claude-code-deepseek-thinking-proxy

One command to fix "content[].thinking in the thinking mode must be passed back to the API" when using DeepSeek V4 with Claude Code.

What problem does this solve?

DeepSeek V4 models (deepseek-v4-pro / deepseek-v4-flash) always run in thinking mode. When the model makes tool calls, it requires the thinking blocks from previous turns to be preserved in the conversation history. If you switch models mid-session (e.g. to Anthropic Sonnet and back), Claude Code strips these blocks. Switching back to DeepSeek then fails with a 400 error.

This proxy sits between Claude Code and DeepSeek, caching thinking blocks and re-injecting them when missing — transparent to Claude Code.

Quickstart

npm install -g claude-code-deepseek-thinking-proxy
deepseek-thinking-proxy

Then set your Claude Code base URL to point to the proxy:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:8787",
    "ANTHROPIC_AUTH_TOKEN": "<your-deepseek-api-key>"
  }
}

That's it. Switch models freely — the proxy handles the rest.

How it works

Claude Code ──→ localhost:8787 ──→ api.deepseek.com/anthropic
                   │
                   ├─ Inbound: scan assistant messages
                   │   ├─ Cache hit → re-inject real thinking blocks
                   │   └─ Cache miss → inject placeholder
                   │
                   └─ Outbound: extract thinking blocks → cache by fingerprint

• Cache: FNV-1a hash of conversation prefix, in-memory, TTL 30 min, max 500 entries
• Placeholder: Falls back to "(thinking omitted)" with empty signature when cache is cold. DeepSeek's Anthropic endpoint does not validate signatures.

Auto-start with launchd (macOS)

cat > ~/Library/LaunchAgents/com.deepseek-thinking-proxy.plist << 'EOF'
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN"
  "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.deepseek-thinking-proxy</string>
    <key>ProgramArguments</key>
    <array>
        <string>/path/to/node</string>
        <string>/path/to/deepseek-thinking-proxy</string>
    </array>
    <key>RunAtLoad</key><true/>
    <key>KeepAlive</key><true/>
    <key>StandardOutPath</key>
    <string>/tmp/thinking-proxy.log</string>
    <key>StandardErrorPath</key>
    <string>/tmp/thinking-proxy.log</string>
</dict>
</plist>
EOF

launchctl load ~/Library/LaunchAgents/com.deepseek-thinking-proxy.plist

Configuration

All via environment variables:

| Variable | Default | Description | |----------|---------|-------------| | PROXY_PORT | 8787 | Listen port | | PROXY_UPSTREAM | https://api.deepseek.com/anthropic | Upstream DeepSeek endpoint | | PROXY_PLACEHOLDER | fallback | off / fallback / always | | PROXY_PLACEHOLDER_TEXT | (thinking omitted) | Placeholder content | | PROXY_DEBUG | 0 | Set to 1 for verbose logging |

Health check

curl http://127.0.0.1:8787/health
# {"status":"ok","cacheSize":3,"upstream":"https://api.deepseek.com/anthropic"}

Use with cc-switch

If you use cc-switch to manage providers, set your DeepSeek provider's base URL to http://127.0.0.1:8787. The proxy runs independently via launchd — cc-switch only needs to point at it.

License

MIT