Reference

Detailed contracts for Compound GPID commands, agents, schemas, configuration, and file structure. Use the focused Commands, Agents, Skills, and Files and Artifacts pages for quicker navigation.

See Workflow for a full explanation of each prompt step. See Modular Guide for choosing between the technical (/cg-*) and research (/cr-*) suites, capability packs, module preferences, and extension rules. See Installation for setup instructions. See Context Files for a detailed guide to copilot-instructions.md, compound-gpid.md, compound-gpid.context.md, and the generated native platform trees (.claude/, .agents/, .opencode/, .kilo/). See Troubleshooting for known issues.

Compound GPID generates native platform trees for Claude Code, Codex, OpenCode, and Kilo from the canonical .github/ source. The trees are committed, release-validated, and distributed through merge-safe per-platform install units. cg-link links all supported platforms by default; use --platforms to narrow the install to a comma-separated list. The adapters/ directory contains legacy source adapters that are superseded by the generated trees.

Native Target Packaging

.github/ is the canonical input. The generator treats every regular skill directory as an atomic skill bundle: it includes by default SKILL.md and all recursively nested regular files. It copies executable files as opaque bytes, records and preserves executable mode, and guarantees they are never executed during inventory, generation, validation, or cleanup.

In short, the regular-file policy is include by default, not an allowlist of known support filenames.

The mapping defines these target-local runtime and support roots:

Target	Commands	Skills	Agents	Instructions	Shared/support
Claude Code	`.claude/commands`	`.claude/skills`	`.claude/agents`	`.claude/instructions`	`.claude/shared`
Codex	`.agents/commands`	`.agents/skills`	`.agents/subagents`	`.agents/instructions`	`.agents/shared`
OpenCode	`.opencode/commands`	`.opencode/skills`	`.opencode/agents`	`.opencode/instructions`	`.opencode/shared`
Kilo	`.kilo/commands`	`.kilo/skills`	`.kilo/agents`	`.kilo/instructions`	`.kilo/shared`

Generation computes a deterministic isolated dependency closure across command support files, complete skill bundles, agents, instructions, and shared contracts. A generated target must resolve without the canonical .github/ tree; unresolved or unsafe canonical runtime references fail generation.

Ownership manifest schema

The fixed manifests are .claude/.compound-gpid-generated.json, .agents/.compound-gpid-generated.json, .opencode/.compound-gpid-generated.json, and .kilo/.compound-gpid-generated.json. Their schema is:

{
  "schemaVersion": 1,
  "target": "claude-code",
  "policyVersion": 1,
  "files": [
    {
      "path": ".claude/commands/cg-setup.md",
      "source": ".github/prompts/cg-setup.prompt.md",
      "kind": "command",
      "sha256": "<64 lowercase hexadecimal characters>",
      "executable": false
    }
  ]
}

Top-level and file-entry fields are fixed; files is sorted by path, and the manifest is deterministic and excludes itself. Before mutation, the generator validates ownership, destinations, and checksums. Stale cleanup is checksum-guarded: only a prior-manifest-owned stale regular file whose current hash equals its recorded sha256 is deleted. Untracked files are preserved. Files are atomically replaced before stale cleanup, and the manifest is written last, so a later rerun provides safe recovery from interruption.

These source-repository ownership manifests are distinct from consumer .compound-gpid/managed-files.json, which controls merge-safe copied install files used by cg-link, cg-update, and cg-unlink.

Verification and release gates

Required evidence is deterministic and does not depend on installed vendor tools: isolated fixtures verify dependency closure without .github/, the generator determinism suite proves repeated byte-identical plans/manifests, and drift tests compare committed trees with canonical inputs. CI runs the native target Python gate on Windows and macOS. Release publication has a separate release gate: create-release.ps1 runs mapping, path-safety, packaging, ownership, closure, determinism, drift, and platform suites before reading credentials or calling GitHub.

A real CLI smoke run with claude, codex, opencode, or kilo is optional, additional runtime evidence only. CLI availability is reported separately; missing real CLI evidence never skips or weakens deterministic isolated closure, and documentation must not imply runtime proof that was not run.

Optional retrieval backend candidates are documented in Retrieval Backend Evaluation and tracked in .github/shared/retrieval-backends.json. The registry is evaluation-only: native-brain-query remains the only active backend.

Snapshot and external-research mode candidates are documented in Snapshot and External-Research Modes and tracked in .github/shared/snapshot-research-modes.json. The registry is evaluation-only: local-workflow remains the only active mode.

Shell Commands

Core install commands are available from PowerShell on Windows and from bash/zsh on macOS. The cg-*-summary wrappers are bash wrappers in bin/; use them from bash/zsh or run python scripts/cg_summary.py <kind> directly on Windows.

Command	Where to run	Purpose
`cg-link [--platforms <list>]`	Project root	Link all supported platforms by default: Copilot `.github/`, Claude Code `.claude/`, Codex `.agents/`, OpenCode `.opencode/`, and Kilo `.kilo/`. Use `--platforms copilot` or another comma-separated list to narrow the install.
`cg-unlink`	Project root	Remove Compound GPID-managed install units and manifest-managed copied files while preserving user-owned platform content.
`cg-update [<version>\|latest\|--list\|--fix]`	Anywhere	Update, pin, unpin, list releases, or repair a Compound GPID installation.
`cg-kilo [<kilo arguments>]`	Project root	Optional contained Kilo launch. Version diagnostics warn only. For projects with Codex/Claude roots, validates containment and disables external skill discovery only in the child Kilo process; direct launches are unsupported.
`cg-brain-init`	Project root	Initialize or configure Team Brain integration and scaffold the central GitHub repository configuration. Usage: `cg-brain-init --repo <owner/name> --manager <github-username>`.
`cg-index`	Project root	Build or query the local `.cg-docs/` Knowledge Brain index.
`cg-index --brain`	Project root	Rebuild generated Brain artifacts such as `BRAIN.md`, topic files, and `brain-index.json`.
`cg-render-artifact <source>`	Project root	Validate and explicitly render one Brainstorm or Plan, even when automatic HTML is disabled.
`cg-render-artifact --automatic <source>`	Project root	Always validate; write HTML only when `artifact-html: true` is configured.
`cg-render-artifact --validate-only <source>`	Project root	Validate one canonical artifact without writing HTML.
`cg-render-artifact --check <source>`	Project root	Print `current` and exit code 0 only for an exact current view; print `missing` or `stale` and exit code 1 otherwise. Input/usage errors return exit code 2.
`cg-publish-markdown <source>`	Project root	Publish one project-contained generic Markdown file to the mirrored `.cg-docs/views/documents/` path.
`cg-publish-markdown --automatic <source>`	Project root	Validate generic Markdown and publish only when `artifact-html: true` is configured.
`cg-publish-markdown --validate-only [--theme reference] <source>`	Project root	Validate generic source, output identity, local resources, and theme without inspecting or writing output.
`cg-publish-markdown --check [--theme reference] <source>`	Project root	Reproduce expected schema-2 bytes; return exit code 0 for `current`, 1 for `missing`/`stale`, and 2 for invalid input.
`cg-publish-markdown --output <documents-view.html> <source>`	Project root	Publish to one portable relative destination under `.cg-docs/views/documents/`.
`cg-token-audit --root . --output-dir .cg-docs/cost --format both --recommendations`	Project root	Generate context-cost reports, token dashboard artifacts, regression checks, and compact recommendations.
`cg-test-summary --root . --format json`	Project root	Summarize `tests/last-run.json` without running tests and store a redacted source artifact.
`cg-diff-summary --root . --format md`	Project root	Summarize changed files, hunks, and risk tags while storing the full redacted diff artifact.
`cg-log-summary --root . --format json`	Project root	Summarize branch-local first-parent commits and notable files.
`cg-tree-summary --root . --max-entries 120 --format md`	Project root	Summarize a bounded repository tree while excluding generated outputs, dependencies, and caches.
`cg-problems-summary --root . --input problems.json --format json`	Project root	Summarize optional diagnostics JSON or text; reports unavailable when no diagnostics input is provided.
`python scripts/cg_project_manifest.py [--root <path>] [--output <path>] [--platforms copilot,kilo] [--validate] [--check-stale <manifest>] [--ensure-state]`	Project root	`cg-project-manifest` — resolve and validate the canonical committed active project manifest (`.compound-gpid/active-manifest.json`) from the strict config plus the versioned module registry. Records config/registry hashes, schema versions (`config-schema-version`), selected suites, derived and explicit capabilities, the resolved module closure, canonical platform ids, platform eligibility, and the projection plan digest. Immutable selection validity is separated from mutable projection ownership. Exit codes: `0` success, `1` resolution/validation failure, `2` missing or invalid project root.
`python scripts/cg_projection_benchmark.py [--root <path>] [--profiles cg-only,cr-only,mixed,capability-python] [--validate]`	Project root	`cg-projection-benchmark` — emit deterministic before-state profile baseline matrices (`.cg-docs/cost/skill-loading-baseline.json` + `.md`) for projection. Per profile: requested command/capability, expected route, expected hard-stop or catalog summary, expected inventory digest, and a supported-host procedure. Token estimates are heuristic (chars/4) and never claim savings; unavailable required host evidence is a blocking `unavailable`, never a zero. Exit codes: `0` success, `1` validation/oracle failure, `2` missing or invalid project root.

Windows installs include matching .cmd wrappers for the core install commands where platform-specific launch behavior is required.

Generic publishing uses the frozen reference theme version 1 and provenance schema 2. Every destination has one source owner identified by sourcePath, documentType, and outputPath. Existing corrupt or differently owned output fails without mutation. Brainstorms and Plans retain strict validation and must use cg-render-artifact; generic publishing cannot act as a schema bypass.

Version Management

Compound GPID supports pinning to specific GitHub Releases so you can choose between stability and bleeding-edge.

Command	Effect	Persists?
`cg-update`	Use current preference (default: latest)	—
`cg-update v0.2.0`	Pin to release `v0.2.0`	Yes — writes to `.cg-version`
`cg-update latest`	Unpin and track `main`	Yes — writes to `.cg-version`
`cg-update --list`	Browse available releases	No
`cg-update --fix`	Repair broken installation (clean + pull)	No

How it works: the version preference is stored per-machine in .cg-version inside your global install directory. This file is gitignored. Pinned users see a yellow hint when a newer release is available at the end of every cg-update run.

Full details: see the Version Management page.

Copilot Chat Prompts

Prompt	Model	Purpose
`/cg-brain-rebuild`	Copilot model picker	Rebuild the project knowledge brain (BRAIN.md + indexes).
`/cg-brainstorm`	Copilot model picker	Brainstorm answers about what to build and how. Use when requirements are fuzzy.
`/cg-commit-push-pr`	Copilot model picker	Stage changes into logical commits, push, and open a PR with plan-driven description.
`/cg-compound-gpid-rd`	Copilot model picker	Research public GitHub repos for features to integrate into Compound GPID and manage the review registry. Developer-only.
`/cg-compound-refresh`	Copilot model picker	Audit and refresh .cg-docs/solutions/ for staleness, drift, and consolidation opportunities.
`/cg-compound`	Copilot model picker	Capture a solved problem as reusable knowledge. Offers canonical .github/ updates; the user applies them manually after fixing a non-trivial issue.
`/cg-devtag`	Copilot model picker	Create a dev tag (v<MAJOR>.<MINOR>.<PATCH>.9000+) on the current branch and push it to origin. Enables end-to-end installation testing via cg-update before an official release. Developer-only.
`/cg-diagnose`	Copilot model picker	Diagnose VS Code crashes. Inspects logs, classifies the crash category, checks for uncommitted work, and recommends recovery steps.
`/cg-fix-problems`	Copilot model picker	Interactive VS Code diagnostics fixer. Scans all workspace files for errors, warnings, and info diagnostics, lets the user select scope and severity, then applies fixes. Dispatches @cg-fix-problems agent.
`/cg-fix-triage`	Copilot model picker	Apply review findings from a saved review report. Fixes all findings or a subset by ID/priority.
`/cg-fixbug`	Copilot model picker	Structured bug-fix workflow: establish the expected-behavior source in Step 1.5, perform test-gap classification in Step 2.5, and require red-green proof.
`/cg-ideate`	Copilot model picker	Generate, critique, and filter improvement ideas for the project. Use before /cg-brainstorm when you want to discover what to work on next.
`/cg-issues`	Copilot model picker	Manage GitHub Issues linked to roadmap work items. Modes: status (default, read-only), backfill, link, adopt, setup.
`/cg-light-work`	Copilot model picker	Qualify and execute one small technical task with bounded discovery, light review, and explicit compounding consent.
`/cg-plan-review`	Copilot model picker	Review an implementation plan for risks, over-engineering, missing edge cases, and flawed assumptions. Use after /cg-plan or on any existing plan.
`/cg-plan`	Copilot model picker	Create a structured implementation plan with research. Use after brainstorming or when requirements are clear.
`/cg-release`	Copilot model picker	Create a GitHub Release for compound-gpid. Detects the next semver tag from git history, drafts curated release notes, checks SCHEMA_VERSION, confirms with the user, and publishes. Developer-only — guarded to the compound-gpid repo; Step 0 stops execution in consumer projects.
`/cg-render-doc`	Copilot model picker	Render a workflow artifact or generic Markdown document to curated HTML. Routes typed artifacts to cg-render-artifact and generic documents to cg-publish-markdown. Supports --theme selection (reference or editorial).
`/cg-resume`	Copilot model picker	Load context and resume interrupted work. Use at the start of a session to pick up where you left off.
`/cg-review`	Copilot model picker	Run multi-agent code review on recent changes. Produces prioritized P0/P1/P2/P3 findings.
`/cg-roadmap-view`	Copilot model picker	Visualize the project roadmap in chat. Supports flags: --milestone, --tasks, --detail, --status, --wip, --plan, --help. Dispatches @cg-roadmap-view agent for rendering.
`/cg-setup`	Copilot model picker	Configure Compound GPID for this project and load context for returning projects.
`/cg-skill`	Copilot model picker	Discover, import, validate, activate, update, audit, deprecate, and remove skills through one lifecycle command.
`/cg-strategy`	Copilot model picker	Strategic project visioning and direction-setting. Use when you have a full project in mind to structure, or when you need to rethink direction mid-project. Dispatches @cg-roadmap for all roadmap writes.
`/cg-token-audit`	Copilot model picker	Analyze Compound GPID token/context usage and suggest cost-efficient workflow choices.
`/cg-verify-pr`	Copilot model picker	Check CI status on current PR, classify failures, and auto-fix with review agents. Use --propose for observe-only diagnosis.
`/cg-wiki`	Copilot model picker	Manage the project wiki: initialize, rebuild pages, restructure sections, check status, or convert to GitHub Wiki format.
`/cg-work`	Copilot model picker	Implement a /cg-plan plan. Supports /cg-work [phaseX], review, and deviate controls.

`/cg-commit-push-pr` Base And Preflight Contract

Use --base <branch> when the intended PR target is not the repository default. The prompt resolves one $baseBranch before generation or staging with this precedence: existing PR baseRefName, explicit --base, then the repository default branch. If the existing PR base conflicts with explicit input, it reports both values and uses the actual existing PR base.

The prepare gate runs cg_pr_preflight.py --phase prepare --base <branch> --run-native-target before staging. After commits, the committed gate runs the same preflight with --phase committed --base <branch> --run-native-target before push. Nonzero or partial results block the operation; a successful generic-not-applicable Kilo result is a neutral capability outcome for generic behavior and does not claim certified-host integration.

gh creation always receives --base <branch>. The VS Code GitHub Pull Request extension must resolve and honor the same baseBranch; if it cannot, the prompt halts with a manual gh pr create --base <branch> --body-file <file> route rather than silently selecting a different base.

`/cg-verify-pr` Exact Diagnosis And Repair Contract

/cg-verify-pr requests the open PR's actual baseRefName in the same statusCheckRollup metadata query and halts if that base is unavailable. It uses that $baseBranch for every fetch, merge-base, rebase, changed-file comparison, preflight, and trailer-history operation; it never infers a base from a remote symbolic ref.

For each failed check, the prompt reads its detailsUrl and accepts only a GitHub Actions job URL containing both identifiers. It retrieves the exact failed job with:

gh run view <run-id> --job <job-id> --log-failed

Missing, non-Actions, unparseable, or unavailable URLs/logs use a manual route: open the check provider's details page, obtain the exact run/job IDs and failed step output, and do not select a latest run by workflow name or recency.

Auto-fix first runs git status --porcelain and stops on any pre-existing staged, unstaged, or untracked change. After a clean baseline, the prompt runs scripts/cg_pr_preflight.py with $baseBranch and the PR changed files to select the exact focused local reproduction. A certified-host Kilo failure confirmed by the exact job log may use the certified-host remediation path; Kilo generic-not-applicable is neutral capability evidence, and generic linker failure is never Kilo integration proof.

Only post-baseline targeted paths are staged. One verification pass creates exactly one fix(ci) commit with one unique CI-Fix-Round: <PR>/<N> trailer in $mergeBase..HEAD; historical fix(ci): subjects do not count. A rebase uses the resolved PR base, and --force-with-lease is used only after that rebase.

Optional certified-host Kilo evidence is separate from generic CI. Configure the protected repository variables CG_KILO_CERTIFIED_RUNNER, CG_KILO_CERTIFIED_VERSION, and CG_KILO_CERTIFIED_SHA256, and require maintainer approval for the cg-kilo-certified environment. The certified job runs only on a protected default-branch push or an explicitly requested workflow dispatch for that same default branch, checks out that trusted ref, compares the preflight-reported executable version and SHA-256 before launch, and uploads kilo-preflight.json and inventory evidence. Missing configuration produces a neutral generic-not-applicable summary; generic CI never claims real-host integration. These CI identity checks are not minimum-version or certification requirements for install, update, link, or cg-kilo. The launcher's old or unreadable version diagnostics produce warnings only; projection, command, schema, and containment failures remain separate outcomes.

Research Suite Commands

These commands are owned by suite-cr and are available when suites: in compound-gpid.local.md includes cr. The research suite composes shared language, review, knowledge, and publication capabilities without depending on the technical command suite.

Prompt	Model	Purpose
`/cr-brainstorm`	Copilot model picker	Research brainstorm — clarify fuzzy research requirements. Classifies task type (theory, EDA, implementation, ML, writing, etc.) and guides methodology decisions. Use for economics and econometrics research tasks.
`/cr-compound`	Copilot model picker	Research compound — capture a solved research problem for future reuse. Extends /cg-compound with research-specific categories: identification, specification, derivation, ml-methodology, reproducibility.
`/cr-plan`	Copilot model picker	Research plan — structured implementation plan for research tasks. Use after /cr-brainstorm to create concrete steps.
`/cr-review`	Copilot model picker	Research review — multi-agent code and methodology review. Orchestrates cg-* agents (code quality, testing, reproducibility) and cr-* agents (research integrity, mathematical verification, identification audit, econometric reasoning). Produces prioritized P0/P1/P2/P3 findings.
`/cr-work`	Copilot model picker	Research work — implement a research plan step by step. Supports /cr-work [phaseX]. Enforces P0 seed, provenance, and specification logging requirements.

The research lifecycle is Scope -> Evidence -> Theory -> Method -> Execute -> Verify -> Communicate -> Maintain. Use /cr-review, not /cg-review, for research-domain agent routing.

`cg-index --brain` — Diagnostic Warnings

cg-index --brain writes scan-pass warnings to stderr during execution:

Message	Meaning
`[cg-index] WARNING: Skipping <file>: ...`	File could not be read and was excluded from the Brain index.
`[cg-index] WARNING: Skipping <file>: no frontmatter found`	File lacks a `---` frontmatter block and was excluded from the index.
`[cg-index] WARNING: <file>: missing required field(s): ...`	Frontmatter is missing required metadata such as `title` or `date`.
`[cg-index] WARNING: Duplicate frontmatter key <key>`	Frontmatter repeats a key; the last value is used.
`[cg-index] WARNING: roadmap feature ... has no id`	A roadmap feature lacks an `id` and is skipped for Brain linking.

To capture warnings: cg-index --brain 2>brain-warnings.txt.

`cg-index query` — Budgeted Knowledge Brain Retrieval

Use cg-index query when a workflow needs prior project knowledge without opening generated Brain partitions by hand:

cg-index query --intent plan --query "workflow token baseline" --budget 600 --format md
cg-index query --intent review --query "Pester safe runner" --changed-file tests/Run-Tests.ps1 --budget 600 --format json

The query mode returns a short answer, selected artifact paths, snippets, selection and exclusion reasons, stale or conflict flags, confidence, and a heuristic token estimate. It is local and deterministic; it does not use vector search, external services, or optional retrieval backends.

Command Output Summary Wrappers

Use the cg-*-summary wrappers when a workflow needs compact evidence from noisy local command surfaces while retaining the full source output on disk:

cg-test-summary --root . --format json
cg-diff-summary --root . --format md
cg-log-summary --root . --format json
cg-tree-summary --root . --max-entries 80 --format md
cg-problems-summary --root . --input diagnostics.json --format json

The wrappers are local stdlib tooling. They do not call external services, mutate GitHub, or replace required validation commands. cg-test-summary only reads existing tests/last-run.json; it does not run Pester, pytest, R, or Stata. Full raw/source outputs are redacted for common secret-looking patterns and written under .cg-docs/token/outputs/YYYYMMDD-HHMMSS-<kind>/. Keep that directory for short-lived validation evidence, not durable project knowledge; record final decisions in plans, reviews, work reports, and solutions instead.

`cg-token-audit` / `cg-audit-context` — Context and Model-Governance Audit

cg-token-audit --root . --output-dir .cg-docs/cost --format both --recommendations
python scripts/cg_audit_context.py [--root PATH] [--output-dir PATH] [--format json|md|both] [--baseline context-audit.json] [--recommendations] [--token-output-dir PATH] [--no-token-artifacts]

Inventories context-contributing files, estimates token burden with a chars/4 heuristic, checks executable model metadata and advisory provenance, detects duplicate paragraph blocks, and benchmarks the tracked /cg-* workflows plus Knowledge Brain/context lookup behavior.

Use --baseline with a previous context-audit.json to render before/after benchmark deltas. Use --recommendations to also write .cg-docs/cost/token-advice.md. The token regression check reports baseline when no previous comparable audit is supplied, pass when a comparable run has no deterministic guardrail failures, and fail when guardrail failures are present.

Workflow baseline artifacts:

Artifact	Purpose
`.cg-docs/token/TOKEN-BUDGET.md`	Human-readable workflow baseline, observability boundaries, and no-savings-claim policy.
`.cg-docs/token/TOKEN-DASHBOARD.md`	Compact maintainer dashboard with regression status, highest workflow budgets, and warning/context summaries.
`.cg-docs/token/token-audit.json`	Canonical JSON baseline payload with workflow telemetry, benchmarks, guardrails, and warning classifications.
`.cg-docs/token/context-map.json`	Workflow-to-context map of deterministic file, skill, agent, tool, and context-loading signals.
`.cg-docs/token/regression-check.json`	Machine-readable token regression status derived from deterministic guardrails and optional baseline comparison.
`.cg-docs/token/workflow-costs.csv`	Spreadsheet-friendly workflow rows for tracked workflows.
`.cg-docs/token/large-context-warnings.md`	Large prompt/instruction/skill and repeated-context warnings without copying large bodies.

Model-advisory guardrails report executable model metadata, invalid advisory provenance or effort labels, missing user-control language, and stale stage coverage. Runtime-only quantities such as command-output size, picker availability, and summary size remain explicit observed/not_observed fields until instrumentation exists.

Exit codes: 0 success, 1 fatal error, 2 missing or invalid project root.

Active-State Handoff Records

Long-running workflows may write a compact restart aid at .cg-docs/active-state/current.json. The schema is defined in .github/shared/active-state.contract.md. Records contain artifact paths, current phase, evidence status, unresolved decisions, and an exact nextCommand; they must not copy transcripts, raw command output, full review findings, or full report bodies. /cg-resume reads the record when present and validates referenced paths before using it. /cg-diagnose may include the same compact pointers in crash recovery handoffs, but remains read-only.

For token-optimization release candidates, complete .cg-docs/cost/token-optimization-release-checklist.md after generating the audit. Keep non-blocking issues in .cg-docs/cost/token-optimization-follow-ups.md so release blockers and future cleanup stay separate.

Model guidance: See Model Guide for stage capability profiles, effort suggestions, provenance, and user-controlled selection.

Project Charter: All /cg-* prompts automatically read compound-gpid.md at session start (if it exists). If missing, prompts remind you to run /cg-setup to optionally create one. Prompts work without a charter — the reminder is advisory.

Prior-work awareness: /cg-brainstorm checks .cg-docs/brainstorms/ and /cg-plan checks .cg-docs/plans/ for related prior work before starting. If a match is found, you can continue from it, follow up, or start fresh.

Task routing: Use /cg-light-work only for a qualified small technical task. It requires Plan approval before source edits, writes a Plan, Work Report, and mandatory fixed light Review Report, permits at most an initial Review pass plus one verification pass after Review fixes, and compounds only after explicit opt-in. Reproducible bugs use /cg-fixbug; research, statistical, and publication work uses /cr-*; larger, ambiguous, security-sensitive, schema, dependency, or destructive work uses /cg-brainstorm -> /cg-plan -> /cg-work. /cg-work executes approved saved Plans and redirects unmatched inline tasks without dispatching them.

Plugin Development (developer-only)

Consumer project users: The prompts below are for compound-gpid maintenance only. /cg-release and /cg-compound-gpid-rd appear in your autocomplete because they are distributed via junctions, but they will not run outside the compound-gpid repo — Step 0 stops them immediately. Do not use these prompts in consumer projects.

Prompt	Purpose	Distribution
`/cg-release [vX.Y.Z[.build]]`	Create a stable release from `main` or a four-component prerelease from `dev`. Detects the next tag unless an exact tag is supplied, drafts release notes from `.cg-docs/`, checks `SCHEMA_VERSION`, and publishes to GitHub Releases.	Distributed via junctions to consumer projects, but Step 0 stops execution immediately if not run inside compound-gpid.
`/cg-compound-gpid-rd`, `/cg-compound-gpid-rd --full`, `/cg-compound-gpid-rd --add <URL>`, `/cg-compound-gpid-rd --remove <id>`	Research public GitHub repositories for Compound GPID. Delta and full modes create reviews. Add and remove modes safely manage the registry without starting a review. `rd` means `research-development`; the current scope is public GitHub repository research for Compound GPID maintainers.	Distributed via junctions to consumer projects, but Step 0 requires the exact Compound GPID charter and stops before registry access, network access, utility calls, or writes in other projects.

Competitive Review System

/cg-compound-gpid-rd uses .cg-docs/competitive-reviews/repos.json to track monitored repositories and review state. Its four mutually exclusive invocation forms are:

Mode	Invocation	Behavior
Delta	`/cg-compound-gpid-rd`	Review releases newer than `lastReviewedRelease`.
Full	`/cg-compound-gpid-rd --full`	Perform a deep README and release review and create or refresh baselines.
Add	`/cg-compound-gpid-rd --add <URL>`	Validate and register one public GitHub repository without starting a review.
Remove	`/cg-compound-gpid-rd --remove <id>`	Remove one exact registry entry after exact case-sensitive confirmation without deleting assessment history.

Missing values, duplicate or combined mode flags, unknown flags, and extra values stop before work starts.

Adding a repository: Use --add; do not edit repos.json manually. The raw URL first passes a 1-164 character ASCII lexical allowlist and remains one process argument. A Python 3.8+ deterministic utility then validates the complete registry, normalizes the public URL to https://github.com/<owner>/<repository>, rejects duplicates, and returns a non-writing plan with exact before/after SHA-256 values. The prompt fetches only that returned canonical URL and requires a non-empty public repository page with the same final resolved URL. Redirects, private or deleted repositories, sign-in pages, 404 responses, empty content, and ambiguous responses stop before mutation.

After accessibility passes, the utility writes one entry with a derived unique id, canonical url, derived releasesUrl, derived unique shortName, and lastReviewedRelease: null. It omits lastReviewDate until a review succeeds. Add mode does not create an automatic baseline or assessment. The next review must be:

/cg-compound-gpid-rd --full

Removing a repository: Use --remove <id>; do not edit repos.json manually. The command runs a non-writing plan and shows the complete entry, ID, URL, and exact hashes. It writes only when the maintainer types the complete exact ID. Apply also requires the displayed URL and the plan's exact source SHA-256, so same-ID URL replacement and unrelated stale state are rejected. The utility removes only the registry entry. It retains all full-review assessments, delta reports, and other historical review files.

Add mode accepts an empty registry. Remove mode can remove the final entry. Full and delta modes stop on an empty registry and direct the maintainer to /cg-compound-gpid-rd --add <URL>. Delta mode also skips entries with null review state and directs the maintainer to run a full review.

The registry root must include "schemaVersion": "compound-gpid-competitive-reviews-v1".

Three-way schema coupling: Keep this exact v1 value synchronized across .cg-docs/competitive-reviews/repos.json, the validation step in .github/prompts/cg-compound-gpid-rd.prompt.md, and EXPECTED_SCHEMA_VERSION in scripts/cg_compound_gpid_rd_registry.py. Update and verify all three together for any future schema migration.

Review cadence: Run /cg-compound-gpid-rd (delta mode) every 1-2 weeks to check for new releases. Run /cg-compound-gpid-rd --full after adding a repository, when a baseline is needed, or for a periodic deep audit.

Outputs: Per-repo full-review files (.cg-docs/competitive-reviews/YYYY-MM-DD-<id>-full-review.md) and delta reports (.cg-docs/competitive-reviews/YYYY-MM-DD-delta-review.md). All four modes get ordered repository scope from the utility's bounded read-only state command. State and mutation responses expose before/after source SHA-256 values and scope digests. The scope digest covers ordered ID-to-URL identities, repository release/date projections, and root review state. Per-repository review state is changed only by deterministic review-repo check-only/apply calls. Each check-only call requires the last accepted chain SHA and exact prior release/date null, absent, or value state. Full root state is changed only by review-full check-only/apply calls; finalization requires the last accepted chain SHA and scope digest. No review mode directly writes repos.json. After a --full run, lastFullReview at the root of repos.json is set to today's date (YYYY-MM-DD), recording the last complete audit across all repos. On partial failure, lastFullReview is set to null and lastFullReviewNote is the deterministic ASCII value partial - <failed IDs in registry order>. lastFullReviewNote is removed on the next successful full run. Full-review and delta-report files are the durable review history. Entry-local lastReviewDate and lastReviewedRelease values describe only repositories that remain in the registry and are removed with an entry. lastFullReview reflects only the most recent successful full-suite run.

Transaction outcomes: Exit 0 means a valid response was flushed; fixed warning codes are committed success. Exit 1 is a definite precommit rejection with no writer dispatch. Exit 2 is invalid CLI syntax. Exit 3, timeout, missing or invalid output, stdout delivery failure, or unexpected post-dispatch stderr is ambiguous. The prompt invokes read-only state, compares exact before/after hashes and identity/review postconditions plus the scope digest, and never retries an ambiguous mutation automatically. Fetched release tags are accepted only when they match the bounded 1-128 character ASCII allowlist ^[A-Za-z0-9][A-Za-z0-9._+/-]{0,127}$; each release is passed as one quoted process argument, and the utility applies the same rule to new and expected release values.

Distribution note: /cg-compound-gpid-rd is distributed to consumer projects via junctions (it lives in .github/prompts/ along with all other prompts). It will appear in the Copilot Chat autocomplete for any project using compound-gpid. The Step 0 guardrail requires the exact Compound GPID charter and stops before registry access, network access, utility calls, or writes when invoked outside the development repo.

Review Agents

Agent	Focus
`cg-code-quality`	Style, linting, DRY, naming
`cg-testing`	Coverage, edge cases, test quality
`cg-documentation`	roxygen2/docstrings/do-file headers, README, comments
`cg-version-control`	Commit hygiene, branching, secrets
`cg-reproducibility`	Lockfiles, relative paths, seeds, repkit
`cg-performance`	Vectorization, memory, algorithm complexity
`cg-architecture`	Project structure, modularity, dependencies
`cg-data-quality`	Input validation, types, missing values
`cg-learnings-researcher`	Cross-reference past solutions (`full` / `thorough` alias only)
`cg-adversarial`	Adversarial testing: edge cases, data corruption, security (`full` / `thorough` alias only)

Review agents are primarily dispatched by /cg-review. /cg-verify-pr also dispatches @cg-testing (test failure analysis) and @cg-code-quality (build error analysis) as part of CI triage. Agents are NOT user-invokable and do not appear in the Copilot Chat agent dropdown.

Research Review Agents

Research agents are owned by suite-cr and dispatched conditionally by /cr-review. They are not imported into /cg-review.

Agent	Focus
`cr-research-integrity`	P0 silent research errors and integrity gates
`cr-provenance-audit`	Claim-evidence traceability and citation provenance
`cr-mathematical-verification`	Derivation-to-code consistency
`cr-identification-audit`	Identification strategy and required diagnostics
`cr-econometric-reasoning`	Structural and econometric model logic
`cr-ml-methodology`	Validation design, leakage, inference, and interpretation
`cr-specification-analysis`	Theory-data implications and specification discipline
`cr-measurement-integrity`	Indicator, threshold, clustering, and comparability integrity
`cr-academic-writing`	Economics-paper structure, notation, and citations
`cr-publication-output`	Publication tables, figures, notes, and deterministic output
`cr-replication-package`	Replication archive completeness, safety, and portability

ℹ️ For stage capability guidance and user-controlled effort selection, see Model Guide.

CR ML Skill

The cr-skill-ml-economics skill is a compact router with eight demand-loaded references. See the Research Skills catalog for ESL-led foundations, high-dimensional methods, split and evaluation guidance, causal ML, survey and panel qualifications, and R tidymodels/Python scikit-learn implementation references.

Review Routing Rules

/cg-review uses staged routing from changed-file risk signals:

Trigger	Resolved mode
Small docs/prompt wording/metadata-only or low-risk test changes	`light`
Ordinary implementation, prompt, or test changes without high-risk signals	`standard`
Pipeline/extract/load scripts, statistical functions (`fmean`, `fsum`, `fgini`, `svymean`, `reghdfe`, `lm`, etc.), summary tables, survey/poverty/welfare/weights/joins/aggregation, or reproducibility-sensitive changes	`data-risk`
Architecture, dependencies, module boundaries, performance, memory, API contracts, or large refactors	`architecture`
Authentication, secrets, credentials, release automation, publishing, install/update paths, linking/unlinking paths, schema changes, or destructive filesystem behavior	`full`
≥ 50 non-test lines changed with otherwise low risk	raises `light` → `standard`
≥ 200 non-test lines changed without higher-risk trigger	recommends `full` to user (does not auto-apply unless explicitly requested)

When any route fires, the prompt tells you the reason, resolved mode, and mandatory emphasis. Explicit review modes win when present: /cg-review light resolves to light, and /cg-review full resolves to full. Auto risk-class routing applies only when no explicit mode is requested. Agents should still mention high-risk signals in their review focus. thorough remains a backward-compatible alias for full.

Per-Finding Status Tracking

Review reports saved to .cg-docs/reviews/ include YAML frontmatter that tracks each finding's status:

---
plan: .cg-docs/plans/2026-04-01-my-feature.md
findings:
  P0.1: open
  P1.1: fixed
  P2.1: skipped
---

Status	Meaning
`open`	Not yet addressed — will appear in the next `/cg-fix-triage` run
`fixed`	Applied by `/cg-fix-triage`; excluded from future sessions
`skipped`	Deliberately deferred; `/cg-resume` still counts them as pending

Legacy review files (from before v0.4.3) do not have the findings: key. Run /cg-fix-triage --migrate once to backfill status tracking on all legacy files.

Used by /cg-review, /cg-fix-triage, and all review agents. Each finding gets a compound ID (e.g., P0.1, P1.2) for selective fixing.

Level	Label	Meaning	Action
P0	BLOCKING	Exploitable security vulnerability, PII/credential exposure, silent data corruption, incorrect statistical results	Immediate remediation required — must fix before anything else
P1	CRITICAL	Bugs causing incorrect behavior, missing critical validation, error handling gaps	Must fix before merge
P2	IMPORTANT	Performance problems, missing tests, poor documentation	Should fix
P3	MINOR	Style improvements, minor refactors, suggestions	Nice to have

Use /cg-fix-triage P0 to fix all blocking findings, /cg-fix-triage P0 P1 to fix blocking and critical, or /cg-fix-triage P1.2 P2.3 to fix specific IDs.

Plan Review Agent

Agent	Focus	Model	User-invocable
`@cg-plan-critic`	Plan review: assumptions, over-engineering, missing edge cases, scope creep, dependency accuracy, and phase structure (logical ordering, independent testability, completion criteria, cross-phase handoffs)	Sonnet 4.6	No

@cg-plan-critic is dispatched exclusively by /cg-plan-review. It is not user-invokable directly. It reads the plan and actual codebase to verify assumptions, checking for over-engineering, missing edge cases, scope creep, flawed dependencies, and (for phased plans) phase structure quality.

Project Scanner Agent

Agent	Focus	Model	User-invocable
`@cg-project-scanner`	Scans project file structure to detect languages, frameworks, project type, and charter-relevant content. Returns structured analysis for `/cg-setup` and other prompts	Claude Haiku 4.5	No

@cg-project-scanner is dispatched by /cg-setup (Phase 2) and other prompts that need project analysis. It is not user-invokable directly. It reads the project file tree, matches signals against the cg-skill-project-scanner catalog, extracts charter-draft content from README and DESCRIPTION, and returns a structured markdown report with language detection, project type, and per-question setup recommendations.

Release Scanner Agent

Agent	Focus	Model	User-invocable
`@cg-release-scanner`	Classifies commits by conventional commit prefix, lists relevant `.cg-docs/` entries within the scan window, and returns a structured categorized report for `/cg-release`	Claude Haiku 4.5	No

@cg-release-scanner is dispatched exclusively by /cg-release. It is not user-invokable directly. It receives the pre-collected git commit log and window parameters from the orchestrating prompt, classifies commits (feat/fix/docs/breaking), matches .cg-docs/ plan and solution entries by keyword, and returns a structured markdown report with Semver Impact recommendation and SCHEMA_VERSION signals.

Roadmap Agent

Agent	Focus	Model	User-invocable
`@cg-roadmap`	Manages `roadmap.json`: add/remove milestones and features, link plans, update statuses	Haiku 4.5	Yes

@cg-roadmap is the only agent users interact with directly (via @cg-roadmap in Copilot Chat) to manage your project roadmap. Prompts like /cg-issues, /cg-plan, and /cg-work dispatch it automatically for roadmap updates when roadmap.json exists.

`roadmap.json` Schema

Field	Type	Values
`milestones[].objective`	required string	One sentence describing why the milestone exists. `@cg-roadmap` validates this field is present before every write.
`milestones[].status`	derived	`planned`, `in-progress`, `done`
`features[].status`	set	`idea`, `planned`, `active`, `done`
`githubIssues`	object	Optional top-level GitHub Issues config block (see below)
`features[].github`	object	Optional per-feature GitHub issue linkage (see below)

Milestone status is computed by @cg-roadmap from feature statuses (never set directly by users). Feature active maps to milestone in-progress. After all features in a milestone are marked done, /cg-work dispatches @cg-roadmap to mark the milestone as done (see Step 3.8). IDs are kebab-case and immutable after creation. features[].plan is a nullable path to a .cg-docs/plans/ file.

`githubIssues` block (optional)

Stored as a top-level key in roadmap.json. All sub-fields are optional.

Field	Type	Default	Description
`enabled`	bool	`false`	Enable GitHub Issues integration for this project
`repo`	string	—	`owner/repo` identifying the GitHub repository
`labelPrefix`	string	`—`	Prefix for auto-created labels (e.g. `"cg:"`) — absent/null means no prefix
`autoCreate`	bool	`false`	If `true`, `/cg-issues backfill` may offer batch creation (still requires per-issue confirmation)

Configure with /cg-issues setup or /cg-setup. @cg-roadmap is the only agent that writes this block.

GitHub Issues are supplementary tracking handles for roadmap work items. The roadmap remains authoritative for feature status and plan links. Workflow prompts should only recommend issue setup/backfill/linking when useful; they do not create issues automatically.

`features[].github` block (optional)

Per-feature GitHub issue linkage. All sub-fields are optional; omit the entire block when no issue is linked.

Field	Type	Default	Description
`repo`	string	—	`owner/repo` — overrides top-level when issue lives in a different repo
`issueNumber`	integer	—	Positive integer GitHub issue number
`issueUrl`	string	—	Full URL: `https://github.com/owner/repo/issues/<number>`
`createdAt`	string	—	ISO date `yyyy-MM-dd` when the link was created

Attach with /cg-issues link or /cg-issues backfill. Adding github metadata never changes features[].status.

After an initial backfill, normal maintenance is delta-based: /cg-strategy, /cg-plan, /cg-work, and /cg-resume should surface missing issue links for newly relevant work and hand off to /cg-issues, while /cg-commit-push-pr uses linked issue metadata for Refs # or confirmed Closes # PR references.

Skills

The Skills Catalog is the public, goal-oriented inventory and documents availability labels and source-of-truth maintenance. The table below is retained as a compact technical summary.

Skill	Contents
`cg-skill-setup`	Project configuration wizard
`cg-skill-r-collapse`	collapse statistical computing: `fmean`/`fsum`/`fmedian`/`fnth` and all Fast Statistical Functions, GRP objects, TRA transformation types, `fwithin`/`fbetween`/`fscale`, `flag`/`fdiff`/`fgrowth`, `collap()`, `fsummarise`/`fmutate`. Dialect-neutral: works on data.table, tibble, and data.frame.
`cg-skill-r-datatable`	data.table manipulation: `DT[i,j,by]` syntax, `:=` in-place mutation, `fread`/`fwrite`, joins, `melt`/`dcast` reshaping, `.SD`/`.SDcols`, `fifelse`/`fcase`/`fcoalesce`, keys and indices.
`cg-skill-r-tidyverse`	tidyverse patterns: dplyr 1.2+ (`.by`, `join_by`, `across`/`pick`/`reframe`), native pipe `\|>`, `pivot_longer`/`pivot_wider`, `readr` I/O, `stringr`, `purrr`. Load for `r-syntax: "tidyverse"` projects.
`cg-skill-r-visualization`	ggplot2 + wbplot: World Bank visualization conventions, `theme_wb()`, `WBCOLORS`, `scale_color_wb_d()`, `scale_fill_wb_c()`, GPID chart types.
`cg-skill-r-analytical`	Analytical domain patterns: `haven` (Stata migration), `fixest` (econometrics), `modelsummary` (tables), welfare/poverty measurement, FGT indices, survey analysis. Syntax-neutral — works with any dialect.
`cg-skill-r-technical`	Infrastructure & packages: roxygen2, package dev, `plumber` APIs, `shiny`, `targets` pipelines, `httr2` clients, `renv`/`pak`. Syntax-neutral.
`cg-skill-r-shared`	Base R style rules universal to all dialects: `<-` assignment, `snake_case`, `TRUE`/`FALSE`, `rlang`/`cli` error handling.
`cg-skill-r-testing`	testthat 3+ patterns: `test_that()`, `describe()`/`it()`, fixtures, mocking (`local_mocked_bindings()`), snapshots, BDD-style testing. Dialect-aware: data.table examples for collapse/data.table projects, tibble examples for tidyverse projects.
`cg-skill-python-best-practices`	polars, numpy, pytest, type hints, `uv`/`poetry`
`cg-skill-stata-best-practices`	Comprehensive Stata reference: universal coding principles (compound quotes, macro expansion traps, stored results, `subpop()` vs `if`, clustering), data management, econometrics, causal inference, graphics, Mata, reproducibility (`repkit`: `repado`, `reproot`, `reprun`, `repscan`, `lint`), and 21 community packages (`reghdfe`, `estout`, `did`, `rdrobust`, etc.). ALWAYS load when writing or reviewing `.do` or `.ado` files.
`cg-skill-stata-testing`	Stata testing & reproducibility patterns: inline assertions (`assert`, `capture`, `_rc`, exit codes), data validation (`isid`, `duplicates`, `misstable`), econometric result verification (`_b[]`, `reldif`, `test`), `reprun`/`repkit` reproducibility workflows, test scaffolding (`foreach`, `preserve`/`restore`), and 9 testing-specific anti-patterns. Load when writing, reviewing, or debugging test blocks in `.do`/`.ado` files.
`cg-skill-git-workflow`	Branching, commits, PR templates, `.gitignore`
`cg-skill-brainstorming`	Requirement elicitation and decision capture
`cg-skill-compound-docs`	Knowledge capture and categorization system
`cg-skill-wb-report-writing`	Progressive-disclosure World Bank institutional report-writing router with source-pack preflight, marker safety, and operation/type routing across PRWP, policy briefs, executive summaries, flagship sections, country narratives, technical methodology, internal memos, and data blogs.
`cg-skill-fix-triage-migrate`	Migration mode for `/cg-fix-triage --migrate`: backfills `findings:` tracking frontmatter on legacy review files. Does NOT apply fixes.
`cg-skill-project-scanner`	Project scanner signal catalog for `/cg-setup`: language/framework detection (Tier 1), project type signals (Tier 2), charter-draft content extraction (Tier 3). Dispatched by `@cg-project-scanner`.
`cg-skill-brain-query`	Selective Knowledge Brain query protocol for relevance, stale/conflicting evidence, and source citation. Loaded by Consult Brain steps.
`cg-skill-pester-safety`	Compound GPID workspace safety rules for Pester execution and the canonical test runner. Internal and environment-specific.
`cg-skill-wiki`	Wiki manifest, ownership, managed-section, conflict, template, and conversion rules. Loaded before `@cg-wiki` operations.
`cg-skill-windows-cmd-python-detection`	Safe Python candidate detection and Windows Store stub rejection for `bin/*.cmd` launchers. Internal and Windows-specific.

Configuration

Run /cg-setup in Copilot Chat after running cg-link. The prompt asks:

Language: R, Python, Stata, or any combination
R syntax dialect (if R is selected): data.table-collapse (default) or tidyverse
Project type: Package, analysis, dashboard, API, tool
Review depth: Light, standard, or thorough (thorough is treated as the full route)
Project charter (optional): project name, objective, deliverables, constraints

This creates compound-gpid.local.md (gitignored, user-specific config), optionally compound-gpid.md (committed, shared project charter), and optionally compound-gpid.context.md (committed, growing project knowledge base) in your project root, and scaffolds the .cg-docs/ directory.

Configuration Fields

All fields are stored as YAML frontmatter in compound-gpid.local.md:

Field	Values	Description
`language`	`"r"`, `"python"`, `"stata"`, `"both"`, or combination	Language(s) used in the project
`r-syntax`	`"data.table-collapse"` (default), `"tidyverse"`	R dialect for skill routing. Determines which R syntax skills are loaded for `.R` files. Use `"tidyverse"` for projects with external coauthors who only know dplyr.
`project-type`	`"package"`, `"analysis"`, `"dashboard"`, `"api"`, `"tool"`	Project type
`review-depth`	`"light"`, `"standard"`, `"thorough"`	Legacy depth default for `/cg-review`; `thorough` maps to the `full` route. Explicit routed modes can be passed at invocation time.
`suites`	`[cg]` (default/absent), `[cr]`, `[cg, cr]`	Active suite configuration for the modular architecture. Selects which workflow prompts/skills are loaded into routine sessions; the generator context-budget filter and `cg_migrate_config.py` use this field.
`artifact-html`	`true`, `false`	Explicit opt-in for automatic Brainstorm/Plan and generic Markdown HTML writes. Missing or invalid values default disabled (invalid values warn). Validation, explicit render, and `--check` remain available.
`cg-schema-version`	date string	Auto-managed by `cg-update`. Do not edit manually.

`compound-gpid.context.md`

A committed, growing knowledge base for project-specific context. Created by /cg-setup. Extended by /cg-compound after each significant task. Ordinary prompts load targeted headings or snippets when tactical facts are relevant instead of reading the whole file by default.

Typical contents: data source locations and caveats, domain vocabulary, workspace folder descriptions, variable-level notes, recurring gotchas. Unlike the charter (compound-gpid.md), compound-gpid.context.md has no fixed structure — organise it by topic.

Directory Structure

After linking and configuring, your project will contain:

your-project/
├── .github/
│   ├── prompts/              → junction to C:\WBG\.compound-gpid\.github\prompts\
│   ├── skills/               → junction to C:\WBG\.compound-gpid\.github\skills\
│   ├── agents/               → junction to C:\WBG\.compound-gpid\.github\agents\
│   ├── instructions/         → junction to C:\WBG\.compound-gpid\.github\instructions\
│   ├── shared/               → junction to C:\WBG\.compound-gpid\.github\shared\
│   ├── copilot-instructions.md  # generated from template (managed marker); regenerated by cg-link/cg-update
│   └── workflows/            # your own GitHub Actions (untouched by cg-link)
├── .claude/                  # Claude Code install units: commands/skills/agents linked; root files copied if managed
├── .agents/                  # Codex install units: commands/skills/subagents linked; root files copied if managed
├── .opencode/                # OpenCode install units: commands/skills/agents linked; config copied if managed
├── .kilo/                    # Kilo install units: commands/skills/agent linked; config copied if managed
├── .compound-gpid/managed-files.json  # sidecar checksums for copied strict config/root files
├── compound-gpid.md          # Project charter (4 sections: Objective, Key Deliverables, Constraints, Current Focus). YAML: project-name, team, created, last-reviewed. Committed -- shared.
├── compound-gpid.context.md  # Growing project knowledge base (data sources, domain vocab, workspace notes). Committed -- institutional memory.
├── compound-gpid.local.md    # Your user config (gitignored)
├── roadmap.json              # Milestone & feature tracker (committed)
├── c-research/               # CR research outputs, created when the cr suite is active
│   ├── evidence/             # Source provenance and claim-evidence artifacts
│   ├── scoping/              # Research scoping memos
│   ├── normative-decisions/  # Per-study decision registers
│   ├── derivations/          # Mathematical derivations
│   ├── specifications/       # Specification and ML search ledgers
│   ├── results/              # Result manifests and outputs
│   ├── manuscripts/          # Working research drafts
│   └── replication/          # Replication materials
└── .cg-docs/                 # Compound GPID knowledge base (committed -- institutional memory)
    ├── archive/              # Archived charter sections removed by the user (not loaded at session start)
    ├── brainstorms/          # /cg-brainstorm outputs
    ├── competitive-reviews/  # /cg-compound-gpid-rd registry (repos.json) and assessment outputs
    ├── cost/                 # context/model audit reports and release-readiness checklists
    ├── inbox/                # unprocessed strategy ideas; not approved roadmap items until promoted via /cg-strategy
    ├── plans/                # /cg-plan outputs
    ├── reviews/              # /cg-review outputs (review reports for /cg-fix-triage)
    ├── strategy/             # /cg-strategy session records
    └── solutions/            # /cg-compound outputs
        ├── build-errors/
        ├── bugs/
        ├── performance-issues/
        ├── testing-patterns/
        ├── data-quality/
        ├── environment-issues/
        └── git-workflows/

.cg-docs/inbox/ is only a holding area. Do not treat files there as approved roadmap items until a separate strategy or roadmap session promotes them.

Archive file format (.cg-docs/archive/charter-history.md): Content removed from the charter is appended with a date heading and source section label:

## Archived YYYY-MM-DD
**Removed from**: <section name>
<removed content>

Something not working? See Troubleshooting.

`.cg-docs/` Document Frontmatter Schema

Each document type in .cg-docs/ uses a defined set of status enum values. These are enforced by convention now and will be validated automatically in a future evals milestone.

Document type	Path	Valid `status` values
Brainstorm	`.cg-docs/brainstorms/`	`open`, `decided`, `abandoned`
Plan	`.cg-docs/plans/`	`draft`, `active`, `completed`, `abandoned`
Solution	`.cg-docs/solutions/`	`draft`, `applied`
Review	`.cg-docs/reviews/`	Per-finding status in `findings:` frontmatter key: `open`, `fixed`, `skipped`
Verify Review	`.cg-docs/reviews/` (filename: `<stem>-verify-review.md`)	Same `findings:` map as Review, plus `parent-review: <path>` (prior review file) and `type: verification`

Plan Frontmatter: Phase Fields

Phased plans (created with /cg-plan when the user requests phases) carry additional frontmatter fields for tracking cross-session execution:

Field	Type	Written by	Read by	Notes
`phases`	integer	`/cg-plan` Step 3.5	(not read at runtime — informational only for human readers)	Convenience hint — may become stale if phases are restructured. The authoritative phase count is always derived by counting `## Phase` headers in the document body. Never use this field as the source of truth for validation.
`completed-phases`	YAML flow sequence of unquoted integers, e.g. `[1, 2]`	`/cg-work` Step 2.5	`/cg-work` Step 1.2, `/cg-resume` Step 2a	Authoritative completion record. Written first at phase boundary (before `current-phase`). A plan with a non-empty list and `status: active` is "paused between phases" — this is the normal cross-session state.
`current-phase`	integer	`/cg-work` Step 2.5	(informational only)	Written after `completed-phases`. Set to N+1 after completing phase N; removed when the final phase completes.

Navigation Menu

Repository navigation

reference.md

reference.md

Reference

Native Target Packaging

Ownership manifest schema

Verification and release gates

Shell Commands

Version Management

Copilot Chat Prompts

`/cg-commit-push-pr` Base And Preflight Contract

`/cg-verify-pr` Exact Diagnosis And Repair Contract

Research Suite Commands

`cg-index --brain` — Diagnostic Warnings

`cg-index query` — Budgeted Knowledge Brain Retrieval

Command Output Summary Wrappers

`cg-token-audit` / `cg-audit-context` — Context and Model-Governance Audit

Active-State Handoff Records

Plugin Development (developer-only)

Competitive Review System

Review Agents

Research Review Agents

CR ML Skill

Review Routing Rules

Per-Finding Status Tracking

Plan Review Agent

Project Scanner Agent

Release Scanner Agent

Roadmap Agent

`roadmap.json` Schema

`githubIssues` block (optional)

`features[].github` block (optional)

Skills

Configuration

Configuration Fields

`compound-gpid.context.md`

Directory Structure

`.cg-docs/` Document Frontmatter Schema

Plan Frontmatter: Phase Fields

FilesExpand file tree

reference.md

Latest commit

History

reference.md

File metadata and controls

Reference

Native Target Packaging

Ownership manifest schema

Verification and release gates

Shell Commands

Version Management

Copilot Chat Prompts

/cg-commit-push-pr Base And Preflight Contract

/cg-verify-pr Exact Diagnosis And Repair Contract

Research Suite Commands

cg-index --brain — Diagnostic Warnings

cg-index query — Budgeted Knowledge Brain Retrieval

Command Output Summary Wrappers

cg-token-audit / cg-audit-context — Context and Model-Governance Audit

Active-State Handoff Records

Plugin Development (developer-only)

Competitive Review System

Review Agents

Research Review Agents

CR ML Skill

Review Routing Rules

Per-Finding Status Tracking

Plan Review Agent

Project Scanner Agent

Release Scanner Agent

Roadmap Agent

roadmap.json Schema

githubIssues block (optional)

features[].github block (optional)

Skills

Configuration

Configuration Fields

compound-gpid.context.md

Directory Structure

.cg-docs/ Document Frontmatter Schema

Plan Frontmatter: Phase Fields

`/cg-commit-push-pr` Base And Preflight Contract

`/cg-verify-pr` Exact Diagnosis And Repair Contract

`cg-index --brain` — Diagnostic Warnings

`cg-index query` — Budgeted Knowledge Brain Retrieval

`cg-token-audit` / `cg-audit-context` — Context and Model-Governance Audit

`roadmap.json` Schema

`githubIssues` block (optional)

`features[].github` block (optional)

`compound-gpid.context.md`

`.cg-docs/` Document Frontmatter Schema