Audit or design a status/health surface so it does not share fate with the system it describes and cannot quietly attest to itself: split read plane (independently deployed reporter + dead-man switch), three-valued verdicts (healthy/degraded/unknown — never green-by-default), external anchoring of any "tamper-evident" status chain, and the availability-inversion check (no client may hard-gate on the monitor). Use when adding a status page, deep-health endpoint, reachability verdict, or signed incident ledger to a service — especially the service every client depends on — or when reviewing a design where the system signs, serves, and stores its own health claims. Keywords: shared fate, self-attestation, status page dies with the service, dead-man switch, three-valued health, external anchor, availability inversion, tamper-evident theater. NOT for finding missing telemetry emit sites (use observability-absences-audit), building dashboards or alert rules (use grafana-dashboard-builder / monitoring-stack-deployer
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Audit or design a status/health surface so it does not share fate with the system it describes and cannot quietly attest to itself: split read plane (independently deployed reporter + dead-man switch), three-valued verdicts (healthy/degraded/unknown — never green-by-default), external anchoring of any "tamper-evident" status chain, and the availability-inversion check (no client may hard-gate on the monitor). Use when adding a status page, deep-health endpoint, reachability verdict, or signed incident ledger to a service — especially the service every client depends on — or when reviewing a design where the system signs, serves, and stores its own health claims. Keywords: shared fate, self-attestation, status page dies with the service, dead-man switch, three-valued health, external anchor, availability inversion, tamper-evident theater. NOT for finding missing telemetry emit sites (use observability-absences-audit), building dashboards or alert rules (use grafana-dashboard-builder / monitoring-stack-deployer), or log/metric implementation mechanics (use logging-observability, structured-logging-design).
license
Apache-2.0
allowed-tools
Read,Grep,Glob,Write,Edit
metadata
{"category":"Observability & Reliability","tags":["status-page","health-checks","shared-fate","attestation","dead-man-switch","slo"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"observability-absences-audit","reason":"That skill finds the signals a service should emit and doesn't; this skill decides where the resulting status surface may live and who is allowed to believe it."},{"skill":"pd-relay-zero-trust","reason":"Supplies the Merkle-chain and external-anchoring primitives (ADR-0049 I2) that upgrade a status ledger from self-attested to externally verifiable."},{"skill":"circuit-breakers-and-retries","reason":"Once a verdict exists, breaker behavior on `degraded`/`unknown` must degrade rather than brick — that skill covers the client-side mechanics."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]"],"produces":["[Truncated]"]}}
Status Attestation Split Plane
A status surface exists to be believed when the system is broken. Most are built so
they can only be believed when the system is fine. This skill audits the four traps
that make a health surface a lie exactly when it matters, and prescribes the shapes
that survive.
Use This For
Adding a status page, /health-beyond-liveness endpoint, reachability verdict
(e.g. "remote harbors: possible/impossible"), or signed incident ledger to any
service — above all, to the shared service every client hits.
Reviewing a design where the same process signs, serves, and stores its own health
claims ("the Merkle chain proves we didn't edit the outage history" — chained by
whom, anchored where, verified by whom?).
Deciding what a client (mobile app, daemon, CLI) is allowed to do with a health
verdict — gate, degrade, or ignore.
Writing the SLO page for a product whose sales pitch includes its own reliability.
Do Not Use This For
Hunting missing emit sites, dead enums, or absent requestIds — observability-absences-audit.
Dashboard/alert-rule construction — grafana-dashboard-builder, monitoring-stack-deployer.
Log line and metric implementation mechanics — logging-observability, structured-logging-design.
The Four Traps
Trap 1 — Shared fate: the status page dies with the patient
If the reporter runs in the same process, deployment, database, or control plane as
the monitored system, its most informative output — silence — is indistinguishable
from "nobody looked." A hospital ship moored to the sinking vessel.
The fate ladder. Classify every component of the status path (probe, storage,
renderer, alert delivery) by what it shares with the monitored system:
Level
Shares
Example
What its silence means
F0
same process
/health handler in the app worker
nothing — silence is ambiguous
F1
same deployment unit
cron in the same Worker/service
nothing during deploys/outages of that unit
F2
same platform/control plane
second Worker on the same cloud, same D1
platform-wide outage still mutes it
F3
independent platform, dead-man wired
external health checker + PagerDuty dead-man
silence itself pages
Rule: the probe may live at F0–F2 (it must touch the real dependencies, so
proximity is a feature), but the report renderer must be ≥F2 with its own deploy
lifecycle, and at least one silence detector must be F3. A dead-man switch —
"absence of heartbeat is the page" — is the only honest answer to "what if the
monitor is down"; anything else is an unmonitored monitor.
Audit greps (locators, not verdicts — read the surrounding code):
grep -rn "handleHealth\|/health" — does it touch any dependency, or return a
constant? grep -rn "scheduled(\|cron" — do probes share the monitored deployment?
Does any config reference an external checker with a dead-man/heartbeat contract?
Trap 2 — Green-by-default: two-valued verdicts
A boolean health verdict has a forced-error problem: when the probe itself cannot
run, the implementation must pick ok or down, and both are lies. Every verdict in
the system must be three-valued: healthy | degraded | unknown, where unknown
means "the examination did not happen," is rendered distinctly (never as green, never
as red), and is produced by construction whenever a probe times out, throws, or is
skipped — not by a default branch that falls through to ok.
Per-dependency verdicts compose upward with unknown as absorbing for confidence,
not for severity: healthy ∧ unknown = unknown-leaning-healthy (show last-known-good
with age), never plain healthy.
Audit: find every place a health struct is initialized. If the zero value is
ok/true, the surface is green-by-default. Find every catch around a probe: does
it record unknown, or swallow into the previous value?
Trap 3 — Self-attestation: tamper-evident theater
A hash-chained, signed incident ledger stored and served by the vendor it describes
is tamper-evident only to someone holding an external copy of the chain head.
Without an anchor, "we cannot quietly edit our status history" is a claim the vendor
verifies about itself — theater.
The attestation ladder — label every status artifact with its rung, publicly:
A0 self-reported: JSON from the service. Trust: full trust in the service.
A1 self-signed: signed by the service's key. Adds nothing against the service
editing its own history; protects only against third-party tampering in transit.
A2 chained: hash-chained ledger. Detects internal inconsistency; still
rewritable wholesale.
A3 externally anchored: chain head periodically published to a place the
vendor cannot silently rewrite (public git repo commit, DNS TXT, transparency log,
a customer-side subscriber that stores heads). Rewrites now require detectable
divergence.
A4 externally verified: some party actually checks anchors against served
history (a CI job in the OSS repo, customer daemons comparing heads — e.g. gossip
chain-head "beacons" where clients republish the signed heads they saw).
Rule: anchor (A3) ships in v1 of any chained status ledger, not "later" — an
unanchored chain accumulates history that can be rewritten right up until the day
anchoring starts, which is precisely the history an incident would motivate
rewriting. And name a verifier (A4) or admit the rung is A3.
Second self-attestation channel: quorum for outside-in probes. If clients report
their observed vitals (every daemon a probe), one lying reporter must not fake or
mask an outage — weight by quorum across independent reporters, and keep reporter
aggregation k-anonymous if reporters are tenants (see derived-index-consent-boundary
for the tenancy side of that pipe).
Trap 4 — Availability inversion: the monitor becomes the SPOF
The moment a client hard-gates on the health verdict ("app shows nothing until the
status service answers"), the status plane's availability bounds the product's
availability — the hospital ship now runs the port. Inversion smells: a splash screen
blocking on the status fetch; unknown treated as impossible; retry storms against
the status endpoint during its own outage; a kill-switch read that fails closed for
reads (kill switches should fail closed for dangerous writes, open for reads).
Rule: verdicts inform degradation, never gate existence. On unknown:
cached last verdict + age + retry with jitter, full local functionality preserved.
Hard gates are permitted only on impossiblewith machine-readable reasons the
client renders — and even then only for the specific capability that is impossible,
not the app. Corollary (Goodhart): if a verdict gates anything commercially visible,
there is pressure to keep it green — which is why Trap 3's anchoring and quorum
measurement must already be in place before any verdict gains gating power.
Design Prescription (greenfield shape)
Deep probe endpoint in the monitored service (F1): exercises each real
dependency (DB SELECT 1 + hot-table head, KV/config read, a synthetic
end-to-end round-trip through the real delivery path with latency measured), and
any SLO the architecture has already committed to in writing — search ADRs for
promised SLOs (e.g. "revocation propagation ≤ 5s") and probe them; a committed,
unmeasured SLO is the highest-signal gap.
Sample store with retention (7d ephemeral tier) + incident ledger
(durable, chained, anchored A3 on day one).
Split reporter (≥F2): separate deploy unit rendering the report; signed;
pushed to client surfaces over the channel clients already hold (don't add a
poll).
Three-valued verdicts end to end, unknown distinct in every renderer.
Client contract written down: per verdict value, what each client does —
with "hard-gate" appearing only under impossible, per-capability.
The monitor's own footprint budgeted and self-monitored (probe cost, sample
write amplification, dedup governor so a sustained breach logs once per window).
Review Checklist (emit as the audit)
For each item: PASS / FAIL / N/A + evidence (file:line or design-doc quote).
Fate ladder mapped for probe, store, renderer, alerting; renderer ≥F2; one F3 silence detector exists.
No verdict is boolean; unknown is constructible, rendered distinctly, and produced on probe failure by construction.
Every "tamper-evident" claim carries an attestation rung; chains are A3-anchored from v1; a named verifier exists or the rung is stated as A3.
Outside-in reporter input is quorum-weighted; a single reporter cannot flip the aggregate.
No client hard-gates on unknown; hard gates only on impossible, per-capability, with rendered reasons.
Every SLO already promised in an ADR/contract has a probe; the SLO page lists owner + review cadence (unowned SLOs rot).
Kill/pause flags: fail-closed for writes, fail-open (cached) for reads — and which is which is written down.
The monitor's own cost and failure story are stated (what melts first at 10x reporters, and what sheds).
