v0.12.18 - Stable release (Bilibili demo video)

Stable release corresponding to the Bilibili demo video (2026-07-31).

Highlights

• Multi-department adversarial debate with cross-debate moderation
• Citation-grounded verification: abstract-level fact check, references never LLM-translated
• Resume-from-checkpoint across all phases
• Dual pipelines: academic literature review + general requirements discussion
• 24k lines, pure-lightweight deps (no local models, runs on any 8GB machine)

Live demo: https://consensus-pipeline.streamlit.app

Note: v2 roadmap (divergence metrics / soft-deadlock detection) is tracked separately and will land on a dev branch first; main stays stable.

v0.9 — Market Dept Filter + Fact Check Fix

Changes since v0.8.1

• Market mode department filter: Select which departments to run in market mode (reduces memory on Streamlit Cloud)
• Market dept config-aware: Department list follows current config (academic/animation), not hardcoded animation defaults
• Widget cache fix: Clear stale dept filter when config changes
• Defensive DEPARTMENTS.get(): Prevent KeyError when switching configs
• Fact check claim extraction: Extract from final report (not debate consensus), filter meta-discussion patterns
• Fact check source indicator: Show whether claims came from final report or dept consensus

Known limitations

• Fact check verification still uses naive paragraph extraction (citation-grounded verification coming in v1.0)
• Streamlit Cloud OOM possible with all departments + multiple market candidates

v0.8 - Streamlit Cloud 适配修复

v0.8 - Streamlit Cloud 适配修复 + 辩论防重跑

新功能

• 学术校对自动修正功能
• 自动修正循环 + 校对轮次滑块 (0-5轮)
• Keepalive 保活 (10分钟HEAD请求, 防Streamlit Cloud 15分钟空闲休眠)

Bug修复

• 5个学术模式bug修复
• 多用户隐私漏洞修复 (autosave改session_state)
• 应用按钮消失bug修复
• latin-1编码错误修复 (2轮, 含_safe_filename)
• 截断限制提升 20k→30k字符

辩论防重跑 (v1→v3)

• v1: try/finally兜底 + 防重入守卫 + auto-start状态清理
• v2: 进入守卫立刻清flag防中途rerun重入
• v3: 错误跨rerun保留 + stale flag补回rerun

已知问题

• 跑到某个进度所有数据清空 (疑似容器OOM重启)
• try/except Exception可能吞Streamlit内部异常

完整commit链

1388767 → 7013a72 → 175ca27 → 06333b5 → 29c787b → f6bdb99 → 81aa24d → 67a641d → a5c251b → e418a4b → ccd2077 → 4fcae9e

v0.7.5 — Cross-Debates Fallback Fix

Bug Fix

• Phase 6 cross-debates empty output (P0): Fixed config generation leaving p2_cross_debates / p5_cross_debates empty for non-energy domains, causing Phase 6 to produce empty JSON.

Changes

• requirement/config_recommender.py: Added _ensure_cross_debates() auto-fallback + _auto_generate_cross_debates() utility + integrity check in _adjust_with_llm() to protect cross-debates from LLM removal
• run_pipeline.py: Added belt-and-suspenders fallback in phase6_cross_debate() — auto-generates pairs from dept_order if p2_cross_debates is empty at runtime
• _generate_default_config() now auto-generates cross-debate pairs for custom domains

v0.7.4 — UI Overhaul + README Rewrite

🎨 Streamlit UI Overhaul

• 9 Tab → 4 Tab 重组：需求与配置 / 辩论 / 产出 / 工具，按工作流顺序展现
• 子区域嵌套：每个Tab内用tabs/expander组织原9个功能模块
• 需求调研→辩论自动跳转：Phase4确认后toast提示切换
• easyScholar降级：从sidebar主区移至高级设置折叠区（仅国内用户可选）
• .env生成器：sidebar一键保存API配置为.env文件，CLI用户免手动创建

📖 README Complete Rewrite

• 新增「Why Not Just Ask ChatGPT?」段落
• 新增完整Quick Start指南（CLI参数表 + Streamlit分步教程）
• 新增FAQ/故障排除（7个实用QA）
• 新增Screenshots占位区
• 版本号同步至v0.7.3
• 中英文双语同步更新

v0.7.3 — Multi-Round Debate Fix

🔧 v0.7.3 — Multi-Round Debate Fix

Bug Fix

rounds parameter was dead code in _debate_department(). The function accepted a rounds argument but never used it — every debater only spoke once regardless of the configured debate_rounds value.

What Changed

• _debate_department() now loops rounds times
• Round 1: each debater gives initial analysis
• Round 2+: debaters can see previous rounds' arguments, respond to challenges, and add new evidence
• Previous rounds' arguments are summarized and injected as context
• Each debater's output accumulates across rounds (separated by round headers)
• debate_rounds from config now actually controls debate depth

Other Fixes (from v0.7.2)

• --lang en parameter for English report output
• Sanitized P1/P4 example content in presets and quickstart
• Cleaned up 22 legacy tags and 21 legacy releases
• README improvements: Phase numbering, domain qualifier, maturity label

v0.7.2 — Bilingual Report Output

🌐 v0.7.2 — Bilingual Report Output

New Feature: --lang Parameter

Run the academic pipeline in English or Chinese:

# Chinese report (default, backward compatible)
python run_pipeline.py --topic "Machine Learning in Energy Economics"

# English report
python run_pipeline.py --topic "Machine Learning in Energy Economics" --lang en

Changes

• Added --lang parameter to run_pipeline.py (choices: zh/en, default: zh)
• report_generator.py now supports English output:
  • Full English system prompt with 26 writing rules
  • English user prompt with Data Card and section templates
  • Bilingual section headers (一→1. Core Findings, etc.)
  • Localized date format (%B %d, %Y vs %Y年%m月%d日)
• All 7 LLM call sites in run_pipeline.py use language helper functions
• Backward compatible: --lang zh produces identical output to previous versions

Other Changes

• Sanitized P1/P4 example content in animation_debate preset and quickstart guide
• Cleaned up 22 legacy tags and 21 legacy releases
• README comparison table renamed from "v5.1.8" to "Legacy Version"
• Created v0.7.1 formal GitHub Release

v0.7.1 — Academic Pipeline Production Release

🔬 v0.7.1 — Academic Pipeline Production Release

Major Features

• QC Department (3-layer filter): hard_filter → LLM_classify → tag_layer. Reduced 219 retrieved papers to 77 relevant ones (64% exclusion rate, 0% off-topic in output)
• Dynamic Domain Config: LLM generates domain-specific exclusion signals, query rotation, and tier definitions. Change topic without code changes.
• Citation Validation: Auto-verify all [N] references against CSV; remove dangling citations
• Confidence Annotation: Every conclusion tagged with (N/M papers, confidence level) — no more unsupported claims
• OpenAlex Priority: Abstract backfill uses OpenAlex first (no 429 rate limits), falls back to Semantic Scholar

Quality Metrics vs Legacy Version

Metric	Legacy	v0.7.1
Off-topic papers in output	49/56 (88%)	0/77 (0%)
Placeholder citations	Multiple	0
Dangling references	Present	0
Confidence annotations	0	15
Domain switchability	Hardcoded	Dynamic
Self-evaluation score	4.8/10	7.4/10

Bug Fixes

• Fixed debater protection in Phase 5 — debaters no longer skip turns
• Fixed rounds parameter — multi-round debates now work correctly
• Fixed cross_validator in Phase 6 — no more empty validation passes
• Removed 5 Chinese residuals in codebase comments and print statements

Migration Note

Version numbering changed from v5.x to v0.7.x to better reflect project maturity. All v3.0–v5.1.8 releases are legacy and have been archived.