research-copilot 0.2.20 → 0.2.21

package/lib/skills/builtin/scientific-schematics/SKILL.md

@@ -1,603 +1,506 @@
 ---
 name: scientific-schematics
-description: Create publication-quality scientific diagrams using OpenRouter API with smart iterative refinement. The skill defaults to Gemini 3 Pro Image for generation plus Gemini 3 Pro for review, and only regenerates if quality is below threshold for your document type.
+description: Prompt guidance for the generate_diagram tool — how to describe different scientific diagram types (flowcharts, pathways, architecture, circuits, networks, conceptual) so the tool produces publication-grade output on the first iteration.
 category: Visualization
 tags: [science, visualization, diagrams]
-triggers: [diagram, schematic, flowchart, architecture diagram, system diagram, 示意图, 流程图, generate diagram]
-allowed-tools: [Read, Write, Edit, Bash]
+triggers: [diagram, schematic, flowchart, architecture diagram, pathway, circuit, 示意图, 流程图]
+allowed-tools: [Read, Write, Edit]
 license: MIT license
 metadata:
-    skill-author: K-Dense Inc.
+    skill-author: Dong Dai
 ---
 
-# Scientific Schematics and Diagrams
+# Scientific Schematics
 
-## Overview
+## Scope of this skill
 
-Scientific schematics and diagrams transform complex concepts into clear visual representations for publication. We want the diagram a textbook style. **This skill uses OpenRouter API for image generation with Gemini quality review.**
+This skill **does not generate images**. Image generation is owned by the
+`generate_diagram` tool, which runs a provider-backed generate → review →
+(optional) edit loop. The tool is always available; this skill is only
+loaded to sharpen the prompt the tool receives.
 
-**How it works:**
-- Describe your diagram in natural language
-- Gemini image model generates publication-quality diagrams via OpenRouter
-- **Gemini reviews quality** against document-type thresholds
-- **Smart iteration**: Only regenerates if quality is below threshold
-- Publication-ready output in minutes
-- No coding, templates, or manual drawing required
+Use this skill to:
+1. Pick the right `diagram_type` parameter for the request.
+2. Write a description that names components, labels, quantities, and
+   flow direction unambiguously (LLM image models hallucinate when these
+   are vague).
+3. Choose an appropriate `doc_type` so the reviewer applies the right
+   quality threshold.
+4. Embed the resulting image in the document correctly.
 
-**Quality Thresholds by Document Type:**
-| Document Type | Threshold | Description |
-|---------------|-----------|-------------|
-| journal | 8.5/10 | Nature, Science, peer-reviewed journals |
-| conference | 8.0/10 | Conference papers |
-| thesis | 8.0/10 | Dissertations, theses |
-| grant | 8.0/10 | Grant proposals |
-| preprint | 7.5/10 | arXiv, bioRxiv, etc. |
-| report | 7.5/10 | Technical reports |
-| poster | 7.0/10 | Academic posters |
-| presentation | 6.5/10 | Slides, talks |
-| default | 7.5/10 | General purpose |
-
-**Simply describe what you want, and AI generates it.** All diagrams are stored in the @ws/figures/ subfolder and referenced in papers/posters.
-
-### Embedding Figures in Documents (MANDATORY)
-
-**After generating every figure, you MUST embed it in the document using markdown image syntax.**
-
-The image path must be **relative to the markdown file you are writing**, not the workspace root. This ensures the renderer resolves the path correctly regardless of where the markdown file lives.
-
-**Example — markdown at workspace root (`paper.md`):**
-```bash
-python @skill/scripts/generate_schematic.py "CONSORT flowchart..." -o @ws/figures/consort.png
-```
-Embed in `paper.md`:
-```markdown
-![Figure 2: CONSORT participant flow diagram](figures/consort.png)
-```
-
-**Example — markdown in a subdirectory (`workspace/paper_draft.md`):**
-```bash
-python @skill/scripts/generate_schematic.py "CONSORT flowchart..." -o @ws/figures/consort.png
-```
-Embed in `workspace/paper_draft.md` — use `../` to go up:
-```markdown
-![Figure 2: CONSORT participant flow diagram](../figures/consort.png)
-```
-
-**Rule of thumb:** count how many directories deep your markdown file is from the workspace root, and prepend that many `../` segments to `figures/filename.png`.
-
-Alternatively, you can generate the image alongside the markdown file:
-```bash
-# If markdown is at workspace/paper_draft.md, save figures next to it:
-python @skill/scripts/generate_schematic.py "CONSORT flowchart..." -o @ws/workspace/figures/consort.png
-```
-Then embed simply as:
-```markdown
-![Figure 2: CONSORT participant flow diagram](figures/consort.png)
-```
-
-Do NOT only cite figures as text ("see Figure 1"). Always include the `![...](...)` embed so the figure renders visually in the output.
-
-**Default model path:** this skill intentionally uses one default generation model, `google/gemini-3-pro-image-preview`, with Gemini review via `google/gemini-3-pro-preview`. Agents should not pick among multiple image backends during normal operation.
-
-### Path Conventions
-
-- Shell command examples use scoped paths:
-  - `@skill/...` for files shipped with this skill (for example scripts)
-  - `@ws/...` for workspace files (inputs/outputs)
-- Python API and JSON log examples show raw filesystem paths (typically workspace-relative, such as `figures/...`).
-
-## Quick Start: Generate Any Diagram
-
-Create any scientific diagram by simply describing it. AI handles everything automatically with **smart iteration**:
-
-```bash
-# Generate for journal paper (highest quality threshold: 8.5/10)
-python @skill/scripts/generate_schematic.py "CONSORT participant flow diagram with 500 screened, 150 excluded, 350 randomized" -o @ws/figures/consort.png --doc-type journal
-
-# Generate for presentation (lower threshold: 6.5/10 - faster)
-python @skill/scripts/generate_schematic.py "Transformer encoder-decoder architecture showing multi-head attention" -o @ws/figures/transformer.png --doc-type presentation
-
-# Generate for poster (moderate threshold: 7.0/10)
-python @skill/scripts/generate_schematic.py "MAPK signaling pathway from EGFR to gene transcription" -o @ws/figures/mapk_pathway.png --doc-type poster
-
-# Custom max iterations (max 2)
-python @skill/scripts/generate_schematic.py "Complex circuit diagram with op-amp, resistors, and capacitors" -o @ws/figures/circuit.png --iterations 2 --doc-type journal
-```
-
-**What happens behind the scenes:**
-1. **Generation 1**: Gemini image model creates an initial image following scientific diagram best practices
-2. **Review 1**: **Gemini** evaluates quality against the document-type threshold
-3. **Decision**: If quality >= threshold → **DONE** (no more iterations needed!)
-4. **If below threshold**: Improved prompt based on critique, regenerate
-5. **Repeat**: Until quality meets threshold OR max iterations reached
-
-**Smart Iteration Benefits:**
-- Saves API calls if first generation is good enough
-- Higher quality standards for journal papers
-- Faster turnaround for presentations/posters
-- Appropriate quality for each use case
-
-**Output**: Versioned images plus a detailed review log with quality scores, critiques, and early-stop information.
-
-### Configuration
-
-Set your OpenRouter API key:
-```bash
-export OPENROUTER_API_KEY='your-openrouter-api-key'
-```
-
-Optional model overrides:
-```bash
-export SCHEMATIC_IMAGE_MODEL='google/gemini-3-pro-image-preview'   # default
-export SCHEMATIC_REVIEW_MODEL='google/gemini-3-pro-preview'         # default
-```
-
-### AI Generation Best Practices
-
-**Effective Prompts for Scientific Diagrams:**
-
-✓ **Good prompts** (specific, detailed):
-- "CONSORT flowchart showing participant flow from screening (n=500) through randomization to final analysis"
-- "Transformer neural network architecture with encoder stack on left, decoder stack on right, showing multi-head attention and cross-attention connections"
-- "Biological signaling cascade: EGFR receptor → RAS → RAF → MEK → ERK → nucleus, with phosphorylation steps labeled"
-- "Block diagram of IoT system: sensors → microcontroller → WiFi module → cloud server → mobile app"
-
-✗ **Avoid vague prompts**:
-- "Make a flowchart" (too generic)
-- "Neural network" (which type? what components?)
-- "Pathway diagram" (which pathway? what molecules?)
-
-**Key elements to include:**
-- **Type**: Flowchart, architecture diagram, pathway, circuit, etc.
-- **Components**: Specific elements to include
-- **Flow/Direction**: How elements connect (left-to-right, top-to-bottom)
-- **Labels**: Key annotations or text to include
-- **Style**: A textbook style plus Any specific visual requirements
-
-**Scientific Quality Guidelines** (automatically applied):
-- Clean white/light background
-- High contrast for readability
-- Clear, readable labels (minimum 10pt)
-- Professional typography (sans-serif fonts)
-- Colorblind-friendly colors (Okabe-Ito palette)
-- Proper spacing to prevent crowding
-- Scale bars, legends, axes where appropriate
-
-## When to Use This Skill
-
-This skill should be used when:
-- Creating neural network architecture diagrams (Transformers, CNNs, RNNs, etc.)
-- Illustrating system architectures and data flow diagrams
-- Drawing methodology flowcharts for study design (CONSORT, PRISMA)
-- Visualizing algorithm workflows and processing pipelines
-- Creating circuit diagrams and electrical schematics
-- Depicting biological pathways and molecular interactions
-- Generating network topologies and hierarchical structures
-- Illustrating conceptual frameworks and theoretical models
-- Designing block diagrams for technical papers
-
-## How to Use This Skill
-
-**Simply describe your diagram in natural language.** AI generates it automatically:
-
-```bash
-python @skill/scripts/generate_schematic.py "your diagram description" -o @ws/output.png
-```
-
-**That's it!** The AI handles:
-- Layout and composition
-- Labels and annotations
-- Colors and styling
-- Quality review and refinement
-- Publication-ready output
-
-**Works for all diagram types:**
-- Flowcharts (CONSORT, PRISMA, etc.)
-- Neural network architectures
-- Biological pathways
-- Circuit diagrams
-- System architectures
-- Block diagrams
-- Any scientific visualization
-
-**No coding, no templates, no manual drawing required.**
+Do not write custom Python, shell, or SVG — the tool produces the image.
 
 ---
 
-# AI Generation Mode (OpenRouter + Gemini Review)
-
-## Smart Iterative Refinement Workflow
-
-The AI generation system uses **smart iteration** - it only regenerates if quality is below the threshold for your document type:
-
-### How Smart Iteration Works
-
-```
-┌─────────────────────────────────────────────────────┐
-│  1. Generate image with Gemini via OpenRouter        │
-│                    ↓                                │
-│  2. Review quality with Gemini                      │
-│                    ↓                                │
-│  3. Score >= threshold?                             │
-│       YES → DONE! (early stop)                      │
-│       NO  → Improve prompt, go to step 1            │
-│                    ↓                                │
-│  4. Repeat until quality met OR max iterations      │
-└─────────────────────────────────────────────────────┘
-```
-
-### Iteration 1: Initial Generation
-**Prompt Construction:**
-```
-Scientific diagram guidelines + User request
-```
+## Tool usage
 
-**Output:** `diagram_v1.png`
-
-### Quality Review by Gemini
-
-Gemini evaluates the diagram on:
-1. **Scientific Accuracy** (0-2 points) - Correct concepts, notation, relationships
-2. **Clarity and Readability** (0-2 points) - Easy to understand, clear hierarchy
-3. **Label Quality** (0-2 points) - Complete, readable, consistent labels
-4. **Layout and Composition** (0-2 points) - Logical flow, balanced, no overlaps
-5. **Professional Appearance** (0-2 points) - Publication-ready quality
-
-**Example Review Output:**
-```
-SCORE: 8.0
-
-STRENGTHS:
-- Clear flow from top to bottom
-- All phases properly labeled
-- Professional typography
-
-ISSUES:
-- Participant counts slightly small
-- Minor overlap on exclusion box
-
-VERDICT: ACCEPTABLE (for poster, threshold 7.0)
-```
-
-### Decision Point: Continue or Stop?
-
-| If Score... | Action |
-|-------------|--------|
-| >= threshold | **STOP** - Quality is good enough for this document type |
-| < threshold | Continue to next iteration with improved prompt |
-
-**Example:**
-- For a **poster** (threshold 7.0): Score of 7.5 → **DONE after 1 iteration!**
-- For a **journal** (threshold 8.5): Score of 7.5 → Continue improving
-
-### Subsequent Iterations (Only If Needed)
-
-If quality is below threshold, the system:
-1. Extracts specific issues from Gemini's review
-2. Enhances the prompt with improvement instructions
-3. Regenerates via OpenRouter
-4. Reviews again with Gemini
-5. Repeats until threshold met or max iterations reached
-
-### Review Log
-All iterations are saved with a JSON review log that includes early-stop information:
-```json
-{
-  "user_prompt": "CONSORT participant flow diagram...",
-  "doc_type": "poster",
-  "quality_threshold": 7.0,
-  "iterations": [
-    {
-      "iteration": 1,
-      "image_path": "figures/consort_v1.png",
-      "score": 7.5,
-      "needs_improvement": false,
-      "critique": "SCORE: 7.5\nSTRENGTHS:..."
-    }
-  ],
-  "final_score": 7.5,
-  "early_stop": true,
-  "early_stop_reason": "Quality score 7.5 meets threshold 7.0 for poster"
-}
-```
-
-**Note:** With smart iteration, you may see only 1 iteration instead of the full 2 if quality is achieved early!
-
-## Advanced AI Generation Usage
-
-### Supported Automation Interface
-
-Use the CLI wrapper as the stable interface:
-
-```bash
-python @skill/scripts/generate_schematic.py \
-  "Transformer architecture diagram" \
-  -o @ws/figures/transformer.png \
-  --iterations 2
-```
+Call the `generate_diagram` tool with:
+
+```
+prompt:         <description crafted using the guidance below>
+output:         figures/<name>.png
+doc_type:       journal | conference | thesis | grant | preprint |
+                report | poster | presentation | default
+diagram_type:   flowchart | architecture | pathway | circuit |
+                network | conceptual | auto
+iterations:     1 | 2 | 3   (default 2)
+aspect:         auto | square | landscape | portrait   (default auto)
+quality:        low | medium | high | auto   (default derived from doc_type)
+format:         auto | png | svg   (default auto — inferred from output extension)
+```
+
+**aspect guidance** — the default `auto` lets the model pick, but
+when you already know the figure's shape, set it explicitly:
+- `landscape` for wide architecture diagrams, left-to-right pipelines,
+  multi-panel (3+ columns) layouts
+- `portrait` for CONSORT/PRISMA flows, top-to-bottom pathway cascades,
+  tall hierarchies
+- `square` for single-concept schematics, cycles, small callouts
+- `auto` when the shape is genuinely ambiguous or the prompt already
+  describes the layout strongly
+
+**quality guidance** — gpt-image-2 exposes four tiers; the cost and
+render time roughly follow low < medium < high. Omitting the field
+selects a sensible default from `doc_type`:
+
+| doc_type                             | default quality |
+|--------------------------------------|-----------------|
+| journal / conference / thesis / grant| high            |
+| preprint / report / poster           | medium          |
+| presentation                         | low             |
+| default                              | medium          |
+
+The loop automatically bumps one tier on a `needs_edit` verdict, so the
+first iteration runs cheap and refinement only spends more compute when
+the reviewer signals the draft was close-but-not-there. Override with
+`quality: "low"` for explicit exploration / drafts, or `"high"` to force
+camera-ready on the first pass.
+
+The tool returns the final image path, per-iteration review scores,
+the quality tier used for each iteration, the verdict trail
+(acceptable / needs_edit / needs_regen), and a JSON review log at
+`<name>_review_log.json`.
+
+When `diagram_type: auto`, the tool infers from keywords in the prompt.
+Explicit types beat inference when you know what you want.
+
+### Choosing PNG vs SVG output
+
+The `format` parameter (or the `output` extension) selects the output
+shape. The two formats have different quality envelopes and use cases:
+
+**PNG (default)** — gpt-image-2 native raster, drives the verdict-driven
+review loop directly. Best raw visual quality. Use when:
+- The figure is final and won't be edited further
+- Embedding in slides, posters, web, or as a paper raster figure
+- The user said "image", "PNG", "图片"
+
+**SVG** — same gpt-image-2 PNG verdict loop, then a vision-capable chat
+model transcribes the finalized PNG into editable SVG markup. The .png
+anchor is preserved as a sibling file (`<name>.png` next to `<name>.svg`)
+so the user can diff visual drift later. Use when:
+- The user wants to **edit labels / colors** in Inkscape, draw.io, or
+  by hand
+- Embedding in LaTeX as `\includegraphics{*.svg}` or in HTML/Markdown
+  for crisp scaling
+- The user said "SVG", "vector", "矢量图", "向量图"
+
+Quality tiers depending on the user's configuration:
+
+| Configuration                                | SVG path active           | Quality |
+|----------------------------------------------|---------------------------|---------|
+| `OPENAI_API_KEY` + vision-capable chat model | PNG-anchored transcription | High (~7+/10, matches PNG) |
+| `OPENAI_API_KEY` + non-vision chat model     | Tool returns `SVG_REQUIRES_VISION_MODEL` error | — (recover by switching model or using PNG) |
+| No `OPENAI_API_KEY`                          | Chat-model-only synthesis  | Limited (~6/10) |
+
+All catalog chat models in this app currently support vision input, so
+the second row is rare in practice — Path B exists as a safety net for
+future non-vision models.
+
+### Internal prompt structure
+
+Under the hood the tool converts your `prompt` into a fixed-slot
+production brief that gpt-image-2 parses reliably:
+
+```
+【SCENE / USE】  publication-grade scientific <diagram_type>
+【SUBJECT】      <your prompt verbatim>
+【COMPOSITION】  <derived from diagram_type and aspect>
+【KEY DETAILS】  <type-specific rules + typography + colour>
+【TEXT】          render every quoted literal verbatim
+【MUST KEEP】    <auto-extracted: quoted strings, n=X, numeric+unit>
+【AVOID】         no figure numbers, titles, captions, 3D, mascots
+```
+
+Two practical consequences for **how you write the prompt**:
+
+1. **Put every label, number, and identifier in quotes.** The tool
+   auto-extracts quoted strings (and `n=...`, `1kΩ` / `10µF` / `128
+   nodes` style numeric+unit tokens) into a `MUST KEEP` checklist the
+   model must render verbatim. `"EGFR"` survives better than just
+   EGFR; `"n=350"` survives better than just n=350.
+2. **Describe in fixed-slot order when you can** — scene/use → subject
+   → key details → composition → text → must-keep → avoid. The tool
+   will structure whatever you write, but pre-structured prose rewrites
+   less aggressively.
+
+### Reference images (iterative refinement and style boards)
+
+Two modes of `reference_mode` are supported; a third is reserved.
+
+**`reference_mode: "revise_layout"`** (default) — the reference is the
+existing draft to polish. The tool treats iteration 1 as a surgical
+edit: gpt-image-2 keeps layout, colour, and positions unchanged unless
+specific blocking issues require otherwise. Use this when the user
+says "improve this figure" or "fix X in this figure".
+
+**`reference_mode: "style_only"`** — the reference is a **style board**,
+not a layout. The tool redraws the subject from scratch in the
+reference's visual idiom (palette, typography, line weights, geometry)
+while ignoring its layout and content. Use this when:
+- The user points at an older figure and says "use this style" but
+  wants a completely different diagram.
+- A lab or institution has a house-style exemplar you want the tool
+  to lift from without copying content.
+- Reference and subject are semantically different (e.g., reference is
+  a CONSORT flowchart, subject is an architecture diagram) but you want
+  them to look like siblings.
+
+`style_only` works with both raster and SVG outputs. In raster mode the
+reference is passed through `/v1/images/edits` with an explicit
+"style only, do not copy layout" instruction. In SVG mode the
+reference SVG source is embedded in the prompt as a style exemplar.
+
+**`reference_mode: "local_edit"`** — reserved; masked-region edit is
+not yet implemented and the tool rejects it with a clear error.
 
-Do not rely on importing private skill modules directly from Python; this skill ships as
-scripts and references, not as an installable Python package.
+---
 
-### Command-Line Options
+## Diagram types and how to describe them
+
+### flowchart
+Use for CONSORT, PRISMA, study-design flow, decision trees, swimlanes.
+
+Required in prompt:
+- Every node's label (exact text)
+- Every arrow's source and target
+- Exact counts (n = …), conditions, yes/no labels
+- Flow direction (top-to-bottom default)
+
+Example:
+> CONSORT participant flow, vertical top-to-bottom. Boxes:
+> "Assessed for eligibility (n=500)" → split into
+> "Excluded (n=150): age<18 n=80, declined n=50, other n=20"
+> (right branch) and "Randomized (n=350)" (down). "Randomized" splits to
+> "Treatment (n=175)" and "Control (n=175)". Each arm shows
+> "Lost to follow-up (n=15 / n=10)" then "Analyzed (n=160 / n=165)".
+> Colour: blue for process, orange for exclusion, green for analysis.
+
+### architecture
+Use for system diagrams, microservices, data pipelines, block diagrams.
+
+Required in prompt:
+- Every block's label and role
+- Layering or grouping (what belongs together)
+- Connection direction + protocol/interface labels on edges
+- Any shared boundaries (VPC, cluster, device)
+
+Example:
+> IoT monitoring architecture, three layers stacked vertically.
+> Bottom layer (sensors): "Temperature", "Humidity", "Motion" in green
+> boxes. Middle layer: "ESP32 microcontroller" in blue; connects up via
+> "WiFi" and sideways to "Local display". Top layer: "Cloud server" in
+> grey, connected to "Mobile app" in light blue. Edges labelled with
+> protocols (I2C, UART, WiFi, HTTPS).
+
+### pathway
+Use for signalling cascades, metabolic pathways, gene regulation,
+protein-protein interaction maps.
+
+Required in prompt:
+- Every molecule by its symbol
+- Arrow type: activation (→) vs inhibition (⊣)
+- Exact order and compartments (cytoplasm, nucleus, membrane)
+- Post-translational modifications if relevant
+
+Example:
+> MAPK signalling pathway, top-to-bottom. Cell membrane at top with
+> "EGFR receptor" embedded. Below: "RAS-GTP" (oval) → "RAF" → "MEK" →
+> "ERK" (all kinases, rectangles). "ERK" arrow crosses nuclear envelope
+> (dashed line) to "Transcription factors" in the nucleus. Label each
+> arrow "phosphorylation". Okabe-Ito palette.
+
+### circuit
+Use for analogue/digital circuits, signal chains, power systems.
+
+Required in prompt:
+- Every component type with value + unit (1kΩ, 10µF, 5V)
+- Ground and supply nets
+- Wire crossings: dot = connection, no dot = jump
+
+Example:
+> Non-inverting op-amp amplifier. Input signal on left connects via
+> 1kΩ resistor (R1) to the + input of an op-amp. Feedback: output
+> through 10kΩ (R2) to − input, and 1kΩ (R3) from − input to ground.
+> Supply rails ±12V. Output on the right. Standard IEEE symbols.
+
+### network
+Use for neural network architectures, graphs, trees, org charts.
+
+Required in prompt:
+- Node types and dimensions where applicable (Dense 128, Conv 3×3)
+- Edge semantics (data flow vs attention vs skip)
+- Hierarchy direction
+
+Example:
+> Transformer encoder-decoder, two stacks side by side. Encoder (left,
+> light blue): "Input embedding" → "Positional encoding" → 6× {Multi-head
+> self-attention, Add & Norm, Feed-forward, Add & Norm}. Decoder (right,
+> light red): "Output embedding" → "Positional encoding" → 6× {Masked
+> self-attention, Add & Norm, Cross-attention (dashed edge from encoder
+> top to decoder cross-attention), Add & Norm, Feed-forward, Add & Norm}.
+> → Linear → Softmax. Label every block.
+
+### conceptual
+Use for frameworks, theoretical models, idea maps, whiteboard-style
+diagrams. Looser rules, more artistic licence.
+
+Required in prompt:
+- Core concepts grouped by category
+- Relationships between categories
+- Any visual metaphor you want (layers, concentric rings, Venn)
 
-```bash
-# Basic usage (default threshold 7.5/10)
-python @skill/scripts/generate_schematic.py "diagram description" -o @ws/output.png
+---
 
-# Specify document type for appropriate quality threshold
-python @skill/scripts/generate_schematic.py "diagram" -o @ws/out.png --doc-type journal      # 8.5/10
-python @skill/scripts/generate_schematic.py "diagram" -o @ws/out.png --doc-type conference   # 8.0/10
-python @skill/scripts/generate_schematic.py "diagram" -o @ws/out.png --doc-type poster       # 7.0/10
-python @skill/scripts/generate_schematic.py "diagram" -o @ws/out.png --doc-type presentation # 6.5/10
+## doc_type thresholds
 
-# Custom max iterations (1-2)
-python @skill/scripts/generate_schematic.py "complex diagram" -o @ws/diagram.png --iterations 2
+| doc_type      | Behaviour |
+|---------------|-----------|
+| journal       | Strictest threshold; accept only near-camera-ready output. |
+| conference    | Strict; small cosmetic issues allowed. |
+| thesis, grant | Same bar as conference. |
+| preprint, report | Standard threshold. |
+| poster        | Tolerates larger labels and bolder colours. |
+| presentation  | Lowest threshold; optimised for first-pass speed. |
+| default       | Middle of the range. |
 
-# Verbose output (see all API calls and reviews)
-python @skill/scripts/generate_schematic.py "flowchart" -o @ws/flow.png -v
+Exact numeric thresholds depend on the review provider and are not
+directly comparable across providers (see the tool's review log for the
+applied threshold on each run).
 
-# Combine options
-python @skill/scripts/generate_schematic.py "neural network" -o @ws/nn.png --doc-type journal --iterations 2 -v
-```
+---
 
-### Prompt Engineering Tips
+## Iteration strategy
+
+Default is **2 iterations** for almost every case. Do not bump to 3
+without concrete evidence — the third iteration frequently *regresses*
+a fix from the second, and the tool now detects this and stops early
+anyway (see "Regression detection" below).
+
+- `iterations: 1` — only when you want the first-pass output without
+  any review-driven rework.
+- `iterations: 2` (default) — one draft plus one review-driven revision.
+  The second pass uses image-to-image editing when the reviewer judges
+  issues are cosmetic (needs_edit). When the reviewer flags structural
+  problems (needs_regen) the second pass redraws from scratch.
+- `iterations: 3` — only when iter 2 is expected to still be below
+  threshold. Good signals: `doc_type: journal | conference` AND the
+  prompt asks for something dense (many labels, multi-panel, intricate
+  relationships). Bad signals: `doc_type: presentation | poster`,
+  single-concept diagrams, simple flowcharts.
+
+The tool stops early the moment a review comes back with verdict
+`acceptable`, so higher `iterations` costs nothing when not needed —
+but each *unnecessary* iteration risks the model undoing earlier
+corrections.
+
+### Relaxed acceptance
+
+When the reviewer returns `needs_edit` but the score is comfortably
+above threshold (**≥ threshold + 1.0**) and no critical issues remain
+(`wrong_content`, `missing_element`, `illegible_text`), the tool
+promotes the verdict to `acceptable` and stops. Cosmetic notes
+(`layout_collision`, `style_mismatch`) do not force another round
+when the diagram is already well past the bar.
+
+### Regression detection
+
+The tool tracks which blocking issues the reviewer stops complaining
+about between iterations — those are the corrections we want to keep.
+Each subsequent edit prompt explicitly reminds the model of the
+already-fixed items and forbids regressing them. If the reviewer in a
+later draft nevertheless re-introduces an issue from that fixed set,
+the loop terminates early with `stoppedReason: "regression_detected"`
+and the current best draft is returned. Two common causes of
+regression:
+
+- Too many iterations on a diagram that is already good enough (ships
+  a new cosmetic flaw while fixing an old cosmetic flaw).
+- Prompt overspecification — the original request listed dozens of
+  tiny details, and the model keeps dropping a different one each
+  round.
+
+When you see `regressedIssues` or `stoppedReason: "regression_detected"`
+in the review log, prefer the current output over re-running with
+more iterations; re-running usually makes it worse.
 
-**1. Be Specific About Layout:**
-```
-✓ "Flowchart with vertical flow, top to bottom"
-✓ "Architecture diagram with encoder on left, decoder on right"
-✓ "Circular pathway diagram with clockwise flow"
-```
-
-**2. Include Quantitative Details:**
-```
-✓ "Neural network with input layer (784 nodes), hidden layer (128 nodes), output (10 nodes)"
-✓ "Flowchart showing n=500 screened, n=150 excluded, n=350 randomized"
-✓ "Circuit with 1kΩ resistor, 10µF capacitor, 5V source"
-```
+---
 
-**3. Specify Visual Style:**
-```
-✓ "Minimalist block diagram with clean lines"
-✓ "Detailed biological pathway with protein structures"
-✓ "Technical schematic with engineering notation"
-```
+## Embedding the result in documents
 
-**4. Request Specific Labels:**
-```
-✓ "Label all arrows with activation/inhibition"
-✓ "Include layer dimensions in each box"
-✓ "Show time progression with timestamps"
-```
+The tool saves the final image to the exact `output` path you specified
+(plus versioned `_v1.png`, `_v2.png`, … siblings for debugging, and a
+`_review_log.json` next to them).
 
-**5. Mention Color Requirements:**
-```
-✓ "Use colorblind-friendly colors"
-✓ "Grayscale-compatible design"
-✓ "Color-code by function: blue for input, green for processing, red for output"
-```
+Embed with markdown image syntax. **The path is relative to the markdown
+file you are writing, not the workspace root.**
 
-## AI Generation Examples
-
-### Example 1: CONSORT Flowchart
-```bash
-python @skill/scripts/generate_schematic.py \
-  "CONSORT participant flow diagram for randomized controlled trial. \
-   Start with 'Assessed for eligibility (n=500)' at top. \
-   Show 'Excluded (n=150)' with reasons: age<18 (n=80), declined (n=50), other (n=20). \
-   Then 'Randomized (n=350)' splits into two arms: \
-   'Treatment group (n=175)' and 'Control group (n=175)'. \
-   Each arm shows 'Lost to follow-up' (n=15 and n=10). \
-   End with 'Analyzed' (n=160 and n=165). \
-   Use blue boxes for process steps, orange for exclusion, green for final analysis." \
-  -o @ws/figures/consort.png
+Markdown at workspace root (`paper.md`) and tool output at `figures/consort.png`:
+```markdown
+![Figure 2: CONSORT participant flow](figures/consort.png)
 ```
 
-### Example 2: Neural Network Architecture
-```bash
-python @skill/scripts/generate_schematic.py \
-  "Transformer encoder-decoder architecture diagram. \
-   Left side: Encoder stack with input embedding, positional encoding, \
-   multi-head self-attention, add & norm, feed-forward, add & norm. \
-   Right side: Decoder stack with output embedding, positional encoding, \
-   masked self-attention, add & norm, cross-attention (receiving from encoder), \
-   add & norm, feed-forward, add & norm, linear & softmax. \
-   Show cross-attention connection from encoder to decoder with dashed line. \
-   Use light blue for encoder, light red for decoder. \
-   Label all components clearly." \
-  -o @ws/figures/transformer.png --iterations 2
+Markdown in a subdirectory (`workspace/paper_draft.md`):
+```markdown
+![Figure 2: CONSORT participant flow](../figures/consort.png)
 ```
 
-### Example 3: Biological Pathway
-```bash
-python @skill/scripts/generate_schematic.py \
-  "MAPK signaling pathway diagram. \
-   Start with EGFR receptor at cell membrane (top). \
-   Arrow down to RAS (with GTP label). \
-   Arrow to RAF kinase. \
-   Arrow to MEK kinase. \
-   Arrow to ERK kinase. \
-   Final arrow to nucleus showing gene transcription. \
-   Label each arrow with 'phosphorylation' or 'activation'. \
-   Use rounded rectangles for proteins, different colors for each. \
-   Include membrane boundary line at top." \
-  -o @ws/figures/mapk_pathway.png
-```
+Rule of thumb: prepend `../` per directory the markdown file is below
+the workspace root.
 
-### Example 4: System Architecture
-```bash
-python @skill/scripts/generate_schematic.py \
-  "IoT system architecture block diagram. \
-   Bottom layer: Sensors (temperature, humidity, motion) in green boxes. \
-   Middle layer: Microcontroller (ESP32) in blue box. \
-   Connections to WiFi module (orange box) and Display (purple box). \
-   Top layer: Cloud server (gray box) connected to mobile app (light blue box). \
-   Show data flow arrows between all components. \
-   Label connections with protocols: I2C, UART, WiFi, HTTPS." \
-  -o @ws/figures/iot_architecture.png
-```
+Always include the `![...](...)` embed — do not only reference the
+figure in prose ("see Figure 1"), or the rendered document will not
+show the image.
 
 ---
 
-## Command-Line Usage
-
-The main entry point for generating scientific schematics:
-
-```bash
-# Basic usage
-python @skill/scripts/generate_schematic.py "diagram description" -o @ws/output.png
-
-# Custom iterations (max 2)
-python @skill/scripts/generate_schematic.py "complex diagram" -o @ws/diagram.png --iterations 2
-
-# Verbose mode
-python @skill/scripts/generate_schematic.py "diagram" -o @ws/out.png -v
-```
-
-**Note:** The AI generation system includes automatic quality review in its iterative refinement process. Each iteration is evaluated for scientific accuracy, clarity, and accessibility.
-
-## Best Practices Summary
-
-### Design Principles
-
-1. **Clarity over complexity** - Simplify, remove unnecessary elements
-2. **Consistent styling** - Use templates and style files
-3. **Colorblind accessibility** - Use Okabe-Ito palette, redundant encoding
-4. **Appropriate typography** - Sans-serif fonts, minimum 7-8 pt
-5. **Vector format** - Always use PDF/SVG for publication
-
-### Technical Requirements
-
-1. **Resolution** - Vector preferred, or 300+ DPI for raster
-2. **File format** - PDF for LaTeX, SVG for web, PNG as fallback
-3. **Color space** - RGB for digital, CMYK for print (convert if needed)
-4. **Line weights** - Minimum 0.5 pt, typical 1-2 pt
-5. **Text size** - 7-8 pt minimum at final size
-
-### Integration Guidelines
-
-1. **Include in markdown** - Use `![Figure N: caption](path/to/figures/filename.png)` for every generated image (path must be relative to the markdown file, not the workspace root)
-2. **Include in LaTeX** - Use `\includegraphics{}` for generated images in LaTeX documents
-3. **Caption thoroughly** - Describe all elements and abbreviations
-4. **Reference in text** - Explain diagram in narrative flow
-5. **Maintain consistency** - Same style across all figures in paper
-6. **Version control** - Keep prompts and generated images in repository
+## What the tool does NOT do
 
-## Troubleshooting Common Issues
+- It does not embed figures into your document — you must add the
+  markdown image syntax yourself.
+- It does not cite figures in text — write the narrative reference too.
+- It does not deduplicate across iterations; the `_v1.png` and `_v2.png`
+  intermediates are left on disk for inspection.
 
-### AI Generation Issues
-
-**Problem**: Overlapping text or elements
-- **Solution**: AI generation automatically handles spacing
-- **Solution**: Increase iterations: `--iterations 2` for better refinement
-
-**Problem**: Elements not connecting properly
-- **Solution**: Make your prompt more specific about connections and layout
-- **Solution**: Increase iterations for better refinement
-
-### Image Quality Issues
-
-**Problem**: Export quality poor
-- **Solution**: AI generation produces high-quality images automatically
-- **Solution**: Increase iterations for better results: `--iterations 2`
-
-**Problem**: Elements overlap after generation
-- **Solution**: AI generation automatically handles spacing
-- **Solution**: Increase iterations: `--iterations 2` for better refinement
-- **Solution**: Make your prompt more specific about layout and spacing requirements
-
-### API Issues
-
-**Problem**: Authentication error
-- **Solution**: Verify `OPENROUTER_API_KEY` is set correctly
-- **Solution**: Check your OpenRouter account has sufficient credits
-
-**Problem**: Model not available
-- **Solution**: Check OpenRouter model availability at openrouter.ai/models
-- **Solution**: Try alternative models via `--image-model` or `--review-model`
-
-## Resources and References
-
-### Detailed References
-
-Load these files for comprehensive information on specific topics:
-
-- **`@skill/references/best_practices.md`** - Publication standards and accessibility guidelines
-- **`@skill/references/README.md`** - Extended usage guide and troubleshooting
-- **`@skill/references/QUICK_REFERENCE.md`** - Condensed command cheat sheet
-
-### External Resources
-
-**Python Libraries**
-- Schemdraw Documentation: https://schemdraw.readthedocs.io/
-- NetworkX Documentation: https://networkx.org/documentation/
-- Matplotlib Documentation: https://matplotlib.org/
-
-**Publication Standards**
-- Nature Figure Guidelines: https://www.nature.com/nature/for-authors/final-submission
-- Science Figure Guidelines: https://www.science.org/content/page/instructions-preparing-initial-manuscript
-- CONSORT Diagram: http://www.consort-statement.org/consort-statement/flow-diagram
-
-## Integration with Other Skills
-
-This skill works synergistically with:
-
-- **Scientific Writing** - Diagrams follow figure best practices
-- **Scientific Visualization** - Shares color palettes and styling
-- **Research Grants** - Methodology diagrams for proposals
-- **Scholar Evaluation** - Evaluate clarity, completeness, and communication quality
-
-## Quick Reference Checklist
-
-Before submitting diagrams, verify:
-
-### Visual Quality
-- [ ] High-quality image format (PNG from AI generation)
-- [ ] No overlapping elements (AI handles automatically)
-- [ ] Adequate spacing between all components (AI optimizes)
-- [ ] Clean, professional alignment
-- [ ] All arrows connect properly to intended targets
-
-### Accessibility
-- [ ] Colorblind-safe palette (Okabe-Ito) used
-- [ ] Works in grayscale (tested with accessibility checker)
-- [ ] Sufficient contrast between elements (verified)
-- [ ] Redundant encoding where appropriate (shapes + colors)
-- [ ] Colorblind simulation passes all checks
-
-### Typography and Readability
-- [ ] Text minimum 7-8 pt at final size
-- [ ] All elements labeled clearly and completely
-- [ ] Consistent font family and sizing
-- [ ] No text overlaps or cutoffs
-- [ ] Units included where applicable
-
-### Publication Standards
-- [ ] Consistent styling with other figures in manuscript
-- [ ] Comprehensive caption written with all abbreviations defined
-- [ ] Referenced appropriately in manuscript text
-- [ ] Meets journal-specific dimension requirements
-- [ ] Exported in required format for journal (PDF/EPS/TIFF)
-
-## Environment Setup
-
-```bash
-# Required
-export OPENROUTER_API_KEY='your-openrouter-api-key'
-
-# Optional model overrides
-export SCHEMATIC_IMAGE_MODEL='google/gemini-3-pro-image-preview'
-export SCHEMATIC_REVIEW_MODEL='google/gemini-3-pro-preview'
-```
+---
 
-## Getting Started
+## House style — what gives every figure its shared identity
+
+Every diagram produced by this tool is drawn against a fixed **house
+visual style** — a single profile bundled with the system that pins:
+
+- **Theme** — editorial institutional, off-white background, graphite
+  text/strokes, restrained contrast, disciplined whitespace. Not
+  "AI-aesthetic", not startup-whitepaper.
+- **Typography** — Inter / SF Pro Text / Helvetica Neue / Arial stack,
+  with fixed size tokens for section labels, node labels, edge
+  annotations, and foot labels.
+- **Palette roles** — text/stroke, primaryStructure, secondaryStructure,
+  contextFill, resultAccent, warningAccent, grid — each mapped to a
+  specific hex. Not "Okabe-Ito preferred" — a concrete editorial blue/red
+  identity with accessibility as fallback, not default.
+- **Geometry tokens** — stroke widths (1.5px / 1px), corner radius 8px,
+  filled-triangle arrowheads, dashed-border group containers, 32px min
+  gutter, 3-step box height scale.
+- **Motifs** — labels outside boxes, arrow stroke matches source colour,
+  at most two accent colours per figure.
+
+The reviewer is explicitly told to grade "house-style adherence" as the
+fifth rubric dimension, so iterations that drift from the profile get
+blocked from acceptance. You'll see `houseProfile: editorial-institutional-v1`
+in the review log to confirm which profile was applied.
+
+### Overriding the house style via prompt
+
+There is **no** user-facing theme selector — consistency across figures
+is the whole point. If a user needs to deviate for a specific figure
+(e.g., matching a publication's brand colours), express the overrides
+in the `prompt` itself using quoted tokens or hex codes:
+
+```
+prompt: >
+  Workflow diagram, three stages.
+  Use "#E30613" for the final "Result" node (publication brand red);
+  rest of the figure follows the default house palette.
+```
+
+The tool's literal extractor picks up hex codes (`#RRGGBB`,
+`#RRGGBBAA`), rgb()/rgba(), hsl()/hsla() tokens from plain prose and
+pins them in MUST KEEP, so the override survives rendering.
+
+## Configuration
+
+Raster image generation (gpt-image-2) requires `OPENAI_API_KEY`
+(set under Settings → API Keys). Review uses either OpenAI or
+Anthropic based on Settings → Diagrams. The `auto` setting prefers
+Claude when both keys are available, so the generator does not grade
+its own family.
+
+### Output format selection (png vs svg)
+
+Format intent flows through two redundant channels — either is enough,
+and when they disagree the explicit `format` parameter wins:
+
+1. **Explicit `format` parameter** — use this when the user said
+   something format-specific in the prompt:
+   - "SVG / 矢量图 / vector / 向量图" → `format: "svg"`
+   - "PNG / 图片 / raster / bitmap" → `format: "png"`
+   - Otherwise omit (defaults to `auto`).
+2. **Output filename extension** — `output: figures/foo.svg` implies
+   `format: "svg"`, `output: figures/foo.png` implies `format: "png"`.
+   The tool will rewrite the extension to match `format` if they
+   disagree (and report `extensionChanged` in the result so you know
+   to update any Markdown embed).
+
+Behavioural summary:
+
+- `format: "svg"` (or `.svg` extension) → **always** synthesises SVG
+  via the chat model, even when `OPENAI_API_KEY` is present. Choose
+  this for vector output (scales infinitely, small filesize, editable).
+- `format: "png"` (or `.png` extension, the default) → raster via
+  `gpt-image-2` when `OPENAI_API_KEY` is configured. Falls back to
+  SVG-via-chat-model automatically when the key is missing; the file
+  is renamed `.svg` and `extensionChanged` is reported.
+
+### SVG fallback
+
+The same verdict-driven iteration loop runs for both formats. In SVG
+mode, generation goes through the chat model, and review picks the
+strongest option available at runtime:
+
+1. **Rasterise-then-vision (preferred)** — when an offscreen renderer
+   is available (Electron main process) and a real vision reviewer
+   (OpenAI or Anthropic) is configured, the generated SVG is
+   rendered to PNG and passed to the vision model. This catches text
+   overflow, element overlap, and other problems invisible at the
+   SVG-source level.
+2. **Source-level fallback** — when rasterisation is unavailable (e.g.
+   running outside Electron) or no vision reviewer auth is present,
+   the reviewer reads the SVG markup as text. Structural checks work;
+   visual-layout checks do not.
+
+Other SVG-mode notes:
+
+- `mode: "svg_fallback"` appears in the tool result and review log.
+  The provider id distinguishes the two review paths:
+  `rasterize+openai:…` or `rasterize+anthropic:…` vs
+  `svg-fallback:…`.
+- Quality of the SVG itself depends on the chat model's spatial
+  reasoning. Claude Opus and GPT-4o / GPT-5 produce usable output for
+  flowcharts, simple architecture, and box-and-arrow schemas. Pathway
+  illustrations and complex circuits will be noticeably weaker than
+  gpt-image-2.
+- Self-grading bias only applies in the source-level path (same model
+  reads back its own SVG). The rasterise-then-vision path sends the
+  rendered image to a different model, so bias is negligible.
+- Scores across raster and SVG modes are still not directly
+  comparable — thresholds are calibrated per reviewer and per path.
+
+To embed an SVG in Markdown:
 
-**Simplest possible usage:**
-```bash
-python @skill/scripts/generate_schematic.py "your diagram description" -o @ws/output.png
+```markdown
+![Figure 1: workflow](figures/workflow.svg)
 ```
 
----
-
-Use this skill to create clear, accessible, publication-quality diagrams that effectively communicate complex scientific concepts. The AI-powered workflow with iterative refinement ensures diagrams meet professional standards.
+Most Markdown renderers (GitHub, typical preview extensions) and
+Milkdown render SVG inline just like PNG.