admin/Dockerfile: 2-stage monorepo Next.js standalone (outputFileTracingRoot added to next.config so the bundle traces workspace deps). scheduler/Dockerfile: single-stage Bun runtime, validated by a local docker build. .dockerignore excludes node_modules/.next/.turbo/.git. Build context = monorepo root; runtime env DATABASE_URL + REDIS_URL.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
X/Twitter (live posting gated on an X developer app, registered separately):
- schema: add 'twitter' to OutletSchema (sanitize 280-char limit already present).
- packages/twitter-client: X API v2 client (POST /2/tweets, OAuth2 user Bearer),
injectable fetch, 280-char guard; 4 contract tests.
- apps/scheduler/src/dispatch.ts: makeOutletDispatcher routes by outlet
(linkedin.* -> linkedin, twitter -> twitter), fail-closed for unconfigured
outlets; 4 tests. main.ts wires both outlets fail-closed until creds exist.
Monorepo fix: add a local tsconfig.json (extends root, include src) to each
package + non-admin app. Without it, every package's `tsc --noEmit` compiled the
WHOLE repo via the root config (no JSX/DOM) and choked on the admin .tsx — masked
until now by turbo build caching.
Tests: 135 passing + 13 gated across 12 tasks.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- apps/scheduler/src/queue.ts: PUBLISH_QUEUE, BullMQ-compatible Redis connection
factory, queue factory, default retry/backoff job opts.
- apps/scheduler/src/worker.ts: makePublishWorker factory drives the publish-loop
state machine (queued -> dispatching -> published | failed/dlq) and persists
each transition to publications. Dispatcher is injected (fake in tests; the
LinkedIn client swaps in once Gate 0.9 / CMA lands).
- apps/scheduler/src/main.ts: runnable entry; boots the worker; LinkedIn dispatch
is fail-closed until CMA is approved.
- apps/scheduler/src/worker.integration.test.ts: gated behind INTEGRATION=1;
verifies enqueue -> consume -> row published within 5s, and failure -> dlq with
error recorded. Run against real Redis + Postgres.
Closes the Stage 4.1 DoD (enqueue -> worker-consume -> DB row updated within 5s).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pure-function pieces that need no Redis/DB to verify (11 new tests):
- apps/scheduler/src/cadence.ts: Beta posterior model, propose(history, config)→[Proposal]; threshold gates (≥8 samples, ≥4 weeks, ≥20% ratio, disjoint 95% CIs); never auto-applied (RBR)
- apps/scheduler/src/publish-loop.ts: state-machine transition function for pending→queued→dispatching→published with retry-vs-DLQ branch on failure and cancellation path
What defers to live Redis + Postgres + LinkedIn:
- 4.1 BullMQ queue + Redis enqueue→consume integration test
- 4.2 End-to-end publish loop trace (POST /api/content → DB → BullMQ → LinkedIn fake → publication row)
- 4.3 Chaos test for idempotency on retry
117/117 tests pass cumulative across all packages and apps.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Angelo B. J. Luidens