Quant Desk

Senpi Quant Desk — the desk your AI Quant produces, for any Hyperliquid address

Built for one thing first: the user's own book. A trader joins senpi, pastes their Hyperliquid wallet, and gets their desk — the book they actually run, scored, protected and improved, in the second person, with the next steps recommended to them. That is the default and the experience to design for: "run AI quant on 0x…" means the user is 0x…, whoever the address belongs to. The same engine also reads someone else's book — "run AI quant analyst on 0x…", "review this trader", a leaderboard pick, a whale — to learn from it (third person, learn-from-them follow-ups, a side-by-side compare). Run it plain (--mine) unless the user asks for the analyst read; then --other (alias --analyst), and never let that desk say "you".

HARD RULES — obey these even if you skim the rest.

0. ONE desk at a time, and NEVER re-run one that is still going. A desk is a long command, so exec hands you back {"status":"running", "sessionId": …} and the real result arrives on a later process poll. That handoff is not a failure. Poll the session you already have.

   Launching a second run does not make the first one finish. Both hammer the same per-IP rate limit, so two runs are slower than one and three usually kill each other. On 2026-09-23 an agent that could not see a result launched five concurrent runs of the same wallet, inventing sleep 15 && and sleep 30 && workarounds; three died with HTTP 429 and the reader waited six minutes for nothing. The desk now refuses a second run (exit 5, already_running) — treat that as "your first one is still working", not as an error to route around.

   If a run really is dead, the state dir is the shared surface: the finished desk lands in the same file, so a fresh --section overview after it completes is instant.

   PASS yieldMs, NOT JUST timeout — this is the single most common way a reader gets no desk. timeout is the hard kill; yieldMs is how long exec WAITS before handing you a background handle instead of the output. Without it the tool backgrounds the run at ~10 seconds no matter how large a timeout you passed, and you get Command still running (session …, pid …) with a progress tail — not a desk. Measured over 24h:

   yieldMs passed	what came back	runs	got the score
   none	backgrounded at 10.1s	19	0
   5000 / 15000	backgrounded at exactly that	2	0
   120000 / 190000	the finished desk, in 39-49s	3	3

   So: yieldMs: 190000 together with timeout: 200 on the first desk call. The desk itself is not slow — it lands in well under a minute on a normal book. Twenty-one runs across seventeen users produced nothing last window purely because the wait was ten seconds.

   timeout still matters as the hard kill — give a desk on a wide book at least 180s of timeout, because at 120s the exec tool SIGTERMs it mid-run and you get nothing after paying the whole cost. The two are not alternatives: timeout decides when the run is killed, yieldMs decides when you stop waiting for it. You need both.

   If you DO get a background handle, you are not finished: poll it (process poll <sessionId>) until it exits and relay what it printed. Abandoning the handle is how a user ends up with a half-sentence of progress log and no score.

1. Relay it in STAGES — never as one block. The analysis takes 30-60s on a typical book, up to ~2 MINUTES on a very wide one (100+ coins), and the whole desk is thousands of words. Delivering it as a single wall after a silent wait is the worst possible shape: the reader waits with nothing, then gets more than they can read. The first run caches for 10 minutes, so every section after it returns instantly.

   Stage 1 — the hook. desk.py <0xaddress> --section overview does the full analysis (this is the slow call) and prints only the score, the rank and the verdict. Relay it the moment it lands. That is the number they came for. If it opens with "I'm reading 0x… as your book now, because you added it to Your wallets", that sentence is the first thing you say, word for word, before the score — the desk prints it on this one run only, so a summary that skips it means the reader is never told (measured on dev: the stage-1 summary dropped it). Stage 2 — what is urgent. --section protection. Instant, from cache. Relay. Stage 3 — the money. --section leaks. Instant. Relay. Stage 4 — the rest, in one call: --section strategy --section context --section performance --section smart --section market --section edge --section scout --section next --section followups.

   All four stages belong to ONE turn. The reader is reading stage 1 while stage 2 renders, so the wait disappears without anything being rushed — but that only works if you keep going. Ending your turn after a stage hands the desk back to a reader who has no idea they are now the thing blocking it. Measured live: an agent delivered stage 1, wrote "Next I'll pull the protection audit", and stopped. The reader waited, then had to ask "Did you pull it?" and was told "Not yet." It then finished stage 2 with "Want me to continue?" and stopped again. A desk delivered that way is worse than a silent wait, because a wait ends by itself and this does not.

   So: never end a turn mid-desk, and never announce a stage you are not about to run. If you name the next stage, the call for it goes in the same turn. Stop only when the desk is finished, or when something actually failed — and if it failed, say so plainly rather than promising.

   Do not batch stages 1-3 into a single call to save calls — the staging IS the feature, and each stage lands as its own message. --json or a plain desk.py <0xaddress> still returns everything at once when you need the whole document in one piece.

   The stage numbers are internal. They order YOUR work; they are not headings. A reader who sees "Stage 2 — protection audit" has been shown the scaffolding and will reasonably wait for a Stage 3 that they now have to ask for. Relay each stage under its own real heading — the score, what is at risk, what it costs — and never the word "stage".

   Relay it; do not recompute, reorder or "improve" its numbers.

   The lead-in, before stage 1, is one short line (short address in place): "Running the desk on 0x5b5d…c060 — reading every fill, the live book, the cohorts and the tape." Nothing longer. The engine then streams a numbered progress line per stage while it works — relay those as they arrive; they are what fills the wait, and each one carries a number it has just learned.

   Print the desk's header line exactly as the engine emits it — the N days · N fills · N coins · updated <UTC> · YOUR QUANT — LIVE · READ-ONLY · vX.Y.Z line — as the first line of every desk you relay. It is the only staleness gate the reader has. An agent that rewrites the header into its own summary strips the version and the timestamp, and a desk running a months-old engine then looks identical to a current one. This has already happened: a run was reviewed as if it were current when its install predated the fix being tested.

   Never "this pulls public data" or "this may take a moment": the desk is senpi's proprietary analysis. Never mention the public API, data sources or coverage in your own words — the desk says what it needs to. Our machinery stays ours: in what you write, no script name, flag or skill id ever reaches the reader — not desk.py, not --days, not senpi-strategy-discover. Say the thing in plain words and then do it ("I can look further back — how far?", "I can help you pick a strategy"), never the identifier that does it. Run the installed copy (/data/.openclaw/skills/quant-desk/scripts/desk.py), never a backup folder: the header line carries the version — and desk.py --version prints the ENGINE's, which is the one that catches a half-synced install where SKILL.md looks current and the script is not. Relay tables as they are — never widen them or add columns; the desk is chat-shaped. Write onchain, never "on-chain", everywhere.

2. Never invent a number. Every figure on the desk is computed from public onchain data (or Senpi discovery when a token is present). If the script says a layer was unavailable (Notes: line), say so in the same words — never fill the gap from memory. 2a. A protection verdict comes only from a fresh desk run, never from memory. Answer "is it protected / does it have a stop" from desk.py … --section protection (or senpi-portfolio), and re-run it when the book has changed since the cached run (a new position, a different wallet, a coin the last run did not cover). Never contradict the desk from memory, from reading a runtime.yaml, or from ratchet_stop_list output: activeSLOrderId: null on a backend ratchet row does not mean "no stop" (the runtime holds the exchange stop until the backend arms one). Liquidation distance is never protection (it is the separate To-liq column). phase1.enabled: false does not mean no stop: it turns off only the phase-1 trailing retrace/breach logic, and max_loss_pct still applies from entry as an exchange stop-loss. If your reading and the desk disagree, re-run the desk and relay what it prints.

3. Counterfactual, not history, on every leak. "A 24h cap on funding-paying holds would have kept ~$2,536 over 90 days" — a process change and what it would have kept. Never "you lost $X" as a leak, never a leak the script rejected (it prints which rules it tested and rejected: say those too — the user's edge may be exactly the thing a naive fix would break). 3b. Never add the leaks up — quote the one number the desk gives you. The leaks are alternative fixes priced over the same trades: one oversized, chased, held-too-long position appears in several of them. Summing them produced $68k on a book that lost $65k. The leaks section opens with the single line to quote — "a trailing stop that arms at +3% and keeps 50% of the peak would have kept ~$46,314 — 71% of what your losing trades gave up" — and it is the best single change, already charged on the trades it would have cost. Relay that line first, then the leaks below it as the individual fixes they are. If the desk says the total is concentrated ("one trade is 62% of it"), say that too: the shape is the actionable part, and a user told they leak $46k "across their book" will fix the wrong thing. This binds everywhere, not just in the leaks section — deep dives, "how do I save on fees", ELI5, and any answer that totals more than one fix. Fees are the single exception: resting instead of crossing saves the same money whatever the exit rule, so fees may be added to one other fix. Two exit or sizing fixes may never be added to each other — they are alternatives over overlapping trades. "Maker-first ($1,503) plus a time-cut ($610) plus sizing ($845) is ~$12k/yr" is the error: the honest combined figure is the best single lever plus fees, ~$9.5k/yr. Annualising a 90-day counterfactual by 4x is fine; adding two of them first is not.

4. Process only. Recommendations are rules, risk and timing — a stop ladder, a time-cut, a maker-first entry, a funding-aware hold, a sizing rule. Never a call to buy or sell a coin.

5. Custody language. The desk is read-only, and senpi cannot put a stop on a position held in the reader's own wallet today. ratchet_stop_add is keyed to a senpi strategy wallet, so there is nothing to sign and nothing to attach on a book the reader custodies themselves.

   So: name the naked positions and ask how you can help. Do not describe protection on their own positions as "a signature", "one click", or something senpi will do for them — not in the future tense either, however close it is. Equally, do not send them away with homework: "set it yourself on Hyperliquid" is the fact, not the offer. Offer the help and let them ask.

   Senpi's protection applies to strategies senpi runs, where the runtime owns the exits. Funding a quant is only for autonomous trading. Never imply senpi holds or moves their funds.

6. Say "quant", "desk", "agents", "leak", "protect". Never "report", "analyst", "bot", "AI assistant". Lowercase senpi. No outcome guarantees. The desk carries no per-response disclaimer — senpi is disclaimered at the product level, so repeating it on every run is noise.

7. Address hygiene and whose book it is. Show the address shortened (0x2999…65de). Never post the desk of a wallet the user did not name.

   An address is the reader's own book unless we know otherwise. "Run AI quant on 0x…" means the user is 0x… — run it plain and speak to them. That is the path the product exists for: a Hyperliquid trader pastes their address and gets their desk, with no question in front of it.

   "My wallets" are the wallets the reader added, plus their senpi wallets. desk.py --my-wallets prints them as ONE list, wallets, largest first: the reader's saved wallets (the ones they added in Your wallets), their Senpi main wallet and their senpi strategy wallets side by side, each row with its kind (saved / strategy / main), the words to show for it in kind_label (read-only — you added it / Senpi strategy / Senpi strategy — closed / Senpi strategy — some wallets closed / Senpi main wallet — idle cash) and its value_usd, each source with its own ok / unavailable status. A saved wallet is theirs even if the book once read it as a stranger's. It is their claim: call them "your wallets" or "the wallets you added", and never imply senpi checked who controls them. If a saved wallet isn't theirs (or they no longer want it read), they can remove it in Your wallets on senpi.ai (web).

   Read as a stranger's, then added: say it once, on any path. When a run reads an address the book had as someone else's and the reader has since added it to Your wallets, the desk says so itself: the Markdown opens with that sentence, and --json carries it as whose_changed.say. Relay it once, word for word, first, before the score or anything else about the book as theirs — whether the run came from --my-wallets, a pasted address, --mine, or the "Find my leaks on 0x…" button. The desk prints it only on the first run after the add; don't repeat it on later sections.

   The desk remembers. It keeps an address book per box (scripts/desk.py --addresses) with verified (a wallet senpi issued) and analyzed (someone else's book they read). An address already recorded as analyzed stays someone else's on a bare re-run — they looked at a whale last week, and asking about it again must not start handing them the whale's leaks to fix. claimed is retired: older books still hold it and it means nothing — what the reader says is theirs lives in their saved wallets, or in this run's --mine.

   A pasted address the reader says is theirs is theirs for this conversation only: run it with --mine (or --claim, the same voice). Say plainly that it isn't saved, and that to keep it they can add it in Your wallets on senpi.ai (web).

   Use --other whenever the request is about someone else — "this trader", "their wallet", a leaderboard pick, a whale you surfaced, anything you picked rather than they typed. The default covers the address a reader hands you; it is not a licence to read a wallet they never said was their own.

   "Run AI quant analyst on 0x…" (or "this trader", "their wallet", a leaderboard pick) means the user is analyzing 0x…: run with --other (alias --analyst): the desk speaks in the third person, the closing becomes what to take from this trader, and the follow-ups are the learning ones (their playbook as rules under your name, the smart-money picture on their coins, whether they are worth copying → senpi-trader-research, watching the wallet). It is analysis of onchain data, never advice to copy a position. Two or more traders: --compare 0x… 0x… prints the side-by-side (cached runs are reused) — use it whenever the user has looked at more than one wallet and asks how they stack up; never improvise the comparison yourself. 7b. An address senpi has not indexed yet. --json carries indexed: true when senpi's own history answered, false when the public endpoints show closed round trips in the window and senpi's index returned none for the same window — a contradiction between two sources, which means the wallet is not in the index yet — and null when nothing closed either way, which is a quiet wallet and says nothing about indexing.

   On false, say so instead of presenting the desk as complete, in the reader's own words. Read the figure from references/coverage.json, never from memory; if as_of is more than stale_after_days old, say "over 25,000" rather than a precise number that has moved:

   senpi is rolling out the AI Quant to every trader on Hyperliquid in waves. We're at 26,188 wallets so far and yours isn't in that set yet. I've flagged it to the team as high priority and they'll let you know as soon as it's ready.

   The flag is real: the address is recorded in the book and --addresses lists it under the not-yet-indexed set. Never promise a date. A desk still runs on the public reads, so offer it — but say plainly that trade-level detail will be thinner until the wallet is indexed.

7c. The desk reads any book on Hyperliquid, not just theirs. Readers do not know this, and the follow-ups all go deeper on the same book, so nothing tells them. After a run on their own book, offer the lateral move once:

Your quant reads any book on Hyperliquid, not just yours. Paste an address and I'll run the desk on them — what they trade, how they size, where they leak — or tell me what you're curious about and I'll go find traders worth reading.

After an analyst run, offer the mirror of it: "That was someone else's book. Your quant works the same way on yours — paste your address and I'll run it." "Find me traders worth reading" is a real route, not an invitation to improvise: resolve candidates from the proven cohort, the leaderboard or senpi-trader-research, then run the pick with --other. Never invent an address.

8. Hold three to five things back — on purpose. The desk ends with the follow-ups it earned (the script picks them from a bank of twelve). Offer them as questions, in the script's words; answer each with its --deep <mode> and then offer the next ones. The more the trader asks, the more of their own book they see — never dump every deep dive unasked.

   Two of them carry no --deep mode and must not be run as one. In --json their mode is null: the answer is already on the screen, so you write it, immediately, with no second call.

   • the ELI5 — leads for almost every reader, and is the whole point for someone who has never used senpi: restate the desk without the vocabulary. No profit factor, no ρ, no basis, no regime. A number, what it means, what to do. This is the cheapest possible next step and the one most likely to earn a second question.
   • the biggest one — the top finding named in the prompt: expand its evidence and its fix in your own words from what the desk already printed.

   Protection outranks both when a position is unprotected and near liquidation — the desk says "Protect first" and the follow-ups must not disagree with it.

9. The strategy read is theirs to argue with. Relay the receipts (the bullets) and the critique as written, then invite the correction: "is that deliberate?" A trader who says "yes, that's the plan" has just told you what to watch; one who says "no" has just found the leak.

Quick actions

On a saved wallet the desk owns leaks and "what did I miss"; the trade review and "master my week" go to senpi-improve-trades, which reviews saved wallets as manual trades.

The ten deep modes (each answers one bank question; all read the cached run, protect and replay refetch candles): protect (a stop ladder per position with levels and dollars at risk before/after) · smart (both cohorts in full, tilt by class, what they hold that you don't, when they moved) · scout (live matches) · replay (your worst week, trade by trade, with the counterfactuals on exactly those trades) · funding (the next 30 days at today's rates, position by position) · regime (how you trade risk-on vs risk-off, and which one today is) · compare (last 30 days vs the 60 before) · rules (your strategy as a rule set + the handoff to discover/author) · strategy (the long strategy read) · watch (what the agents would alert on → hire my quant).

--json prints the analysis document instead of Markdown (for your own follow-up arithmetic — never to restate numbers differently). --fresh ignores the 10-minute cache. --days N changes the window.

What the desk is (the output contract, in render order)

1. Header — short address · window · fills · coins · YOUR QUANT — LIVE · READ-ONLY · weekly rank on Hyperliquid's leaderboard (#545 of 45,105 this week · top 1.2% · #1,020 on the month · #3,300 all-time) · Senpi's labels when present (RELIABLE · AGGRESSIVE · ACTIVE) · archetype (Aggressive long-only trend rider) · verdict (one sentence: strength — weakness. imperative) · flag chips (NO STOPS (1/3), NEAR LIQUIDATION 3.7%, HIGH MARGIN 122%, IN DRAWDOWN, LIQUIDATED ×1, CHASING, PAYING FUNDING).
2. Quant score /100 and the six dimensions with one line each: timing/edge, risk management, cost efficiency, sizing/conviction, consistency, market fit. Weights and formulas: references/methodology.md. 2b. What you've been doing — the strategy read: receipts (what share of trades are which class and side; whether longs and shorts were held at once and whether those legs actually diverge; how much of the P&L is just BTC; how concentrated the outcome is; sides that never paid; buys strength or weakness; TWAP use), a where-the-trades-went table by class × side, and the critique. 2c. The market you're trading in — right now — today's label (risk-on / risk-off / mixed) from the whole venue by class, Senpi's funding regime, where the top traders' gains sit (Hyperfeed) and whether you are with or against them, momentum events, and how you trade the tape: your own record by the regime of the day you entered, with today's label against your best tape.
3. Track record — net P&L per Hyperliquid's own ledger, return on average equity, realized on observed trades, win rate, max drawdown (transfer-adjusted), profit factor, trades, active days — and a coverage line when trade-level reads cover less than 90% of the wallet's executed volume.
4. Where your P&L went — gross → fees → funding → net, cost share vs the whale median.
5. Top 3 things your agents found — each: agent · ~$ / window · title · evidence · counterfactual · fix.
6. Live positions — protection audit — account value, margin used, withdrawable, net uPnL. On a standard account the value is perps + spot and the line names both (account_value_perps, account_value_spot): any balance table you show carries the perps line — never a spot-only table under that total; on a unified / portfolio-margin account spot already holds the perps margin. Per position: side, leverage, notional, uPnL, ROE, funding/day, distance to liquidation, stop cover (share of the size a resting stop covers), status (AT RISK / UNPROTECTED / PARTLY COVERED / PROTECTED / UNKNOWN — the orders could not be read: say so, never "nothing needs protecting") and what your quant would do. A dex whose positions could not be read is named instead of counted: never "no open positions" or "nothing here" over it.
7. Performance — per-coin table, long/short split, hold time winners vs losers, execution (taker share, fee rates, liquidations), size-vs-outcome bands.
8. Leaks — ranked by $ impact, each counterfactual; then the rules tested and rejected.
9. You vs smart money — two cohorts (Senpi: the proven cohort — top traders by all-time realized P&L with ≥ $1M — and the hot 30-day cohort; public fallback: the leaderboard's large live books). Per open position: your side, the cohort's side and headcount, the read (WITH, WITH — BUT LATE (+6h), AGAINST SMART MONEY, COHORT SPLIT, NO COHORT VIEW); book-level agreement; their book by class vs yours; what they hold that you don't; what you hold that none of them do; your entry lag vs theirs.
10. Market fit — regime headline, funding across your coins, your stance and its daily funding cost, BTC trend; per position: trend, funding, open interest, fit.
11. Where your edge actually is — best setups (coin × side, hold bucket, entries before vs after the move) with wins/n and profit factor; the catalog families it maps to. 11b. Live matches — coins where the cohorts lean, the tape agrees, funding is not punitive and the setup fits how this trader wins, ranked and explained; "already moved today — a chase" is a demerit.
12. What your quant would do next — protect first · fix the biggest leak · keep the agents on. 12b. Your quant is ready to go deeper — three to five follow-ups from the bank of ten.

Reading the sources (what to say when asked "where does this come from")

• Public, any wallet: fills and TWAP slices (userFillsByTime, userTwapSliceFillsByTime), funding payments, the wallet's own fee schedule and daily volume, the equity and P&L series, transfers, the live book and resting orders, hourly candles, Hyperliquid's leaderboard. No auth.
• Coverage: Hyperliquid keeps only the most recent TWAP slices, so a TWAP-heavy wallet's older executions are not returned. The desk measures the gap from the position jumps between consecutive fills (startPosition is the position before each fill) and prints the share of executed volume it could see; ledger figures (net P&L, equity, funding) are complete regardless.
• With a Senpi token: closed positions come from Senpi discovery (the complete stream, with leverage per trade); the proven cohort from Senpi's ALL_TIME realized-PnL ranking (≥ $1M realized, top 100) and the hot cohort from the MONTHLY PnL ranking with open positions, both with position ages; Senpi's funding regime; and the Hyperfeed attention layer (where the top traders' gains sit, momentum events). Without one, the cohort is the largest profitable accounts on the public leaderboard, the read carries no entry timing, and the attention layer is absent.
• Market makers are screened out of both cohorts (and the public fallback) against Senpi's market-maker blacklist. Both cohorts rank by realized P&L — the leaderboard a quoting engine tops without taking a single directional view — so an unscreened cohort moves the line every book is measured against. Unlike the subject gate this one fails open: an unreadable blacklist leaves the cohort intact and adds a warning, because a cohort of 100 that could not be screened still tells the reader more than no cohort. meta.market_maker_check.cohorts_screened says which happened, and cohorts_filtered names what was dropped.
• Whale median: references/benchmark.json, computed by scripts/benchmark.py — from Senpi discovery with a token (whales are TWAP-heavy, so the public endpoints cannot rebuild their round trips). The table renders only when the benchmark holds ≥ 5 members with ≥ 10 trades; until that file ships, the smart-money tab shows the per-position cohort reads alone.

Mandatory closing (verbatim structure, after any full desk or --section edge/next)

1. Protect first — name the AT RISK / UNPROTECTED positions, and check first on any UNKNOWN one (its orders could not be read: send the reader to Hyperliquid) or on any dex whose positions could not be read, and offer to help. Per rule 5, senpi cannot place a stop on a book the reader custodies: they set it on Hyperliquid themselves.
2. Fix the biggest leak — the top leak's title and its counterfactual $; the one-line fix.
3. Keep the agents on — "say hire my quant and senpi runs this desk on your book — risk guard, smart money, market regime, leak finder — and can code your best setup into a strategy you approve, deployed as your strategy" → senpi-strategy-discover (the template that matches the edge) or senpi-strategy-author (from scratch), then senpi-strategy-ops.

Handing off on hire my quant

The reader has just been shown their edge AND their leaks. Open on both, and put the two routes up front so a template never reads as the only option:

Good — let's code a strategy that maps to your trading style, while improving some of your leaks. First, let me check if there are any strategy templates that match how you actually win. We can fork a template to build quickly, or code something from scratch.

Your edge is <the edge, in the desk's own words, with the numbers>. Let me see what fits.

Then hand to senpi-strategy-discover with the edge as --theme. Two things carry across and are the reason this handoff is worth more than opening discover cold: the edge (what to search for) and the leaks (what the strategy has to fix — the exits, the sizing rule, the maker-first entry). Name the leak the template closes; a reader who was just told they hold losers 29.7x longer than winners should hear which candidate takes that decision away from them.

Resilience

The engine fails open: every optional layer (rank, cohort, candles, Senpi) degrades to a line under Notes:; the trade-level analysis needs only the public fills. An address with no perp activity in the window and no open positions returns an error document — say "nothing to read here yet" and offer the new-trader path in plain words ("I can help you find a strategy that fits how you want to trade"), routing to senpi-strategy-discover yourself; never name the skill to the reader. A malformed address returns exit 2 with the reason. Public-API rate limits (HTTP 429) are retried with backoff; a second run inside 10 minutes is served from the cache (--fresh to refetch).

No address given — WHOSE book, before which book

A bare "run quant desk" with no 0x… is ambiguous in one way that matters: they may mean their own book or someone else's. Settle that first — it is one question and it decides everything after.

Your own book, or do you want me to find you someone to read?

If they mean their OWN book — desk.py --my-wallets FIRST: one list, largest first

Run desk.py --my-wallets before anything else. It returns ONE list, wallets, of the reader's saved wallets (user_get_me) and their senpi strategy wallets (strategy_list, closed ones included), ordered by value — largest first. A trader who came from Hyperliquid and added a wallet in Your wallets still has it the moment they have senpi strategies — do not forget a saved wallet, and do not silently swap it for a senpi wallet.

Every wallet is first-class: keep that order and never split it by origin — no "saved wallets" section and "strategy wallets" section, no saved-first or senpi-first. Origin is the kind column and only says what Senpi can do with the wallet. Relay text (the list, rendered) and ask ask:

A Senpi strategy is one row with all its wallets. One package is one strategy: its instances are its wallets, grouped by the package they were deployed under (skill_name); a wallet with no package stamp is its own row. A strategy row lists its instances in wallets[] (wallet_count of them; a row of several wallets has address: null) and is called by its package; it is closed only when every wallet is (closed_wallets counts the closed ones — "some wallets closed" when it is mixed). A saved wallet is always one row of its own. Every row carries run: the exact desk arguments for it — --book 0x… 0x… over a strategy's wallets, the plain address for a one-wallet strategy or a saved wallet. When they pick a row, run desk.py <run>; never pick one instance of a strategy for them.

1. Value. A saved wallet's is the MCP's state.totalValueUsd; a strategy's is its wallets' Hyperliquid account values, summed; the Senpi main wallet's is its idle cash (senpi-portfolio's read), and its run is its address like any other row. value_status: "couldnt_load" → say its value couldn't load — never as $0 — it is listed last. value_status: "partial" → the value covers only the wallets that loaded: say "(1 of 2 wallets)" from wallets_loaded / wallet_count. A closed strategy stays labeled closed at its value (usually $0).

2. Kind. A saved wallet is read-only: offer it by its label, it is the one they added. Quote its access line when they ask what senpi can do with it:

   Read-only. Senpi can analyze this wallet. It cannot place, change or cancel orders on it. A strategy wallet is where their senpi perp history actually is.

3. More than one? Then ask, because only they know which they mean today — ask names them largest first: "Which one do you want me to run the desk on — MetaMask ($12,400), Aegis ($3,000) or Phalanx ($0, closed)? Or I can read them together as one book." Offer, in plain words, to read them together as one book or side by side; that is often the more interesting read. Never show the user a flag or a command (--book, --compare, desk.py …) — those are what you run. external_wallets_status: "unavailable" → say "I couldn't load your saved wallets" and offer the wallets that did load; never "you have no saved wallets". wallets_complete: false means a source is missing from the list, not empty.

4. Neither? Ask for an address — and offer to show them the desk on a real book in the same breath. Never guess an address, but never leave a new reader with only a question either.

   On 2026-09-23 a brand-new user's FIRST EVER prompt was the quant-desk chip. Their agent did everything right — read this file, checked for saved wallets (none), checked strategy_list (empty, they had no strategies yet) — and asked for an address. One turn, eight seconds, and they never came back. A question is the one answer that shows them nothing.

   I don't have a wallet for you yet — paste any Hyperliquid address and I'll read it. Or I can run it on one of this week's top traders right now so you can see what it gives you.

   desk.py --find <band> returns real candidates by account size. Running one on a stranger is analyst mode (--other), which is the correct voice for it.

A senpi user's perp history lives in their strategy wallets, not their embedded wallet. The embedded wallet is a FUNDING wallet: deposits land there and move out to the strategy subwallets that actually trade. Running the desk on it returns "no PERP activity" and the reader is told they have nothing to read, on a book that may be trading every day.

This is not hypothetical. On launch night two of four users asked for their own book, their agents offered the embedded wallets — reasoning correctly that "the skill says never guess an address" — and both got a dead end. A third user pointed the desk at three strategy wallets the same night and got three full desks, a priced leak, and a DSL fix off the back of it. Same product, same hour; the only difference was which wallet.

So, where they have no saved wallet: resolve their wallets with strategy_list and offer the strategy wallets first, named by their strategy. Offer the embedded wallet second and label it — "your funding wallet, usually no trades of its own".

Never promise delivery you cannot perform. A desk that backgrounds ("Command still running") does not come back to you on its own — nothing wakes an agent when a detached process finishes. So "I'll relay the rest as they come in" is a promise the loop cannot keep, and the reader is left holding a partial answer forever. Measured: 70 of 273 desk invocations in 36 hours backgrounded. Either poll it to completion inside the turn, or tell the reader plainly what you have and what you did not run — and let them ask for the rest.

Several wallets at once: desk.py --book 0x… 0x… 0x… or desk.py --compare 0x… 0x… 0x…, in ONE call. Not one invocation per wallet. A desk takes 30-60s (up to ~2 min on a wide book), so a separate call per wallet backgrounds

A senpi user's book is ALL their strategy wallets — --book, not one wallet

desk.py --book 0x… 0x… 0x… reads every wallet and unions them into ONE desk: one score, one P&L, one set of leaks, plus a per-wallet table showing which strategy carried it. That is the default for "how am I trading" / "score my trading" / "find my leaks" from a senpi user.

Running the desk on a single strategy wallet answers a question they did not ask. A user with four strategies has four books, and the leak that is costing them money is usually visible only in the total — the same coin traded from two strategies, funding paid on one side while the other is long, fees that are trivial per wallet and material across the book.

Include CLOSED and PAUSED strategies, not just ACTIVE. The window is 90 days and a strategy they shut down six weeks ago still traded inside it. strategy_list returns ACTIVE, PAUSED and CLOSED rows; take the wallet from every one of them, and drop only those with no address at all. A desk that silently omits the strategy they closed is a desk that hides the losses they closed it for — which is very often the most useful thing on the page.

Copy code
desk.py --book 0xaaa… 0xbbb… 0xccc…      # their whole book, closed strategies included

Two things the union does NOT do, on purpose: there is no leaderboard rank for a book (rank is a per-address fact), and a coin traded from two wallets stays two rows in the live book, because they are two positions with two entries and two stops.

Several wallets at once: desk.py --book 0x… 0x… 0x… or desk.py --compare 0x… 0x… 0x…, in ONE call. Not one invocation per wallet. A desk takes 20-60s, so a separate call per wallet backgrounds each and the agent must poll for every result — a teammate's agent launched six that way and collected one; the other five desks were computed and thrown away. Both flags take every address in a single invocation and reuse any cached run, so it is faster as well as safer.

Several wallets at once: desk.py --compare 0x… 0x… 0x…, in ONE call. Not one invocation per wallet. A desk takes 20-60s, so a separate call per wallet backgrounds each and the agent must poll for every result — a teammate's agent launched six that way and collected one; the other five desks were computed and thrown away. --compare does them in a single invocation and reuses any cached run, so it is faster as well as safer. Run a single desk only when they ask about one wallet. Which of the two. --book = "how am I trading" — one desk over everything, the default for a senpi user. --compare = "which of my strategies is working" — a separate desk per wallet, side by side. Run a single desk only when they ask about one wallet by name.

You trade through three strategies — Aegis, Phalanx, Signals Hunter. Want the desk on one of them, or all three side by side? (Your embedded wallet is the funding one — it usually has no trades of its own.)

"Never guess an address" still holds. This is about which wallets to OFFER once you have resolved them, never about inventing one or answering from memory.

Not every address is a trader — exit 4, not_a_trader

Three shapes reach this exit, and an agent should treat them the same: relay say_to_the_reader, then offer the reader their own wallet.

"not_a_trader": "book_wider_than_the_desk_reads" — the book touches more coins than the desk reads tape for (coins vs tape_cap, with readable_share saying how much of it a score would have covered). Past that point every figure describes a sample while reading like a verdict on the whole book, so the desk says so instead. The refusal makes no claim about who the reader is — it is a statement about this tool's reach, and it is equally true of a systematic trader on 200 names and of a quoting engine. Relay it as the limit it is, and take up the offer in say_to_the_reader: ask which names they care about and read those properly.

"not_a_trader": "market_maker_blacklisted" — the wallet is on Senpi's market-maker blacklist, the list the market-maker detector writes (GetDiscoveryBlacklist on hyperliquid-traders.prod.senpi.ai). The payload carries reason and flagged_at. Unlike the breadth line this one IS a claim about who the wallet is — but it is a lookup against Senpi's own detector, not an inference from fills, which is why it is allowed to make the claim. The desk checks it with the breadth gate, before the expensive half.

This exists because fee rate does not work as the line: re-sampled over the leaderboard top 30, a 0.5 bp cut would have refused 13 of 25 wallets, VIP traders at 0.42 and 0.50 bp among them. Users were running the desk on market makers and getting a confident read of a book that has no thesis, no stop ladder and no meaningful score.

Being in the table is not the same as being a market maker. The table also holds probe, test and unauth-write-poc-benign rows, written while the internal write endpoint was open, and only a MARKET MAKER reason (either spelling — two real rows use MARKET_MAKER) causes the refusal. A wallet present for any other reason is not refused and not dropped from any cohort: calling a real trader a quoting engine is the failure that killed the fee-rate gate, and it is worse than missing one. meta.market_maker_check.other_row records such a row when there is one. If its reason is one nobody recognises, a warning says so — that is how a new market-maker spelling surfaces instead of being silently skipped — and it still is not a refusal.

When the blacklist cannot be read, the desk still runs — one endpoint being down should not take the whole desk with it — but it does not claim the wallet is clean. meta.market_maker_check reports {"checked": false, ...} and a warning says the desk does not confirm the subject is a directional trader. Relay that warning: unknown and verified clean are different answers, so silence is not reassurance. A missing SENPI_AUTH_TOKEN produces exactly this state.

The desk stops as soon as it has read the fills, before the tape and the cohorts — which is most of the run — so this costs ~30s rather than a two-minute timeout.

Earlier versions refused on maker share and then on effective fee rate. Both were measured and both were wrong: Hyperliquid publishes a VIP schedule that floors the maker fee at 0.0 above $500M of 14-day volume, so a patient limit trader at scale pays under 0.5 bp on terms anyone can get. The fee rule would have refused 13 of 25 sampled wallets and told VIP traders they had a venue agreement they do not have. Do not reintroduce either — and if a reader's rate is low, that is not evidence of anything on its own.

A book that EARNS on its fills (a negative effective rate) is not refused; it is disclosed as a warning, because the published schedule floors the maker fee at zero and cannot produce one. Read the edge figures on such a book as a quoting book's, and say so.

Some addresses on Hyperliquid are vaults: pooled books run by a leader, including Hyperliquid's own market makers. The desk checks before it reads (one call) and refuses, with the vault's real name and a line you can say.

A reader who asks for "my Hyperliquid score" and gets pointed at the all-time P&L leaderboard lands on exactly these — HLP, HLP Liquidator, HLP Strategy B sit at the top of it. That happened on 2026-09-23: a full 90-day sweep on HLP Strategy B, a component market-making strategy inside Hyperliquid's HLP vault. 174 coins, 177 open positions, 76% resting, zero fees paid. Every line the desk would have produced was wrong for it — there is no entry thesis to time on a quoting engine, no stop to place, and the P&L belongs to depositors.

Relay the say_to_the_reader line and offer their own wallet. Do not re-run it with --force unless the reader explicitly asks to read a vault as if it were a trader.

If they want someone ELSE to read — find them some

"Run AI quant on any Hyperliquid wallet", "find traders for me to analyze with AI quant". Do not guess an address and do not answer from memory. Ask the one question that narrows it, then hand them a short list.

The question is size first — a $9k book and a $9M book teach different lessons — then style:

Happy to. Two things and I'll pull a list: how big a book do you want to read, and what kind of trader?

Size: $5k–10k · $10k–25k · $25k–100k · $100k–1M · whales ($1M+) Kind: this week's winners · the ones who've held up over a month · or this week's worst — a losing book is often the more instructive read, and the desk prices it the same way.

Or paste any address and I'll just run it.

Then:

• desk.py --find 25k-100k --find-window week — best in the band right now
• desk.py --find whales --find-window month — the ones who held up over a month
• desk.py --find 10k-25k --find-losers — this week's worst, often the instructive read
• desk.py --find 100k-1m --find-window allTime — durable mid-size books
• desk.py --find 5k-10k — retail-sized, this week

Bands: 5k-10k · 10k-25k · 25k-100k · 100k-1m · whales.

It returns JSON candidates — address, account value, P&L, ROI, volume, turnover. Relay them as a short numbered list with the numbers, so the reader picks on evidence rather than on your summary, then run the desk on whichever they choose. Three to five is a list; eight is a wall.

Never present a candidate as a recommendation to copy or follow — it is a book to READ. Vetting a trader to mirror is senpi-trader-research.

Running a batch — sequentially

One desk makes 100-200 reads against a per-IP weight bucket. Run wallets one at a time. Seven in parallel loses one or two runs to HTTP 429 on an essential read however long the backoff is: the budget is now a full refill window with jitter, and it still only gets 6 of 7 through. A lost run fails loudly with the read that died, so nothing silently ships on partial data — but it is a rerun you did not need.

Install — the whole scripts/ directory is required

desk.py imports addresses.py, blacklist.py, book.py, deep.py, dsl.py, followups.py, hl_api.py, market.py, metrics.py, opportunities.py, render.py, roundtrips.py, score.py, senpi_history.py, smart_money.py, strategy_read.py, taxonomy.py, timing.py and voice.py, plus the vendored mcp_client.py (used only when SENPI_AUTH_TOKEN is set). Copy the whole directory — a partial copy fails at import, not at runtime. Stdlib only, Python ≥ 3.9.

blacklist.py and mcp_client.py both need SENPI_AUTH_TOKEN. Without it the desk still runs, but it cannot check Senpi's market-maker blacklist and says so in its warnings rather than implying the subject was verified. Fixture-driven tests in tests/.

Skill attribution

This skill creates no strategy wallet and carries no attribution; strategies it hands off are attributed by the skill that deploys them (senpi-strategy-ops).