- Backend: ChatRequest.attachments; image files read as base64 data URLs (8MB cap) and passed to run/run_stream as image_urls; last user message converted to multimodal parts only on the LLM-bound copy (history stays text); vision whitelist (kimi-k3, gpt-4o, claude, etc.), non-vision models keep text/OCR path
- Frontend: send attachments metadata with chat body (also fixes attachment context never being sent — body now uses fullText); user bubble shows image thumbnails
- Backend: token-level streaming via on_delta callback + asyncio.Queue bridge, answer_chunk SSE events; final event carries token_usage with cost_yuan; kimi pricing in cost_estimator
- Frontend: typewriter rendering of answer chunks; token/cost in message meta; abort marks message '已停止' and no longer triggers duplicate non-stream fallback; session pin/delete in dropdown (pinned first); date dividers across days
- preset questions now generated by LLM from agent name+description and
cached in localStorage (keyed by agent id+updated_at); fixes keyword
heuristic mismatch (e.g. 购房专家 got data-analysis prompts)
- add regenerate button on all assistant messages (was error-only retry)
- clear chat now asks for confirmation, notes session history is kept
- remove duplicate final-answer render in orchestrate result
(bubble body already shows it)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previously the voice map referenced 10+ non-existent Edge TTS voice names
(xiaohan, xiaochen, xiaoshuang, etc.), causing NoAudioReceived errors and
fallback to browser default TTS. Now only maps to the 6 actually available
Chinese voices (xiaoxiao, xiaoyi, yunxi, yunyang, yunjian, yunxia).
Added voice speed adjustments and fixed --rate argument format for Windows.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2.1 Orchestrator in workflow:
- New run_orchestrator_node() in workflow_integration.py loads agents from DB,
supports route/sequential/debate/pipeline modes
- New 'orchestrator' node type in workflow_engine.py execute_node dispatch
2.2 Tool-level human approval:
- AgentToolConfig extended with require_approval, approval_timeout_ms,
approval_default fields
- New ApprovalManager (approval_manager.py) with asyncio.Event-based
create/wait_for_decision/resolve pattern
- AgentRuntime run() and run_stream() intercept tool execution,
wait for approval decision before executing
- New POST /api/v1/approval/{id}/resolve REST endpoint
- Frontend: approval_required SSE event handling, approval dialog UI
with approve/deny/skip buttons
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
