fix: magi_clarify now emits magi_done (prevent UI freeze), Caspar R1 3-attempt retry, all join(timeout=5)->30

feat: add debate_rounds param (1-round or 2-round), fix dynamic rounds selection

feat: 2-round debate + PE upgrade + token limits expanded

fix: CASPAR all modes use background thread + keepalive to prevent Nginx timeout

fix: raise all agent max_tokens to 32000 (DeepSeek/Grok/Claude), remove artificial output token limits

fix: max_tokens 8000->16000 to prevent CASPAR truncation; fix mode routing for admission probability queries

fix: replace SSE comment ping with real data events (magi_keepalive), reduce interval to 5s to prevent Nginx proxy_read_timeout cutting Caspar stream and causing missing magi_verdict

feat: activate ps_distillation_kb_v4 RAG + add essay_review mode

fix: add UChicago/MIT/Stanford hardcoded round rules + dynamic mode UI labels

fix: stream Caspar content delta + local accumulation + remove double-buffering

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

fix: replace _drain_queue with _drain_queue_with_keepalive - prevent Nginx idle timeout during Balthasar/Melchior calls (root cause of Agent 3 not showing)

fix: add magi_done event to all modes; prevent spurious terminated error display

P1 Grok Judge: KB fact-consistency audit after Caspar in planning mode + P0-1/P0-2 prompt upgrades

P0: KB policy audit in Balthasar + round decision micro-scaffold in Caspar (Claude Opus 4.7 recommendation)

fix: Caspar keep-alive - _stream_claude_et_with_keepalive in all 4 modes, prevents HF Space idle timeout

fix: universal system-level MAZE data injection for all 6 agents (evaluation+planning)

feat: Universal dual-track fusion injector - all modes/agents now fuse Agentic Tool data + pretrained knowledge via _dispatch_stream + _call_claude_et hooks

fix: dual-track prompt engineering - force MELCHIOR to fuse Agentic Tool data + pretrained knowledge in evaluation+planning modes

feat: add Agentic Tool Prefetch (方案A) - MAZE/school data fetched before 3-agent deliberation in evaluation+planning modes

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

fix: Claude max_tokens 8000→16000, budget_tokens 4000→6000 to fix Caspar truncation

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

feat: upgrade Caspar system prompts (planning+evaluation) - remove word limit, require full MAZE data citation, detailed school list, deep strategy

fix: 消除 planning/arbiter system prompt 中「你女儿/你孩子」幻觉——改为第三方称谓禁令

feat: Grok Multi-Agent review - upgrade planning mode MAZE analysis

update agent/magi_system.py

update agent/magi_system.py

update agent/magi_system.py

update agent/magi_system.py

update agent/magi_system.py

feat: remove all KB retrieval limits - return all matching results

feat: add school_rounds_kb (678 schools ED/EA/ED2 data) + increase KB retrieval limits

fix: broaden KB retrieval triggers - each KB gets independent budget

Upload agent/magi_system.py with huggingface_hub

feat: inject public uni rubric into Caspar/Arbiter for public school questions

Upload agent/magi_system.py with huggingface_hub

Upload agent/magi_system.py with huggingface_hub

Update agent/magi_system.py: multi-turn dialog + clarify + new KB + planning mode

fix: remove sat/score keyword fallback that caused misrouting to sat_tolerance

fix: replace hardcoded intent detection with Grok-based classifier; add planning mode to MAGI

Add feeder_school_info + summer_program_recommendation tools (v1.1)

feat: _tool_ps_distill uses v4 three-tier kb (T10+藤/T15/T20-30)

v3.1: serial workflow + model_config routing + simplified Chinese fixes

feat: add agentic_engine.py v2.0 - true LLM tool-calling loop with 10 tools

agentic_planner.py: ps_distill now uses ps_distillation_kb.json (MAZE 158 cases) + blog_cards

Add Agentic Planning Engine v1.0 with activity recommend + PS distill tools