Add Store Query

Add a query (and optionally a migration) to the store

pkg/store/ is the SQLite layer. Queries are SQL string constants in queries.go; methods on *Store execute them through s.Tx(...) for writes or s.db.QueryContext(...) for reads. Schema lives in migrations/000N_*.sql, embedded via embed.FS and applied at startup.

There are two workflows here. Pick the right one before you start.

Need new SQL only?  ──> Workflow A (query)
Need a new column / table / index?  ──> Workflow A + Workflow B (migration)

Workflow A — add a query and method

Three files. Same pattern every time.

Query const

Constants are grouped at the bottom of queries.go next to related queries. Name them q<Verb><Noun> — e.g., qSelectColdNodes, qBoostTemperature, qIncrementAccess. No package prefix; the file is the namespace.

const qListSourceFiles = `SELECT DISTINCT source_file FROM nodes ORDER BY source_file`

For batch queries with a variable IN list, follow the qBoostTemperatureBatchPrefix pattern: define the prefix as a const, append placeholders + ) in the calling method.

Method shape

Two flavors based on read vs write.

Reads call s.db.QueryContext directly, defer rows.Close(), and return through collectRows (or a dedicated row-scanner if the columns differ from nodeColumns). Reads do not wrap in s.Tx(...).

func (s *Store) ListSourceFiles(ctx context.Context) ([]string, error) {
    rows, err := s.db.QueryContext(ctx, qListSourceFiles)
    if err != nil {
        return nil, err
    }
    defer func() { _ = rows.Close() }()

    var out []string
    for rows.Next() {
        var path string
        if err := rows.Scan(&path); err != nil {
            return nil, err
        }
        out = append(out, path)
    }
    return out, rows.Err()
}

Writes wrap in s.Tx(ctx, func(tx *sql.Tx) error { ... }) so the FTS5 triggers fire atomically with the main update. See BoostTemperature / DecayTemperatures in temperature.go.

func (s *Store) MarkVisited(ctx context.Context, id string) error {
    return s.Tx(ctx, func(tx *sql.Tx) error {
        _, err := tx.ExecContext(ctx, qMarkVisited, time.Now().Unix(), id)
        return err
    })
}

For writes that need RowsAffected, capture it inside the closure via a named-scope variable; see DecayTemperatures for the pattern.

Test

store_test.go defines a local unexported openTestDB(t) helper (the public testutil.OpenTestDB is for cross-package callers like integration tests; inside pkg/store itself, use the local helper to avoid an import cycle). Seed rows through other store methods (don't poke SQL directly inside the test body), exercise the new method, assert. Mirror the existing tests for tone — minimal, table-driven where it helps.

Workflow B — add a migration (when schema changes)

Migrations are numbered SQL files in migrations/, loaded via //go:embed *.sql (migrations/migrations.go). They run in lex order at startup, so file naming carries the order.

Naming

000N_<verb>_<noun>.sql — zero-padded to 4 digits, snake_case body. Existing files: 0001_init.sql, 0002_compile_root.sql. Next is 0003_<verb>_<noun>.sql.

Content rules

• IF NOT EXISTS / IF EXISTS everywhere. Migrations may be applied to a partially-initialized DB; idempotency is the only safe assumption.
• One logical change per migration. Adding a column and adding an index = two migrations, two files. Easier to roll forward, easier to read in git log.
• Constraints belong on the column, not in a separate migration. NOT NULL, DEFAULT, REFERENCES go on the ALTER TABLE ADD COLUMN line.

FTS5 trigger sync — the easy-to-miss bit

nodes_fts is a virtual FTS5 table whose content is mirrored from nodes via three triggers (nodes_fts_insert, nodes_fts_delete, nodes_fts_update in 0001_init.sql:52-67). If your migration adds a searchable column to nodes, you must:

1. Add the column to the nodes_fts USING fts5(...) declaration — but FTS5 virtual-table column lists can't be ALTERed. The migration has to drop and recreate nodes_fts, then rebuild by triggering the insert path for every existing row.
2. Update all three triggers to reference the new column.

This is invasive. Reach for it only when the column truly needs full-text matching. For metadata-only columns, leave nodes_fts alone.

If you only need indexable but non-FTS lookup, add a regular CREATE INDEX IF NOT EXISTS idx_nodes_<col> ON nodes(<col>); line — see the existing idx_nodes_temperature, idx_nodes_parent etc. at the bottom of 0001_init.sql.

Verify the migration applies cleanly

go test ./pkg/store/... -run TestMigrate    # must pass
go test ./...                                # full suite picks up schema drift

Quick reference

Workflow A (query only):
1. pkg/store/queries.go        (q<Verb><Noun> const)
2. pkg/store/<domain>.go       (method on *Store)
3. pkg/store/store_test.go     (Test<Method>_<Case>)
4. go test ./pkg/store/...

Workflow B (schema):
1. migrations/000N_<verb>_<noun>.sql   (one logical change, IF NOT EXISTS)
2. If touching nodes searchable cols: rebuild nodes_fts + update 3 triggers
3. Then Workflow A for the read/write methods that consume the new shape
4. go test ./...

Common mistakes

• s.db.ExecContext for writes instead of s.Tx(...). FTS5 triggers run inside the implicit transaction; raw Exec may still work for simple updates but breaks atomicity guarantees and is inconsistent with the codebase. Use s.Tx.
• Naming the const qInsertSomething when it's an UPDATE. The verb has to match the SQL operation. Reviewer-grep relies on it.
• Adding a column without an index, then querying it. SQLite does table scans on un-indexed columns. If your new method's WHERE clause filters on the column, add CREATE INDEX IF NOT EXISTS in the same migration.
• Touching nodes_fts without rebuilding. A schema change to nodes searchable columns leaves nodes_fts stale until a full rebuild. Search results will silently drop the new column from ranking.
• Forgetting IF NOT EXISTS on CREATE TABLE / CREATE INDEX. Re-running migrations on a partially-applied DB will fail without it. Every existing line in 0001_init.sql uses it.
• Using time.Now() in the SQL string. Use SQLite's unixepoch() function (see qBoostTemperature) or pass time.Now().Unix() as a bind arg. String concatenation is both wrong and the start of an injection bug.

Cross-references

• .claude/rules/go-concise.md — error wrapping (action errors get failed to <verb>:), method receiver consistency, no premature interface
• .claude/rules/git-versioning.md — feat(store): add MarkVisited query is one commit; feat(migration): 0003 add visited_at column is a separate commit if it ships independently. Migration files are append-only — never edit a committed migration; write a new one
• .claude/skills/add-mcp-tool/SKILL.md — when a new tool needs a new query, do this skill first