protoc-gen-go-jet

name

protoc-gen-go-jet

description

Author Postgres migrations when a persisted proto changes, manage the protoc-gen-go-jet generation pipeline, wire new resources, and diagnose drift. The `protoc-gen-go-jet diff` subcommand either copies ddl.sql verbatim (first-ever migration) or emits an LLM brief (the caller authors the ALTER SQL beneath it); `protoc-gen-go-jet check` re-runs the whole chain and fails if it diverges from ddl.sql. Trigger when a proto with `(gojet.v1.table)` changes, a new persisted resource is added, or `protoc-gen-go-jet check` reports drift.

protoc-gen-go-jet — full workflow

Proto is the API + schema declaration. The plugin turns annotated messages into a complete Postgres storage layer: DDL, jet model/table, mapper, pagination schema, aliases with embedded DDL. Migrations are authored by the human/LLM operator, scaffolded by protoc-gen-go-jet diff with a structured brief. protoc-gen-go-jet check is the machine-verified guardrail.

Files you care about

File	Owner	Regenerated	Lifecycle
Your annotated *.proto files with (gojet.v1.table)	Human	never	Edit during API design
<proto_pkg>/storage/<resource>_ddl.sql	Plugin	every buf generate	Canonical SQL; reference, not applied
<proto_pkg>/storage/<resource>_{mapper,schema,aliases}.go	Plugin	every buf generate	<Prefix>Model, <Prefix>Table, <Prefix>ListSchema, <Prefix>DDL (go:embed), <Prefix>UpdateAll()
<proto_pkg>/storage/jet/public/{model,table}/*.go	Plugin (internal)	every buf generate	Emitted by jet during plugin run
Your migrations directory: NNNN_*.up.sql	Human (with scaffold)	only when proto changes shape	Forward-only deploy mechanism

One hand-written consumer tree lives next to your generated storage package: repository.go, repository_pg.go, repository_memory.go, *_test.go — they import the generated package under a gen alias for everything.

When to trigger this skill

1. Proto field added/removed/retyped on a message with (gojet.v1.table). Regenerate, scaffold a migration with diff, and review it.
2. New annotated message added. The diff scaffold detects the missing table and emits the CREATE.
3. New proto package with annotated messages (rare). No tooling changes — the plugin discovers everything from proto. You only point --ddl-roots at the new storage package.
4. protoc-gen-go-jet check reports drift. The failure is a path-qualified structural diff — translate it into the missing migration.

The workflow

There is no task runner. Three explicit steps drive the whole loop: regenerate, scaffold the migration, drift-check.

# 1. Edit the proto, then regenerate everything (runs the plugin as a buf plugin).
buf generate

# 2. Pick the next sequence number yourself, scaffold the migration into a new file.
protoc-gen-go-jet diff \
    --ddl-roots=<dir-with-*_ddl.sql> \
    --migrations=<your-migrations-dir> \
    [--tenant-role=<role>] \
    > <your-migrations-dir>/NNNN_<short_snake_desc>.up.sql

# 3. Author the SQL (brief mode only — see below), then drift-check.
protoc-gen-go-jet check \
    --ddl-roots=<dir-with-*_ddl.sql> \
    --migrations=<your-migrations-dir> \
    [--tenant-role=<role>]

buf generate runs protoc-gen-go-jet as a buf plugin (see the repo's buf.gen.*.yaml templates and cmd/protoc-gen-go-jet/e2e/buf.gen.yaml for a worked template). One pass emits ddl.sql + mapper + schema + aliases + the go-jet model/table packages. Docker is required — the plugin boots an ephemeral Postgres internally to drive go-jet's schema introspection.

--ddl-roots is one or more comma-separated directories containing the plugin-managed *_ddl.sql files. --migrations is your forward-only migrations directory. --tenant-role names the runtime PG role your RLS policies attach to; drop it if the schema isn't tenant-scoped. Both diff and check boot ephemeral Postgres containers per invocation — nothing persists between runs.

diff has two modes

The diff subcommand dispatches on whether your migrations directory already has non-empty migrations:

Initial mode — no non-empty migrations under the target directory. The plugin concatenates every plugin-managed *_ddl.sql (strip header, preserve order) into the scaffold. Applying it reproduces ddl.sql 1:1. No LLM authoring needed; check passes automatically.

Brief mode — existing migrations present. The plugin boots one ephemeral Postgres, applies the existing migration chain, dumps the resulting schema via pg_dump --schema-only, and writes a structured brief to stdout:

• Migration preamble (boilerplate)
• LLM BRIEF banner pointing back to this skill
• Active PLUGIN-RENAME directives (one line per rename)
• CURRENT section — full pg_dump output of the migration-chain-produced schema, SQL-commented
• TARGET section — concatenated *_ddl.sql, SQL-commented
• TODO(llm): marker

The LLM (or human) authors forward-only ALTER/CREATE/DROP SQL beneath the TODO marker. check then spins up two cold PGs, applies every migration to one and every ddl.sql to the other, and fails if any plugin-managed table diverges. That diff is the authoritative verdict — if it's clean, the authored SQL is correct.

Leaving the TODO marker in place will fail check because the migration body is still effectively empty relative to the schema delta.

A handy preview before writing any migration: git diff HEAD -- '*_ddl.sql' shows the raw SQL delta since HEAD — confirms the proto edit landed the shape you intended before a migration carries it to the schema.

Editing the proto — the annotation checklist

import "gojet/v1/options.proto";

// Thing is a widget. Leading comments flow to COMMENT ON TABLE / COLUMN
// and into go-jet model doc comments.
message Thing {
  option (gojet.v1.table) = {
    name: "things"
    primary_key: ["tenant_id", "name"]
    indexes: [{ columns: ["tenant_id", "created_at"], order: ["ASC", "DESC"] }]
    tenancy: { column: "tenant_id", runtime_role: "app-tenant" }
    default_order_by: "created_at desc"
    tie_breaker: "name asc"
  };

  // Resource name.
  string name        = 1 [(gojet.v1.column) = { immutable: true, orderable: true, check: "name <> ''" }];
  string display     = 2 [(gojet.v1.column) = { orderable: true }];
  Type   type        = 3 [(gojet.v1.column) = { immutable: true }];  // enum → TEXT + auto CHECK
  repeated string tags = 4;                                          // persists by default; TEXT[]
  google.protobuf.Timestamp created_at = 10 [(gojet.v1.column) = { immutable: true, orderable: true }];
  google.protobuf.Timestamp updated_at = 11;                          // persists by default
  string url = 20 [(gojet.v1.column) = { skip: true }];               // computed at runtime; opt out
}

Rules:

• Persist-by-default. Every field on a (gojet.v1.table) message persists with inferred options. Annotate only to override inferred options or opt out with skip: true for computed fields (values derivable at runtime).
• skip vs OUTPUT_ONLY are orthogonal. (google.api.field_behavior) = OUTPUT_ONLY is about write-path semantics (server populates it); skip: true is about storage (don't store at all). created_at is usually both OUTPUT_ONLY and persisted.
• orderable: true enables cursor pagination. Non-orderable types (JSONB, bytes) fail the plugin if set.
• immutable: true is a hint to the repo layer. The generated UpdateAll() excludes these columns automatically. (google.api.field_behavior) = IDENTIFIER or IMMUTABLE auto-sets this — AIP-203 defines both as "server preserves through Update" which matches the storage contract. Keep the explicit storage-side annotation if you want defence-in-depth against an API refactor that drops the behavior annotation (the two agree and the storage side stays pinned).
• allow_zero_enum: true when the UNSPECIFIED value is a legitimate stored state (e.g. CREDENTIAL_MODE_UNSPECIFIED = "use default"). Otherwise excluded from auto-CHECK.
• check: "..." — arbitrary CHECK body. Overrides the auto-generated enum CHECK if both set.
• rename_from: "old_name" on a field declares a column rename. The plugin writes a PLUGIN-RENAME directive into the emitted ddl.sql; the diff subcommand surfaces it in the brief's "Active PLUGIN-RENAME directives" list. Author the matching ALTER TABLE ... RENAME COLUMN in the migration body. Drop the annotation on the next release once every environment has applied the rename.
• unique: true on a column auto-scopes to (tenant_id[, user_id], col) to match the RLS boundary. Explicit Table.indexes entries with unique: true must include the scope columns or use where: for a deliberately-global partial index.
• foreign_key = { target: "<Message>.<field>" } declares a single-column FK to a field on another persisted message. See "Foreign keys" below.
• Oneof with (gojet.v1.oneof_column) = {} → emits <name>_kind TEXT + pair with auto CHECK on variants.

Cross-tenant uniqueness guards (all caught at generate time)

• Tenant-scoped primary_key must include tenancy.column — otherwise the PK enforces uniqueness globally across tenants and a duplicate-key error at INSERT leaks the existence of another tenant's row.
• User-scoped tenancy additionally requires tenancy.user_column in the PK — same leak, one scope level deeper.
• Table.indexes entries with unique: true (and no where: partial) must include the tenant/user scope columns for the same reason.
• skip: true still applies: opting a field out of persistence also opts it out of every guard above.

Authoring a migration from a brief

When the scaffold comes back in brief mode, the body you author goes directly beneath the TODO(llm): marker. Drift-check (check) catches structural mistakes but doesn't audit your intent.

The brief has three signal sources — read all three before writing any SQL:

1. Active PLUGIN-RENAME directives (top block, only present when rename_from was declared) — these are ground truth about authorial intent. The column WILL show up in DELTA as a - old + + new pair because the generator diff is deliberately mechanical. It's your job to reconcile: a matching directive means write one ALTER TABLE ... RENAME COLUMN line instead of a DROP+ADD. A directive without a matching DELTA pair means the rename already landed; drop the annotation.
2. DELTA — the mechanical structural diff. Each line is + kind t.name, - kind t.name, or ~ kind t.name. The line tells you WHAT changed and which name, not WHY or HOW. Kinds: column, check, index, policy, plus the whole-table primary_key, rls, table_comment.
3. TARGET — full concatenated ddl.sql. This is where you look up the new column's type / default / comment / constraint body whenever DELTA says something was added or modified. Never hand-type a column definition; copy from TARGET.

Intent inference heuristics (the generator cannot make these calls for you):

• - column + + column, both with similar type, and a PLUGIN-RENAME directive connecting them → RENAME, not DROP+ADD. (DROP+ADD loses data.)
• - column without a matching +, and no rename directive → a real drop. Consider whether a two-stage deprecation (skip: true first, real DROP later) applies before emitting ALTER TABLE t DROP COLUMN c;.
• ~ rls t with target disabled → STOP. RLS off means every tenant sees every row. Confirm the proto actually dropped tenancy and the drop is intentional.
• ~ primary_key t → schedule carefully. The DROP CONSTRAINT / ADD CONSTRAINT pair leaves the table without a PK for an instant; concurrent writes during that window can lose uniqueness invariants.

Common deltas and their idioms:

Change	SQL
Column added	ALTER TABLE t ADD COLUMN c TYPE [NOT NULL DEFAULT ...];
Column dropped (forward-only = destructive)	ALTER TABLE t DROP COLUMN c; Prefer two-stage: annotate skip: true, deploy, then drop.
Column renamed (PLUGIN-RENAME directive present)	ALTER TABLE t RENAME COLUMN old TO new;
Column type changed	ALTER TABLE t ALTER COLUMN c TYPE new_type USING ...; — rewrites every row under AccessExclusive on a populated table.
Column SET NOT NULL on populated table	ALTER TABLE t ADD CONSTRAINT t_c_not_null CHECK (c IS NOT NULL) NOT VALID; then VALIDATE CONSTRAINT in a follow-up migration.
CHECK added on populated table	ALTER TABLE t ADD CONSTRAINT name CHECK (expr) NOT VALID; then VALIDATE CONSTRAINT later — avoids full-table scan under AccessExclusive.
Index added on populated table	CREATE INDEX CONCURRENTLY ix ON t (...); — but golang-migrate wraps each file in a transaction. CONCURRENTLY only works if the file is standalone (one statement) or if you drop CONCURRENTLY and accept the brief lock.
Enum CHECK variant added	Drop old CHECK, add new one including the new variant. Do it in one migration; the CHECK is short-lived between statements.
RLS policy added	Same as ddl.sql: CREATE POLICY n ON t TO "app-tenant" USING (...) WITH CHECK (...); (quote the hyphenated role).
RLS policy body changed	DROP POLICY n ON t; CREATE POLICY n ON t TO "..." USING (new_expr) ...; — PG has no ALTER POLICY USING (...) on all versions.
Table added	Paste the CREATE TABLE from TARGET, plus its CHECKs, RLS, indexes, and COMMENT ON.
Table dropped	DROP TABLE t; — data loss; coordinate with the team.
COMMENT ON changed	COMMENT ON TABLE t IS $$new body$$; — COMMENTs don't lock; safe to batch.

Hazards to watch:

• DROP COLUMN is irreversible and runs under AccessExclusive; the row-remove phase is fast (catalog only) but rewriting pages happens lazily — still queue behind it.
• ALTER TABLE ... SET DEFAULT doesn't backfill existing rows. If you need existing rows to take the new default, follow with an explicit UPDATE.
• DISABLE ROW LEVEL SECURITY is a tenancy leak — every tenant sees every row from that statement forward. Never.
• Primary-key swap: DROP CONSTRAINT t_pkey; ADD CONSTRAINT t_pkey PRIMARY KEY (...). The window between the two is a window with no PK; schedule carefully.
• GIN / BRIN indexes on large tables can build for hours — always CONCURRENTLY.
• Stale rename_from (the directive points at a column that already has the new name) is an author bug. Drop the annotation and regenerate.

Workflow — delegate to a subagent, don't author inline.

Migration authoring is a tight mechanical loop: scaffold, edit, drift-check, read diff, edit again, drift-check again, until green. Context-heavy steps (reading a 100-line brief, staring at a cmp.Diff dump) waste the parent conversation's context window. Spawn a subagent via the Agent tool and hand it the whole loop.

Template prompt for the subagent:

Author a forward-only Postgres migration. You are the last-mile author
in the protoc-gen-go-jet flow.

Proto change: <one-sentence summary of what the user just edited,
e.g. "added map<string,string> labels on Widget", or "renamed
description -> summary via rename_from">.

Tools: Edit, Read, Bash, Grep.

Variables: DDL_ROOTS=<dir-with-*_ddl.sql>, MIGRATIONS=<your-migrations-dir>,
TENANT_ROLE=<role or omit>.

Procedure:
1. buf generate (skip if already done).
2. Pick the next NNNN sequence number by looking at MIGRATIONS, then:
   protoc-gen-go-jet diff --ddl-roots=$DDL_ROOTS --migrations=$MIGRATIONS \
     [--tenant-role=$TENANT_ROLE] > $MIGRATIONS/NNNN_<snake_desc>.up.sql
   This writes a scaffold containing an LLM brief.
3. Read the scaffolded file. The brief contains:
   - Active PLUGIN-RENAME directives (ground truth for renames)
   - A structural DELTA (+/-/~ column|check|index|policy, terse)
   - Concatenated TARGET ddl.sql (source of truth for shapes)
4. Write forward-only ALTER/CREATE/DROP beneath the `TODO(llm):`
   marker. Rules:
   - PLUGIN-RENAME directive + matching -/+ in DELTA => RENAME,
     never DROP+ADD (data loss).
   - A `~ check X` that DELTA shows with a "NOT VALID" body on the
     migrations side means the preceding migration added it NOT
     VALID. Add a sibling migration (next NNNN, e.g.
     NNNN_validate_X.up.sql) that issues `ALTER TABLE t VALIDATE
     CONSTRAINT X;` — both land in the same PR.
   - Copy column / policy / index definitions from TARGET verbatim
     (quote hyphenated role names like "app-tenant").
5. protoc-gen-go-jet check --ddl-roots=$DDL_ROOTS --migrations=$MIGRATIONS \
     [--tenant-role=$TENANT_ROLE]. Read the output.
6. If green, STOP. Report the migration file path(s).
7. If red, the diff output names exactly which column / check /
   index / policy / body differs. Fix — either edit the most
   recent migration, or add another sibling. Go to 5.

Up to ~5 iterations is normal. If after 5 you're still red,
stop and report the last diff output to the caller — it is
probably a constraint that pg_get_constraintdef renders
differently from ddl.sql (custom type, policy body, RLS
behavior) and needs a human judgment call.

See .claude/skills/protoc-gen-go-jet/SKILL.md for the full
pattern table (rename, drop, type-change, CHECK NOT VALID,
CONCURRENTLY-in-tx, RLS policy replacement).

Report back to the parent with just:

• The authored migration path(s)
• The final drift-check verdict
• Anything non-obvious about the resulting SQL the reviewer should double-check

The subagent's detailed iteration stays in its own context; the parent sees a clean summary.

Do not paste the brief's CURRENT/TARGET comment blocks into the migration body — they're -- prefixed context, not executable. Write SQL beneath TODO(llm):.

Common migration patterns

Add a column

buf generate
protoc-gen-go-jet diff --ddl-roots=<dir> --migrations=<dir> [--tenant-role=<role>] \
    > <dir>/NNNN_add_description.up.sql
# → ALTER TABLE ... ADD COLUMN ... + COMMENT ON

Hand-written repos using gen.<Prefix>UpdateAll() pick up the new field automatically on regen — the column list is generated, not hand-maintained. No repo edit needed unless the field has special handling (e.g. server-computed values).

Rename a column (declarative)

1. In the proto, rename the field and add rename_from:

   string label = 2 [(gojet.v1.column) = { rename_from: "display_name" }];
   

2. buf generate — ddl.sql now carries a -- PLUGIN-RENAME: <table> display_name -> label directive.
3. protoc-gen-go-jet diff ... > NNNN_rename_display_name.up.sql — the brief surfaces the PLUGIN-RENAME directive under Active PLUGIN-RENAME directives. Author the migration as ALTER TABLE t RENAME COLUMN display_name TO label;.
4. Run protoc-gen-go-jet check .... If it's green, commit.
5. After every environment has applied the migration, drop the rename_from annotation on the next release. Stale rename_from (matches the current column name) is rejected at generate.

If the rename also needs a value transform (e.g. replace(display_name, '-', ' ')), append the UPDATE after the RENAME in the same migration file.

Drop a column (destructive)

Forward-only means this deletes data in prod. Prefer a two-stage deprecation (annotate skip: true to stop writing; deploy; drop in a later release). When you do drop, the brief's CURRENT-vs-TARGET delta tells you exactly which column disappears; author a single ALTER TABLE t DROP COLUMN c;.

Stop persisting a field (non-destructive)

Annotate with (gojet.v1.column) = { skip: true } and regenerate. The field stays on the API surface (service layer populates it at runtime from whatever computed source replaces it) but leaves the table. Author the matching ALTER TABLE t DROP COLUMN c; beneath the brief's TODO marker.

Bootstrap a new resource

Plugin emits <resource>_ddl.sql into the storage package. If your migrations directory is still empty, the next diff lands in initial mode and copies every ddl.sql verbatim — including the new resource. If migrations already exist, you're in brief mode; paste the new resource's CREATE TABLE (and its CHECKs, RLS, indexes, COMMENT ONs) from the brief's TARGET section into the migration body.

Storage is only half of a resource. The RPC service, authz wiring, and tenancy-scope plumbing are application-specific and live outside the plugin's territory. The end-to-end harness at cmd/protoc-gen-go-jet/e2e/ shows the full round-trip — fixture protos under cmd/protoc-gen-go-jet/e2e/proto/jet/e2e/v1/*.proto and the integration tests in cmd/protoc-gen-go-jet/e2e/e2e_integration_test.go exercise the generated mapper + go-jet table + aipjet pagination against a Postgres testcontainer.

Field-kind inference (proto kind → SQL type) beyond the example above: see the full matrix in cmd/protoc-gen-go-jet/README.md > Field-kind inference.

Foreign keys

Declare a FK at the field level — target is a proto reference, not a SQL column name:

message MCPServer {
  option (gojet.v1.table) = {
    name: "mcp_servers"
    primary_key: ["tenant_id", "id"]
    tenancy: { column: "tenant_id", runtime_role: "app-tenant" }
  };
  string id = 1 [(google.api.field_behavior) = IDENTIFIER];
  string provider_id = 2 [(gojet.v1.column) = {
    foreign_key: { target: "LLMProvider.id" }
  }];
}

The plugin resolves LLMProvider.id against the registry of messages in the current buf generate pass, then emits:

• ALTER TABLE mcp_servers ADD CONSTRAINT mcp_servers_provider_id_fkey FOREIGN KEY (tenant_id, provider_id) REFERENCES llm_providers (tenant_id, id) ON DELETE RESTRICT ON UPDATE NO ACTION;
• CREATE INDEX idx_mcp_servers_provider_id_fk ON mcp_servers (provider_id); (auto-suppressed when another declared index already covers)
• A supplemental UNIQUE (tenant_id, id) on the target table, when the target's PK doesn't already cover the composite lookup.

Rules:

• Target form. <Message>.<field> in the current proto package; .fq.pkg.Message.field (leading dot) for cross-package. The referenced message must carry (gojet.v1.table) and the referenced field must be either the table's single-column primary key or carry unique: true.
• Tenant auto-composite. When both the referring and referenced tables carry tenancy, the FK expands to (tenant_id, col) REFERENCES t (tenant_id, target) so cross-tenant references are structurally impossible. One-sided tenancy (referrer has it, target doesn't, or vice-versa) is rejected at generate time — align both sides or use Table.custom_sql if the cross-scope reference is intentional.
• Actions. on_delete defaults to ACTION_RESTRICT; on_update defaults to ACTION_NO_ACTION. SET NULL needs nullable: true on the column; SET DEFAULT needs default_expr. CASCADE is never the default — name it explicitly with on_delete: ACTION_CASCADE when you want it.
• Backing index. Emitted automatically and auto-suppressed when the column is already the leading column of the primary key, a Table.indexes entry, or unique: true (on non-tenant tables). Set index: false to opt out when a predicate / expression index covers the column but the plugin can't statically see it.
• Not in IDL: NOT VALID. Proto declares end state ("this FK exists"). Whether the migration applies it one-shot or two-step is a migration concern: the diff scaffold emits a HAZARD: ADD CONSTRAINT takes ACCESS EXCLUSIVE + full table scan line when an FK is added to an already-populated table, with the NOT VALID + follow-up VALIDATE CONSTRAINT idiom spelled out. For brand-new tables the FK lands inline with the CREATE TABLE's first migration; no hazard.

When NOT to use the annotation:

• Composite FKs other than the tenancy auto-composite — use Table.custom_sql.
• Deferrable / INITIALLY DEFERRED constraints — custom_sql.
• FK into a table that isn't persisted by the plugin — custom_sql.
• FK into a column that isn't PK or unique: true (and isn't reachable via the tenant-composite path) — custom_sql, or promote the target to unique: true first.

Drift check extensions: pg_constraint rows of contype='f' are diffed via pg_get_constraintdef() (normalises action spellings + column ordering), convalidated is part of the snapshot so a forgotten VALIDATE CONSTRAINT surfaces as drift, and a FK without a backing btree leading on the FK column is itself reported as drift — an unindexed FK is a latent perf bug, not a config choice.

See cmd/protoc-gen-go-jet/README.md (foreign-key section) for the full FK option vocabulary.

List filtering — AIP-160 is pushed into SQL

List methods take an aip.Params value directly. The Filter field is a plain AIP-160 expression string; callers don't build jet conditions themselves. The PG repository compiles the string into a jet BoolExpression via aipjet.FilterToCondition(filter, gen.<Prefix>FilterFields()) and passes it to aipjet.ExecuteWithCondition so the cursor predicate ANDs with the filter and pagination stays correct.

The plugin generates <Prefix>FilterFields() in aliases.go with every column the translator can produce a WHERE clause for — synth columns, JSONB, and bytea stay out. Durations bind as BIGINT nanoseconds; accept bare ints or duration("5m"). Repeated-text columns (TEXT[]) bind via containment — zones = "us-west-1" compiles to 'us-west-1' = ANY(zones), != wraps in NOT, ordering is rejected. Adding a new proto field automatically appears in the filter surface on regen; removing one removes it. No hand-maintained map.

// Repository
func (r *PGRepository) List(ctx context.Context, params aip.Params) (...) {
    cond, err := aipjet.FilterToCondition(params.Filter, gen.LLMProviderFilterFields())
    if err != nil { return ..., fmt.Errorf("%w: %w", storageerr.ErrInvalidInput, err) }
    // stmt := postgres.SELECT(...).FROM(...); pass cond as baseCondition:
    rows, next, err := aipjet.ExecuteWithCondition(ctx, schema, params, stmt, cond, tx)
}

// Service — encode the proto filter as AIP-160 string
filter := ""
if f := req.Msg.GetFilter(); f != nil {
    filter = aip.MakeNameFilter(f.GetNameContains())  // `name:"VALUE"`
}
res, err := repo.List(ctx, aip.Params{PageSize: ps, PageToken: t, Filter: filter})

Supported AIP-160 constructs today: =, !=, <, <=, >, >=, : (has/substring on scalar strings, containment on TEXT[]), NOT/-, AND, OR, parens. Per-column-type dispatch rejects invalid combinations (e.g. < on bool, ordering on TEXT[]) with ErrUnsupportedFilter — callers map that to InvalidArgument. Never filter in-process after pagination: a filter that eliminates most rows from a page looks "done" to the caller even when later pages still have matches.

See the filter translator + security guards (LIKE metacharacter escaping, 4 KiB filter-length cap) and the AIP-160 reference at https://google.aip.dev/160. The narrow aip.MakeNameFilter / aip.ParseNameFilter helpers are what memory repos use (they can't build jet SQL).

Writing Update in a hand-written PGRepository

The plugin emits gen.<Prefix>UpdateAll() in aliases.go — an UPDATE statement pre-populated with every mutable column. Spread into go-jet's chain:

err = gen.LLMProviderUpdateAll().
    MODEL(updated).
    WHERE(gen.LLMProviderTable.Name.EQ(postgres.String(name))).
    RETURNING(gen.LLMProviderTable.AllColumns).
    QueryContext(ctx, tx, &written)

Never hand-maintain a UPDATE(table.Col1, table.Col2, ...) column list — a new persisted proto field won't land there, and updates to that field silently drop. The generator excludes primary-key members, synthesized columns, and immutable: true fields; includes updated_at and oneof kind/json pairs. Regen is the only thing that touches the list.

Commit shape

Every proto-shape change touches all three layers. Commit them together:

# proto + regenerated files
<proto_pkg>/thing.proto
<proto_pkg>/storage/<thing>_{ddl.sql, mapper.go, schema.go, aliases.go}
<proto_pkg>/thing.pb.go
<proto_pkg>/storage/jet/public/{model,table}/things.go

# migration (under your migrations directory)
<your-migrations-dir>/NNNN_<desc>.up.sql

Run your formatter over the proto + Go before committing; the generated files carry DO NOT EDIT. headers and are not hand-formatted.

Don'ts

• Don't edit plugin output (<resource>_ddl.sql, <resource>_mapper.go, <resource>_schema.go, <resource>_aliases.go, jet/public/**). DO NOT EDIT. headers. Next buf generate overwrites.
• Don't write .down.sql. Forward-only is the tool's convention — check only verifies the forward chain reproduces ddl.sql, and rollbacks are unsafe (data loss) and handled manually with review.
• Don't paste the brief's CURRENT/TARGET comment blocks into the migration body as executable SQL — they're -- prefixed context, not statements. The migration body goes beneath the TODO(llm): marker.
• Don't ship a migration without reading the "Hazards to watch" list. Destructive operations (DROP COLUMN, DROP TABLE, DISABLE ROW LEVEL SECURITY) run without a prompt.
• Don't bypass drift-check when check is red. Fix the migration.
• Don't cram multiple unrelated schema changes into one migration. One logical change per NNNN_*.up.sql.

References

• Plugin README (full options vocabulary, foreign-key section, field-kind inference matrix, what-won't-work list): cmd/protoc-gen-go-jet/README.md
• Plugin source: cmd/protoc-gen-go-jet/
• Options proto: proto/gojet/v1/options.proto (import as gojet/v1/options.proto)
• End-to-end worked example: cmd/protoc-gen-go-jet/e2e/ — fixture protos under cmd/protoc-gen-go-jet/e2e/proto/jet/e2e/v1/*.proto, integration tests in cmd/protoc-gen-go-jet/e2e/e2e_integration_test.go, generation template in cmd/protoc-gen-go-jet/e2e/buf.gen.yaml.