feat(mist): add a daily data digest skill for Mist app databases

[AaronIngersoll]

What

Adds mist/daily-data-digest — a skill that reads a Mist app's own database and reports what changed in the last 24 hours: per-table volume trends, spikes, drops, and stalls. It delivers the result twice, as a written digest in chat and as a show_dashboard dashboard that overwrites itself each day, so there's always exactly one current view per app.

Two files: the skill body and a manifest.json entry alongside the other mist/ skills.

Why

Mist can already query a Mist app's database (db_schema / db_query resolve to the active project's Neon or PGlite side), but nothing turns that capability into a recurring read on whether an app is actually healthy. Answering "is this thing still getting traffic, and did anything break overnight?" means someone remembering to ask. Paired with a recurring schedule from skill-schedules, this makes that answer arrive on its own — the run's final message becomes the schedule summary and the notification the user reads.

How it works

Every Mist app has its own arbitrary schema, so the skill is schema-agnostic: db_schema overview to rank tables, then tables mode on a shortlist to read real column names and types. It never assumes a table or column exists.

Design decisions worth reviewing:

• Rolling 24-hour windows, not calendar days. A cron firing at 08:00 compares an eight-hour partial "today" against full-day averages and manufactures a drop every single morning. Rolling windows are always complete and always comparable.
• Anomaly rules carry absolute floors — a spike needs >=2x baseline and >=5 rows. Without the floor, 1 → 3 rows reads as a "200% spike" and the digest becomes noise nobody opens after a week.
• Stable dashboard title. The dashboard id derives from the title and the same title overwrites in place (dashboards.service.ts:221), so the skill mandates an app-specific, date-free title. A date in the title would mint a new dashboard file daily and bury the panel.
• timestamptz vs timestamp is called out explicitly, because comparing the wrong one against now() silently slides the window boundaries by the server's timezone offset.
• Aggregates only, never row-level PII. These are real production tables, and the output lands in both a chat log and a saved dashboard file on disk.
• Non-blocking by construction — no human_in_the_loop, never ends a turn waiting for input, since the whole point is the unattended scheduled path.
• A mandatory Coverage section lists tables that were skipped and why. A silent skip reads as "all clear" when the table was never checked.

Known limitation

db_query and db_schema resolve against the active project only (tools/index.ts:11968) — there is no cross-app query tool — so one run covers one Mist app, and covering five apps means five daily schedules. The skill says this plainly rather than pretending to sweep. A genuine roll-up would need a target-slug parameter on db_query; the plumbing already exists, since the discovery service caches every app's connection as mist-db:<slug>:<side> and exposes resolveSide().

How to test

1. python3 -m unittest discover -s scripts -p 'test_*.py' → 16 tests pass.
2. python3 scripts/validate_catalog.py → catalog validation passed: 42 skills, 6 workflows (was 41 skills). Both were run locally on this branch and pass; they're the same two steps validate-catalog.yml runs in CI.
3. End-to-end, against a real Mist app: copy mist/daily-data-digest/ into ~/Fluid/skills/, relaunch Mist, open a Mist app project with a provisioned production database, and run the skill. Expect a digest naming real tables with real counts, plus a saved dashboard. Then add a daily recurring schedule captured against that project and confirm the headless run produces the same digest as its run summary.

Step 3 has not been run — the skill is unexercised against a live database so far, and that's the review risk worth weighing: the SQL templates and the show_dashboard card shapes are derived from the tool contracts and the descriptor schema rather than from a real execution.

Review guide

Focus on: mist/daily-data-digest/SKILL.md — specifically the roll-up SQL template and the anomaly thresholds.
Skim: manifest.json — a mechanical nine-line catalog entry.

🤖 Generated with Claude Code