project-graph-prg

name	project-graph-prg
description	Generate, inspect, validate, edit, and geometry-check Project Graph `.prg` files from natural-language requirements or reference diagrams. Use when Codex needs to create a new `.prg`, modify an existing `.prg`, inspect a `.prg` archive, validate required nodes/edges/sections, enforce no-overlap layout rules, or route edges around blocks.

Project Graph PRG

Use this skill to turn user requirements or a reference image into a Project Graph .prg file and verify the result with deterministic local scripts.

Use Unicode text in generated content. Write JSON, Markdown, and script I/O as UTF-8.

Quick Workflow

1. Read the user request and decide whether the task is:
   • inspect an existing .prg
   • generate a new .prg
   • edit an existing .prg
   • validate content expectations
   • check layout overlaps
2. If generating from prose or an image, translate the request into an intermediate JSON spec.
   • First produce graphIntent with primaryType, confidence, evidence, extractionFocus, and fallbackType.
   • Use graphIntent to guide typed extraction, grouping, edge semantics, and layoutPlan.
   • Do not build objects first and infer layout later.
   • Use references/spec-format.md for the supported object schema.
3. After generating a spec, run prg_cli.py classify spec.json.
   • If metrics such as hasCycles, density, component_count, or root_count contradict graphIntent, update graphIntent/layoutPlan before generate.
4. If editing an existing .prg, translate the request into a UTF-8 JSON edit patch.
   • Use references/edit-patch-format.md for selectors and edit operations.
   • Use references/prompt-templates.md for prompt wording and output expectations.
5. Prefer the unified CLI scripts/prg_cli.py with uv run --with msgpack --python 3.12 ....
   • The older single-purpose scripts are still available for compatibility.
6. After generation or edit, run validation in this order:
   • validate_prg.py
   • validate_operation_links.py when the graph contains logic/operator blocks such as #ADD# or #DIV#
   • check_overlap.py last
7. Inspect the validation results and decide whether repair is appropriate.
8. Repair logic before layout.
   • Run auto_fix_operation_links.py only for illegal operator-link problems.
   • Run auto_fix_layout.py only after logic is settled, because earlier repairs may change layout.
9. If the final overlap/spacing check passes, decide whether edge routing should insert ConnectPoint detours.
   • Run route_edge_crossings.py or prg_cli.py route-edge-crossings only when the user asks for routed edges or routingPlan.insertConnectPoints is true.
   • If ConnectPoint insertion is disabled, report that straight edges may pass through blocks and do not run the routing command.
10. If validation still fails after repair, adjust the JSON spec or edit patch and retry instead of hand-editing the binary archive.

Model Workflow Template

When using this skill, prefer to make the model follow this explicit workflow in the prompt instead of relying on a one-shot command:

1. Build or inspect the UTF-8 JSON/spec context first.
2. For generation, classify the graph type before extracting nodes and edges.
3. Generate graphIntent, typed extraction, layoutPlan, and objects in that order.
4. Run prg_cli.py classify on generated specs and reconcile metric conflicts before generation.
5. Run prg_cli.py generate, prg_cli.py edit, or prg_cli.py inspect as appropriate.
6. Run prg_cli.py validate.
7. If the graph contains operator blocks such as #ADD#, #DIV#, or #SIN#, run prg_cli.py validate-op.
8. Read the validation output and decide whether logic repair is appropriate.
9. If needed, run prg_cli.py fix-op-links, then re-run prg_cli.py validate-op.
10. Run prg_cli.py overlap last for geometry validation.
11. Read the overlap/spacing output and decide whether layout repair is appropriate.
12. If needed, run prg_cli.py fix-layout --min-gap 200, then re-run prg_cli.py overlap.
13. After overlap passes, decide whether to run prg_cli.py route-edge-crossings based on user preference or routingPlan.insertConnectPoints.
14. Report whether ConnectPoint insertion was enabled, what passed, what failed, and whether the final .prg is acceptable.

Do not auto-repair blindly. The model should inspect validation results first, then decide whether repair is safe.

Scripts

The skill bundles these scripts in scripts:

• prg_cli.py: unified CLI entry point for inspect/generate/edit/validate/overlap/fix/route flows
• graph_layout.py: graph classification, graphIntent/layoutPlan recommendation, native coordinate layout, and strategy dispatch
• layout_prg.py: single-purpose compatibility wrapper for re-layout of existing .prg files
• edit_prg.py: apply a semantic JSON patch to an existing .prg
• generate_prg.py: create a .prg from a JSON spec
• inspect_prg.py: summarize archive contents, nodes, edges, sections, and drawings
• validate_prg.py: verify archive structure and optional expected texts/sections/edges
• check_overlap.py: detect overlapping rectangular blocks and minimum-gap violations using recomputed text bounds and resolved Section bounds
• auto_fix_layout.py: shift overlapping blocks until no overlap remains or iteration limit is reached
• validate_operation_links.py: enforce that operator blocks only reach other operator blocks through exactly one blank intermediate register node
• auto_fix_operation_links.py: repair direct operator-to-operator links by inserting one blank intermediate register node
• route_edge_crossings.py: insert ConnectPoint detours when straight LineEdge segments pass through a block
• geometry_prg_tools.py: geometry-aware helpers for recomputed sizes, resolved bounds, overlap checks, and edge routing

Validation scripts and repair scripts are intentionally separate. Use the validation output to decide whether a repair script should run. Do not assume repair is always correct. Keep overlap detection and layout repair at the end of the workflow to avoid wasted layout work caused by earlier logic repairs.

Run them from the skill directory with relative paths.

Commands

Generate:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py generate spec.json out.prg

Generate without automatic layout:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py generate spec.json out.prg --layout off

Classify a spec before generation:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py classify spec.json

Re-layout an existing .prg:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py layout input.prg output.prg --strategy auto --direction right --min-gap 200 --seed 7

Edit:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py edit input.prg patch.json edited.prg

Inspect:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py inspect out.prg --limit 10

Validate:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py validate out.prg

Validate expected content:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py validate out.prg --expect expect.json

Check overlap after logic validation and any logic repair:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py overlap out.prg --exclude-sections

The default clearance rule is 200px. To make it stricter, for example 300px:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py overlap out.prg --exclude-sections --min-gap 300

Auto-fix layout at the end of the workflow:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py fix-layout out.prg fixed.prg --min-gap 200

Route edge crossings after overlap passes:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py route-edge-crossings fixed.prg routed.prg

Validate logic operator links:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py validate-op logic.prg

Auto-fix logic operator links:

uv run --with msgpack --python 3.12 .\scripts\prg_cli.py fix-op-links logic.prg fixed-logic.prg

Do not ignore containment overlap. If one block fully contains another block, that still counts as overlap.

Generation Rules

• Build the .prg through the JSON spec, not by mutating msgpack manually.
• For generation, decide graphIntent before extracting objects; graphIntent must guide extraction, grouping, edge semantics, and layoutPlan.
• Use tree, mindmap, dag, cyclic_dependency, dense_network, pipeline, or clustered as first-class graph types.
• Use native layout strategies tree, mindmap, layered, scc-layered, force, pipeline, and clustered; do not depend on Mermaid, ELK, Dagre, or NetworkX at runtime.
• Run prg_cli.py classify after spec generation and reconcile conflicts before prg_cli.py generate.
• Use Unicode for node labels, section titles, and inserted helper text. Store JSON/reference files in UTF-8.
• Generated files should rely on recomputed text-node bounds, not placeholder width/height guesses.
• Keep text blocks separated enough that check_overlap.py returns no overlaps or spacing violations.
• Leave overlap detection and layout repair until the end of the workflow, after logic and structural checks.
• If overlap exists at the final layout stage, prefer auto_fix_layout.py before manually repositioning objects.
• After layout passes, check whether the user wants ConnectPoint-based edge routing. If yes, run route_edge_crossings.py; if no, skip it and state that edge-through-block detection was intentionally not enforced.
• Prefer TextNode, Section, and LineEdge unless the user clearly needs UrlNode, ImageNode, SvgNode, ConnectPoint, MultiTargetUndirectedEdge, or PenStroke.
• If the user provides a reference image, preserve the logical structure first, then approximate layout.
• After generating from an image, explicitly tell the user the .prg is a reconstructed approximation, not a pixel-perfect import.

Validation Rules

• Treat validate_prg.py failure as a hard failure.
• Treat any reported overlap as a layout failure unless the user explicitly allows overlap.
• Treat containment overlap as overlap. Do not disable that behavior.
• By default, treat 200px edge-to-edge clearance as required. If the user specifies a different clearance rule such as "300px around each unit must stay empty", run check_overlap.py with --min-gap 300 and treat any spacing violation as a failure.
• Treat any direct operator-to-operator link as invalid. Two operator blocks are only legal when they are separated by exactly one blank intermediate register node.
• Let the model decide whether to run auto_fix_layout.py or auto_fix_operation_links.py after reading validation output. Keep detection and repair as separate steps.
• Treat a straight LineEdge that enters one side of a block and exits another side as invalid only when ConnectPoint-based edge routing is enabled. Route it by inserting ConnectPoint detours after overlap/spacing checks pass.
• When the user cares about exact content, create an expect.json file and run content validation.

References

• references/spec-format.md: supported JSON generation schema
• references/edit-patch-format.md: supported JSON edit patch schema
• references/prompt-templates.md: reusable prompt patterns for image-to-PRG and requirement-to-PRG tasks

Portable Ideas From The App's Built-In AI Tools

The app's runtime AI tools live in app/src/core/service/dataManageService/aiEngine/AITools.tsx. The most useful ideas to port into this skill are the ones that do not depend on live UI state.

Good candidates for this headless .prg skill:

• regex search over text nodes
• parent/child graph queries
• direct connection checks between nodes
• breadth expansion from a node
• depth expansion from a node
• sort/reorder nodes by X or Y axis

Lower-value or non-portable runtime-only tools:

• current selection queries
• viewport queries
• tools that depend on what the user has selected in the open GUI

If you extend this skill further, prefer adding the portable graph-query and graph-edit capabilities before any GUI-state-dependent tool.