feat: switch to Bedrock Mantle endpoint, new model lineup, fix tool-use bugs

- Switch from native Bedrock SDK to Mantle (OpenAI-compatible endpoint)
  which gives all models uniform tool-use support including GLM 5
- New 3-tier lineup: MiniMax M2.1 (budget), GLM 4.7 (mid), GLM 5 (premium)
- Fix NadirClaw bug: empty string content on assistant tool_call messages
  became None, rejected by Bedrock (sed patch in Dockerfile)
- Disable NADIRCLAW_OPTIMIZE — the optimizer strips tool_calls and
  tool_call_id from messages, breaking all tool-use conversations
- Load Mantle API key from ~/.nadirclaw/mantle.env via docker-compose

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: eugenio

Dockerfile

@@ -5,6 +5,11 @@ WORKDIR /app
 # Install nadirclaw with dashboard extra
 RUN pip install --no-cache-dir "nadirclaw[dashboard]>=0.13" boto3>=1.35
 
+# Fix NadirClaw bug: empty string content on assistant tool_call messages
+# becomes None, which Bedrock/Mantle rejects. Preserve empty string instead.
+RUN sed -i 's/content = text if text else message.content/content = text if text is not None else message.content/g' \
+    /usr/local/lib/python3.11/site-packages/nadirclaw/server.py
+
 # Pre-download the sentence-transformers model so first startup is fast
 RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('all-MiniLM-L6-v2')"
 

config/nadirclaw.env

@@ -1,49 +1,47 @@
-# NadirClaw configuration for AWS Bedrock (eu-west-2 London)
-# eu-west-2 has the most coding models of any EU region
+# NadirClaw configuration for AWS Bedrock Mantle (eu-west-2 London)
+# Uses Bedrock Mantle (OpenAI-compatible endpoint) for all models
 # Copy to ~/.nadirclaw/.env or set as environment variables
 
-# AWS Bedrock credentials (use AWS_PROFILE or explicit keys)
-AWS_PROFILE=default
-AWS_DEFAULT_REGION=eu-west-2
-# AWS_ACCESS_KEY_ID=
-# AWS_SECRET_ACCESS_KEY=
+# Bedrock Mantle endpoint (OpenAI-compatible, all models support tool-use)
+NADIRCLAW_API_BASE=https://bedrock-mantle.eu-west-2.api.aws/v1
+# API key loaded from ~/.nadirclaw/mantle.env at container startup
+# OPENAI_API_KEY is set via docker-compose env_file
 
-# Three-tier routing with Bedrock models (verified available in eu-west-2)
-#   Budget:  Qwen3 Coder 30B  - fast, cheap, good for simple tasks
-#   Mid:     Qwen3 Coder 480B - code-specialized, strong benchmarks
-#   Premium: Kimi K2.5        - top coding scores (SWE 76.8%, HE 99%, LCB 85%)
-NADIRCLAW_SIMPLE_MODEL=bedrock/qwen.qwen3-coder-30b-a3b-v1:0
-NADIRCLAW_MID_MODEL=bedrock/qwen.qwen3-coder-480b-a35b-v1:0
-NADIRCLAW_COMPLEX_MODEL=bedrock/moonshotai.kimi-k2.5
+# Three-tier routing with Bedrock models (all available in eu-west-2)
+#   Budget:  MiniMax M2.1     - 10B active params, fast, cheap, good tool calling
+#   Mid:     GLM 4.7          - best tool-call reliability, 200K context, strong SWE
+#   Premium: GLM 5            - top-tier coding, 200K context, 128K output, best consistency
+NADIRCLAW_SIMPLE_MODEL=openai/minimax.minimax-m2.1
+NADIRCLAW_MID_MODEL=openai/zai.glm-4.7
+NADIRCLAW_COMPLEX_MODEL=openai/zai.glm-5
 
 # Tier thresholds (score 0-1: <= simple_max -> budget, >= complex_min -> premium)
 NADIRCLAW_TIER_THRESHOLDS=0.35,0.65
 
 # Fallback chain (try next model on failure)
-# Qwen3 30B is the final fallback — it handles large system prompts that 480B/Kimi reject
-NADIRCLAW_FALLBACK_CHAIN=bedrock/moonshotai.kimi-k2.5,bedrock/qwen.qwen3-coder-480b-a35b-v1:0,bedrock/qwen.qwen3-coder-30b-a3b-v1:0
-NADIRCLAW_SIMPLE_FALLBACK=bedrock/qwen.qwen3-coder-480b-a35b-v1:0
-NADIRCLAW_MID_FALLBACK=bedrock/moonshotai.kimi-k2.5,bedrock/qwen.qwen3-coder-30b-a3b-v1:0
-NADIRCLAW_COMPLEX_FALLBACK=bedrock/qwen.qwen3-coder-480b-a35b-v1:0,bedrock/qwen.qwen3-coder-30b-a3b-v1:0
+# All models support tool-use via Mantle endpoint
+NADIRCLAW_FALLBACK_CHAIN=openai/zai.glm-5,openai/deepseek.v3.2,openai/zai.glm-4.7,openai/minimax.minimax-m2.1
+NADIRCLAW_SIMPLE_FALLBACK=openai/zai.glm-4.7-flash,openai/zai.glm-4.7
+NADIRCLAW_MID_FALLBACK=openai/zai.glm-5,openai/deepseek.v3.2
+NADIRCLAW_COMPLEX_FALLBACK=openai/deepseek.v3.2,openai/zai.glm-4.7,openai/minimax.minimax-m2.1
 
 # Budget controls (USD)
 NADIRCLAW_DAILY_BUDGET=5.00
 NADIRCLAW_MONTHLY_BUDGET=80.00
 NADIRCLAW_BUDGET_WARN_THRESHOLD=0.8
 NADIRCLAW_BUDGET_STDOUT_ALERTS=true
 
-# Context optimization (reduces token usage on long conversations)
-NADIRCLAW_OPTIMIZE=safe
-NADIRCLAW_OPTIMIZE_MAX_TURNS=40
+# Context optimization DISABLED — the optimizer strips tool_calls and
+# tool_call_id from messages, breaking tool-use conversations.
+# See: optimize_messages() converts messages to {"role","content"} only.
+NADIRCLAW_OPTIMIZE=off
 
 # LiteLLM: merge consecutive user/tool blocks for Bedrock compatibility
-# Without this, OpenCode tool-use messages cause BadRequestError on Bedrock models
 LITELLM_MODIFY_PARAMS=true
 
 # Server (localhost only)
 NADIRCLAW_PORT=4000
 
 # Logging
 NADIRCLAW_LOG_DIR=~/.nadirclaw/logs
-NADIRCLAW_LOG_RAW=true
-LITELLM_LOG=DEBUG
+NADIRCLAW_LOG_RAW=false

docker-compose.yml

@@ -8,6 +8,7 @@ services:
       - "127.0.0.1:4000:4000"
     env_file:
       - config/nadirclaw.env
+      - "${HOME}/.nadirclaw/mantle.env"
     volumes:
       # Mount AWS credentials (read-only)
       - "${HOME}/.aws:/root/.aws:ro"

opencode.json

@@ -10,13 +10,13 @@
       },
       "models": {
         "auto": {
-          "name": "Auto (3-tier routing: Qwen3 30B / Mistral Large 3 / Kimi K2)"
+          "name": "Auto (3-tier: MiniMax M2.1 / GLM 4.7 / GLM 5)"
         },
         "eco": {
-          "name": "Budget (Qwen3 Coder 30B)"
+          "name": "Budget (MiniMax M2.1)"
         },
         "premium": {
-          "name": "Premium (Kimi K2 Thinking)"
+          "name": "Premium (GLM 5)"
         }
       }
     }