ENGINEERING 01

A failed run resumes where it stopped

Per-run state machine. Every step emits a typed artifact validated with zod, holds a lease so two runs can never touch the same channel, and retries only on transient errors. You re-run one step instead of the whole night.

researchscript visualsvoice assemblepublish

ENGINEERING 02

The policy gate spends nothing on a bad trend

A deterministic gate rejects real people, brands, music and gore before any generation call is made. The policy reviewer agent runs on top of it. Rejections cost you nothing but the research step.

ENGINEERING 03

You can see what it did at 3am

Unattended software you cannot inspect is a liability. Every job, step, agent call, vendor request and subprocess emits an OpenTelemetry span, so a run that went wrong is something you read rather than guess at. yca telemetry show --last 1 reads the last run from disk with no backend at all; point OTEL_EXPORTER_OTLP_ENDPOINT at the bundled Grafana stack in ops/ if you want dashboards. Ten modules in src/telemetry/, its own vitest file, docs/OBSERVABILITY.md and a /yca-telemetry skill.

Spans carry structure and timing, not your scripts: content is deliberately kept out of telemetry.

ENGINEERING 04

The expensive failures are tested, not hoped for

Unit tests catch a broken function. What costs money overnight is an agent doing something reasonable-looking and wrong, so four eval cases assert the behaviour instead: a billing failure stops the run rather than retrying into a dead subscription; a trend depicting a real person or a brand is held before a single generation credit is spent; a step left running by a dead coordinator resumes from the run's own state rather than the model's account of it; and a publisher reporting in-progress is resolved against YouTube rather than treated as a missing video and posted twice.

Behind them sit thirteen regression tasks: four golden runs and nine mutations that each inject one real past failure — a wrong market figure, a drifted bookend, a real-person prompt, an unapproved clip, a silent clip, stale trends, a Shorts slot collision. They run inside npm test and spend no credits. Four capability tasks and five judge rubrics sit alongside them, calibrated against human labels.

ENGINEERING 05

It can improve its own prompts, and show its working

yca eval build assembles a suite out of your runs, preferring real production transcripts over anything synthetic, and hands you a review page before it counts. grader-check then checks the graders themselves; it found two real grader bugs in our own suite. diagnose reports the infrastructure error rate, warns when the baseline is already above 95% and there is no headroom left to measure, and derives a noise floor so you know which gains are real.

yca eval hillclimb runs the loop: a seeded train/test split whose test half the proposer never sees, one targeted patch per round to a role's instruction file, and a linter that rejects a patch for pasting case content, hardcoding a path or a date, or touching more than forty lines. A patch is kept only if the training gain clears the noise floor, the held-out set does not drop, and no case regresses. Everything else is reverted with the reason recorded. It writes an accepted.diff for you to read. It never edits your repo.

We ran it on our own scriptwriter. The baseline was already 99.6%, so nothing was accepted, which is the correct outcome and the reason we are not claiming it improved anything. An example suite ships so you can run the whole loop offline on the mock engine before spending a credit.

ENGINEERING 06

It checks the machine before it trusts it

yca setup writes .env from the example without ever overwriting one, checks the machine, and prints the next command. yca doctor is the one you run when something is wrong at 6am: node, ffmpeg, python and Pillow, free disk, the power settings, whether the machine is set to wake for the run, systemd linger on Linux, which engine is selected, whether its CLI is actually on PATH and answers a probe, every channel config, and the scheduler itself.

Most overnight failures we have had were not code. They were a machine that dozed off, an expired token, or a vendor out of credit. doctor exists because finding that out at 6am is worse than finding it out at 6pm.