Storage Schema Change

Keep persistence portable across SQLite and PostgreSQL, keep every stored value typed, and put each piece of data in the column whose readers need it.

Authoritative inputs

Apply the shared engineering preparation first; reuse any completed discovery.

1. Read AGENTS.md and TERMINOLOGY.md.
2. Read the storage/ section of ARCHITECTURE.md, including "Schema, migrations, and where data goes". Treat it as the source of truth; this skill does not replace it.
3. Read crates/agentic-server-core/migrations/ in order, storage/schema.rs, and the affected storage/models/, storage/types/, and store files.

Checklist

1. Layering. SQL lives only in storage/models/, as one function per operation over one row struct per table. Rows convert to Rust types only in storage/types/, through one conversion pair per column. Stores take and return typed values; only executor/modes/ calls them.
2. No backend in queries. Model SQL is portable over sqlx::Any. Backend checks belong in pool.rs (connection tuning) and schema.rs (migration and verification), plus the existing lock_in_tx. Reject dialect functions in queries; restructure into portable queries or move the work into Rust.
3. Migration. A schema change is a new numbered SQL file that runs on both backends. Released migrations are never edited. Prefer ALTER TABLE ADD/DROP COLUMN to rebuilding a table, and drop a column's indexes first. New columns on tables that may hold rows are nullable or have a default; new integer columns are BIGINT.
4. Incompatible data. If existing rows can no longer be read, raise MINIMUM_SUPPORTED_MIGRATION instead of adding decoders for the old format, and document that the default SQLite database is recreated and other databases are rejected.
5. Column or metadata. responses.metadata holds only what a continuation needs and stays small, because every continuation decodes it and session checkpoints retain it. items.metadata holds storage facts about an item, never content, and items.data is the item's plain serde JSON. Data read by a different path (retrieval, for example), large data, data used in SQL predicates, ordering, joins, or indexes, data that needs a constraint, and data updated on its own each get a dedicated column. Reject a second copy of data another column holds.
6. Typed values. Stored JSON round-trips through a typed Rust struct or enum. No serde_json::Value at a store boundary, no hidden keys inside wire-shaped JSON, and no placeholder arguments.
7. Everything that names the column. Check the row struct, its SQL (including explicit column lists), the types/ conversion, the supervisor verification column list and counts in schema.rs, the readiness probe, ARCHITECTURE.md, and the deployment docs when operators are affected.
8. Tests. Cover a fresh SQLite database, the ignored PostgreSQL suites (postgres_* tests, run with TEST_POSTGRES_URL and --ignored), supervisor-mode verification, and the outdated-database path when the minimum migration changes. Re-run the storage benchmarks when a hot read or write path changes.