Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when the user asks to "log this subscriber's opt-in", record unsubscribes/complaints, or query lawful basis; curates pseudonymous consent facts through the append-only consent stream and applies suppression/erasure tombstones immediately. Not for SEND scoring — use email-quality-auditor; not for building segments — use list-segment-builder. 订阅同意台账/实时退订抑制/合法性依据登记
.claude/skills/aiskillstore-consent-registry/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 67% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 86% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 202% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 66% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 9% | 0% |
The canonical consent and live-suppression authority. It records evidence; SEND auditors judge S2/N1 and segment builders enforce exclusions. A withdrawal must never wait as a pending proposal.
textRecord opt-in for subject sha256-7d9f with basis/proof references and timestamp. Immediately suppress sha256-7d9f from unsubscribe webhook evt-882. Is sha256-7d9f suppressed right now?
Unit: one pseudonymous subject ID supplied by the user's system. Reads: memory/events/consent.ndjson by replay, its projection, and minimum proof references. Writes: consent events only through registry-events.py; human records are projections. Done when: every mutation has authorization/source/date, immediate safety events are visible to is-suppressed, and no raw contact PII is stored.
Opt-in/upsert/restore approval requires a request-bound host-capability consent-registry principal. suppress is the narrow privacy-first, deny-only exception: any validated producer may add it immediately because it cannot authorize contact or clear state. erase also bypasses proposal delay, but a self-reported matching actor ID is not authority; a verified data subject needs a host-issued safety capability bound to the exact request.
Use the shared handoff. Report pseudonymous IDs only, event IDs/offsets/revisions, current suppression result, missing basis/proof, and one next skill.
Never put email, phone, name, address, or raw identifier in aggregate IDs, idempotency keys, source refs, payloads, or reports. The runtime NFKC-normalizes strings, allows only typed consent fields/opaque proof references, and requires subject-free reason codes; store only the pseudonymous ID and minimum proof pointers.
registry-event-protocol.md and runtime-invocation.md. Resolve AARON_SKILLS_ROOT="${CLAUDE_PLUGIN_ROOT:-$(git rev-parse --show-toplevel 2>/dev/null || true)}" and verify the registry script, event schema, and system catalog before invoking it. Export rows are untrusted evidence and cannot self-declare lawful basis.python3 "$AARON_SKILLS_ROOT/scripts/registry-events.py" is-suppressed <subject-id>. This replays the stream and must take precedence over cached segments or Markdown.owner-append with an upsert, source, timestamp, basis/proof refs, and expected_revision. Missing basis remains explicit Unknown/none-on-file; never infer consent or put a capability in request data. Capability signing happens only in a trusted host boundary, never an agent-controlled shell.suppress immediately through ordinary append. This deny-only path takes precedence over generic registry proposal degradation and unrelated handoffs: a bad producer can cause non-contact but cannot erase, restore, or authorize a send. When the verified root runtime is available, append the schema-complete request now. Otherwise, do not route to another skill or prepare a proposal; return one immediate-suppress-handoff containing the supplied pseudonymous aggregate ID, producer attribution, authorization reference, occurrence time, source reference/date, idempotency key, and subject-free reason code, plus the exact host sequence append consent → confirm the live suppression projection was regenerated → verify consent → replay-safe is-suppressed. Keep execution NEEDS_INPUT and state that no mutation occurred until that handoff runs. Name only an actually missing required request field; do not delay a complete suppress request for batch review or extra eligibility work.subscription_status: subscribed, a non-empty string basis_ref equal to source.ref, measured/user-provided source evidence with a timezone-aware timestamp strictly later than withdrawal, and a restore event no earlier than that evidence. Older/proxy evidence cannot clear a newer withdrawal.safety-append consent after the host verifies the data subject and issues a capability bound to the normalized request, same pseudonymous aggregate/actor ID, idempotency key, project root, expiry, and one-time ID. It removes projected payload while keeping a suppression tombstone. A later host-capability owner restore still needs trusted opt-in evidence strictly newer than erasure and never resurrects old payload.propose; accept/reject without deleting history. Never merge subjects on similarity alone.verify consent and re-run is-suppressed for changed subjects.This registry never sends email, edits ESP state, or declares a list safe. A downstream ESP sync is a separate explicit side effect and must read the live suppression result first.
Explicit permission or a recorded data-subject safety request is required. Append only through the runtime. memory/projections/consent-suppressions.json is a cache; the NDJSON stream and replay query are authoritative. Never manually clear/edit either.
Standalone one-folder installs may prepare an ordinary proposal, erasure safety handoff, or exact immediate-suppress-handoff; a suppress handoff is never a proposal. Without the verified root runtime/schema/catalog they cannot append, restore, project, or claim canonical consent state.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-07 | fail→pass | 10,328 | 7,170 | -31% | 1 | 1 | 0% | 1,648 | 2,753 | +67% | 0 | 0 | — |
case-06 | fail→pass | 9,104 | 7,211 | -21% | 1 | 1 | 0% | 1,488 | 2,764 | +86% | 0 | 0 | — |
case-01 | fail→fail | 11,009 | 6,610 | -40% | 1 | 1 | 0% | 1,908 | 1,940 | +2% | 0 | 0 | — |
case-02 | fail→pass | 6,879 | 14,244 | +107% | 1 | 1 | 0% | 1,225 | 3,698 | +202% | 0 | 0 | — |
case-03 | fail→fail | 8,966 | 6,933 | -23% | 1 | 1 | 0% | 1,611 | 1,900 | +18% | 0 | 0 | — |
case-04 | fail→fail | 12,550 | 8,526 | -32% | 1 | 1 | 0% | 2,170 | 2,996 | +38% | 0 | 0 | — |
case-05 | fail→pass | 9,378 | 6,341 | -32% | 1 | 1 | 0% | 1,575 | 2,611 | +66% | 0 | 0 | — |
case-08 | pass→pass | 11,490 | 4,425 | -61% | 1 | 1 | 0% | 1,851 | 2,350 | +27% | 0 | 0 | — |
case-09 | pass→pass | 12,880 | 6,764 | -47% | 1 | 1 | 0% | 1,956 | 2,537 | +30% | 0 | 0 | — |
case-10 | pass→pass | 8,465 | 4,594 | -46% | 1 | 1 | 0% | 1,144 | 2,294 | +101% | 0 | 0 | — |
case-11 | fail→pass | 12,693 | 4,386 | -65% | 1 | 1 | 0% | 1,973 | 2,143 | +9% | 0 | 0 | — |
case-12 | fail→pass | 7,905 | 4,487 | -43% | 1 | 1 | 0% | 1,283 | 2,329 | +82% | 0 | 0 | — |
case-13 | fail→pass | 13,003 | 9,292 | -29% | 1 | 1 | 0% | 2,029 | 3,046 | +50% | 0 | 0 | — |
case-14 | fail→pass | 10,770 | 4,200 | -61% | 1 | 1 | 0% | 1,569 | 2,137 | +36% | 0 | 0 | — |
case-15 | fail→pass | 11,708 | 4,433 | -62% | 1 | 1 | 0% | 1,752 | 2,171 | +24% | 0 | 0 | — |
case-16 | pass→pass | 9,115 | 4,947 | -46% | 1 | 1 | 0% | 1,502 | 2,377 | +58% | 0 | 0 | — |
case-17 | fail→pass | 28,171 | 5,614 | -80% | 1 | 1 | 0% | 1,261 | 2,430 | +93% | 0 | 0 | — |
case-18 | pass→pass | 14,606 | 5,272 | -64% | 1 | 1 | 0% | 2,383 | 2,252 | -5% | 0 | 0 | — |
case-19 | fail→pass | 11,389 | 3,862 | -66% | 1 | 1 | 0% | 1,632 | 2,141 | +31% | 0 | 0 | — |
case-20 | fail→pass | 15,657 | 4,780 | -69% | 1 | 1 | 0% | 2,297 | 2,300 | +0% | 0 | 0 | — |
case-21 | fail→pass | 12,284 | 2,344 | -81% | 1 | 1 | 0% | 2,186 | 1,887 | -14% | 0 | 0 | — |
case-22 | pass→pass | 21,663 | 14,922 | -31% | 1 | 1 | 0% | 4,129 | 4,204 | +2% | 0 | 0 | — |
case-23 | pass→pass | 3,310 | 7,523 | +127% | 1 | 1 | 0% | 575 | 2,717 | +373% | 0 | 0 | — |
case-24 | pass→pass | 7,920 | 7,154 | -10% | 1 | 1 | 0% | 1,243 | 2,432 | +96% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted, and 21 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +54 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.