Garelier control library

Garelier Library Control

This skill turns the currently running AI into a project librarian. It uses the same curated knowledge trees, registries, provenance/import-export rules, templates, and graph as full Garelier's Librarian, without enabling roles, shelf branches, Dock, or driver.

It composes with garelier-control-project: project control manages durable project authority under __garelier/<pm_id>/control/; garelier-control-library manages curated knowledge in two tracked layers (DEC-077) — the shared, project-wide __garelier/__atmos/knowledge/ tier plus the additive per-pm __garelier/<pm_id>/knowledge/ layer.

Together they form the standalone Garelier Control management plane. Full Garelier can be added later, but is not required to keep using this plane.

Activation

1. Read ../garelier-librarian/knowledge_contract.md.
2. Read the knowledge.toml knowledge marker and the relevant indexes and registries when present.
3. Resolve the staging pm_id:
   • use the id explicitly named by the user;
   • otherwise use the sole evident control/full-PM namespace;
   • if multiple namespaces exist, list them and ask which staging/management context to use; never silently choose one;
   • otherwise default to _workshop.
4. Treat the shared __garelier/__atmos/knowledge/ tier and the per-pm __garelier/<pm_id>/knowledge/ layer as tracked curated knowledge, and __garelier/<pm_id>/runtime/librarian/ as local-only working data.

Knowledge has two tracked layers (DEC-077): the shared, project-wide __garelier/__atmos/knowledge/ tier and an additive per-pm __garelier/<pm_id>/knowledge/ layer. So choosing a pm_id selects both the staging/management context AND that pm's additive knowledge layer — resolution is shared-priority + per-pm-additive (same knowledge_id → shared wins by default; the per-pm layer ADDs ids absent from shared, and overrides a shared topic only via an explicit, auditable override_shared: true opt-in). This skill remains usable from a separately launched AI after full Garelier starts.

Initialize

garelier library-init --project <project-root> --pm-id _workshop

Initialization is no-overwrite. It creates the knowledge marker, registry templates, an empty project category/index, and gitignored staging directories. It does not seed all full-Garelier policies unless the user asks for them.

Retrieve

• Never bulk-read the knowledge trees. Full-tree loading is a context-budget defect.
• Start from role_index.toml when the question belongs to a known role.
• Otherwise start from the relevant category index.md.
• Use graph/registry metadata to narrow candidates, search headings/terms, then read only the necessary section of the smallest candidate set.
• Open a complete topic only when the full rule set is needed.
• Follow source/routine relationships and return compact path:line pointers with one-line conclusions. Stop once authoritative pointers answer.
• If curated knowledge does not answer, say not covered and identify the correct next step. Do not invent or silently adopt external information.

Curate

• Use garelier-librarian/templates/knowledge_document.md for new/materially updated topics and knowledge_index.md for category indexes.
• Register external sources before adoption and keep original project wording.
• Keep index, role, source, routine, and runbook references consistent.
• The user/knowledge owner decides rule meaning and exceptions; this AI organizes and applies those decisions.
• Read references/library-management.md for the detailed workflow.

Import / export

Read references/import-export.md.

• Clean bundles use garelier-librarian/scripts/knowledge_import and knowledge_export.
• Messy input is staged, inventoried for provenance/rights/conflicts, normalized into canonical documents, graph-validated, and only then promoted into tracked knowledge.

Validate and graph

garelier knowledge-graph --project <root> --pm-id <pm_id> --validate

Use --format mermaid or --format json for the derived graph. Full Garelier's Status Web Knowledge page uses the same graph. The graph contains metadata and pointers only; it deliberately does not embed topic bodies.

Boundaries

• This skill curates knowledge only. If the request actually needs code execution or project-management state, route per ../garelier-core/references/entry_routing.md.
• Do not treat unregistered external content as authoritative.
• Do not change a rule's meaning or approve an exception without the owner.
• Do not commit/export raw data, secrets, PII, unknown-rights content, or unresolved conflicts.
• Do not write durable curated knowledge into runtime/librarian/.