First-touch radar

1. Run pnpm --filter analytics radar:watch --window daily and read both the JSON and Markdown under .tmp/first-touch-radar/runs/.
2. Treat exit 0 as clear, exit 1 as a fired, unavailable, or insufficient check needing attention, and exit 2 as a run that could not complete. Keep unavailable and insufficient metrics explicit; never turn missing evidence into zero or a pass.
3. On a fire, run pnpm --filter analytics radar:triage and inspect the outcome-ranked sessions. Use metric-watch and replay-triage for details.
4. For a weekly review, run pnpm --filter analytics radar:report --week <Monday> and read the coverage flags, denominators, campaign annotations, previous-issue recovery, and A/B maturity before drawing conclusions. The campaign-followup check runs in daily watch and weekly report; compare identified recipients' post-send activity with the seven days before each send, split by recipient app and activity app. Weekly reports cap cumulative BigQuery dry-run estimates at 100 GB; use --max-total-gb to set a different report-wide cap, and treat a reached cap as incomplete evidence.
5. Run feedback-ingest for Slack themes. Searches are read-only and begin at the saved watermark.
6. Send one Telegram to Steve only when a check fired, is unavailable or is insufficient, or the weekly report is ready. Include counts and the private run-file path. Send nothing on a quiet day. Never send Slack or email.
7. Keep recipient emails, HMAC keys, and prompt text out of logs, reports, Telegram, and commits. The campaign check reads its mode-0600 recipient file in memory, binds keyed email digests as query parameters, disables local and BigQuery result caching, and emits aggregate counts only. BigQuery matching requires ANALYTICS_RADAR_CAMPAIGN_HMAC_KEY and a configured ANALYTICS_RADAR_CAMPAIGN_HMAC_FUNCTION backed by the same secret server-side. The function accepts a normalized email and returns lowercase hex HMAC_SHA256; keep its key in Secret Manager and never pass the key to BigQuery. When either setting is missing, keep the campaign check unavailable. Unknown signup campaign labels are grouped as other. Anonymous landings cannot be assigned to a recipient; do not interpret identified counts as a complete click total.
8. This radar changes scripts and skills only. Do not make user-visible UI changes. If UI work becomes necessary, stop that scope and request Steve's screenshot approval before implementing it.
9. A fired metric may create a local brief through fix-brief; no fix starts until Steve approves it.
10. Interpret the first-touch journeys separately: Design web excludes /mcp/oauth% signup paths; Design MCP is signup → agent_connected → first $mcp_tool_call → design_created. Clips activation is signup → recording_ready within 24 hours. Treat 2026-10-21 as a definition break: recording_started before then includes local-only recordings.
11. Read the weekly paid-acquisition notes alongside signup counts: Slides ads stopped after 2026-10-16, Bing CPC ended around 2026-10-20 21:00 UTC for Clips/Design, and Google paid spend was about 45% lower from 2026-10-21.