Custom Provider (9router) Returns 0 Output Tokens via Telegram

[caothuyen]

GoClaw Bug Report: Custom Provider (9router) Returns 0 Output Tokens via Telegram

Environment

• GoClaw Version: v2.4.0 (protocol 3)
• OS: Windows 10.0.26200 (x64)
• Provider: 9router (custom OpenAI-compatible provider)
• Base URL: http://host.docker.internal:20128/v1
• Model: kr/claude-sonnet-4.5
• Channel: Telegram
• Agent Type: Predefined (shared context)

Issue Description

When using a custom provider (9router) with Claude models via Telegram channel, the agent receives 0 output tokens and does not call any tools. The same configuration works perfectly via Web Chat (ws).

Reproduction Steps

1. Configure custom provider "9router" with base URL http://host.docker.internal:20128/v1
2. Create agent with:
   • Provider: 9router
   • Model: kr/claude-sonnet-4.5
   • Tool Policy: full
   • Skills: enabled
3. Bind agent to Telegram channel
4. Send message via Telegram requesting tool usage (e.g., "check weather in HCM")

Expected Behavior

Agent should:

• Call appropriate tools (web_search, web_fetch, etc.)
• Return complete response with output tokens
• Work the same way as Web Chat

Actual Behavior

Via Telegram:

• Input tokens: 28.9K - 29.3K
• Output tokens: 0 (or very few: 14-129)
• Spans: 2 (1 LLM, 0 tool)
• No tool calls
• Empty or truncated response

Via Web Chat (same agent, same provider):

• Input tokens: 25.8K - 152.1K
• Output tokens: 96 - 360
• Spans: 2 - 10 (includes tool calls)
• Tools called successfully
• Complete responses

Trace Evidence

Failed Telegram Request (03:11:34 PM)

Duration: 4.1s
Tokens: 29.3K / 0
Spans: 2 (1 LLM, 0 tool)
Status: completed
Output: (empty)

Successful Web Chat Request (03:09:51 PM)

Duration: 27.5s
Tokens: 100.3K / 360
Spans: 10 (multiple LLM + tool calls)
Status: completed
Output: (complete response with tool results)

Successful Telegram Request with Different Model (02:40:30 PM)

Model: openrouter/stepfun/step-3.5-flash:free
Duration: 42.6s
Tokens: 77.0K / 983
Spans: 8 (4 LLM, 3 tool)
Tools: web_search, web_fetch (x2)
Status: completed
Output: (complete weather response)

Logs

time=2026-03-23T07:46:11.287Z level=INFO msg="resolved agent from DB" agent=jarvis-qd model=kr/claude-sonnet-4.5 provider=9router
time=2026-03-23T07:46:12.120Z level=INFO msg="system prompt built" mode=full contextFiles=6 hasMemory=true hasSpawn=true isBootstrap=false promptLen=24132
time=2026-03-23T07:46:17.195Z level=INFO msg="stream: ended, handing off to Send()" chat_id=5233603014 message_id=274

Note: No "tool call" log entries for Telegram requests with 9router/Claude.

Analysis

What Works

• ✅ 9router provider is accessible and responding
• ✅ Telegram channel integration is functional
• ✅ Web Chat with 9router/Claude works perfectly
• ✅ Telegram with OpenRouter/stepfun model works perfectly
• ✅ System prompt is built successfully (24KB)
• ✅ Stream ends without errors

What Fails

• ❌ 9router + Claude + Telegram = 0 output tokens
• ❌ No tool calls via Telegram with 9router/Claude
• ❌ Response appears to be cut off or not streamed properly

Hypothesis

The issue appears to be specific to the combination of:

1. Custom provider (9router)
2. Claude models (kr/claude-sonnet-4.5)
3. Telegram channel

Possible causes:

• Streaming response handling differs between Telegram and Web Chat
• Claude response format may not be compatible with Telegram streaming handler
• Timeout or buffer issue specific to Telegram integration
• Tool definitions may not be sent correctly to Claude via Telegram
• Response size or format incompatibility with Telegram message limits

Workarounds

1. Use Web Chat instead of Telegram with 9router/Claude
2. Use OpenRouter provider instead of 9router for Telegram
3. Use different model (e.g., stepfun/step-3.5-flash:free) via OpenRouter

Additional Context

• Provider 9router is a custom OpenAI-compatible API gateway
• Same API key and configuration works for Web Chat
• No errors in GoClaw logs
• Container is healthy and running normally
• All dependencies (Python 3.12.12, Node.js v22.22.0, pip3) are installed

Request

Please investigate why custom providers with Claude models return 0 output tokens via Telegram while working correctly via Web Chat. This appears to be a channel-specific integration issue with streaming responses from custom providers.

---

Reporter: Cao Thuyên
Date: 2026-03-23
GoClaw Version: v2.4.0
Issue Type: Bug - Channel Integration