paradigm-data-discovery
Catalog and query launcher for raw market data available to Dime in s3://dt-exchange-venue-data, s3://dt-paradigm-data, and s3://dt-paradex-data, read with the pod's AWS access. Use for data inventory, coverage, schema and historical analysis, for current venue market data (marks, IV, greeks, OI, funding, spot) on Deribit, OKX, Bybit and Bullish, and for recent block trades when the user has no Paradigm credentials enrolled (with them, paradigm-api reads the live tape). Explains the per-venue raw files and lets the model choose the bounded reads needed for the question. Does not cover account state, positions, vaults, or order placement.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- paradigm-data-discovery
- metadata
- { "author": "tradeparadigm", "version": "2.0" }
- description
- Catalog and query launcher for raw market data available to Dime in s3://dt-exchange-venue-data, s3://dt-paradigm-data, and s3://dt-paradex-data, read with the pod's AWS access. Use for data inventory, coverage, schema and historical analysis, for current venue market data (marks, IV, greeks, OI, funding, spot) on Deribit, OKX, Bybit and Bullish, and for recent block trades when the user has no Paradigm credentials enrolled (with them, paradigm-api reads the live tape). Explains the per-venue raw files and lets the model choose the bounded reads needed for the question. Does not cover account state, positions, vaults, or order placement.
02 资源文件 5
- 参考文档references/datasets.md8.2 KB
- 参考文档references/exchange-raw.md16.4 KB
- 参考文档references/s3-access.md5.5 KB
- 脚本scripts/execution_tape.py9.5 KB
- 脚本scripts/s3_async.py5.0 KB
03 SKILL.md 正文
Paradigm Data Discovery
Use the raw data that is available, then choose the query plan that best fits the user's question. This skill is a catalog and a set of correctness constraints, not a fixed workflow.
Hard rules
- Do not read
s3://dt-exchange-venue-data/hot/. Do not use an object whose name startshot__, even if another skill or script suggests it. Read the direct inputs that replace them, in references/exchange-raw.md. - Check the catalog before declaring data unavailable. Venue and asset assumptions are not evidence that a dataset is absent.
- Bound every read. Select only the venues, data types, currencies, levels, and date/hour partitions needed for the request.
- Respect native schemas and units. Read instrument metadata before combining venue-native volume, premium, IV, or OI; otherwise keep results separated and label their units. Historical conversions require event-time-applicable metadata; the default harmonized-history support is at most 30 days, not the full raw retention.
- Check record timestamps for freshness. Never infer freshness from S3 modification time.
- Fail visibly. Do not invent values, simulate a plausible answer, silently substitute a different window, or describe missing data as a quiet market.
Available data
Read references/exchange-raw.md whenever the request involves exchange prices, trades, blocks, IV, greeks, DVOL, funding, volume, OI, spot, or perps. It documents:
- raw and normalized per-period Parquet layout;
- Deribit, Deribit USDC, OKX, Bybit, and Bullish feeds;
- event and ticker schemas by venue;
- block identifiers and their limitations;
- instrument metadata and unit conversions;
- freshness and bounded-query patterns;
- the non-hot Paradigm RFQ/trade tapes and Paradex trade tape.
Read references/datasets.md for the detailed Paradigm RFQ and executed-trade tape schemas and the Paradex DEX trade schema.
The three buckets are all in ap-northeast-1:
| Bucket | What to use |
|---|---|
s3://dt-exchange-venue-data |
raw/, normalized/, meta/instruments/, and current daily paradigm_trade_tape/ executions |
s3://dt-paradigm-data |
Current Paradigm RFQ CSV and frozen legacy execution CSV |
s3://dt-paradex-data |
Paradex historical trade tape and Parquet parts |
Choose the reads
The model owns the plan. Start from the requested output and choose the smallest raw inputs that can support it:
- Latest mark, IV, greeks, bid/ask, or OI: latest raw
option_summaryrow per symbol for the requested venue/currency; supplement with a venue's public API when the raw partition is absent or stale. - Trades, volume, put/call, or block flow: raw
option_traderows in the requested event-time window; group only on real venue identifiers. - DVOL: Deribit raw
dvol. - Spot or funding: raw
perp_summary, or Bullish raw spot/perp trades when that venue is relevant. - Cross-venue comparison: read each venue separately, harmonise with the newest instrument metadata, then combine only comparable units.
- Paradigm RFQ lookup: current RFQ tape for request metadata; daily
paradigm_trade_tape/partitions for executed legs and raw exchange trades for corroboration. The legacy executed CSV is frozen at 2026-08-10 and must be labelled historical if used. - Paradex history: the Paradex tape, excluding busted trades.
The model may choose raw S3, normalized rows/per-period aggregates, or direct venue APIs based on which gives the clearest answer. It must not choose the hot surface. Use the catalog's Inputs behind the hot files map to check source coverage.
Query execution
Open DuckDB S3 queries with:
INSTALL httpfs; LOAD httpfs;
INSTALL aws; LOAD aws;
CREATE OR REPLACE SECRET s3_irsa (
TYPE S3, PROVIDER CREDENTIAL_CHAIN, REGION 'ap-northeast-1',
ENDPOINT 's3.ap-northeast-1.amazonaws.com'
);
Keep the credential setup and query in the same DuckDB process. See references/s3-access.md for runtime details.
When the agent can execute the query, answer from the results rather than returning SQL alone. When execution is unavailable, provide a ready-to-run query with explicit paths and time bounds.
For "latest option trade data" questions, name both the venue's raw
option_trade path and its companion raw option_summary path, then verify
max(timestamp) in a narrow recent partition. Never invent example row counts,
timestamps, or records and present them as query output.
Output
For inventory or schema questions, answer concisely with the relevant paths,
coverage, fields, and unit caveats. An unqualified "Paradigm trade tape" schema
question means the current partitioned execution tape: show its lowercase
columns and options filter from datasets.md, with the frozen CSV as a
separately labelled historical alternative. Use the uppercase CSV schema only
when the user specifically asks about that CSV or historical source.
For analysis, include:
- the answer;
- the source paths and exact event-time window;
- material gaps, stale inputs, or unresolved joins;
- the query only when it helps the user reproduce the result.
Do not dump the full catalog unless the user asks for it.
Fixed source facts
- The partitioned execution tape uses
rfq_id,block_trade_idandvenue_block_trade_id; CSV tapes use uppercaseRFQ_ID/BLOCK_TRADE_ID. - Filter partitioned executions to options with
instrument_kind == "OPTION"(case-sensitive). Only the CSV tapes usePRODUCT LIKE '%OPTION%'. - The current RFQ tape does not contain execution price, mark, side, trade id, or block id.
- The non-hot executed Paradigm CSV stopped updating on 2026-08-10.
- A raw venue block id is not automatically a Paradigm RFQ id.
- Paradex historical trades must filter
NOT IS_TRADEBUST. - For single-trade analysis, route to
paradigm-block-analyst; for an explicitly requested full recap, route toparadigm-options-recap. Focused questions remain here: choose only the datasets, fields and time range needed.
04 同仓库更多 Skills
example-api
A skill body that exists so the fixture is a complete app. It teaches nothing and is never published.
telegram-bot
Send messages, photos and files to Telegram through the user's own bot, using a bot token held by the DIME credential proxy. Covers finding the chats the bot can reach, posting to a private chat, group or channel, editing and deleting what it sent, and sending from a data feed. Use for ANY request to message, notify or alert the user on Telegram, "send this to my Telegram", "post alerts to my channel", or a 403 from the credential proxy on api.telegram.org.
paradigm-rfq-trader
Trigger institutional block trades via Paradigm's DRFQv2 flow. The workflow is venue-agnostic — resolve instruments, build the RFQ / order payload, benchmark, run a confirmation gate, submit, verify settlement. Per-venue specifics (fair-value sources, naming conventions, edge syntax, settlement checks) live in references/venues.md. In scope today: PRDX (Paradex, primary focus) and DBT (Deribit). Adding more DRFQv2 venues is a references/venues.md edit, not a skill-body change. Covers takers (build, benchmark, cross) and makers (poll, price, manage). Every state-changing action goes through an explicit confirmation gate. Needs both Paradigm credentials enrolled. Use when the user asks to "send a Paradigm block RFQ", "block-trade X BTC", "send a BTC straddle on Paradex / Deribit", "quote rfq_X", "hit the best bid", "cancel rfq_X". Does NOT cover pricing a structure without trading (its legs' current marks are paradigm-data-discovery's), small central-order-book trades, post-trade analysis (paradigm-block-analyst), historical tape (paradigm-data-discovery).
paradigm-options-recap
Options market recap for a requested or default window, invoked via /recap. Parses "/recap [asset] [options] [window]" (e.g. "/recap btc options 8h") and builds the fixed four-section recap — Snapshot, Biggest Print, Block Flow, Vol Surface — from bounded raw exchange venue files and source tapes, never from Dime hot files, read with the pod's AWS access. Use when the user types /recap or asks for a market recap, an options flow summary, "what happened in BTC options" or "last Xh of flow". A single question is not a recap: the biggest print across venues is paradigm-data-discovery's (paradigm-api reads only Paradigm's own tape), and the vol surface, one IV or one venue's data is paradigm-data-discovery's, as are dataset inventory, schema and historical lookups. The output format is fixed — always the same four sections in the same order.
paradigm-block-analyst
Cross-venue analysis of Paradigm RFQ block trades using live market data from Deribit, OKX, and Bybit. Invoked as `/analyze <rfq_id> <rfq description>`: resolves the rfq_id by searching the Paradigm trade tape via the paradigm-data-discovery skill (the daily execution partitions, keyed by RFQ_ID) for the cleared-block record, then fetches live marks, IVs, and greeks per venue, computes net greeks for multi-leg structures, benchmarks the fill vs mark, reports how much of the structure traded over 24h / 7d / 30d and where else it printed, reads whether the flow moved the vol surface, and outputs a concise analysis. Use when the user runs `/analyze <rfq_id> ...`, pastes a Paradigm block trade JSON, or asks to analyze, benchmark, or get market color on a Paradigm RFQ execution. Covers outright calls/puts (CL/PL), strangles (SN), straddles (ST), butterflies (BF), condors (CO), calendars (CA), risk reversals (RR), covered calls, and custom multi-leg combos (CM). Also handles perp combos. Reads the Paradigm trade tape from S3 with the pod's AWS access.
paradigm-api
Paradigm's DRFQv2 RFQ API: recent block trades on the public block tape (`paradigm__paradigm_drfqv2_trades` with `mode: "tape"`), verifying the credentials, instruments, counterparties, listing and creating RFQs, quotes and crosses, cancels, your trades, leg pricing and the maker circuit-breaker. The tools are `paradigm__*` MCP tools, called by exact name, and need the user's Paradigm credentials enrolled. Use for ANY request to Paradigm's REST API, including "recent block trades", "is my Paradigm key working", "list my Paradigm RFQs" or "cancel rfq_...". The tape here is the latest prints only: volumes over time, comparisons across days and whether a structure traded before are paradigm-data-discovery's, and so is the block tape without credentials. The RFQ workflow and its confirmation gate are paradigm-rfq-trader.