[LittleChimera]

Related issue

Closes #

Round-3 work package R3-03 (schema and vocabulary). Spec: ROUND3-SPEC §16, plus the data shapes of §8–§15.

Summary

All round-3 persistence lands in one forward-only migration, so the later round-3 packages add no migrations and never fight over alembic heads. This is the data layer only: nothing calls the new tables yet, so runtime behaviour does not change.

• Migration fz1a2b3c4d5e (chains after fy1a2b3c4d5e, the single head on staging):
  • 12 new tables: factory_people, factory_chat_messages, factory_chat_reads, factory_proposals, factory_tasks, factory_loops, factory_issues, factory_spans, factory_cost_daily, factory_versions, factory_definitions, factory_feedback. They follow the cockpit conventions: workspace_id-led keys, Uuid16 ids, SMALLINT enums with named CHECKs, compressed JSON, and no foreign keys.
  • New columns:
    • factory_projects: status (default 1), parent_slug, closed_at, def_version, version_id.
    • factory_prs: opened_at, first_ready_at, task_id, version_id.
    • factory_decisions: assignee, claimed_by, claimed_at, thread, subject.
  • ck_factory_decisions_source is dropped and recreated as IN (1..9) through batch_alter_table, so it also works on SQLite. downgrade() reverses everything.
• Models, codecs, entities: ORM models, enum codecs for every new enum (the decision sources gain proposal=6, loop=7, access_request=8, done_check=9) and frozen entity dataclasses.
• Store (FactoryStore, with in-memory and SQL implementations that behave the same). Each method's docstring names the package that will use it.
  • CRUD and queries:
    • people by factory and by user;
    • chat pages by thread and seq, thread stats and next seq;
    • read marks that only move forward;
    • proposals: compare-and-set on status, dedupe key, filters;
    • tasks: upsert by (project, key), an atomic attempt counter;
    • loops: upsert by (kind, subject);
    • issues: re-raise by dedupe_key (bumps the count and reopens resolved issues), numbered per class;
    • spans: insert-or-ignore by dedupe_key, open-span and range queries;
    • cost ledger: atomic upsert-increment;
    • versions: deduped by content hash;
    • definitions: immutable versions;
    • feedback.
  • Per-factory number sequences (C-n, T-n, INC-n/F-n, D-n, chat seq, definition versions) are serialized where the backend can: a transaction advisory lock on PostgreSQL, a GET_LOCK named lock held across the write on MySQL, and BEGIN IMMEDIATE on SQLite. Every write replays UNIQUE collisions and lock conflicts (MySQL 1213/1205, PostgreSQL 40001/40P01) after a jittered exponential pause; sequence writes get a larger budget (12), because CockroachDB serializes nothing and relies on the key.
  • Free-text values (proposal/task titles, issue rule/category/target/summary, span name, loop verdict/reason, ledger agent/model, version tag/note) are cut to their VARCHAR bounds in both stores, so an over-long value never raises DataError on PostgreSQL.
  • append_chat with a message_id that already exists returns the stored message in both stores (a replayed append adds nothing).
• Vocabulary (omnigent/factory.py):
  • labels def_version, version, bundle, mode;
  • ROLE_PLANNER, which is in ROLES but not TOP_LEVEL_ROLES. The planner registers the roster tool only, matching spec §18. The role guard now imports it from here (R3-04 defined it locally with a TODO), and R3-04's planner rules apply;
  • the 8 new decision actions;
  • state, kind and taxonomy tuples;
  • Room thread keys.
  • LABEL_DEFAULT_REPO, LABEL_DEFAULT_BASE, PROJECT_KINDS and the ProjectKind view type come from R3-01 (merged first); this PR reuses them.
• server/factory/actor.py: Actor (§10.1) plus ref parse/format/find for T-n, C-n, D-n, INC-n, F-n, !n, #slug, p:<slug> and task:<slug>/<key>. A project slugged factory or improvements formats as p:<slug>, so it never reads back as a Room thread.
• class alias views (ProposalView, IssueView) use populate_by_name and dump by alias by default, so they construct by attribute name and emit class on any pydantic 2.x (validate_by_name / serialize_by_alias alone need 2.11). Raising the floor instead would touch uv.lock, which trips the security gate's OSV scan.
• Views:
  • schemas.py adds PersonView, ViewerBlock, ChatMessageView, ThreadView, ProposalView, TaskView, LoopView, IssueView, SpanView, NowState and DefinitionView.
  • The snapshot additions are all optional fields: viewer, people_count, presence, counts, def_version, engine_mode, project/PR/member/decision extras and totals.usd_per_h.
  • factoryApi.ts mirrors them. TeamStrip gets the planner label.
• Fixture snapshot.round3.json, derived from the live run:
  • 7 projects, including base-env (integration);
  • 41 agents and 3 humans (creator, operator, viewer);
  • a pending projects proposal C-4 mirrored as D-33;
  • the db-version task graph, with the implement loop on T-4 exhausted at 3/3 (D-31);
  • INC-13 (agent.repeat) and C-14 (tuning) on trial;
  • definition v7 pinned.

fy1a2b3c4d5e (cockpit) ──► fz1a2b3c4d5e (round 3: +12 tables, +14 columns, source CHECK 1..9)
                              ▲ every later round-3 package reads/writes through FactoryStore

Test Plan

• uv run --frozen pytest -q tests/db/test_migration_factory_round3_tables.py tests/stores/test_factory_store_round3.py tests/factory/test_factory_vocabulary.py tests/server/factory/test_contract_fixture.py
  • Migration up/down/up. Checks keys, indexes and UNIQUEs, that the ORM matches the migrated schema, that cockpit-era rows survive with status=1, that each CHECK accepts its last code and rejects the next, and that the downgrade drops round-3 decision sources.
  • Store round trips for every table, run against both the SQL store and the in-memory store. Covers dedupe and upsert semantics, plus 8 threads racing on each numbered sequence and on the span, issue and proposal dedupe keys.
  • Codec coverage, the vocabulary matching the codec tables, ref parser table tests, and Actor.
  • The contract test: both fixtures round-trip through the pydantic models, and the round-3 shape and cross-references hold.
• The same migration and store files also pass on PostgreSQL 16: a local container with OMNIGENT_TEST_DB_URI=postgresql+psycopg://…. The whole tests/db + tests/stores set passes there too (1335 passed), except the Databricks volume store, which needs the databricks extra.
• pytest tests/server/factory tests/runner tests/examples tests/policies tests/test_factory_gates.py tests/factory tests/integration/test_factory_tree.py: everything passes except 7 runner tests that fail locally for environment reasons. 4 need the databricks module, 1 needs linux_bwrap (Linux only), and 2 compare against the local Claude subscription model catalog. None of these touch factory code.
• alembic heads returns fz1a2b3c4d5e only.
• Review fixes, PostgreSQL 16 container: the round-3 and cockpit store suites plus both factory migration tests pass (143). test_definition_numbering_replays_unserialized_racers races 8 auto-numbered put_definition calls with the sequence lock off (the CockroachDB situation); with the old DefinitionVersionExists check it failed 1 run in 3 on PostgreSQL, with the fix it passes every run.
• test_writes_replay_lock_conflicts pins the replay classification: MySQL 1213/1205 and PostgreSQL 40P01 replay, other operational errors raise at once.
• Web: npm run type-check, npm run lint and npm run format:check pass; vitest on src/components/factory, src/lib/factory* and src/pages/factory passes (634 tests); storybook build passes.
• pre-commit run --files <changed>: ruff, the hardcoded-model check, prettier, oxlint and web-tsc all pass. pyrefly shows no errors in the touched files (the existing 24 are missing opentelemetry imports and one launcher override).

Demo

N/A. No visible UI change; the only web edits are types and the planner role label.

Type of change

• Bug fix
• Feature
• UI / frontend change
• Refactor / chore
• Docs
• Test / CI
• Breaking change

Test coverage

• Unit tests added / updated
• Integration tests added / updated
• E2E tests added / updated
• Manual verification completed
• Existing tests cover this change
• Not applicable

Coverage notes

Manual verification: I ran the migration and store suites against a throwaway local PostgreSQL 16 container, which is the same image as CI's stores-postgres lane.

Not verified locally:

• CockroachDB and MySQL (no local images); CI's stores-crdb and stores-mysql lanes run the same store tests.
• alembic upgrade head on a copy of the sandbox database. The Postgres test seeds cockpit-era rows and upgrades them instead.
• The deployed server starting with this migration, which needs the newTag bump.

Updated existing tests:

• the two head pins: fy1a2b3c4d5e → fz1a2b3c4d5e;
• the cockpit CHECK test, whose rejected decision-source code moves from 6 to 10;
• the stable decision-source codes in test_enum_codecs.py;
• the factory tool registration matrix, which gains the planner.

🤖 Generated with Claude Code