Release v0.166.0 · nstfkc/gemi

AI: `DatabaseAgentStore`, threads in the database (#1063)

New API (minor). Nothing to change in apps. See AI Thread Store.

DatabaseAgentStore from gemi/ai implements AgentStore over two tables of the app's database (gemi_agent_threads, gemi_agent_messages; Prisma models in the docs, or createTable()), on SQLite, Postgres and MySQL 8 / MariaDB 10.6+. It supports regenerate (removeMessages) and compaction summaries with lockSummary.
Extra methods for the app's own thread list: listThreads({ userId, limit, offset }), getThread(id) and deleteThread(id). createThread({ userId }) stores the user id, and the first user message becomes the thread's title.
Opt in by setting store on the controller to one instance made at module scope. MemoryAgentStore stays the default.

New API (minor). Nothing to change in apps. See ORM → Describing the schema for a prompt.

describeSchema(options?) from gemi/orm writes the registered models (or the schemas you pass) as one line per model: table, columns (database names under @map), keys and relations. describeSchemaModels(options?) returns the same as data. Options: models, exclude (models or Model.field), sensitive (Model.field), sensitiveColumns: "mark" | "omit".
/// @gemi.sensitive on a Prisma field emits sensitive: true in the generated artifact. Only describeSchema reads it; encrypted columns count as sensitive too. The artifact of a schema that doesn't use it is unchanged, and the artifact version is not bumped. Re-run prisma generate after adding the annotation.

New optional API (minor). Nothing to change in apps. See AI Tool Output for the Model.

New toModel(output) on AgentTool.create (server tools): the model reads its value instead of the whole output, while the client, onMessage and the store still get output. The value is kept on the ok result as the new modelOutput field, computed once when the result is produced (and when a background job settles).
The provider request and compaction summaries send modelOutput when a result has one, and output otherwise, as before.

New optional APIs (minor). Nothing to change in apps. See AI Tool Approvals → Showing a preview and → Who approved a call.

New approvalPreview(input, ctx) on AgentTool.create (needs requiresApproval): computes a dry run to show before the user approves. Its value is the pending call's new preview field, on the awaiting-input frame, in useChat's pending and in the stored AgentMessage.pending. Display only: not signed, so execute re-checks what it relies on. A preview that throws is logged and the call is asked without one.
New ctx.approval (ToolApprovalRecord: { approved, at, by, reason? }) on the ToolContext of a call that runs because a person approved it. by is the run's subject (runOwner).
The ToolResultPart of a call a person approved or refused now carries the same record as approval. Stores that keep the whole message keep it; nothing else changes on results of calls that never asked.
New onToolResult(result, ctx) hook on AgentController and in defineAgentHooks: once per tool result on the run's stream (top-level calls; a background job settling later is not one).

New API (minor). Nothing to change in apps. See ORM → Read-only transactions.

DB.readOnly(options?, fn) and DB.connection(name).readOnly(options?, fn) run fn in a transaction that starts with SET TRANSACTION READ ONLY and the SET LOCALs for statementTimeoutMs, lockTimeoutMs, idleInTransactionTimeoutMs and role. It is always rolled back, and returns what fn returned. Postgres only; refused inside an open transaction.
New exported type ReadOnlyOptions.