Pixeltable Data Apps

Daily Pixeltable DevX stress harness: the main routine fires around 9 AM in the owner's timezone every day, weekends included; a catch-up routine later the same morning stays quiet if already done. A shared runbook is the single source of truth, with a run lock and one team-channel post per day.

Daily reports: post the parent summary and deep-dive threads only in the team channel the owner named. The runbook and both daily routines should match that destination.

Every daily app is published as a new public GitHub repo under the owner's account, named after the product plus a short slug, before the team-channel post. The public scrub must pass: no secrets, machine paths, localhost URLs, internal ticket or chat links, or real cloud URIs. Publishing is idempotent and recorded in a published-repos ledger.

After publishing, set each new repo's GitHub social preview. If GitHub asks for sign-in or 2FA, skip and say so in the handoff.

A hub repo with a public Pages gallery is regenerated from the published-repos ledger. Every app repo gets a devcontainer, a citation file, a social preview image, topics, and a hub footer.

In daily apps, only user-defined functions go in udfs.py. Tables and query functions stay in app.py.

The daily run alternates two kinds of day. Demand days take a real user-job topic from the demand backlog. Coverage days follow a feature and changelog rotation. Every pick goes through a duplicate check against published apps. Demand READMEs use query-style titles with an FAQ. Claimed topics are marked in the owner's demand tracker when they have one. No daily app may duplicate an idea already published.

The daily main run defers to the runbook, has a per-run model spend limit, must always deliver the daily summary to the owner, and does not rebuild the demand backlog itself.

A separate weekly demand-refresh routine runs before the Monday daily run. It rebuilds the demand backlog and only messages the owner if fewer than 4 topics remain open or a source cannot be read.

The daily run rechecks every tracked ticket's reproduction, comments once when a bug clears or regresses (no daily still-reproducing ticket comments), and resolves a ticket after 2 consecutive clear runs on different days, including tickets owned by others. Track this in a local ledger.

Each daily run first repairs local catalog data directories in place (missing empty dirs, correct permissions, stale lock) and only moves a folder aside if repair or startup fails.

After each daily run, the forecast and topic claims are rebuilt from real history. This bot's own stale scheduled claims are moved to the new slot or cleared and reported. The weekly refresh rebuilds the backlog before planning.

When a scheduled demand topic has no tracker row, append a new row with the claim instead of skipping the claim.

The demand backlog refills itself. After every weekly rebuild, and on daily runs when fewer than 4 topics are open, generate new topics so the backlog has at least 8 open topics from the configured sources (search queries, citations, competitor notes, social signals, and outreach).

The later-morning catch-up runs a completeness script first and exits silently if today's run is already complete or still in progress. It does not check whether the daily summary reached the owner. If the script reports gaps, it finishes only those gaps (or sends the owner one line if stuck) and still clears stale run locks.

Deep-dive write-ups create or link a ticket first, then link that ticket in the team-channel thread. On success, delete today's cloud database and stop hosted services; keep them only if the run failed.

Never print provider API keys. Stay within the runbook's per-run provider budget; use the cheapest model that exercises the feature. Before creating a ticket, search existing tickets, including ones filed by other bots.

This workflow uses GitHub to publish each daily app as a public repository, maintain an optional gallery hub, read product release notes as an input, and attach a public SDK reproduction to new issues. Connect GitHub during setup.

Use as the first conversation with a new owner of this daily Pixeltable QA bot: ask what to call them, which Slack, Jira, and GitHub to connect, schedule and Cloud, then save memories and point the routines at their setup.

Use when installing, upgrading or checking Pixeltable: venv install, version check, 1-minute image demo, pxt dashboard, app project layout.

Use when building RAG over docs, video frame search, image captioning, audio transcription or a tool-calling agent with Pixeltable.

Use before writing or debugging Pixeltable code to avoid known traps around UDFs, ordering, views, indexes, media inputs and keys.

Use when answering any Pixeltable API question: check the installed version, read llms.txt and the docs, confirm in source, run it.

Fires every morning so the agent ships a new app, runs it locally and on Cloud, files what broke, and publishes a public repo.

Fires later the same morning to finish only leftover steps if the main run did not complete, and stays silent when the day is already done.

Fires Monday morning to rebuild the backlog of real user-job topics before that day's run, and stays silent unless too few topics remain or a source cannot be read.