Note

⚠️ WORK IN PROGRESS

Core workflows (brainstorm, plan, work, review, fix-triage, compound) are stable. Advanced features (release automation, context layer) are still evolving. Prompts, agents, skills, and conventions may change without notice.

Compound GPID

A GitHub Copilot plugin for data science teams, built on the Compound Engineering Philosophy.

Runtime scope: Compound GPID supports GitHub Copilot, Claude Code, Codex, OpenCode, and Kilo. .github/ is the canonical source; native platform trees (.claude/, .agents/, .opencode/, .kilo/) are generated from it and distributed through per-platform install units. cg-link links all supported platforms by default; use cg-link --platforms copilot or another comma-separated list to narrow the install.

Install, update, and link do not require Kilo, a minimum Kilo version, or host certification. Kilo host checks belong to the optional cg-kilo launcher. Old or unreadable version output produces a warning, not a version-based block. For a Kilo launch with Codex/Claude skill roots, cg-kilo retains projection and containment checks and sets KILO_DISABLE_EXTERNAL_SKILLS=1 only for the child process. See Installation for the launch boundary.

Native targets package each canonical skill as an atomic skill bundle. A bundle contains SKILL.md plus all nested regular files and includes them by default; it is not a SKILL.md-only copy. Executable resources are copied as opaque bytes with their executable bit recorded and are never executed during generation. See Generated Native Platform Trees for ownership, recovery, and verification guarantees.

Each unit of work should make subsequent units easier — not harder.

Why Compound GPID?

Compound GPID enforces a repeatable Brainstorm → Plan → Work → Review → Fix Triage → Compound loop that does more than produce code — it produces knowledge. Every solved problem is captured as a structured document and fed back into future reviews, so your team compounds its expertise with every task.

Key benefits:

• Consistency — coding standards, tests, and documentation are enforced on every PR, for R, Python, and Stata.
• Compounding returns — the cg-learnings-researcher agent cross-references past solutions so the team never solves the same problem twice.
• Project awareness -- Optionally create a compound-gpid.md project charter to give Copilot persistent knowledge of your project's objective, deliverables, constraints, and current focus. Every session then starts in context.
• R dialect selection — set r-syntax: "tidyverse" in your local config to have all R assistance use tidyverse/dplyr patterns instead of data.table/collapse. Ideal for projects with external coauthors who only know the tidyverse. See docs/reference.md for details.
• Knowledge brain — cg-index --brain and /cg-brain-rebuild build a structured knowledge brain (BRAIN.md, BRAIN-NN.md, BRAIN-log.md, brain-index.json) by clustering .cg-docs/ artifacts into topics and mapping typed relationships between artifacts — so every session can surface relevant past work automatically.
• Roadmap tracking — @cg-roadmap manages a roadmap.json milestone and feature tracker. Brainstorm, Plan, and Work prompts hook into it automatically: brainstorms register feature ideas, plans link to features, and work marks them active — so your roadmap stays current without manual updates.
• Modular suites — a validated registry separates the kernel, reusable capability packs, the technical /cg-* suite, and the research /cr-* suite. Projects select suites: [cg], [cr], or [cg, cr] without creating cross-suite dependencies.
• Cross-platform — native support for GitHub Copilot, Claude Code, Codex, OpenCode, and Kilo from a single .github/ source. Generated platform trees are committed, release-validated, and distributed through merge-safe per-platform install units. cg-link links all platforms by default.
• Zero friction — one global clone, per-project links and managed projections, and shell commands (cg-link, cg-unlink, cg-update, cg-index, cg-brain-init, cg-publish-markdown, cg-token-audit) wire everything into VS Code / Positron automatically.
• Secure document views — cg-publish-markdown turns one project-contained generic Markdown file into a deterministic, self-contained reference HTML view while preserving strict Brainstorm/Plan validation and excluding generated bodies from model context.
• Token guidance — /cg-token-audit runs deterministic context/model analysis and returns compact advice on context size, review depth, and model selection without changing project files.
• Team-wide — one global update serves linked projects; rerun cg-link from each project root with its intended platform selection to refresh managed copies and projections.
• Version management — pin to a specific release for stability, or explicitly choose to track main. For any older installation, pinned or tracking latest, follow Upgrade from an older version first and keep the resulting release pin.

Documentation

→ Open the documentation site or start with Getting Started.

Page	Contents
Getting Started	Understand the project, install/configure it, and complete a first workflow
Why Compound GPID?	Institutional focus, upstream inspiration, differences, and tradeoffs
Workflows	Task-oriented paths from strategy through verified knowledge capture
Research Handbook	Newcomer path for the research suite, from activation through a first CR workflow
Skills	Canonical analytical, technical, testing, and institutional skill catalog
Configuration	Context files, platform targets, settings, and managed content
Modular Guide	Suite selection, capability composition, registry ownership, and extension rules
Governance	Data safeguards, review gates, operating constraints, and limitations
Reference	Complete commands, agents, schemas, configuration, and file structure
Help	Recovery routes and complete troubleshooting links

Contributing

See CONTRIBUTING.md for local test setup, cross-platform requirements, commit conventions, and the PR workflow.

License

MIT