Schedule a private weekly funnel review

Why this is different from a weekly prompt

A cron prompt with a private owner key can choose new queries, read raw users, and turn a report into an unreviewed side effect. It also tends to compare moving now() windows, which makes reruns produce different evidence.

[email protected] renders a disabled automation-v2 candidate with a single immutable report source. Agentry anchors the comparison to the trusted schedule receipt, executes the approved aggregate blueprint server-side, and gives the exact runner only that run-bound source. The runner cannot read raw identities or generic analytics.

The template never publishes, sends messages, changes code, creates a draft PR, starts an experiment, merges, deploys, or enables itself.

Configure it

Start with the machine-readable version and filtered OpenAPI schemas:

GET /v1/automation-playbooks/weekly-funnel-review?version=2.0.0
POST /v1/projects/:project_id/automation-playbooks/weekly-funnel-review/render

The strict renderer requires:

• 2–20 distinct ordered event names from the saved signal map;
• a conservative five-field cron expression and IANA timezone;
• a bounded schedule-start window and exact comparison-period length;
• a minimum sample size for both adjacent periods;
• run, turn, runtime, and cost budgets with the PR budget locked to zero;
• the selected runner provider and an escalation destination.

It also checks the saved signal map, current verified onboarding state, live event/property reads, scoped credentials, and saved contract, and ordered_funnel_period_compare blueprint. Missing or synthetic-only events, stale verification, an unavailable blueprint, or missing scheduler readiness remain blocking gates.

Review the normalized definition and binding hash, then save it with:

PUT /v1/projects/:project_id/automation-contract

Saving leaves the automation disabled. An owner reviews the exact ETag, creates one agentry_runner_… credential bound to this automation_id, configures the selected scheduler, and explicitly enables the reviewed contract.

What one run does

1. The scheduler wakes Agentry using only the configured automation_id.
2. Agentry issues a single-use schedule receipt for one eligible occurrence.
3. The exact runner credential starts a run whose policy is snapshotted.
4. Agentry executes ordered_funnel_period_compare using the trusted scheduled time as the exclusive end of the current period.
5. The blueprint returns prior and current half-open windows, ordered funnel counts, conversion and drop-off values, and explicit sample readiness.
6. The runner reads only the run-bound aggregate report source and records a private typed report of evidence, hypotheses, gaps, and escalation needs.
7. Agentry closes or escalates the run according to the saved policy.

Rerunning the same occurrence preserves the same evidence window. The runner cannot replace the event names, move the period boundary, or request arbitrary HogQL during the run.

Review the report

Use the human control aggregate to inspect the schedule, credential health, latest runs, budgets, and attention items:

GET /v1/projects/:project_id/automation-control

A safe report separates exact evidence from interpretation:

Weekly funnel review — <prior bounds> vs <current bounds>

Readiness
- approved events: <entry> → <progress> → <completion>
- prior/current minimum sample: pass | blocked

Observed
- step counts and conversion for both exact periods
- absolute and relative movement
- largest measured drop-off change

Hypotheses
- <explicitly labelled ideas to investigate>

Missing proof / escalation
- <telemetry gap, low sample, late run, or owner decision>

No report row grants authority to contact customers, change a product, or make a causal claim. A human chooses any follow-up.

Stop control

An owner can cancel a nonterminal run with a stable Idempotency-Key:

POST /v1/automation-runs/:run_id/cancel

Cancellation stops Agentry coordination and releases internal claims. It does not undo an external scheduler action or publish anything on the user’s behalf.

Review checklist

• Every selected event is approved, observed, distinct, and in exact order.
• The schedule timezone and boundary match the business reporting calendar.
• The minimum sample is meaningful for both periods.
• The runner credential is bound to this one automation and is not an owner key.
• Generic analytics, raw identities, publication, messaging, and code changes remain forbidden.
• The candidate stays disabled until an owner reviews and enables its exact ETag.