data-digest

Pulls a per-project digest covering, since the last check:

• Users — total distinct user_ids in events.identifies ever, plus new first-seen-in-window count (signups since last check)
• Top user activities — top events by count, with distinct-user fanout
• API / LLM / scraping costs — cost.recorded rolled up by provider/operation
• Top error paths — logs.raw 5xx count by path
• Recent feedback — events.feedback items

It also appends a per-handle x-monitor section. The script reads each subscribed handle's local archive (tweets/<handle>.jsonl) and filters to tweets within the digest window — it does not rely on the latest run file (that file would miss handles subscribed after the run was taken). No re-fetch — costs $0.

The Python script returns structured JSON; the agent (Claude) reformats it into a tight markdown digest.

Setup

# from anywhere
uv run --quiet <skill-dir>/daily_digest.py setup

State lives at ~/.cache/daily-digest/ (projects.json, state.json, runs/<iso>.json). Override with DAILY_DIGEST_HOME if you want it on iCloud Drive for cross-machine sync.

Add a project

uv run --quiet daily_digest.py add my-project \
    --bq-project my-project \
    --group-type household \
    --description "Cooklang recipe iOS app + Next.js + Hono API"

Flags:

list, rm, are also there:

uv run --quiet daily_digest.py list
uv run --quiet daily_digest.py rm my-saas

Daily digest workflow (what /loop fires)

/loop 1d use the data-digest skill to generate today's roundup

When invoked:

1. Run the check. Lookback window = now − last_checked_at per project (default 1 day on first run). Pass --since 24h to override, or --no-state to leave last_checked_at untouched (useful for ad-hoc runs).

   uv run --quiet daily_digest.py check --json

2. Reformat the JSON into a tight bulleted digest. Use proper markdown — # title, ## per project / x-monitor, real bullet lists. Structure (NOT style) stays minimal: just signups, events, costs, optional 5xx/feedback, then x-monitor. No priorities, no initiatives, no takeaway.

   # Daily digest — DD/MM/YYYY
   
   ## <project>
   
   - **Users:** new N, total M _(group: Household Name, …)_   ← group only if new>0 and --group-type set
   - **Events:** <rough total> — `event_a` N×, `event_b` N×, `event_c` N×
   - **Costs:** $X.XX total — `provider/op` $X.XX, `provider/op` $X.XX, …
   - **5xx:** `/path` N×, …                           ← omit line entirely when zero
   - **Feedback:** N items — _[kind] subject_         ← omit line entirely when zero
   
   ## <next project>
   …
   
   ## X-monitor
   
   - No new posts from @a or @b
     OR
   - **@a** — N× new: one-line summary

   Rules:

   • Date format DD/MM/YYYY in the title.
   • ## for each project and for x-monitor — gives clear visual separation when scanning multiple projects.
   • Bold the field name (**Signups:**, **Events:**, …); inline-code the event names / paths / providers.
   • Round large counts — 1865 → 1.9k, 12340 → 12k+. Exact under 1000.
   • Top 3 events / top 4–5 cost ops inline, comma-separated. No nested bullets.
   • Omit 5xx and feedback bullets when zero — always show signups, events, costs.
   • No priorities, no initiatives, no takeaway, no emojis. This is a glance, not a briefing.
3. Append the x-monitor section from digest.x_monitor — the script already filtered each subscribed handle's archive to the digest window; just summarize per-handle new tweets following the x-monitor skill's bullet format.
4. One-line takeaway across projects — what stands out today. Skip if nothing did.
5. State advances automatically. The next check only covers what's happened since this run. Pass --no-state to skip the advance for ad-hoc questions ("what's happened in the last 7 days?").

Smoke test (one-shot, no state)

uv run --quiet daily_digest.py check --since 24h --no-state --json | jq

This runs all configured projects with a fixed 24h window and doesn't advance last_checked_at. Use for development.

What the JSON looks like

{
  "generated_at": "2026-05-07T06:10:27Z",
  "projects": [
    {
      "name": "my-project",
      "since": "2026-05-06T06:10:29Z",
      "description": "...",
      "signups":   { "new_users": 1, "sample": [...] },
      "activities":{ "top": [{event_name, n, users}, ...] },
      "costs":     { "total_usd": 0.94, "total_calls": 179, "breakdown": [...] },
      "errors":    { "top": [{path, errs, total}, ...] },
      "feedback":  { "items": [{ts, kind, subject, message, user_id}, ...] },
      "initiatives": {
         "dir": "/abs/path/to/docs/initiatives",
         "total_files": 45,
         "readme": "...full markdown content of README.md...",
         "files": [
           { "file": "...", "title": "...", "summary": "...",
             "mtime": "...", "size_bytes": N,
             "modified_recently": true }, ...
         ]
      }
    }
  ],
  "x_monitor": { ...latest x-monitor run JSON... }
}

Per-project fields are omitted when their source isn't configured (e.g. no bq_project → no signups/activities/costs/errors/feedback).

Schema assumptions

The bq fetchers assume the bq-analytics layout:

• events.raw — event_name, user_id, ts, JSON properties. Cost events use event_name = 'cost.recorded' with properties.{provider, operation, model, cost_micros}.
• events.identifies — user_id, ts, JSON traits. First-seen-per-user_id = signup.
• events.feedback — kind, subject, message, user_id, ts.
• logs.raw — ts, level, path, status, message (Vercel Log Drain landing zone).

If your project doesn't follow this exact shape, edit bq_signups/bq_costs/etc. in daily_digest.py — they're independent ~15-line functions.

Costs

• bq queries scan a few MB per project → free under the 1 TB/mo BQ on-demand quota.
• x-monitor section reads the already-archived latest run, so $0 in API calls. (The actual fresh fetch happens in the x-monitor skill's own /loop.)
• Total cost of a daily digest across 5 projects with bq-analytics: $0.

Troubleshooting

• Access Denied: User does not have bigquery.jobs.create permission — switch gcloud accounts: gcloud config set account <email>. The skill calls bq via subprocess, so it inherits whatever account is active.
• signups.new_users = 0 but you know there were sign-ups — verify events.identifies is being written; some projects only call track() and never identify(). In that case, treat first-seen-in-events.raw as the signup proxy and add a fetcher.
• No files flagged modified_recently — mtime floor is 7 days. If recent edits don't appear, the files might have been touched only in git history, not on the local filesystem (e.g. checkout from a different branch). Run touch or widen INITIATIVES_LOOKBACK.
• x-monitor section says "no runs yet" — set X_MONITOR_HOME to point at the synced folder, or run x-monitor's setup + check once.

Recurring with /loop

/loop 1d use the data-digest skill to produce a digest across configured projects

For a self-paced cadence (let the model decide when):

/loop use the data-digest skill whenever appropriate to summarize project activity