Problem: callWithRequest/Full/Stream all called client.buildRequestBodyFromRequest
directly (not via hooks), so ClaudeClient could never override it. This meant
tool calling sent OpenAI format to Anthropic (wrong field names, wrong roles).
Changes:
mcp/interface.go
- Add buildRequestBodyFromRequest(*Request) map[string]any to clientHooks
- Improve comments: document what each hook group does and why
mcp/client.go
- All three paths (callWithRequest, callWithRequestFull, CallWithRequestStream)
now call client.hooks.buildRequestBodyFromRequest — ClaudeClient picks up
mcp/claude_client.go
- Full rewrite with format comparison table in package doc
- buildRequestBodyFromRequest: produces correct Anthropic wire format
* system prompt → top-level "system" field
* tools: parameters → input_schema, no "type:function" wrapper
* tool_choice "auto" → {"type":"auto"} object
* assistant tool calls → content[{type:tool_use, id, name, input}]
* role=tool results → role=user content[{type:tool_result,...}]
* consecutive tool results merged into single user turn
- convertMessagesToAnthropic: handles all three message types
- parseMCPResponseFull: extracts text + tool_use blocks
- parseMCPResponse: delegates to parseMCPResponseFull
All mcp and agent tests pass.
Migrate the Telegram bot agent from an XML tag hack (<api_call>) to
OpenAI-native function calling via CallWithRequestFull.
Key changes:
- mcp/interface.go: add parseMCPResponseFull to clientHooks interface
- mcp/client.go: route callWithRequestFull through hooks for overridability
- mcp/claude_client.go: override parseMCPResponseFull for Claude response
format (tool_use blocks instead of choices[].message.tool_calls)
- telegram/agent/agent.go: rewrite Run() to use CallWithRequestFull;
define api_request tool with JSON Schema; implement tool-call loop
with role="tool" result messages; remove XML parsing entirely
- telegram/agent/apicall.go: remove parseAPICall (dead code)
- telegram/agent/prompt.go: simplify — remove XML format instructions,
replace with concise api_request tool usage instructions
- telegram/agent/agent_test.go: rebuild all tests using LLMResponse
objects; add TestNarrationStructurallyImpossible, TestOnChunkCalledWithFinalReply,
TestToolCallIDPropagated; remove XML-specific tests
Architecture advantage: with native function calling, the LLM returns
EITHER ToolCalls OR Content — never both. Narration is now structurally
impossible at the protocol level, not just enforced by prompt rules.
All 11 agent tests pass. mcp package tests pass.
- Add Telegram bot with long-polling and AI agent loop (api_call tool)
- SSE streaming with real-time message editing and ⏳ placeholder
- Account state injection at conversation start (models, exchanges,
strategies, traders, per-trader PnL and statistics)
- Lane semaphore per chat serializes concurrent messages (60s timeout)
- Idle timeout watchdog (60s) prevents hung streaming connections
- Look-ahead buffer prevents partial <api_call> tag leaking to user
- Fix PUT /strategies/:id to merge config (read-then-merge pattern)
- Add route registry with full API schema for LLM documentation
- Add TelegramConfig store and Web UI config modal
- Add GetAnyEnabled to AIModel store for bot LLM client selection
* fix(trader): get peakPnlPct using posKey
* fix(docs): keep readme at the same page
* improve(interface): replace with interface
* refactor mcp
---------
Co-authored-by: zbhan <zbhan@freewheel.tv>
## Problem
AI responses were being truncated due to a hardcoded max_tokens limit of 2000,
causing JSON parsing failures. The error occurred when:
1. AI's thought process analysis was cut off mid-response
2. extractDecisions() incorrectly extracted MACD data arrays from the input prompt
3. Go failed to unmarshal numbers into Decision struct
Error message:
```
json: cannot unmarshal number into Go value of type decision.Decision
JSON内容: [-867.759, -937.406, -1020.435, ...]
```
## Solution
- Add MaxTokens field to mcp.Client struct
- Read AI_MAX_TOKENS from environment variable (default: 2000)
- Set AI_MAX_TOKENS=4000 in docker-compose.yml for production use
- This provides enough tokens for complete analysis with the 800-line trading strategy prompt
## Testing
- Verify environment variable is read correctly
- Confirm AI responses are no longer truncated
- Check decision logs for complete JSON output
This update enables users to configure any OpenAI-compatible API endpoint,
allowing the use of:
- OpenAI official API (GPT-4, GPT-4o, etc.)
- OpenRouter (access to multiple models)
- Local deployed models (Ollama, LM Studio, etc.)
- Other OpenAI-format compatible API services
Changes:
- config: Add custom_api_url, custom_api_key, custom_model_name fields
- mcp: Add SetCustomAPI function and ProviderCustom constant
- trader: Update AI initialization logic to support custom API
- manager: Pass custom API config to trader instances
- Add CUSTOM_API.md documentation with usage examples
- Update config.json.example with custom API sample
Co-Authored-By: tinkle-community <tinklefund@gmail.com>
Architecture improvements:
- Extract AI decision engine to dedicated `decision` package
- Create `mcp` package for Model Context Protocol client
- Separate market data structures into `market/data.go`
- Update trader to use new modular structure
New packages:
- `decision/engine.go` - AI decision logic and prompt building
- `mcp/client.go` - Unified AI API client (DeepSeek/Qwen)
- `market/data.go` - Market data type definitions
Benefits:
- Better separation of concerns
- Improved code organization and maintainability
- Easier to test individual components
- More flexible AI provider integration
- Cleaner dependency management
Updated imports:
- trader/auto_trader.go now uses decision and mcp packages
- Consistent API across different AI providers
Co-Authored-By: tinkle-community <tinklefund@gmail.com>
