tres-rollup-review

bởi anthropic

Tính toán chính xác tác động của từng giao dịch con đối với một quy tắc rollup trong TRES Finance — mọi bộ lọc được áp dụng — và gắn cờ các vấn đề thực sự (không có kết quả khớp, trùng lặp, cấu hình…

npx skills add https://github.com/anthropics/claude-plugins-community --skill tres-rollup-review

TRES Finance — Rollup Review

Take a fully-specified rollup rule and compute exactly which sub-transactions it will claim — with every one of its filters applied — then flag genuine problems. You do NOT design new rollups and you do NOT execute mutations.

All GraphQL runs through the TRES Finance MCP connector (execute tool). All variable keys and nested input fields MUST use camelCase, never snake_case.


What this adds over tres-rollup-rules

The tres-rollup-rules discovery step groups sub-txs only at the (wallet, asset, direction) level — it never applies a single candidate's full filter combination (fees + method_ids + subtx_type + amount bounds + counterparty + excludeRollups together). This skill does exactly that, per rule, via the subTransactionRollupRulePreview query (which runs the engine's real query_sub_transactions), giving:

  • the exact match count → the zero-match gate (a rule that matches nothing must not be created), and
  • pre-create overlap awareness, so the user only ever sees viable proposals.

It is also usable standalone to review a rule someone submitted (Mode B) — a case the discovery skill doesn't cover.


Two modes

Mode A — pre-create gate (invoked by tres-rollup-rules)

When tres-rollup-rules invokes this skill (via the Skill tool), these instructions load into the same context that already holds the candidate rule objects — they are not passed as arguments, they are already in the conversation above. Iterate those candidates one at a time; for each, run the checks below and emit exactly one verdict:

  • PASS — clean, ready to propose (include the exact sub-tx + parent-tx count).
  • DROP — matches 0 sub-txs; must not be created.
  • FIX — a config problem with the exact change (e.g. "minAmount > maxAmount", "recipientIdentifier on an INFLOW rule — should be senderIdentifier").

Overlaps are reported as notes, not DROP/FIX (see Check 2). Produce one verdict line per candidate, then control returns to tres-rollup-rules Step 2.5.

Mode B — standalone review (when asked to review submitted/pending rules)

Fetch pending rules and review each; produce a per-rule findings summary.

query PendingRollupRules {
  subTransactionRollupRule(status: "Pending", limit: 50) {
    totalCount
    results { id name interval startDate endDate status rule createdBy createdAt }
  }
}

The status filter is title-case ("Pending", "Active", "Disabled"). If totalCount is 0, say "No pending rollup requests in this org" and stop.


Ground truth — how a rule selects sub-transactions

A rule claims sub-transactions matching ALL of (this mirrors the engine's query_sub_transactions):

  • belongsTo_In: [internal_account_id], asset_In: [asset_id] (the asset KEY), platform matches, balanceFactor matches balance_factor
  • already-rolled-up (excludeRollups) and derived/locked/synced sub-txs excluded
  • fees: EXCLUDE drops GAS/FEE, ONLY keeps only GAS/FEE, INCLUDE keeps all
  • optional narrowing: method_ids, min_amount/max_amount, subtx_type, sender_identifier/recipient_identifier, original_sender_prefix/original_recipient_prefix

The rule JSON (as stored / returned) uses snake_case. balance_factor = -1 is OUTFLOW, +1 is INFLOW — get this right or every direction check is backwards.

Verified subTransaction filter names

IntentFilter
WalletbelongsTo_In: [ID]
Asset by keyasset_In: [ID]
DirectionbalanceFactor: Float (-1 / 1)
Min / max amountamount_Gte / amount_Lte
Sender / recipientsender_Identifier_In: [String] / recipient_Identifier_In: [String]
Method idtx_MethodId: String (single) / tx_Classification_MethodId_In: [String]
Sub-tx typetype: String / type_In: [String] (lowercase FinancialAction value)
Exclude already-rolled-upexcludeRollups: true
Exclude gas/feeexcludeGasFees: true
Datestimestamp_Gte / timestamp_Lte (DateTime)

type / type_In take the lowercase FinancialAction value ("gas", "fee", "reward", …) — opposite casing from the rule's UPPERCASE subtxType enum. So subtx_type is countable exactly (type: "<value>") and fees: ONLYtype_In: ["gas", "fee"].

Gotcha: fees: ONLY already selects GAS+FEE, so a gas/fee rollup needs fees: ONLY alone — adding subtxType: GAS on top is redundant (and narrows to GAS, dropping FEE).

typeId is NOT a rule filter here. type_id is only the rollup rule's id when paired with type_In: ["rollup", "rollup_fee"]. On other types it means something else entirely. To inspect what an existing rollup rule produced, query type_In: ["rollup","rollup_fee"], typeId: "<rule_id>" — never typeId alone.

Sentinel dates (operational note): if startDate == "0001-01-01" or endDate == "9999-12-31", omit that bound (don't pass the sentinel — it may not parse). Unbounded dates are normal and preferred; they are NOT a finding.

Evidence: the on-chain hash is tx.identifier (sub-select tx { identifier }), never typeId. For examples also pull amount, sender { identifier }, recipient { identifier }, asset { symbol key }.


Scope — DISABLED rules are out of scope

Only Active and Pending rules are live. Never query, compare against, or flag DISABLED rules (they're superseded versions).


The checks

Check 1 — Exact impact (the core; always run)

Get the ground-truth count straight from the engine via the subTransactionRollupRulePreview query — it runs the real query_sub_transactions with all of the rule's filters applied server-side, so you never reconstruct filters by hand:

query RollupPreview($rule: RollupRuleInputType!, $startDate: Date, $endDate: Date, $interval: Interval) {
  subTransactionRollupRulePreview(rule: $rule, startDate: $startDate, endDate: $endDate, interval: $interval) {
    subTransactionCount
    transactionCount
  }
}

Pass the candidate's rule object verbatim as $rule — the same camelCase shape as the create mutation (internalAccountId, assetId, platform + balanceFactor + fees as enum NAMES like OUTFLOW/ONLY, plus any methodIds/subtxType/minAmount/maxAmount/identifiers). Always pass $interval = the candidate's interval (DAY or MONTH); it governs how the eligibility-window boundary is normalized, so a MONTH rule and a DAY rule over the same data can return different counts. startDate/endDate are optional YYYY-MM-DD — omit for the rule's full unbounded range.

The count is what the rule would actually roll up on its next run, not raw all-time matches: the engine funnels the window through its own eligibility logic, so the current incomplete period and the unsynced/buffer tail are excluded. transactionCount is the distinct parent-tx count. (Already-rolled-up sub-txs are also excluded — for what an existing rule has already produced, see the existing-rule section below.)

  • subTransactionCount == 0 → DROP (Mode A) / blocking flag (Mode B): "matches 0 sub-txs as configured" — name the filters that zero it out (common culprits: minAmount > maxAmount, an identifier on the wrong direction, or an inverted date range). The reason is the useful part.

If subTransactionRollupRulePreview is not in this org's MCP schema (confirm with validate_query or introspect), fall back to a fully-filtered count query: a single subTransaction(... excludeRollups: true ...) that applies every one of the rule's filters using the verified filter-name table above, selecting only totalCount. This is slightly less precise than the engine preview (it does not exclude the incomplete period / buffer tail) — say so when you report the number, and keep the ~ prefix.

Check 2 — Overlap (informational, NOT a hard limit)

Another Active/Pending rule (or another candidate) on the same internalAccountId + assetId + platform + balanceFactor with intersecting dates and non-disjoint filters can claim some of the same sub-txs. This is not blocking: the engine processes rules one-by-one ordered by id (PK), so a sub-tx matching multiple rules is always claimed by the same rule consistently. Surface it as a note (other rule name + id, up to 3 example tx hashes) so the user is aware — never DROP or force a split for overlap alone.

No separate config-validation check. Date-order, identifier-on-wrong-direction, mutually-exclusive prefixes, and bad cutoffTime are enforced by the create mutation (returned as blocking validationIssues) — don't re-validate them here. A start date set to clear a locked period is a good use, not a flag (leave the end date open).

Existing-rule impact (Mode B — reviewing an Active/Pending rule)

Do NOT call an existing rule useless because Check 1 returns a low count. A running rule has already rolled up its pool, and those rollups are excluded from Check 1 — so a rule that has been doing its job shows only the few recent un-rolled sub-txs. Also count what it produced: subTransaction(type_In: ["rollup","rollup_fee"], typeId: "<rule_id>") { totalCount }. Report both — "currently matches N new sub-txs; has already produced M rollups". A rule is only ineffective if it has produced ~nothing and matches ~nothing now. (If its filter looks stale — e.g. a methodId that no longer matches new txs — say that, with the produced-vs-current numbers, rather than "useless".)


Output

Mode A: one line per candidate back to tres-rollup-rules: <#> <name>: PASS (<subtx>/<tx> sub-txs) / DROP — <reason> / FIX — <change>, plus any overlap notes.

Mode B: one findings block per rule. Flat bullet list, each bullet with concrete evidence; omit a check that found nothing. Always include the impact line. If clean: "Reviewed (id ) — matches ~ sub-txs across ~ txs. No issues."


Rules

  • Never query, compare against, or flag DISABLED rules.
  • Never propose new rollups or optimizations — that's tres-rollup-rules. You only compute impact and flag problems.
  • Never execute mutations.
  • Every flag carries evidence (a count, a date, a config value, or a tx hash).

Thêm skills từ anthropic

analyzing-financial-statements
anthropic
Kỹ năng này tính toán các tỷ số và chỉ số tài chính chính từ dữ liệu báo cáo tài chính để phân tích đầu tư.
applying-brand-guidelines
anthropic
Kỹ năng này áp dụng nhận diện thương hiệu và phong cách doanh nghiệp nhất quán cho tất cả tài liệu được tạo ra, bao gồm màu sắc, phông chữ, bố cục và thông điệp.
creating-financial-models
anthropic
Kỹ năng này cung cấp bộ công cụ lập mô hình tài chính nâng cao với phân tích DCF, kiểm tra độ nhạy, mô phỏng Monte Carlo và lập kế hoạch kịch bản cho đầu tư…
board-minutes
anthropic
Soạn thảo biên bản cuộc họp hội đồng hoặc ủy ban theo định dạng của bạn. Tự động phát hiện các cuộc họp hội đồng và ủy ban sắp tới từ lịch của bạn, yêu cầu chương trình họp và…
crm-cleanup
anthropic
Quét HubSpot để tìm các deal đã cũ, danh bạ trùng lặp và các trường bị thiếu, sau đó sửa những gì chủ sở hữu phê duyệt. Chấp nhận đối số phạm vi tùy chọn cho deals, contacts,…
redshift-api
anthropic
Chạy SQL trên Amazon Redshift — gửi câu lệnh, kiểm tra trạng thái, phân trang kết quả và duyệt cơ sở dữ liệu/lược đồ/bảng. Sử dụng công cụ này bất cứ khi nào người dùng muốn…
ticket-deflector
anthropic
Đọc email hoặc ticket của khách hàng được chuyển tiếp, lấy trạng thái đơn hàng/hoàn tiền từ PayPal và lịch sử tài khoản từ HubSpot, soạn thảo phản hồi phù hợp với giọng điệu của chủ sở hữu…
reg-feed-watcher
anthropic
Kiểm tra các nguồn cấp dữ liệu quy định ngay bây giờ và báo cáo những gì mới kể từ lần kiểm tra trước, được lọc theo ngưỡng trọng yếu của bạn. Sử dụng khi người dùng nói "kiểm tra các nguồn cấp dữ liệu",…