- Add _extract_output_str() to unwrap MCP ToolMessage content objects
instead of calling str() on the raw object
- Handle read_url url param when LLM passes a list instead of a string
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- Add langchain-community dependency for GoogleSerperAPIWrapper
- Add serper_api_key to config with default key
- Create app/tools/serper.py with async serper_search tool
- Register "serper" key in tools/__init__.py (independent from "search"/Jina MCP)
- Add tool title and input/output summaries in chat.py
- Set SERPER_API_KEY on Azure App Service (Operation resource group)
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- Add has_tool_activity flag to distinguish first agent pass (thinking)
from post-tool agent pass (generating)
- Replace on_chain_start debug logging with status SSE emission
- Filter to only graph-level agent nodes via graph:step: tag prefix
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- tool_start/tool_end/tool_error SSE payloads now include call_id field
derived from LangGraph run_id for reliable tool event correlation
- tool_start_ts dict keyed by call_id instead of tool_name to handle
concurrent calls to the same tool
- Added on_chain_start debug logging to observe chain names and metadata
(no SSE emission yet, observation only)
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Add CoT execution trace to the SSE stream: status events for stage
transitions, enriched tool_start with input_summary, tool_end with
output_summary and duration_ms, and tool_error for failed tool calls.
Helper functions _sse, _summarize_input, _summarize_output provide
human-readable summaries for each tool type.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
When a tool call crashes (e.g. httpx timeout in kb_search), the LangGraph
checkpoint retains an AIMessage with tool_calls but no corresponding
ToolMessage. Subsequent requests to the same conversation_id fail with:
ValueError: Found AIMessages with tool_calls that do not have a
corresponding ToolMessage
Now the except block in _stream_response detects this specific ValueError
by checking for "tool_calls" and "ToolMessage" in the error string, then
calls checkpointer.adelete_thread() to purge the corrupted thread state.
The frontend receives {"type":"error","content":"对话状态异常，已自动重置..."}
followed by {"type":"done"}, so the user can simply resend their message.
API confirmed: AsyncPostgresSaver.adelete_thread(thread_id) deletes from
checkpoints, checkpoint_blobs, and checkpoint_writes tables for the thread.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
- Add web_search tool with Jina Search/Reader/Rerank
- Flash mode: top 3, 8s timeout, no Reader/Rerank
- Pro mode: top 10, 20s timeout, concurrent Reader + Rerank top 5
- Add Redis async cache (TTL=300s) for search results
- Register "search" in ALL_TOOLS mapping
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add complete Python backend (Litestar + LangGraph) with chat, conversations, tickets APIs
- Add GitHub Actions workflow for auto-deploying backend to Azure Web App (soc-backend)
- Add gunicorn to requirements.txt for production serving
- Update CLAUDE.md and EXTERNAL_SERVICES.md with latest config
- Remove obsolete claudehd.md (merged into gpthd.md)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
