scientify 2.1.0 → 3.1.0

package/skills/figure-standardize/SKILL.md

@@ -0,0 +1,110 @@
+---
+name: figure-standardize
+description: "Use this when the user wants to improve chart quality, standardize plotting style, regenerate release figures, or add captions/protocol notes. Normalizes fonts, colors, legends, units, and scope notes across Scientify figures."
+metadata:
+  {
+    "openclaw":
+      {
+        "emoji": "📊",
+      },
+  }
+---
+
+# Figure Standardization
+
+**Don't ask permission. Just do it.**
+
+Use this skill to turn one-off Scientify charts into release-ready figures.
+
+**Do not run new experiments here.** Work from existing results, plotting scripts, and figure bundles. If the source data is missing or inconsistent, report that explicitly instead of smoothing it over.
+
+## Required Outputs
+
+1. Updated plotting script(s) or a shared style helper
+2. Regenerated `.png` and `.pdf` files when the pipeline supports both
+3. A figure spec file:
+   - prefer `reports/figures/figure_spec.md`
+   - otherwise `project/figures/figure_spec.md`
+4. `paper/figures_manifest.md` when the figure family is paper-facing or a `paper/` workspace already exists
+
+## Workflow
+
+### Step 1: Inspect Inputs
+
+Read:
+
+- existing figures
+- the generator script(s)
+- the result tables / JSON / Markdown that feed the figures
+- any surrounding README or release notes that explain the figure family
+
+Prefer improving an existing generator over creating a second one-off script.
+
+### Step 2: Standardize the Figure Family
+
+Normalize the full family, not just one chart:
+
+- font family and title hierarchy
+- semantic color mapping
+- axis labels and units
+- legend order and naming
+- decimal precision and tick formatting
+- line widths / marker sizes
+- caption structure
+- protocol note wording
+- callout wording
+- paper placement intent
+
+Use:
+
+- `references/figure-style-guide.md`
+- `references/caption-template.md`
+- `references/figure-placement-template.md`
+
+### Step 3: Write the Figure Spec
+
+Create or update `figure_spec.md` with one section per figure:
+
+- figure filename
+- source files
+- metrics shown
+- baseline or comparison family
+- quality guard / evaluation constraint
+- simulator/runtime note
+- intended takeaway
+
+If the figure is used in a paper or paper-facing report, also create or update the matching entry in `paper/figures_manifest.md` with:
+
+- `figure_id`
+- `file_path`
+- `latex_label`
+- `section`
+- `placement_hint`
+- `caption_short`
+- `caption_long`
+- `takeaway_sentence`
+- `callout_sentence`
+- `baseline`
+- `evidence_type`
+- `source_metrics`
+- `source_files`
+- `supports_claim_ids`
+- `must_appear_before_claim_ids`
+
+Keep `figure_spec.md` and `paper/figures_manifest.md` aligned. The spec is the release-facing summary; the manifest is the paper-facing contract.
+
+### Step 4: Re-render and Verify
+
+Re-render the figures after script updates.
+
+Keep filenames stable unless the user explicitly asked for a new release bundle.
+
+## Figure Rules
+
+1. Keep metric semantics identical across a figure family.
+2. Always show units explicitly.
+3. If a result comes from simulator or proxy evaluation, state that in the caption or protocol note.
+4. Do not hide failing or quality-guard-breaking baselines; mark them clearly.
+5. Do not change the scientific claim. This skill improves packaging, not evidence.
+6. If a figure is paper-facing, produce both a long caption and a first-use callout sentence.
+7. If a figure supports a claim, the manifest must name that claim in `supports_claim_ids`.

package/skills/figure-standardize/references/caption-template.md

@@ -0,0 +1,12 @@
+# Caption Template
+
+```text
+caption_short:
+Figure X. {Short statement of what is compared}.
+
+caption_long:
+Figure X. {What the figure shows and the main takeaway}. Baseline: {baseline family}. Metric: {metric + unit}. Evidence type: {simulator/local_runtime/runtime}. Protocol: {guardrail or evaluation scope}. Boundary: {scope note if needed}.
+
+callout_sentence:
+Figure \ref{{latex_label}} shows {takeaway_sentence} under {protocol or evaluation frame}.
+```

package/skills/figure-standardize/references/figure-placement-template.md

@@ -0,0 +1,30 @@
+# Figure Placement Template
+
+Use this guide when assigning `section` and `placement_hint` in `paper/figures_manifest.md`.
+
+Recommended fields:
+
+```yaml
+section: "main_results"
+placement_hint: "figure[t]"
+must_appear_before_claim_ids:
+  - "claim-001"
+```
+
+Placement hint options:
+
+- `figure[t]`
+  - standard single-column top placement
+- `figure[b]`
+  - standard single-column bottom placement
+- `figure* [t]`
+  - wide figure for two-column layouts
+- `inline_reference_only`
+  - no figure block in the current section; the text only points to a figure placed elsewhere
+
+Placement rules:
+
+- Put the figure in the same section that carries its main supported claim whenever possible.
+- If a figure supports a headline result, the first callout should appear before or immediately adjacent to the corresponding claim.
+- Do not place a figure so late that the reader sees the claim well before the figure is introduced.
+- Use `must_appear_before_claim_ids` to mark claims that should not appear before the figure callout.

package/skills/figure-standardize/references/figure-style-guide.md

@@ -0,0 +1,36 @@
+# Figure Style Guide
+
+Use one consistent style across a figure family:
+
+- one font family
+- one semantic color per method family
+- one stable baseline line style
+- explicit units in axis labels
+- compact legend names
+- protocol note under the figure or in the caption
+- one stable caption and callout structure across the family
+
+Required paper-facing figure contract fields:
+
+- `caption_short`
+- `caption_long`
+- `callout_sentence`
+- `placement_hint`
+- `supports_claim_ids`
+
+Required caption fields:
+
+- what is compared
+- which baseline is used
+- what metric is shown
+- what evidence type supports the figure
+- what protocol or guardrail defines the evaluation regime when that matters for the claim
+
+Paper-facing figures should stay aligned across four surfaces:
+
+1. `paper/figures_manifest.md`
+2. the first prose callout
+3. the figure caption
+4. the eventual LaTeX figure block
+
+Do not let these four surfaces drift apart.

package/skills/idea-generation/SKILL.md

@@ -19,27 +19,25 @@ Generate innovative research ideas grounded in literature analysis. This skill r
 
 **Core principle:** Ideas MUST be grounded in actual papers, not generated from model knowledge.
 
-**Workspace:** `$W` = working directory provided in task parameter. Outputs go to `$W/ideas/`.
+Outputs go to `ideas/`.
 
 ---
 
 ## Step 1: Check Workspace Resources
 
-First, check what resources already exist in `$W`:
+First, check what resources already exist:
 
 ```bash
-ls $W/papers/ 2>/dev/null | head -20
-ls $W/papers/_meta/ 2>/dev/null | head -10
-ls $W/survey/ 2>/dev/null
+ls papers/ 2>/dev/null | head -20
+ls knowledge/ 2>/dev/null
 ```
 
 ### Assess Available Resources
 
 | Resource | Location | Status |
 |----------|----------|--------|
-| Papers | `$W/papers/` | Count: ? |
-| Survey clusters | `$W/survey/clusters.json` | Exists: Y/N |
-| Repos | `$W/repos/` | Count: ? |
+| Papers | `papers/` | Count: ? |
+| Knowledge | `knowledge/_index.md` | Exists: Y/N |
 
 ---
 
@@ -77,7 +75,7 @@ This will:
 - Search 100+ papers systematically
 - Filter by relevance (score ≥4)
 - Cluster into research directions
-- Save to $W/papers/
+- Save to papers/
 
 After survey completes, run /idea-generation again.
 ```
@@ -86,31 +84,25 @@ After survey completes, run /idea-generation again.
 
 For fast iteration, do minimal search:
 
-1. **ArXiv search:**
+1. **Search papers:**
 ```
-Tool: arxiv_search
-Arguments:
-  query: "{user_topic}"
-  max_results: 10
+arxiv_search({ query: "{user_topic}", max_results: 10 })
+openalex_search({ query: "{user_topic}", max_results: 10 })
 ```
 
-2. **Clone 3-5 reference repos:**
-```bash
-mkdir -p $W/repos
-git clone --depth 1 {repo_url} $W/repos/{name}
-```
+2. **Download papers:** 按 /paper-download 的方式下载到 `papers/`
 
-3. **Download paper sources:**
+3. **Clone reference repos (optional):**
 ```bash
-mkdir -p $W/papers/{arxiv_id}
-curl -L "https://arxiv.org/src/{arxiv_id}" | tar -xz -C $W/papers/{arxiv_id}
+gh search repos "{paper_title} implementation" --limit 5 --sort stars
+git clone --depth 1 {repo_url} repos/{name}
 ```
 
 ---
 
 ## Step 4: Analyze Literature
 
-**Prerequisites:** At least 5 papers in `$W/papers/`
+**Prerequisites:** At least 5 papers in `papers/`
 
 ### 4.1 Read Papers
 
@@ -130,7 +122,7 @@ Look for:
 - Scalability issues
 - Assumptions that could be relaxed
 
-Document gaps in `$W/ideas/gaps.md`:
+Document gaps in `ideas/gaps.md`:
 ```markdown
 # Research Gaps Identified
 
@@ -146,7 +138,7 @@ Document gaps in `$W/ideas/gaps.md`:
 
 ## Step 5: Generate 5 Ideas
 
-Create `$W/ideas/idea_1.md` through `idea_5.md` using template in `references/idea-template.md`.
+Create `ideas/idea_1.md` through `idea_5.md` using template in `references/idea-template.md`.
 
 **Requirements:**
 - Each idea cites ≥2 papers by arXiv ID
@@ -175,28 +167,12 @@ Create `$W/ideas/idea_1.md` through `idea_5.md` using template in `references/id
 
 ### 6.2 Enhance Selected Idea
 
-Create `$W/ideas/selected_idea.md` with:
+Create `ideas/selected_idea.md` with:
 - Detailed math (loss functions, gradients)
 - Architecture choices
 - Hyperparameters
 - Implementation roadmap
 
-### 6.3 (Optional but recommended) OpenReview Evidence Check
-
-For the top 1-2 shortlisted ideas, validate novelty/positioning risk with `openreview_lookup`:
-
-- Query using core title keywords or representative baseline paper title
-- Extract evidence:
-  - decision (if available)
-  - average rating/confidence
-  - reviewer weakness patterns
-- Add a short "submission risk note" section per idea:
-  - likely reviewer concern
-  - mitigation experiment to add
-  - positioning adjustment
-
-Do not claim accept/reject predictions as facts. Report evidence-backed risk signals only.
-
 ---
 
 ## Step 7: Code Survey
@@ -205,13 +181,13 @@ Map idea concepts to reference implementations.
 
 See `references/code-mapping.md` for template.
 
-**Output:** `$W/ideas/implementation_report.md`
+**Output:** `ideas/implementation_report.md`
 
 ---
 
 ## Step 8: Summary
 
-Create `$W/ideas/summary.md`:
+Create `ideas/summary.md`:
 - All 5 ideas with scores
 - Selected idea details
 - Next steps: `/research-pipeline` to implement

package/skills/idea-generation/references/code-mapping.md

@@ -9,13 +9,13 @@ From `selected_idea.md`, list concepts needing implementation.
 ## Step 2: Survey Codebases
 
 ```bash
-grep -r "class.*Attention" $WORKSPACE/repos/
-grep -r "def forward" $WORKSPACE/repos/
+grep -r "class.*Attention" repos/
+grep -r "def forward" repos/
 ```
 
 ## Step 3: Implementation Report
 
-Write to `$WORKSPACE/ideas/implementation_report.md`:
+Write to `ideas/implementation_report.md`:
 
 ```markdown
 # Implementation Report

package/skills/idea-generation/references/idea-template.md

@@ -1,6 +1,6 @@
 # Idea Template
 
-Use this template for each generated idea (`$WORKSPACE/ideas/idea_N.md`).
+Use this template for each generated idea (`ideas/idea_N.md`).
 
 ## Required Sections
 

package/skills/idea-generation/references/reading-long-papers.md

@@ -5,8 +5,8 @@ For papers >50KB or >15k tokens, use chunked reading.
 ## Step 1: Structure Scan
 
 ```bash
-ls -la $WORKSPACE/papers/{arxiv_id}/
-wc -l $WORKSPACE/papers/{arxiv_id}/*.tex
+ls -la papers/{arxiv_id}/
+wc -l papers/{arxiv_id}/*.tex
 ```
 
 ## Step 2: Chunked Reading
@@ -16,7 +16,7 @@ Use Read tool with `offset` and `limit`:
 ```
 Tool: Read
 Arguments:
-  file_path: "$WORKSPACE/papers/2404.04429/main.tex"
+  file_path: "papers/2404.04429/main.tex"
   offset: 1
   limit: 500    # First 500 lines
 ```

package/skills/metabolism/SKILL.md

@@ -1,24 +1,80 @@
 ---
 name: metabolism
-description: "Knowledge metabolism cycle: ingest new papers, update knowledge state, detect cross-topic links, generate hypotheses. Use /metabolism to trigger manually."
-user-invokable: true
+description: "Knowledge metabolism: Day 0 init (discuss + baseline survey) or daily cycle (incremental search + update knowledge + generate hypotheses). Triggered by bootstrap and daily cron."
 ---
 
-# Continuous Knowledge Metabolism — Incremental Cycle
+# Knowledge Metabolism
 
-你正在执行知识新陈代谢循环。严格按以下步骤执行。
+读取 `config.json` 的 `currentDay` 字段判断执行哪个分支：
 
-**前提：** `metabolism/config.json` 必须已存在且 `currentDay >= 1`。如果不存在或 `currentDay` 为 0，提示用户先执行 /metabolism-init 完成初始化。
+- **`currentDay == 0` 或 config.json 不存在** → 执行 Day 0（初始化）
+- **`currentDay >= 1`** → 执行每日增量循环
 
-## 准备
+---
+
+## Day 0：初始化
+
+与用户讨论研究方向后，构建领域基线知识。
 
-1. 读取 `metabolism/config.json` 获取关键词、arXiv 分类、`processed_ids` 和 `currentDay`
-2. 读取 `metabolism/knowledge/_index.md` 获取当前知识状态
+### 准备
 
-## Step 1: Search（增量搜索）
+1. 检查 `config.json` 是否存在
+   - 不存在：询问用户研究方向，创建 `config.json`（包含 `keywords`, `arxivCategories`, `sources`, `currentDay: 0`, `processed_ids: []`）
+   - 已存在且 `currentDay > 0`：跳转到每日循环
 
-用滑动窗口（过去 5 天）搜索，靠 `processed_ids` 去重：
+### Step 1: 宽泛调研
 
+委托 /research-collect 执行宽泛调研（不限日期），构建初始知识：
+
+```
+sessions_spawn({
+  task: "/research-collect\n研究主题: {从 config.json 的 keywords 提取}\n这是 Day 0 基线构建，请做宽泛调研（不限日期），覆盖领域经典工作和近期进展。",
+  label: "Day 0 Baseline Survey",
+  runTimeoutSeconds: 1800
+})
+```
+
+等待完成后，读取 `papers/` 获取论文列表。
+
+### Step 2: 阅读与知识提取
+
+对每篇论文：
+- 读 .tex 源码（优先）或 PDF
+- 提取：核心方法、关键结论、领域现状
+
+将每篇论文的 arXiv ID / DOI 追加到 `config.json` 的 `processed_ids`。
+
+### Step 3: 构建初始知识状态
+
+创建 `knowledge/_index.md`：
+- Research Goal（从 config.json 提取）
+- Topics 表格（按论文主题聚类）
+- Cross-topic Links（如有）
+- Timeline（记录 Day 0）
+
+为每个识别出的主题创建 `knowledge/topic-{name}.md`：
+- 已知方法
+- 关键论文与结论
+- 开放问题
+
+### Step 4: 记录日志
+
+写入 `log/{YYYY-MM-DD}-init.md`，更新 `config.json`：`currentDay` 设为 1。
+
+---
+
+## Day 1+：每日增量循环
+
+### 准备
+
+1. 读取 `config.json` 获取关键词、arXiv 分类、`sources`、`processed_ids` 和 `currentDay`
+2. 读取 `knowledge/_index.md` 获取当前知识状态
+
+### Step 1: Search（增量搜索）
+
+根据 `config.json` 的 `sources` 字段选择搜索来源，用滑动窗口（过去 5 天）搜索，靠 `processed_ids` 去重：
+
+**arXiv**（如果 sources 包含 "arxiv"）：
 ```
 arxiv_search({
   query: "{keywords} AND cat:{category}",
@@ -26,7 +82,10 @@ arxiv_search({
   sort_by: "submittedDate",
   max_results: 30
 })
+```
 
+**OpenAlex**（如果 sources 包含 "openalex"）：
+```
 openalex_search({
   query: "{keywords}",
   filter: "from_publication_date:{5天前 YYYY-MM-DD}",
@@ -37,14 +96,9 @@ openalex_search({
 
 合并结果，按 arXiv ID / DOI 去重，**跳过 `processed_ids` 中已有的论文**。
 
-下载新论文：
+按 /paper-download 的方式下载新论文到 `papers/`（arXiv 优先 .tex 源文件，DOI 通过 Unpaywall 获取 OA PDF）。
 
-```
-arxiv_download({ arxiv_ids: ["{id1}", "{id2}", ...] })
-unpaywall_download({ dois: ["{doi1}", "{doi2}", ...] })
-```
-
-## Step 2: Read（阅读）
+### Step 2: Read（阅读）
 
 对每篇新论文：
 - 读 .tex 源码（优先）或 PDF
@@ -52,9 +106,9 @@ unpaywall_download({ dois: ["{doi1}", "{doi2}", ...] })
 
 将每篇论文的 arXiv ID / DOI 追加到 `config.json` 的 `processed_ids`。
 
-## Step 3: Update Knowledge
+### Step 3: Update Knowledge
 
-读取当前 `metabolism/knowledge/_index.md` 和相关 `topic-*.md`，根据今日阅读的论文更新。
+读取当前 `knowledge/_index.md` 和相关 `topic-*.md`，根据今日阅读的论文更新。
 
 **更新原则：**
 - 新发现 → 添加到相关章节
@@ -64,15 +118,15 @@ unpaywall_download({ dois: ["{doi1}", "{doi2}", ...] })
 
 **篇幅管理：** 每个 topic 文件控制在 200 行以内。接近上限时，压缩早期内容（合并相似结论、删除低价值条目），保留信息密度。不要为了压缩而丢失关键结论和来源引用。
 
-## Step 4: Hypothesize（假设）
+### Step 4: Hypothesize（假设）
 
-更新完 knowledge.md 后，回顾今日新增内容，自问：
+更新完知识后，回顾今日新增内容，自问：
 
 - 有没有反复出现但尚未被验证的模式？
 - 有没有两个独立发现组合后暗示的新方向？
 - 有没有现有方法的明显空白？
 
-**有想法** → 写入 `metabolism/hypotheses/hyp-{NNN}.md`：
+**有想法** → 写入 `ideas/hyp-{NNN}.md`：
 
 ```markdown
 # Hypothesis {NNN}
@@ -94,27 +148,17 @@ unpaywall_download({ dois: ["{doi1}", "{doi2}", ...] })
 
 然后用 `sessions_send` 通知 main session。
 
-**说明：** 发送给用户的内容不要输出“触发条件检查表/列表”，只需解释为什么这个假设合理、哪些已读内容启发了它。
-
 **没有想法** → 跳过，不要硬凑。
 
-## Step 5: Log & Finish
+### Step 5: Log & Finish
 
-写入 `metabolism/log/{YYYY-MM-DD}.md`：
+写入 `log/{YYYY-MM-DD}.md`，更新 `config.json`：`currentDay` +1。
 
-```markdown
-# Day {currentDay} — {YYYY-MM-DD}
-
-新论文: {N} 篇
-知识更新: {简述主要变更}
-假设: {有/无}
-```
-
-更新 `config.json`：`currentDay` +1。
+---
 
 ## 行为约束
 
 1. 不捏造论文中未出现的事实性声明，但可以用自身知识做推理和关联判断
 2. 没有想法时不生成假设
-3. 自主运行，不向人类提问
+3. 自主运行，不向人类提问（Day 0 初始配置除外）
 4. 修改知识文件前必须先读取当前内容

package/skills/paper-download/SKILL.md

@@ -0,0 +1,61 @@
+---
+name: paper-download
+description: "Download academic papers: arXiv source/PDF by ID, DOI papers via Unpaywall open access. Supports batch download."
+---
+
+# Paper Download
+
+将论文下载到当前工作目录的 `papers/` 下。
+
+## arXiv 论文
+
+**优先下载 .tex 源文件**（可读性远优于 PDF）：
+
+```bash
+mkdir -p papers/{arxiv_id}
+curl -L "https://arxiv.org/src/{arxiv_id}" | tar -xz -C papers/{arxiv_id}
+```
+
+如果 tar 解压失败（部分论文只提供 PDF），回退到 PDF：
+
+```bash
+curl -L -o papers/{arxiv_id}.pdf "https://arxiv.org/pdf/{arxiv_id}"
+```
+
+> arXiv 限速：连续下载时每篇间隔 3 秒（`sleep 3`）。
+
+## DOI 论文（通过 Unpaywall）
+
+查询开放获取链接，有则下载，无则跳过：
+
+```bash
+curl -s "https://api.unpaywall.org/v2/{doi}?email=research@openclaw.ai" | \
+  python3 -c "
+import sys, json
+d = json.load(sys.stdin)
+oa = d.get('best_oa_location') or {}
+url = oa.get('url_for_pdf') or oa.get('url')
+if url: print(url)
+else: print('NO_OA', file=sys.stderr)
+" | xargs -I{} curl -L -o papers/{doi_slug}.pdf "{}"
+```
+
+> `{doi_slug}` = DOI 中的 `/` 替换为 `_`，例如 `10.1000/xyz123` → `10.1000_xyz123`。
+> 非开放获取论文静默跳过，不报错。
+
+## 批量下载
+
+```bash
+# 批量 arXiv
+for id in 2401.12345 2403.00001 2405.67890; do
+  mkdir -p papers/$id
+  curl -L "https://arxiv.org/src/$id" | tar -xz -C papers/$id || \
+    curl -L -o papers/$id.pdf "https://arxiv.org/pdf/$id"
+  sleep 3
+done
+```
+
+## 下载后
+
+- 下载的论文 ID 应追加到 `config.json` 的 `processed_ids`（如果存在）
+- 优先读 `.tex` 源码而非 PDF（信息更完整，公式可直接提取）