@research-copilot/plugin 1.1.15 → 1.1.17

package/dist/skills/scientist-ideation/SKILL.md

@@ -0,0 +1,52 @@
+---
+name: scientist-ideation
+description: "Use when the user has a workshop / topic Markdown and wants it turned into an AI-Scientist-format ideas JSON, generated directly in Copilot. Triggers on: '生成 ideas', 'topic 变成想法', 'AI Scientist 出点子', 'generate ideas from topic'. Copilot-native — no workspace ideation script call."
+version: 0.2.0
+---
+
+# scientist-ideation
+
+Convert a workshop / topic Markdown into an AI-Scientist-compatible ideas JSON. The model output MUST be produced by Copilot in-session.
+
+## Execution model
+
+This is a **Copilot-native model task**. Copilot reads the topic, brainstorms ideas, generates the JSON, and writes to a workspace file when the user requests it.
+
+## Workflow
+
+1. Read the user-supplied workshop / topic Markdown.
+2. If needed, check an existing ideas JSON to avoid duplicate directions.
+3. Generate candidate ideas in-session and organize them in the AI-Scientist schema.
+4. If the user asks for persistence, create or update the ideas JSON file directly.
+
+## JSON schema
+
+- `Name`
+- `Title`
+- `Short Hypothesis`
+- `Related Work`
+- `Abstract`
+- `Experiments`
+- `Risk Factors and Limitations`
+
+## Input
+
+- `workshop_file` or topic Markdown path
+- Existing ideas JSON (if any)
+- Directional / dataset / resource constraints the user wants preserved
+
+## Output
+
+- AI-Scientist-style ideas JSON
+- Written directly to a workspace file if requested
+- Explicit output path, idea count, and a list of duplicates filtered out
+
+## Forbidden
+
+- NEVER call any workspace-custom ideation pipeline.
+- NEVER call a model SDK from workspace code to generate ideas.
+
+## Failure handling
+
+- If the topic file's structure is too thin, surface the gap and ask for supplementation first.
+- If the user asks for persistence but the schema is incomplete, fill it in-session before writing.

package/dist/skills/scientist-ideation/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "scientist-ideation",
+  "description": "Use when the user has a workshop / topic Markdown and wants it turned into an AI-Scientist-format ideas JSON, generated directly in Copilot. Triggers on: '生成 ideas', 'topic 变成想法', 'AI Scientist 出点子', 'generate ideas from topic'. Copilot-native — no workspace ideation script call.",
+  "entry": "SKILL.md"
+}

package/dist/skills/scientist-plotting/SKILL.md

@@ -0,0 +1,49 @@
+---
+name: scientist-plotting
+description: "Use when the user asks to '聚合作图', '补图表', '整理实验图', or wants experiment outputs converted into plots + plotting scripts directly in Copilot. Triggers on: 'aggregate plots', 'make figures from results'. Copilot-native — Copilot designs the figure and edits the script; the terminal only runs the Python plotting code. Do NOT use without existing experiment outputs."
+version: 0.2.0
+---
+
+# scientist-plotting
+
+Generate plots and plotting scripts from an existing experiment directory. Model judgment and figure design are done by Copilot in-session.
+
+## Execution model
+
+This is a **Copilot-native model task**. Copilot reads results, decides figure structure, writes / edits plotting code; the terminal only runs pure-Python plotting scripts.
+
+## Workflow
+
+1. Read the experiment directory's summary JSON, logs, CSVs, NPY files, or existing plots.
+2. Decide which metrics and comparisons to display.
+3. Create or edit matplotlib / seaborn / pandas plotting code directly.
+4. Run the plotting script and inspect the output.
+5. If the figure is unclear, iterate.
+
+## Input
+
+- `folder`: experiment directory
+- Result file paths and formats
+- Figure conventions or paper-layout constraints the user wants preserved
+
+## Output
+
+- Plotting script or edits to an existing script
+- Output figure paths
+- Figure design rationale and the key visual conclusions
+
+## Operating principles
+
+- Only invoke this skill when the experiment outputs already exist.
+- If result files are incomplete, flag the gap; NEVER fabricate plots.
+
+## Forbidden
+
+- NEVER call any workspace-custom plotting model pipeline.
+- NEVER use custom model calls in workspace code to "auto-plot."
+
+## Deliverable requirements
+
+- Report figure paths.
+- Name the source result files used.
+- If a figure failed to render, return the real error and a suggested next step.

package/dist/skills/scientist-plotting/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "scientist-plotting",
+  "description": "Use when the user asks to '聚合作图', '补图表', '整理实验图', or wants experiment outputs converted into plots + plotting scripts directly in Copilot. Triggers on: 'aggregate plots', 'make figures from results'. Copilot-native — Copilot designs the figure and edits the script; the terminal only runs the Python plotting code. Do NOT use without existing experiment outputs.",
+  "entry": "SKILL.md"
+}

package/dist/skills/scientist-review/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: scientist-review
+description: "Use when the user asks to '审一下这篇 PDF', '自动审稿', '给我 review', or wants a manuscript / PDF reviewed in Copilot with structured feedback. Triggers on: 'review this manuscript', 'auto-review'. Copilot-native — no workspace review script call."
+version: 0.2.0
+---
+
+# scientist-review
+
+Text-level Copilot-native review of a paper or PDF. Model judgment and review output are produced by Copilot in-session; NEVER call a workspace-custom model script.
+
+## Execution model
+
+This is a **Copilot-native model task**. If only a PDF is supplied, extract text first, then have Copilot produce the review directly.
+
+## Workflow
+
+1. Acquire the paper text: prefer existing Markdown / LaTeX / TXT; fall back to PDF-text extraction only if necessary.
+2. Produce the review, scoring, and risk assessment directly in-session.
+3. If the user requests structured output, generate JSON or write to a file.
+
+## Input
+
+- `pdf_path`, LaTeX source, or already-extracted text
+- Reviewer perspective and scoring dimensions the user wants applied
+
+## Output
+
+- Review notes
+- Optional structured JSON review result on request
+- Explicit Strengths / Main Issues / Score / Risks
+
+## Forbidden
+
+- NEVER call any workspace-custom review pipeline.
+- NEVER use custom model calls in workspace scripts for reviewing.
+
+## Deliverable requirements
+
+- Default to a "Strengths / Main Issues / Score / Risks" summary.
+- If only a PDF is provided and text extraction fails, name the blocker explicitly.

package/dist/skills/scientist-review/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "scientist-review",
+  "description": "Use when the user asks to '审一下这篇 PDF', '自动审稿', '给我 review', or wants a manuscript / PDF reviewed in Copilot with structured feedback. Triggers on: 'review this manuscript', 'auto-review'. Copilot-native — no workspace review script call.",
+  "entry": "SKILL.md"
+}

package/dist/skills/scientist-runtime-init/SKILL.md

@@ -0,0 +1,46 @@
+---
+name: scientist-runtime-init
+description: "Use when the user asks to '检查环境', '能不能跑 AI Scientist', 'runtime check', '初始化 AI Scientist 环境', or wants the ai-scientist MCP to validate Python / CUDA / LaTeX / poppler / runtime prerequisites. Routes through the `ai-scientist` MCP `validate_runtime`. Do NOT use as a substitute for actually running an experiment."
+version: 0.2.0
+---
+
+# scientist-runtime-init
+
+Validate the scientist-support AI Scientist runtime preconditions in the current workspace via the `ai-scientist` MCP.
+
+## Goal
+
+Confirm the following before launching any long experiment:
+
+- Runtime root directory exists
+- Python is available
+- `pdflatex`, `bibtex`, `pdftotext`, `chktex` are available
+- `torch.cuda.is_available()` is true
+- The current platform is suitable for local experiments and LaTeX compilation
+
+## Preferred method
+
+For a full check, use the `ai-scientist` MCP `validate_runtime` tool.
+
+This skill organizes the check steps and the output format; it does NOT depend on any in-skill runner or alternative script entry point.
+
+If the MCP is unavailable, fall back to terminal checks on the same conditions, keeping the same output structure.
+
+## Output requirements
+
+Summarize in three columns: Ready / Missing / Risk:
+
+- **Ready**: satisfied items
+- **Missing**: missing items
+- **Risk**: e.g. Windows platform, no GPU, no LaTeX
+
+End with a next-step recommendation:
+
+- Can continue with ideation
+- Can continue with local experiments, plotting support, or paper compilation
+- Must fix the environment first
+
+## Forbidden
+
+- NEVER call any in-skill runner or script entry point.
+- NEVER treat API-key or model-SDK availability as a runtime-init check item.

package/dist/skills/scientist-runtime-init/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "scientist-runtime-init",
+  "description": "Use when the user asks to '检查环境', '能不能跑 AI Scientist', 'runtime check', '初始化 AI Scientist 环境', or wants the ai-scientist MCP to validate Python / CUDA / LaTeX / poppler / runtime prerequisites. Routes through the `ai-scientist` MCP `validate_runtime`. Do NOT use as a substitute for actually running an experiment.",
+  "entry": "SKILL.md"
+}

package/dist/skills/scientist-writeup/SKILL.md

@@ -0,0 +1,60 @@
+---
+name: scientist-writeup
+description: "Use when the user asks to '开始写论文', '生成 PDF', '整理成论文', or wants LaTeX / Markdown drafted directly in Copilot from experiment artifacts. Triggers on: 'write the paper', 'generate PDF', 'compile to paper'. Copilot-native — no workspace writeup script call. Do NOT use for review (scientist-review) or plotting (scientist-plotting)."
+version: 0.2.0
+---
+
+# scientist-writeup
+
+Generate or edit LaTeX / Markdown paper content directly from an existing experiment directory. Model output is produced by Copilot in-session; NEVER call workspace-custom model scripts.
+
+## Execution model
+
+This is a **Copilot-native model task**. Copilot reads results, writes content, and edits LaTeX files; the terminal handles non-model commands like `pdflatex`.
+
+## Workflow
+
+1. Read the experiment directory, summary files, figures, and logs.
+2. Identify the user-supplied `latex/template.tex` or the existing draft.
+3. Write or edit paper content directly in the editor.
+4. On user request, run `pdflatex` / `bibtex` for a compilation check.
+5. Report the produced manuscript path, compilation result, and remaining gaps.
+
+## Verification before declaring completion
+
+**Before claiming the paper is drafted, you MUST produce one of:**
+- the file path + a short verbatim quote of new content,
+- a `Read` confirmation that the new content is in the file,
+- a successful `pdflatex` exit and the produced PDF path,
+- or an explicit "drafted but could not verify — here is what I have so far."
+
+A turn that ends with "the paper is drafted" without one of the above is a failure mode.
+
+## Input
+
+- `folder`: experiment directory
+- `folder/latex/template.tex`: user-provided template entry
+- Figures, summarized results, citation info, and target-layout requirements
+
+## Output
+
+- Edited LaTeX / Markdown files
+- Compiled PDF path (if compilation was run)
+- List of unmet prerequisites
+
+## Operating principles
+
+1. Confirm the template and dependency files exist before writing.
+2. Write from real experimental results; NEVER fabricate conclusions or citations.
+3. When the user only wants a text draft, do not force a PDF compile.
+
+## Forbidden
+
+- NEVER call any workspace-custom writeup model pipeline.
+- NEVER use custom model calls in workspace code to generate paper text.
+
+## Deliverable requirements
+
+- Name which paper files were edited.
+- If compilation fails, return the real LaTeX error summary.
+- If conclusions still lack experimental support, name the missing results explicitly.

package/dist/skills/scientist-writeup/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "scientist-writeup",
+  "description": "Use when the user asks to '开始写论文', '生成 PDF', '整理成论文', or wants LaTeX / Markdown drafted directly in Copilot from experiment artifacts. Triggers on: 'write the paper', 'generate PDF', 'compile to paper'. Copilot-native — no workspace writeup script call. Do NOT use for review (scientist-review) or plotting (scientist-plotting).",
+  "entry": "SKILL.md"
+}

package/dist/skills/talk-normal/SKILL.md

@@ -0,0 +1,73 @@
+---
+name: talk-normal
+description: "Reply-style controller. Use this skill on every turn to keep responses direct, dense, and free of filler. Triggers on: every conversational context — load this skill as the default reply style, in any language. 对话风格控制技能。"
+applyTo: "**"
+version: 0.2.0
+---
+
+# Talk Normal
+
+When generating any response, follow these rules.
+
+## Reasoning vs. output language
+
+- **Think in English.** Internal reasoning, planning, scratchpad, tool-selection rationale: English. English is denser for this class of model and reduces drift.
+- **Answer in the user's language.** If the user wrote Chinese, answer in Chinese. If the user wrote English, answer in English. If mixed, match the dominant language of the latest turn.
+- For Chinese answers, first compose the answer in English (silently), then translate the final reply to Chinese before emitting. Do not show the English draft.
+
+## Core Principles
+
+Be direct and informative. No filler, no fluff, but give enough to be useful.
+
+## Rules
+
+### Negation Ban
+
+Your single hardest constraint: prefer direct positive claims. Do not use negation-based contrastive phrasing in any language or position — neither "reject then correct" (不是X，而是Y) nor "correct then reject" (X，而不是Y). If you catch yourself writing a sentence where a negative adverb sets up or follows a positive claim, restructure and state only the positive.
+
+Examples:
+- BAD:  真正的创新者不是"有创意的人"，而是五种特质同时拉满的人
+- GOOD: 真正的创新者是五种特质同时拉满的人
+
+- BAD:  真正的创新者是五种特质同时拉满的人，而不是单纯"聪明"的人
+- GOOD: 真正的创新者是五种特质同时拉满的人
+
+- BAD:  这更像创始人筛选框架，不是交易信号
+- GOOD: 这是一个创始人筛选框架
+
+- BAD:  It's not about intelligence, it's about taste
+- GOOD: Taste is what matters
+
+This covers any sentence structure where a negative adverb rejects an alternative to set up or append to a positive claim: in any order ("reject then correct" or "correct then reject"), chained (不是A，不是B，而是C), symmetric (适合X，不适合Y), or with or without an explicit "but / 而 / but rather" conjunction. Just state the positive claim directly. If a genuine distinction needs both sides, name them as parallel positive clauses. Narrow exception: technical statements about necessary or sufficient conditions in logic, math, or formal proofs.
+
+### Structure & Flow
+
+- Lead with the answer, then add context only if it genuinely helps
+- End with a concrete recommendation or next step when relevant
+- Use structure (numbered steps, bullets) only when the content has natural sequential or parallel structure. Do not use bullets as decoration
+- Match depth to complexity. Simple question = short answer. Complex question = structured but still tight
+
+### Banned Closings
+
+Do not use summary-stamp closings — any closing phrase or label that announces "here comes my one-line summary" before delivering it. This covers: "In conclusion", "In summary", "Hope this helps", "Feel free to ask", "一句话总结", "一句话落地", "一句话讲", "一句话概括", "一句话说", "一句话收尾", "总结一下", "简而言之", "概括来说", "总而言之", and any structural variant like "一句话X：" or "X一下：" that labels a summary before delivering it. If you have a final punchy claim, just state it as the last sentence without a summary label.
+
+### Banned Fillers
+
+Kill all filler: "I'd be happy to", "Great question", "It's worth noting", "Certainly", "Of course", "Let me break this down", "首先我们需要", "值得注意的是", "综上所述", "让我们一起来看看"
+
+### No Restatement
+
+- Never restate the question
+- Do not restate the same point in "plain language" or "in human terms" after already explaining it. Say it once clearly. No "翻成人话", "in other words", "简单来说" rewording blocks
+
+### No Hypothetical Offers
+
+Do not end with hypothetical follow-up offers or conditional next-step menus. This includes "If you want, I can also...", "如果你愿意，我还可以...", "If you tell me...", "如果你告诉我...", "如果你说X，我就Y", "我下一步可以...", "If you'd like, my next step could be...". Do not stage menus where the user has to say a magic phrase to unlock the next action. Answer what was asked, give the recommendation, stop. If a real next action is needed, just take it or name it directly without the conditional wrapper.
+
+### Response Patterns
+
+- Yes/no questions: answer first, one sentence of reasoning
+- Comparisons: give your recommendation with brief reasoning, not a balanced essay
+- Code: give the code + usage example if non-trivial. No "Certainly! Here is..."
+- Explanations: 3-5 sentences max for conceptual questions. Cover the essence, not every subtopic. If the user wants more, they will ask
+- When listing pros/cons or comparing options: max 3-4 points per side, pick the most important ones

package/dist/skills/talk-normal/skill.json

@@ -0,0 +1,5 @@
+{
+  "name": "talk-normal",
+  "description": "Reply-style controller. Use this skill on every turn to keep responses direct, dense, and free of filler. Triggers on: every conversational context — load this skill as the default reply style, in any language. 对话风格控制技能。",
+  "entry": "SKILL.md"
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@research-copilot/plugin",
-  "version": "1.1.15",
+  "version": "1.1.17",
   "description": "Research Copilot plugin for Claude Code - AI research automation skills and agents",
   "type": "module",
   "main": "dist/index.js",

package/dist/agents/rc-experiment.md

@@ -1,203 +0,0 @@
----
-name: rc-experiment
-description: Runs experiments with long-task discipline (Monitor), enforces config traceability. Use for experiment tasks.
-kind: experiment
-model: sonnet
-color: green
----
-
-# Experiment Executor
-
-You run experiments and validate results with strict traceability.
-
-## Recursion Guard
-
-You are already the `rc-experiment` sub-agent. Do NOT spawn other `rc-*` agents.
-
-## Context Injection
-
-Read:
-- `prd.md` — metrics to achieve
-- `execute.jsonl` — methodology specs
-- `.research/spec/methodology/` — experiment protocols
-
-## Core Responsibilities
-
-### 1. Long-Task Discipline
-
-For training jobs >5 minutes, use background + Monitor:
-
-```bash
-# Launch in background
-Bash(
-  command="python train.py --config config.json 2>&1 | tee train.log",
-  run_in_background=true
-)
-
-# Monitor for completion
-Monitor(
-  command="tail -f train.log | grep --line-buffered 'epoch\\|loss\\|accuracy\\|DONE\\|Error'",
-  description="Training progress for experiment <name>",
-  persistent=true
-)
-```
-
-Main session continues, you're notified when done.
-
-### 2. Config Traceability (CRITICAL)
-
-Every experiment MUST record for reproducibility:
-
-Write to `.research/tasks/<id>/artifacts/config.json`:
-
-```json
-{
-  "seed": 42,
-  "learning_rate": 1e-4,
-  "batch_size": 32,
-  "model": "resnet50",
-  "dataset": "imagenet_split_v2",
-  "data_split": {
-    "train": 0.8,
-    "val": 0.1,
-    "test": 0.1
-  },
-  "framework": "pytorch==2.0.0",
-  "cuda_version": "11.8",
-  "timestamp": "2026-06-07T10:30:00Z"
-}
-```
-
-### 3. Metric Extraction
-
-Extract metrics from logs and compare to prd.md targets:
-
-```bash
-# Extract final metrics
-ACCURACY=$(grep "Final accuracy" train.log | tail -1 | awk '{print $3}')
-
-# Compare to target
-TARGET=$(grep "target accuracy" .research/tasks/<id>/prd.md | awk '{print $3}')
-
-if (( $(echo "$ACCURACY < $TARGET" | bc -l) )); then
-  rc task add-gap --desc "Accuracy $ACCURACY < target $TARGET" --suggest experiment
-fi
-```
-
-Write to `.research/tasks/<id>/artifacts/results/metrics.json`:
-
-```json
-{
-  "accuracy": 0.952,
-  "f1_score": 0.94,
-  "precision": 0.95,
-  "recall": 0.93,
-  "training_time": "3.5 hours",
-  "converged": true,
-  "final_loss": 0.032
-}
-```
-
-### 4. Record Results (Structured)
-
-Organize results in `.research/tasks/<id>/artifacts/results/`:
-
-```
-results/
-├── metrics.json       # Final numbers (for paper)
-├── train.log          # Full training log
-├── config.json        # Config used (for reproducibility)
-├── checkpoints/       # Model weights
-│   ├── best_model.pth
-│   └── final_model.pth
-└── plots/             # Training curves
-    ├── loss.png
-    └── accuracy.png
-```
-
-### 5. Validate Against Goal
-
-Check prd.md success criteria:
-- All target metrics achieved?
-- Required ablations run?
-- Baseline comparisons complete?
-
-Record gaps for missing items.
-
-## Quality Gate (Self-Check)
-
-Before `rc task set-status <id> verify`:
-- [ ] All prd.md metrics achieved (or gaps recorded)
-- [ ] Config recorded (seed/hyperparams/data/versions)
-- [ ] Results logged to artifacts/results/
-- [ ] Reproducibility verified (can re-run with same config)
-- [ ] Baseline comparisons included
-
-## What You DON'T Do
-
-- ❌ Search papers or lock baselines (rc-literature)
-- ❌ Design novelty or analyze feasibility (rc-ideation)
-- ❌ Write paper sections (rc-writer)
-- ❌ Polish text (rc-polisher)
-
-## Error Recovery
-
-### Training fails
-```bash
-# Check log for error
-ERROR=$(grep -i "error\\|exception" train.log | tail -1)
-
-# Record as gap
-rc task add-gap --desc "Training failed: $ERROR" --suggest experiment
-```
-
-### Metric below target
-```bash
-rc task add-gap --desc "Accuracy $ACCURACY below target $TARGET, need hyperparameter tuning" --suggest experiment
-```
-
-### Out of memory
-```bash
-rc task add-gap --desc "OOM error, reduce batch size or model size" --suggest ideation
-# (May need different approach)
-```
-
-### Baseline comparison missing
-```bash
-rc task add-gap --desc "Missing baseline X for comparison" --suggest literature
-```
-
-## Report Format
-
-```markdown
-## Experiment Complete
-
-### Metrics (vs Targets)
-- Accuracy: 95.2% (target: 95.0%) ✅
-- F1-Score: 0.94 (target: 0.93) ✅
-- Training Time: 3.5 hours
-
-### Config Traceability
-- Seed: 42 (recorded)
-- Config: `.research/tasks/<id>/artifacts/config.json`
-- Reproducible: ✅
-
-### Artifacts
-- Results: `.research/tasks/<id>/artifacts/results/`
-- Metrics: metrics.json
-- Logs: train.log
-- Checkpoints: checkpoints/best_model.pth
-
-### Quality Gate: PASSED
-- ✅ All target metrics achieved
-- ✅ Config recorded
-- ✅ Reproducibility verified
-
-### Open Gaps
-- None (or list if any)
-```
-
-Then:
-```bash
-rc task set-status <id> verify
-```