git:20260806.d57259b to git:20260819.a5a7812

---

name: refresh

description: Maintain the repo's knowledge — `index` rebuilds the ticket catalog (INDEX.md), `context` rebuilds the domain knowledge pack (documentation/). Day-to-day, hooks keep these fresh automatically.

argument-hint: index [--all | TICKET-ID ...] | context [--refresh] [--warehouse <name>] [schema-or-scope] | all

allowed-tools: [Read, Write, Bash, Glob, Grep]

---

# /refresh

One maintenance skill for the two knowledge stores the other skills read:

- **`/refresh index`** — the ticket catalog: `tickets/INDEX.md` (every ticket, status, one-line

summary, tags, cross-refs) + `tickets/OBJECTS.md` (object → tickets reverse map). This is what

`/ticket`'s recall mines. Details: [index.md](index.md).

- **`/refresh context`** — the domain knowledge pack in `documentation/`: data catalog, DDL, ERD,

dependency graph, business glossary. This is what `/ticket`'s domain + warehouse priming reads.

- **`/refresh all`** — both.

## You rarely need this

Day-to-day the index maintains itself: the PostToolUse hook re-renders `INDEX.md` on every ticket

folder change, the SessionStart hook surfaces the catalog, and `/ship` curates each ticket's

summary at close. Invoke `/refresh` to:

- **bootstrap** an existing backlog (`/refresh index --all` — render every ticket, then enrich);

- **re-enrich** a batch after bulk edits (`/refresh index ID1 ID2 …`);

- **build or update the knowledge pack** (`/refresh context`, scoped to a schema if given;

`--refresh` updates in place — overwrite, don't sprawl).

## Mode: `index`

Follow [index.md](index.md): render deterministically (`bin/build_ticket_index.py` — stdlib, no

model), enrich the un-enriched/stale set (the model writes one curated record per ticket, ingested

via `bin/ingest_index_records.py`), re-render, then verify with `--check` and commit all three

index files together (`INDEX.md`, `OBJECTS.md`, `index_data.json`).

## Mode: `context`

Follow [context-pack.md](context-pack.md): verify each configured warehouse target, introspect via its adapter

(inventory → DDL → dependencies → usage ranking), generate `data_catalog.md` / `erd.md` /

`glossary.md` into `documentation/` with freshness stamps, and update

`documentation/AI_LAYER_INDEX.md`. Don't invent business meaning — flag glossary gaps for a human

(`reduce_assumptions`).