}

Conclusion: Masalahnya di OpenClaw parser (response side), BUKAN di Kimi API! 🎯

🔧 SOLUSI: Downgrade ke 2026.3.2

⚡ Quick Fix (One-Liner)

pkill -f openclaw; npm install -g openclaw@2026.3.2

📋 Step-by-Step Detail

# 1️⃣ Stop gateway yang lagi jalan
pkill -f openclaw

# 2️⃣ Downgrade ke versi yang works
npm install -g openclaw@2026.3.2

# 3️⃣ Reinstall gateway service (kalau pakai systemd)
openclaw gateway install --force

# 4️⃣ Restart gateway
openclaw gateway restart

# 5️⃣ Verify version
openclaw version
# Expected: 2026.3.2 ✅

🎨 Visual Step-by-Step

   ┌──────────┐
   │   🛑     │  1. Stop OpenClaw
   │   Stop   │     pkill -f openclaw
   └────┬─────┘
        ▼
   ┌──────────┐
   │   📦     │  2. Downgrade
   │ Install  │     npm i -g openclaw@2026.3.2
   └────┬─────┘
        ▼
   ┌──────────┐
   │   ⚙️     │  3. Reinstall service
   │  Setup   │     openclaw gateway install --force
   └────┬─────┘
        ▼
   ┌──────────┐
   │   🚀     │  4. Start lagi
   │ Restart  │     openclaw gateway restart
   └────┬─────┘
        ▼
   ┌──────────┐
   │   ✅     │  5. Kimi 2.5 works!
   │ Success  │     Tool calling balik normal
   └──────────┘

🧪 Verifikasi: Cek Tool Calling Works

Setelah downgrade, test dengan command sederhana:

User: "execute pwd"

Expected Output:
┌────────────────────────────────────────┐
│  Exec                                  │
│  Command: pwd                          │
│  Status: ✅ Success                    │
│  Output: /home/user/workspace          │
└────────────────────────────────────────┘

Kalau muncul real tool card → ✅ Berhasil downgrade!

Kalau cuma text exec({"command": "pwd"}) → ❌ Masih broken, coba ulang step-nya

⏳ Kapan Bisa Update Lagi?

✅ UPDATE NOW! — 2026.3.11+ Released

🟢 SAFE — 2026.3.11+  →  UPDATE NOW! Kimi 2.5 fixed ✅
🔴 AVOID — 2026.3.7 - 2026.3.10 →  Buggy versions
✅ LEGACY — 2026.3.2   →  Still works if you prefer

Command to update:

pkill -f openclaw && npm install -g openclaw@latest && openclaw gateway restart

Release Notes:v2026.3.11

📜 Historical Timeline

Date	Version	Status	Notes
Mar 11, 2026	2026.3.11	✅ FIXED	Kimi 2.5 tool calling restored
Mar 10, 2026	2026.3.10	❌ BUGGY	Still broken
Mar 8, 2026	2026.3.8	❌ BUGGY	Tool calling broken
Mar 7, 2026	2026.3.7	❌ BUGGY	First buggy release
Mar 5, 2026	2026.3.2	✅ SAFE	Last known good before bug

Monitor progress (archived):

Issue: #41297 — ✅ CLOSED
Issue: #39907 — ✅ CLOSED

💡 Alternatif Sementara (Kalau Mau Stay 2026.3.7+)

Kalau ada alasan kuat harus stay di 2026.3.7+, ada workaround:

Option A: Switch ke Model Lain

Model	Tool Calling Status	Notes
`google/gemini-3-flash`	✅ Works	Fast, cheap, reliable
`zai/glm-4.7`	✅ Works	ByteDance, good for coding
`kimi-coding/k2p5`	❌ Broken	Avoid in 2026.3.7+

// ~/.openclaw/openclaw.json
{
  "agents": {
    "defaults": {
      "model": {
        "primary": "google/gemini-3-flash",  ← Switch sementara
        "fallbacks": ["zai/glm-4.7"]
      }
    }
  }
}

Option B: Use Sub-Agents dengan Ollama

Kalau tasknya lokal/bukan butuh Kimi specifically:

# Run local model via Ollama (free, offline)
ollama run llama3.1

# Sub-agent pakai Ollama untuk task simple
# Main agent tetep bisa pakai model lain

📊 Impact Summary

┌──────────────────────────────────────────────────────┐
│                  IMPACT ANALYSIS                     │
├──────────────────────────────────────────────────────┤
│                                                      │
│  🎯 Affected Users:                                  │
│     • Kimi K2.5 users (kimi-coding/k2p5)            │
│     • Heavy tool users (exec, browser, file ops)    │
│                                                      │
│  📉 Severity: HIGH 🔴                                │
│     • Tool execution unreliable                     │
│     • Bisa kasih fake "success" padahal gak jalan   │
│     • Automation jadi untrustworthy                 │
│                                                      │
│  ✅ Solution:                                        │
│     • Downgrade ke 2026.3.2                         │
│     • Atau switch ke Gemini/GLM sementara           │
│                                                      │
└──────────────────────────────────────────────────────┘

🎓 Lessons Learned

┌─────────────────────────────────────────────────────────┐
│  💡 TAKEAWAYS                                          │
│                                                         │
│  1. Jangan buru-buru update stable release terbaru    │
│     → Wait 1-2 minggu, lihat community feedback       │
│                                                         │
│  2. Kalau tool calling penting buat workflow:         │
│     → Test dulu di dev environment                    │
│     → Backup config sebelum update                    │
│                                                         │
│  3. Know your rollback:                               │
│     → npm install -g openclaw@VERSION                 │
│     → Simpan versi yang works di catatan              │
│                                                         │
└─────────────────────────────────────────────────────────┘

📚 Referensi

Resource	Link
Original Issue #41297	https://github.com/openclaw/openclaw/issues/41297
Original Issue #39907	https://github.com/openclaw/openclaw/issues/39907
Related Issue #40157	https://github.com/openclaw/openclaw/issues/40157
OpenClaw Releases	https://github.com/openclaw/openclaw/releases

🙏 Credits

Thanks to OpenClaw community yang report dan investigate issue ini — especially yang udah trace sampe root cause di response parser! 🎉

Last Updated: March 12, 2026
Author: OpenClaw Sumopod Community
Applies to: OpenClaw 2026.3.7 - 2026.3.10 (BUG) / 2026.3.11+ (FIXED)

💬 Questions?

Ada pertanyaan tentang update atau butuh bantuan? Join:

🌏 OpenClaw Discord: https://discord.com/invite/clawd
💬 Tanya di Telegram: @RaditClaw_bot

Update to 2026.3.11+ and happy coding! 🚀

⚠️ BUG CRITICAL: Jangan Update OpenClaw ke Versi 2026.3.7!