Skip to content

Latest commit

 

History

History
76 lines (58 loc) · 5.17 KB

File metadata and controls

76 lines (58 loc) · 5.17 KB

Changelog

[2.1.0] - 2026-08-09

Added

  • 记忆图谱与自进化:台账之上新增派生结构化索引 .research/graph/nodes.jsonl / edges.jsonl / candidates.jsonl / events.jsonl + 自动生成的人读视图 kg.md),节点/边全部带 provenance 回源台账
  • graph_evolve.py 全链路命令:init / scan / review / edge / node / promote / feedback / profile / venue / digest / query / render / migrate / consolidate / status
  • 确认策略:机械边(evidence_for / derived_from 等)scan 自动入图;判断边 (contradicts / supersedes / complements 等)只生成候选,必须用户确认
  • A-MEM 式演化:supersedes 确认后旧节点自动置 superseded 并记录 superseded_by
  • 共享层经验图谱:promote 项目经验→共享候选→确认后入 lessons/<domain>/feedback add 同一纠正累计 3 次自动生成画像候选;profile add / venue add 沉淀用户规则与期刊画像
  • UserPromptSubmit 记忆注入钩子memory_inject.py):会话开始注入图谱摘要 (active 节点、未回应矛盾、待确认候选、近 14 天变更、相关经验/画像/期刊), 默认 1000 token(MTC_GRAPH_DIGEST_BUDGET 可调),fail-open
  • 门禁消费图谱:S6 新增 graph_matrix_registered(矩阵实验必须登记节点)、 S9 新增 graph_contradictions_responded(内部矛盾边必须标记 responded_in
  • project_context.py 输出增加 project_graph / shared_graph--create 自动建图谱骨架
  • docs/memory-graph.md:两层图谱 Schema、命令、三条自进化闭环、安全边界

Changed

  • research-ledger / academic-workbench:阶段收尾先 scan + review 再布防门禁; S9 写作要求图谱内部矛盾边在正文回应并标记
  • hooks.json 同时注册 UserPromptSubmit(记忆注入)与 Stop(门禁)两类钩子

[2.0.0] - 2026-08-09

Added

  • 多课题隔离projects/<id>/.research/ 项目记忆 + .research/meta/ 共享元记忆 + project_context.py 注册表;跨课题访问必须显式点名(V1 为单课题 .research/
  • 写作保真协议:两级保全不变量、claim-strength 阶梯、evidence/preservation ledger、输出契约(提炼自 Yila-AI/sci-ssci-skills,Apache-2.0)
  • 目标期刊适配建模:Paper Style Card → Style Profile → 动态写作规则 + ML/CV/NLP 默认基础规则(提炼自 WantongC/journal-adapt-writing-skill,MIT)
  • 阶段门禁(Codex Stop hooks):S2/S3/S4/S5→S6/S6/S7/S9/S11/S12/S13 门禁清单 + 自动返工循环(默认最多 3 轮);.codex/hooks.json + gate_check.py + gates.json 随仓库分发,克隆后一次信任即生效;钩子不可用时降级为强流程约束
  • nature-proposal-writer/scripts/check_preservation.py 确定性保全检查(数字/引用/术语/claim 标记)
  • 投稿前机械检查与投稿信 align-check 协议(借鉴 bahayonghang/academic-writing-skills 机制,仅机制不抄文本)
  • 基线策略三态:reproduce / cite_published / excluded,默认引用已发表数字不复现

Changed

  • S6 重设计为工程实现与实验矩阵冻结:协议登记 → 矩阵冻结 → 实现 → 冒烟 → 三轮自检 → 验收报告(全 PASS 才进 S7)
  • S7 合并为实验运行与参数优化闭环:pilot → 基线 → 调参/消融/go-no-go(只在 pilot 数据上)→ 配置锁定 → canonical 正式评估 → 台账回写
  • 技能本地化:删除上游同步脚本(install / update / check / ghfetch / patch-schematic-ua),技能直接在 .codex/skills/ 维护,改文件即生效
  • ARS 更新至上游精简版(V1 捆绑 1746 文件 → V2 747 文件)
  • 学术诚信红线新增第 6 条:写作/润色保真

Removed

  • 11 个与投稿流水线无关的 nature-* 技能:nature-writing、nature-polishing、nature-citation、nature-paper2ppt、nature-paper-to-patent、nature-experiment-log、nature-reader、nature-paper-card、nature-data、nature-response、nature-reviewer

[1.1.0] - 2026-08-07

Added

  • README hero / pipeline / Codex-integration assets (generated by assets/make_assets.py, PNG + SVG)
  • docs/third-party.md: credits & introduction for the 4 upstream projects (nature-skills, scientific-agent-skills, engineering-figure-agent, academic-research-suite), replaces docs/features.md
  • docs/codex-guide.md: Codex-native usage guide (clone via Codex, do not manually copy skills)
  • docs/stages.md: detailed walkthrough of all 13 stages
  • README sections: “Why different”, “How to use with Codex”, FAQ links

[1.0.0] - 2026-08-07

Added

  • 13-stage research routing workbench (academic-workbench entry)
  • 6 first-party skills: academic-workbench, research-ledger, research-tree-search, novelty-sweep, math-verification, narrative-engine
  • 34 third-party skills bundled with license notices (nature-skills, scientific-agent-skills, academic-research-suite, engineering-figure-agent)
  • Install / update / check / distill scripts with parameterized paths
  • .research ledger scaffold templates
  • Dev Container, GitHub Actions CI, Issue/PR templates, Code of Conduct
  • API configuration template (.env.example)