Detect accounts reselling US frontier models through a relay or proxy, using account-level user-agent and egress-network dominance in default.generations, and file the survivors as Sentinel ban candidates. Use when asked to sweep for proxy/relay traffic,…
Implement a new batch provider (e.g. Anthropic, Gemini) from its committed research note — file-by-file recipe for the endpoint-family schema entry, skin selection/creation, adapter lifecycle methods, enum/factory registration, runtime env plumbing, and the…
Write and audit tests for the Batch API — colocated bun:test unit tests in packages/batch and services/batch-api, the vitest e2e happy path in tests/e2e/api/batches against the local Tilt batch stack + fake provider, and the test-quality audit pass. Sub-skill…
Capture real sync-API responses and provider batch output and turn them into committed JSONL fixtures for Batch API tests — batch input files, batch output lines, sync-response golden vectors for parity tests, and the Zod-validated loader pattern. Covers both…
Run and create end-to-end tests after creating a PR AND after every subsequent code push that changes behavior. For API changes, run the e2e test suite and inspect dev-fs-logs. For frontend changes, record a video, review it for bugs, fix any issues found,…
Query the ClickHouse analytics cluster for users whose current (intraday-corrected) balance is negative, using the same balance-correction and eligibility logic as the negative-balance-usage-sync cron. Use when asked who currently has a negative balance, how…
Adjudicate a Sentinel case that groups accounts by a shared mail relay or mail-domain family — decide whether the relay is operator infrastructure or a commercial host, then derive the per-account signals that justify or refuse enforcement.
Challenge the tests you just wrote with StrykerJS mutation testing. Use after adding or changing tests for a pure function, parser, serializer, or other logic-heavy module, to find assertions the suite executes but never checks. Also use it before you open a…