Product heartbeat
Unknown
No verified read yet.
When Baby's HomeSentinel · Founder
Founder-only review. Practice cases are fictional and kept for seven days. The separate Support inbox is read-only: opening a message does not create a case, call a model, post to Slack, approve a reply, or send anything. A real message is kept only when you choose Start a reply case, and even then nothing is sent.
Checking sign-in…
Founder access
Sign in to continue.
This review page is for the When Baby's Home founder account. If someone forwarded you this link, it shows nothing without that sign-in.
Sign-in happens on the When Baby's Home account page, then returns you here. Your session is checked on every request.
Your queue
Checking what needs you
Each item is one issue, what Sentinel looked at, what it thinks happened, and one suggested direction. Sentinel investigates on its own under its weekly cap. Acting on anything is still your call.
Nothing is waiting.
Sentinel keeps this page quiet until it has an issue, a look at the evidence, and a direction to suggest. A quiet queue is not a whole-product all-clear; coverage is still partial.
Everything else: the dashboard, incidents, cases, support and learning
Operational view
Everything else
What Sentinel can verify right now, what is only partial, and where a founder decision is still required. Decisions live on the queue.
Current assessment
Checking evidence
Sentinel is loading bounded operational evidence.
Platform evidence
Unknown
No verified feed yet.
Open signals
—
No signal summary yet.
Practice cases
—
Founder review rehearsal only.
Last 24 hours
Bounded platform activity
- Server request failures
- —
- Worker attention / failures
- —
- Identity rejections, signed-out visits excluded
- —
- Invalid event contracts
- —
Counts are content-free aggregates. Raw logs, paths, people, messages, traces, and provider payloads never appear here.
Right now
Households stuck at a step
| Step | Households |
|---|
A marked row is at or above the level the product calls a finding: two households at one step, or one undeliverable launch invitation. One household is an observation. No household is named here, because the block carries none.
Attention
What needs a decision
No verified item is waiting yet.
Unknown evidence is not treated as healthy, and it cannot close an existing signal. Coverage is partial by design, and a read where every block Sentinel has did answer still counts as quiet and can close one.
Dashboard reads are founder-authenticated and do not trigger a Product check, call a model, or change production.
Agent operations
Incidents
Real Production issues appear first. Sentinel saves the evidence automatically; you decide whether to run the paid investigation and what happens to its proposal.
Production evidence · Read only
Real-evidence preflight
- Heartbeat
- —
- Platform feed
- —
- Open signals
- —
- 24-hour failures
- —
Production evidence is unavailable.
The preflight failed closed. The fictional lesson is still available, but no real-evidence run can start.
Saved Production incidents are unavailable.
Sentinel could not verify the durable incident list. It will not start a paid run or show an unverified proposal.
No incidents need your review.
Sentinel will keep this page quiet until it has a bounded issue and a proposed next step.
Incidents are unavailable.
Sentinel could not verify the investigation service, so it will not start a run or show a decision.
Agent lab · Fixed fictional investigation
The issue
What I looked at
What I think happened
Suggested direction
Other options
Your call
No proposal yet. Start the fixed lesson when you are ready to watch the agent team investigate.
Run one paid fictional investigation?
This uses the fixed lesson under Sentinel's $5 weekly model cap. It cannot read or change Production, contact anyone, or trigger an operational action.
Investigation stopped.
One recovery attempt is available.
This is a separate exception for the fixed specialist evidence-contract correction. It preserves both stopped runs and cannot become another ordinary retry.
Run the one paid recovery attempt?
The fixed bound-specialist-evidence-v3 correction will use the same saved evidence under Sentinel's $5 weekly cap. The two stopped runs stay in the record. Nothing runs automatically, and this cannot change Production or contact anyone.
Fixed kind: one_time_runtime_correction. Reason: specialist_evidence_contract_correction. Runtime: bound-specialist-evidence-v3. Decision contract: real-founder-decision-v1 or real-founder-decision-v2, which differ only in whether the result carries a suggested direction.
Fictional lesson only. Starting or reviewing it cannot change Production, contact anyone, or trigger an operational action.
How Sentinel got here
Why Sentinel raised it
What is still unknown
How the agents reached this
Run record and limits
CEO view
How Sentinel works
A small operating system for watching When Baby's Home, turning bounded evidence into a useful recommendation, and keeping consequential decisions with the founder.
Authority
What runs automatically
- Read fixed aggregate health and operational evidence
- Apply deterministic signal rules
- Retain bounded, content-free operational records
- Send a generic alert for a new signal or missed heartbeat
Founder control
What needs you
- Reading a selected support message
- Approving any customer-facing reply
- Changing prompts, permissions, budgets, or policy
- Deploying, repairing, or writing to production
Privacy boundary
Private content stays on the narrowest path.
Monitoring stores only fixed enums, safe counts, release IDs, and timestamps. Support text is fetched only after the founder chooses it and is not copied into cases, traces, Slack, learning, or memory. Raw platform logs are discarded before storage.
Production path
Today, next, later
Watch and review
Founder sign-in, bounded monitoring, read-only support, fictional case approval, Slack notices, and exact-release source reads.
Understand real trouble
Journey-stall detectors, evidence fusion, durable hosted findings, and a protected real-case investigation boundary.
Act with guardrails
Evaluated reply drafting, exactly-once sending, approval-gated runbooks, and narrowly scoped remediation.
Developer view
Sentinel technical architecture
The runtime topology, trust boundaries, contracts, schedules, failure semantics, and recovery rules behind the founder view.
Topology
Evidence moves inward; authority does not.
When Baby's Home
Strict health DTO, operational events, and selected support message source.
Fixed contracts
Schema validation, exact-project filtering, content removal, and bounded enums.
Separate service
Founder auth, separate Neon database, deterministic monitor, review UI, and guarded workers.
Founder boundary
Protected review, explicit approval, and fixed generic Slack notices.
Trust and content classes
Four different boundaries
- Aggregate
- Counts, fixed enums, safe release and time metadata.
- Operational
- Strict event fields only; raw provider logs are transient and discarded.
- Support
- Selected plain text is browser-transient and never a monitoring input.
- Practice
- Repository-owned fictional cases may enter saved review and model rehearsal.
Runtime contracts
Fail closed, preserve uncertainty
- Every founder API re-authenticates the Clerk session and allowlist.
- Internal workers use separate fixed bearer capabilities.
- Unknown or unavailable evidence cannot resolve an active signal.
- Ambiguous Slack delivery is parked and never auto-retried.
- No Product database URL is accepted by the Sentinel service.
Schedulers
Deterministic cadence
- Four hours
- Product monitor reads health, then the alert worker checks for new unresolved signals or an overdue heartbeat.
- Platform log mode
- When paused, detailed Vercel log coverage is unavailable.
- Four hours, outside Vercel
- GitHub checks the heartbeat at 40 minutes past the cycle; it reports stale at four and a half hours.
- Daily
- Retention removes expired cases and 90-day platform evidence.
Signal semantics
Evidence, not diagnosis
- Fixed policy names a bounded signal and next inspection step.
- Initial feed silence remains unknown rather than an incident.
- A previously active operational feed becoming stale is actionable.
- Release metadata is context, not proof that a release caused a failure.
Approval and recovery
Changes require exact state
- Case decisions bind to one revision; edits invalidate approval.
- Database migrations use named recovery branches and exact schema guards.
- Deployment follows the private main branch; secrets stay encrypted.
- Provider outcomes are read back before a connection is called live.
Known gaps
Still deliberately outside the system
- Individual journey-stall detection and Blob orphan inventory
- Hosted finding memory and founder acknowledge/done controls
- Real-message investigation, evaluated drafting, and sending
- External dead-man monitoring outside the shared Vercel platform
- Automatic production repair, merge, deployment, or permission expansion
This page describes the reviewed architecture. Runtime truth belongs on the Dashboard; a configured capability is not treated as proof that its connection is healthy.
Your review cases.
Open a saved case, or choose a clearly fictional example to practice.
Choose a practice case
Accuracy
Unavailable family link
Keep the helper boundary clear without inventing a cause or offering access Sentinel does not have.
Relevance
Choosing a real date
Answer the scheduling question directly without adding account or privacy advice.
Warmth
Offering laundry help
Acknowledge a thoughtful offer naturally while keeping the answer short and useful.
Saved practice
No saved cases yet. Choose one above when you’re ready.
Founder view
What Sentinel is learning
All examples here are fictional practice. Sentinel does not learn from real customer messages, and prompts never change automatically.
Practice at a glance
- Cases reviewed
- —
- Rewrites requested
- —
- Approved unchanged
- —
- Edited afterward
- —
- Average rewrite time
- —
- Weekly Sentinel spend
- —
Correction themes
Prompt suggestions
No prompt suggestions yet. A theme must appear in three rewrite requests before Sentinel proposes a change.
Private · Read only
Support inbox
Sentinel lists content-free receipts first. A message is fetched from Resend only after you choose it, appears here temporarily, and is not added to cases, traces, learning, Slack, or memory unless you start a reply case for it.
Reply cases
Real messages you chose to keep, stored in a protected space for 90 days. Each one needs your approval before any reply exists.
No reply cases yet.
Recent messages
No current support receipts.
Message preview
Choose a message to read.
Its sender, subject, and plain-text message will appear here. Opening it does not start an investigation or prepare a reply.
Protected · Real support message
A reply case is open.
Customer message
What this case can and cannot do
- It is kept for 90 days, then deleted automatically. You can delete it now.
- A reply can go only to the original sender. Sentinel's model never sees the sender or the address.
- Approval covers one exact version for 15 minutes. Sending is not connected yet, so nothing leaves Sentinel.
- Nothing here enters Slack, traces, learning, or memory.
No reply yet.
Ask Sentinel for a first draft, or write it yourself. Either way, a reply needs your approval, and nothing is sent.
Sentinel is drafting…
This usually takes a few seconds. Nothing will be sent.
Latest Sentinel draft
Before
After
See saved version history
Versions stay in the protected store for 90 days. They are not used for automatic learning.
A reply waits for you.
Fictional support email
Prepared example investigation
What Sentinel can establish
This reply is recorded. Reloading keeps its result. To practice changes, start another fictional case.
Latest Sentinel rewrite
Before
After
See saved version history
These fictional examples stay local for seven days. They are review evidence, not automatic prompt training.