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.
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 THE HARD KILL — this is the single most common way a reader gets no
desk. The hard kill (seconds) is timeoutSeconds when the exec tool lists it, otherwise timeout.
Send only the one the exec tool lists. 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 hard kill 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 a hard kill of 200 seconds 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.
The hard kill still matters — give a desk on a wide book at least 180s of it, because at 120s
the exec tool SIGTERMs it mid-run and you get nothing after paying the whole cost. The two are not
alternatives: the hard kill 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.
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.
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.
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.
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.
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.
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.
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.
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.
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
| User says |
Run |
Then |
| "run AI quant on any Hyperliquid wallet", "find traders for me to analyze", no address given |
desk.py --find <band> |
ask size + kind first, then relay the candidates with their numbers |
| "run AI quant / run quant / run quant desk on 0x…", "score my trading", "find leaks", "what did I miss", "master my week", "how am I doing" |
desk.py 0x… |
relay the full desk — the user is 0x… |
| "are my positions protected", "am I at risk" |
desk.py 0x… --section protection |
relay; the AT RISK rows first |
| "where am I leaking money", "what's costing me" |
desk.py 0x… --section leaks |
relay, biggest first, with the rejected rules |
| "am I with or against smart money", "compare me to whales" |
desk.py 0x… --section smart |
relay both tables |
| "does my book fit this market" |
desk.py 0x… --section market |
relay |
| "where's my edge", "what am I good at" |
desk.py 0x… --section edge |
relay; then the closing (rule 9) |
| "what should I fix first" |
desk.py 0x… --section next |
relay the three steps |
| "run AI quant analyst on 0x…", "review this trader 0x…", a leaderboard pick |
desk.py 0x… --other (alias --analyst) |
relay in the third person; copying → senpi-trader-research |
| "how do they stack up", after two or more desks |
desk.py --compare 0x… 0x… [0x…] |
relay the side-by-side and its "what separates them" |
| a senpi user: "score my trading", "find my leaks", "how am I doing" |
desk.py --book 0x… 0x… 0x… over EVERY strategy wallet, closed ones included |
one desk over the whole book; relay the by-wallet table with it |
| "write their playbook as rules" (another trader) |
desk.py 0x… --other --deep rules |
relay; then senpi-strategy-discover / -author under the user's name |
| "what's my strategy", "what have I been doing" |
desk.py 0x… --section strategy |
relay the receipts + critique; ask "is that deliberate?" |
| "what's the market doing", "does my book fit today" |
desk.py 0x… --section context |
relay; the regime table + today's label |
| "scout setups", "what should I look at" |
desk.py 0x… --section scout |
relay; process only |
| a follow-up the desk offered |
desk.py 0x… --deep <mode> |
relay; then offer the next follow-ups |
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)
- 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).
- 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.
- 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.
- Where your P&L went — gross → fees → funding → net, cost share vs the whale median.
- Top 3 things your agents found — each: agent · ~$ / window · title · evidence · counterfactual · fix.
- 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.
- Performance — per-coin table, long/short split, hold time winners vs losers, execution
(taker share, fee rates, liquidations), size-vs-outcome bands.
- Leaks — ranked by $ impact, each counterfactual; then the rules tested and rejected.
- 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.
- Market fit — regime headline, funding across your coins, your stance and its daily funding
cost, BTC trend; per position: trend, funding, open interest, fit.
- 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.
- 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
- 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)
- 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.
- Fix the biggest leak — the top leak's title and its counterfactual $; the one-line fix.
- 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-author (it forks the closest template to the edge when
one fits), then senpi-strategy-ops.
Handing off on hire my quant
The reader has just been shown their edge AND their leaks. Open on both:
Good — let's code a strategy that maps to your trading style, while improving some of your leaks.
Your edge is <the edge, in the desk's own words, with the numbers>. Let's code your best setup into
your strategy.
Then hand to senpi-strategy-author with the edge as the thesis. Two things carry across: the edge
(what to build) and the leaks (what the strategy has to fix — the exits, the sizing rule, the
maker-first entry). Name the leak each rule closes; a reader who was just told they hold losers 29.7x
longer than winners should hear which rule takes that decision away from them. Author forks the closest
template to the edge when one fits; name none before that.
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 build a strategy around how you want to trade"),
routing to senpi-strategy-author 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.
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).
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.
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.
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
…(truncated)
1---2name: quant-desk3description: >-4 **Quant Desk** — the desk your **AI Quant** produces. Users reach it by either name, spaced or5 hyphenated: "run AI quant", "run ai-quant", "run quant", "run quant desk", "run quant-desk". Paste ANY Hyperliquid address (0x…) and get the desk —6 what the trader has actually been doing (a strategy read with a critique), a quant score with six7 explained dimensions, the market they are trading in right now and how they trade each regime, the8 live book with a protection audit, leaks priced as counterfactual dollars, their book against the9 PROVEN cohort (top traders by all-time realized P&L, ≥ $1M) and the HOT 30-day cohort — side,10 headcount, when they moved, what they hold that the trader doesn't — live matches where the tape,11 the cohorts and the trader's own pattern agree, and a bank of twelve follow-ups the quant is12 prepared to go deeper on. Works for wallets that never touched senpi (public onchain data,13 read-only); with a Senpi token the closed-trade history, both cohorts, the funding regime and the14 Hyperfeed attention layer come from Senpi's own data.15 TRIGGERS — any of these, with or without an address: "run AI quant", "run ai-quant", "run quant-desk", "run AI quant on my Hyperliquid wallet", "run AI quant on any Hyperliquid wallet", "run quant", "run the quant on 0x…", "run quant16 desk on 0x…", "score my trading", "rate my trading", "find leaks on my Hyperliquid wallet", "where am17 I leaking money", "what did I miss" (about a book, a week or a trade), "master my week" (on a18 SAVED wallet the user trades by hand — "my MetaMask", "my own Hyperliquid wallet", "my saved19 wallet" — leaks and "what did I miss" are always this skill; its trade review and "master my week"20 belong to senpi-improve-trades, which also keeps the leaks of senpi strategies), "analyze my21 wallet / my Hyperliquid address", "how am I doing", "what's my strategy", "am I on the right side of22 smart money", "are my positions protected", "what should I fix first", "compare me to the whales",23 "scout setups for me", "find traders for me to analyze with AI quant", "run AI quant on any24 Hyperliquid wallet".25 THE IN-PRODUCT SUGGESTED PROMPTS, verbatim — these are buttons users click, so they must match26 exactly, in the product's own second person ("your wallet", not "my wallet"): "Run quant desk on27 your Hyperliquid wallet", "Run quant desk on any Hyperliquid wallet", "Score my trading", "Find28 leaks on your Hyperliquid wallet", "Find traders for me to analyze with quant desk", "Run quant29 desk", "What did I miss?". Note "find traders for me to analyze" ALONE belongs to30 senpi-trader-research (vetting a trader to COPY); with "quant desk" or "AI quant" it is this skill31 (reading a trader to LEARN from). "your Hyperliquid wallet" in a suggested prompt means the32 READER's own wallet — ask for their address and run the own-book desk, never someone else's.33 PLURAL COUNTS, and it is what a senpi user reaches for first, because they HAVE several: "find34 leaks on my wallets", "score my wallets", "run quant desk on my wallets", "leaks across my35 wallets". senpi-portfolio owns "across all wallets" for HOLDINGS; the leaks and the score on those36 same wallets are THIS skill. A teammate's agent read portfolio's SKILL.md first on "find leaks on37 my wallets" and only reached the desk forty-five seconds later. No address given? `desk.py --find <band>` offers candidates by account size38 ($5k-10k through whales) and by this week's winners, the month's, or this week's worst — ask which,39 never guess an address and never answer from memory.40 The default is the user's OWN book: "run AI quant on 0x…" means the user is 0x… — the desk speaks to41 them and recommends their next steps. "Run AI quant analyst on 0x…" (or "review this trader 0x…")42 means the user is analyzing someone else.43 Hidden engine: scripts/desk.py. NOT for choosing or deploying a strategy (senpi-strategy-discover /44 -ops), reviewing a Senpi strategy's own trades (senpi-improve-trades), or vetting a trader to copy45 (senpi-trader-research). senpi-portfolio resolves the reader's wallets and answers what they46 hold; a request to SCORE or RATE that trading comes here, on those wallets.47license: Apache-2.048metadata:49 source: https://skillmd.com/skills/jiayaoqijia/quant-desk50 author: Senpi51 version: "1.50.0"52 platform: senpi53 exchange: hyperliquid54---5556# Senpi Quant Desk — the desk your AI Quant produces, for any Hyperliquid address5758Built for one thing first: **the user's own book.** A trader joins senpi, pastes their Hyperliquid59wallet, and gets their desk — the book they actually run, scored, protected and improved, in the second60person, with the next steps recommended to them. That is the default and the experience to design for:61**"run AI quant on 0x…" means the user is 0x…**, whoever the address belongs to. The same engine62also reads **someone else's book** — **"run AI quant analyst on 0x…"**, "review this trader", a63leaderboard pick, a whale — to learn from it (third person, learn-from-them follow-ups, a side-by-side64compare). Run it plain (`--mine`) unless the user asks for the analyst read; then `--other` (alias65`--analyst`), and never let that desk say "you".6667**HARD RULES — obey these even if you skim the rest.**68690. **ONE desk at a time, and NEVER re-run one that is still going.** A desk is a long command, so70 `exec` hands you back `{"status":"running", "sessionId": …}` and the real result arrives on a71 later `process` poll. That handoff is not a failure. **Poll the session you already have.**7273 Launching a second run does not make the first one finish. Both hammer the same per-IP rate74 limit, so two runs are slower than one and three usually kill each other. On 2026-09-23 an agent75 that could not see a result launched **five concurrent runs of the same wallet**, inventing76 `sleep 15 &&` and `sleep 30 &&` workarounds; three died with `HTTP 429` and the reader waited six77 minutes for nothing. The desk now refuses a second run (**exit 5**, `already_running`) — treat78 that as "your first one is still working", not as an error to route around.7980 If a run really is dead, the state dir is the shared surface: the finished desk lands in the same81 file, so a fresh `--section overview` after it completes is instant.8283 **PASS `yieldMs`, NOT JUST THE HARD KILL — this is the single most common way a reader gets no84 desk.** The hard kill (seconds) is `timeoutSeconds` when the exec tool lists it, otherwise `timeout`.85 Send only the one the exec tool lists. **`yieldMs` is how long exec WAITS before handing you a86 background handle instead of the output.** Without it the tool backgrounds the run at87 **~10 seconds** no matter how large a hard kill you passed, and you get88 `Command still running (session …, pid …)` with a progress tail — not a desk. Measured over 24h:8990 | `yieldMs` passed | what came back | runs | got the score |91 |---|---|---|---|92 | none | backgrounded at 10.1s | 19 | **0** |93 | 5000 / 15000 | backgrounded at exactly that | 2 | **0** |94 | 120000 / 190000 | the finished desk, in 39-49s | 3 | **3** |9596 So: **`yieldMs: 190000` together with a hard kill of 200 seconds** on the first desk call. The desk97 itself is not slow — it lands in well under a minute on a normal book. Twenty-one runs across98 seventeen users produced nothing last window purely because the wait was ten seconds.99100 The hard kill still matters — give a desk on a wide book **at least 180s** of it, because at 120s101 the exec tool SIGTERMs it mid-run and you get nothing after paying the whole cost. The two are not102 alternatives: the hard kill decides when the run is killed, `yieldMs` decides when you stop103 waiting for it. You need both.104105 If you DO get a background handle, you are not finished: poll it (`process poll <sessionId>`)106 until it exits and relay what it printed. Abandoning the handle is how a user ends up with a107 half-sentence of progress log and no score.1081091. **Relay it in STAGES — never as one block.** The analysis takes 30-60s on a typical book, up to110 ~2 MINUTES on a very wide one (100+ coins), and the111 whole desk is thousands of words. Delivering it as a single wall after a silent wait is the worst112 possible shape: the reader waits with nothing, then gets more than they can read. The first run113 caches for 10 minutes, so every section after it returns instantly.114115 **Stage 1 — the hook.** `desk.py <0xaddress> --section overview` does the full analysis (this is116 the slow call) and prints only the score, the rank and the verdict. Relay it the moment it lands.117 That is the number they came for. **If it opens with "I'm reading 0x… as your book now, because118 you added it to Your wallets", that sentence is the first thing you say, word for word, before the119 score** — the desk prints it on this one run only, so a summary that skips it means the reader is120 never told (measured on dev: the stage-1 summary dropped it).121 **Stage 2 — what is urgent.** `--section protection`. Instant, from cache. Relay.122 **Stage 3 — the money.** `--section leaks`. Instant. Relay.123 **Stage 4 — the rest**, in one call: `--section strategy --section context --section performance124 --section smart --section market --section edge --section scout --section next --section followups`.125126 **All four stages belong to ONE turn.** The reader is reading stage 1 while stage 2 renders, so127 the wait disappears without anything being rushed — but that only works if you keep going. Ending128 your turn after a stage hands the desk back to a reader who has no idea they are now the thing129 blocking it. Measured live: an agent delivered stage 1, wrote "Next I'll pull the protection130 audit", and stopped. The reader waited, then had to ask "Did you pull it?" and was told "Not yet."131 It then finished stage 2 with "Want me to continue?" and stopped again. A desk delivered that way132 is worse than a silent wait, because a wait ends by itself and this does not.133134 So: **never end a turn mid-desk, and never announce a stage you are not about to run.** If you135 name the next stage, the call for it goes in the same turn. Stop only when the desk is finished,136 or when something actually failed — and if it failed, say so plainly rather than promising.137138 Do not batch stages 1-3 into a single call to save calls — the staging IS the feature, and each139 stage lands as its own message. `--json` or a plain `desk.py <0xaddress>` still returns everything140 at once when you need the whole document in one piece.141142 **The stage numbers are internal.** They order YOUR work; they are not headings. A reader who143 sees "Stage 2 — protection audit" has been shown the scaffolding and will reasonably wait for a144 Stage 3 that they now have to ask for. Relay each stage under its own real heading — the score,145 what is at risk, what it costs — and never the word "stage".146147 Relay it; do not recompute, reorder or "improve" its numbers.148149 **The lead-in, before stage 1, is one short line** (short address in place): *"Running the desk on150 `0x5b5d…c060` — reading every fill, the live book, the cohorts and the tape."* Nothing longer. The151 engine then streams a numbered progress line per stage while it works — relay those as they arrive;152 they are what fills the wait, and each one carries a number it has just learned.153154 **Print the desk's header line exactly as the engine emits it — the `N days · N fills · N coins ·155 updated <UTC> · YOUR QUANT — LIVE · READ-ONLY · vX.Y.Z` line — as the first line of every desk you156 relay.** It is the only staleness gate the reader has. An agent that rewrites the header into its157 own summary strips the version and the timestamp, and a desk running a months-old engine then158 looks identical to a current one. This has already happened: a run was reviewed as if it were159 current when its install predated the fix being tested.160161 Never "this pulls public data" or "this may take a moment": the desk162 is senpi's proprietary analysis. Never mention the public API, data sources or coverage in your163 own words — the desk says what it needs to. **Our machinery stays ours: in what you write, no script164 name, flag or skill id ever reaches the reader** — not `desk.py`, not `--days`, not `senpi-strategy-discover`.165 Say the thing in plain words and then do it ("I can look further back — how far?", "I can help you166 pick a strategy"), never the identifier that does it. Run the installed copy167 (`/data/.openclaw/skills/quant-desk/scripts/desk.py`), never a backup folder: the header line carries the168 version — and `desk.py --version` prints the ENGINE's, which is the one that catches a half-synced169 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. Write170 **onchain**, never "on-chain", everywhere.1712. **Never invent a number.** Every figure on the desk is computed from public onchain data (or Senpi172 discovery when a token is present). If the script says a layer was unavailable (`Notes:` line), say so173 in the same words — never fill the gap from memory.1742a. **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.1753. **Counterfactual, not history, on every leak.** "A 24h cap on funding-paying holds would have kept176 ~$2,536 over 90 days" — a process change and what it would have kept. Never "you lost $X" as a leak,177 never a leak the script rejected (it prints which rules it tested and rejected: say those too — the178 user's edge may be exactly the thing a naive fix would break).1793b. **Never add the leaks up — quote the one number the desk gives you.** The leaks are alternative180 fixes priced over the *same* trades: one oversized, chased, held-too-long position appears in181 several of them. Summing them produced $68k on a book that lost $65k. The `leaks` section opens182 with the single line to quote — "a trailing stop that arms at +3% and keeps 50% of the peak would183 have kept ~$46,314 — 71% of what your losing trades gave up" — and it is the best *single* change,184 already charged on the trades it would have cost. Relay that line first, then the leaks below it185 as the individual fixes they are. If the desk says the total is concentrated ("one trade is 62% of186 it"), say that too: the shape is the actionable part, and a user told they leak $46k "across their187 book" will fix the wrong thing.188 **This binds everywhere, not just in the leaks section** — deep dives, "how do I save on fees",189 ELI5, and any answer that totals more than one fix. Fees are the single exception: resting instead190 of crossing saves the same money whatever the exit rule, so fees may be added to one other fix.191 Two *exit or sizing* fixes may never be added to each other — they are alternatives over192 overlapping trades. "Maker-first ($1,503) plus a time-cut ($610) plus sizing ($845) is ~$12k/yr"193 is the error: the honest combined figure is the best single lever plus fees, ~$9.5k/yr.194 Annualising a 90-day counterfactual by 4x is fine; adding two of them first is not.1954. **Process only.** Recommendations are rules, risk and timing — a stop ladder, a time-cut, a196 maker-first entry, a funding-aware hold, a sizing rule. **Never a call to buy or sell a coin.**1975. **Custody language.** The desk is read-only, and **senpi cannot put a stop on a position held in198 the reader's own wallet today.** `ratchet_stop_add` is keyed to a senpi strategy wallet, so there199 is nothing to sign and nothing to attach on a book the reader custodies themselves.200201 So: **name the naked positions and ask how you can help.** Do not describe protection on their202 own positions as "a signature", "one click", or something senpi will do for them — not in the203 future tense either, however close it is. Equally, do not send them away with homework: "set it204 yourself on Hyperliquid" is the fact, not the offer. Offer the help and let them ask.205206 Senpi's protection applies to strategies senpi runs, where the runtime owns the exits. Funding a207 quant is only for autonomous trading. Never imply senpi holds or moves their funds.2082096. **Say "quant", "desk", "agents", "leak", "protect".** Never "report", "analyst", "bot", "AI assistant".210 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.2117. **Address hygiene and whose book it is.** Show the address shortened (`0x2999…65de`). Never post212 the desk of a wallet the user did not name.213214 **An address is the reader's own book unless we know otherwise.** "Run AI quant on 0x…" means the215 user is 0x… — run it plain and speak to them. That is the path the product exists for: a Hyperliquid216 trader pastes their address and gets their desk, with no question in front of it.217218 **"My wallets" are the wallets the reader added, plus their senpi wallets.** `desk.py --my-wallets`219 prints them as ONE list, `wallets`, largest first: the reader's **saved** wallets (the ones they220 added in Your wallets), their Senpi main wallet and their senpi strategy wallets side by side, each221 row with its `kind` (`saved` / `strategy` / `main`), the words to show for it in `kind_label`222 (read-only — you added it / Senpi strategy / Senpi strategy — closed / Senpi strategy — some wallets223 closed / Senpi main wallet — idle cash) and its224 `value_usd`, each source with its own `ok` / `unavailable` status. A saved wallet is theirs even225 if the book once read it as a stranger's. It is their claim: call them "your wallets" or226 "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).227228 **Read as a stranger's, then added: say it once, on any path.** When a run reads an address the book229 had as someone else's and the reader has since added it to Your wallets, the desk says so itself:230 the Markdown opens with that sentence, and `--json` carries it as `whose_changed.say`. Relay it231 once, word for word, first, before the score or anything else about the book as theirs — whether the run came from232 `--my-wallets`, a pasted address, `--mine`, or the "Find my leaks on 0x…" button. The desk prints it233 only on the first run after the add; don't repeat it on later sections.234235 **The desk remembers.** It keeps an address book per box (`scripts/desk.py --addresses`) with236 **verified** (a wallet senpi issued) and **analyzed** (someone else's book they read). An237 address already recorded as *analyzed* stays238 someone else's on a bare re-run — they looked at a whale last week, and asking about it again must239 not start handing them the whale's leaks to fix. **claimed** is retired: older books still hold it240 and it means nothing — what the reader says is theirs lives in their saved wallets, or in this run's241 `--mine`.242243 **A pasted address the reader says is theirs** is theirs for this conversation only: run it with244 `--mine` (or `--claim`, the same voice). Say plainly that it isn't saved, and that to keep it they245 can add it in Your wallets on senpi.ai (web).246247 **Use `--other` whenever the request is about someone else** — "this trader", "their wallet", a248 leaderboard pick, a whale you surfaced, anything you picked rather than they typed. The default249 covers the address a reader hands you; it is not a licence to read a wallet they never said was250 their own.251252 "Run AI quant analyst on 0x…" (or253 "this trader", "their wallet", a leaderboard pick) means the user is analyzing 0x…: run with254 `--other` (alias `--analyst`): the desk speaks in the third255 person, the closing becomes *what to take from this trader*, and the follow-ups are the learning256 ones (their playbook as rules under **your** name, the smart-money picture on their coins, whether they257 are worth copying → `senpi-trader-research`, watching the wallet). It is analysis of onchain258 data, never advice to copy a position. Two or more traders: `--compare 0x… 0x…` prints the259 side-by-side (cached runs are reused) — use it whenever the user has looked at more than one wallet260 and asks how they stack up; never improvise the comparison yourself.2617b. **An address senpi has not indexed yet.** `--json` carries `indexed`: `true` when senpi's own262 history answered, `false` when the public endpoints show closed round trips in the window and263 senpi's index returned none for the same window — a contradiction between two sources, which means264 the wallet is not in the index yet — and `null` when nothing closed either way, which is a quiet265 wallet and says nothing about indexing.266267 On `false`, say so instead of presenting the desk as complete, in the reader's own words. Read the268 figure from `references/coverage.json`, never from memory; if `as_of` is more than `stale_after_days`269 old, say "over 25,000" rather than a precise number that has moved:270271 > senpi is rolling out the AI Quant to every trader on Hyperliquid in waves. We're at272 > **26,188** wallets so far and yours isn't in that set yet. I've flagged it to the team as high273 > priority and they'll let you know as soon as it's ready.274275 The flag is real: the address is recorded in the book and `--addresses` lists it under the276 not-yet-indexed set. Never promise a date. A desk still runs on the public reads, so offer it —277 but say plainly that trade-level detail will be thinner until the wallet is indexed.2782797c. **The desk reads any book on Hyperliquid, not just theirs.** Readers do not know this, and the280 follow-ups all go *deeper on the same book*, so nothing tells them. After a run on their own book,281 offer the lateral move once:282283 > **Your quant reads any book on Hyperliquid, not just yours.** Paste an address and I'll run the284 > desk on them — what they trade, how they size, where they leak — or tell me what you're curious285 > about and I'll go find traders worth reading.286287 After an analyst run, offer the mirror of it: *"That was someone else's book. Your quant works the288 same way on yours — paste your address and I'll run it."* "Find me traders worth reading" is a real289 route, not an invitation to improvise: resolve candidates from the proven cohort, the leaderboard or290 `senpi-trader-research`, then run the pick with `--other`. **Never invent an address.**2912928. **Hold three to five things back — on purpose.** The desk ends with the follow-ups it earned (the293 script picks them from a bank of twelve). Offer them as questions, in the script's words; answer each294 with its `--deep <mode>` and then offer the next ones. The more the trader asks, the more of their own295 book they see — never dump every deep dive unasked.296297 **Two of them carry no `--deep` mode and must not be run as one.** In `--json` their `mode` is298 `null`: the answer is already on the screen, so you write it, immediately, with no second call.299 - *the ELI5* — leads for almost every reader, and is the whole point for someone who has never300 used senpi: restate the desk without the vocabulary. No profit factor, no ρ, no basis, no regime.301 A number, what it means, what to do. This is the cheapest possible next step and the one most302 likely to earn a second question.303 - *the biggest one* — the top finding named in the prompt: expand its evidence and its fix in304 your own words from what the desk already printed.305306 Protection outranks both when a position is unprotected **and** near liquidation — the desk says307 "Protect first" and the follow-ups must not disagree with it.3089. **The strategy read is theirs to argue with.** Relay the receipts (the bullets) and the critique as309 written, then invite the correction: "is that deliberate?" A trader who says "yes, that's the plan" has310 just told you what to watch; one who says "no" has just found the leak.311312## Quick actions313314| User says | Run | Then |315|---|---|---|316| "run AI quant on any Hyperliquid wallet", "find traders for me to analyze", no address given | `desk.py --find <band>` | ask size + kind first, then relay the candidates with their numbers |317| "run AI quant / run quant / run quant desk on 0x…", "score my trading", "find leaks", "what did I miss", "master my week", "how am I doing" | `desk.py 0x…` | relay the full desk — the user is 0x… |318| "are my positions protected", "am I at risk" | `desk.py 0x… --section protection` | relay; the AT RISK rows first |319| "where am I leaking money", "what's costing me" | `desk.py 0x… --section leaks` | relay, biggest first, with the rejected rules |320| "am I with or against smart money", "compare me to whales" | `desk.py 0x… --section smart` | relay both tables |321| "does my book fit this market" | `desk.py 0x… --section market` | relay |322| "where's my edge", "what am I good at" | `desk.py 0x… --section edge` | relay; then the closing (rule 9) |323| "what should I fix first" | `desk.py 0x… --section next` | relay the three steps |324| "run AI quant analyst on 0x…", "review this trader 0x…", a leaderboard pick | `desk.py 0x… --other` (alias `--analyst`) | relay in the third person; copying → `senpi-trader-research` |325| "how do they stack up", after two or more desks | `desk.py --compare 0x… 0x… [0x…]` | relay the side-by-side and its "what separates them" |326| a senpi user: "score my trading", "find my leaks", "how am I doing" | `desk.py --book 0x… 0x… 0x…` over EVERY strategy wallet, closed ones included | one desk over the whole book; relay the by-wallet table with it |327| "write their playbook as rules" (another trader) | `desk.py 0x… --other --deep rules` | relay; then `senpi-strategy-discover` / `-author` under the user's name |328| "what's my strategy", "what have I been doing" | `desk.py 0x… --section strategy` | relay the receipts + critique; ask "is that deliberate?" |329| "what's the market doing", "does my book fit today" | `desk.py 0x… --section context` | relay; the regime table + today's label |330| "scout setups", "what should I look at" | `desk.py 0x… --section scout` | relay; process only |331| **a follow-up the desk offered** | `desk.py 0x… --deep <mode>` | relay; then offer the next follow-ups |332333On a saved wallet the desk owns leaks and "what did I miss"; the trade review and "master my week"334go to `senpi-improve-trades`, which reviews saved wallets as manual trades.335336**The ten deep modes** (each answers one bank question; all read the cached run, `protect` and `replay`337refetch candles): `protect` (a stop ladder per position with levels and dollars at risk before/after) ·338`smart` (both cohorts in full, tilt by class, what they hold that you don't, when they moved) · `scout`339(live matches) · `replay` (your worst week, trade by trade, with the counterfactuals on exactly those340trades) · `funding` (the next 30 days at today's rates, position by position) · `regime` (how you trade341risk-on vs risk-off, and which one today is) · `compare` (last 30 days vs the 60 before) · `rules` (your342strategy as a rule set + the handoff to discover/author) · `strategy` (the long strategy read) · `watch`343(what the agents would alert on → *hire my quant*).344345`--json` prints the analysis document instead of Markdown (for your own follow-up arithmetic — never346to restate numbers differently). `--fresh` ignores the 10-minute cache. `--days N` changes the window.347348## What the desk is (the output contract, in render order)3493501. **Header** — short address · window · fills · coins · **YOUR QUANT — LIVE · READ-ONLY** · weekly rank351 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 ·352 AGGRESSIVE · ACTIVE`) · **archetype** (`Aggressive long-only trend rider`) · **verdict** (one sentence:353 strength — weakness. imperative) · flag chips (`NO STOPS (1/3)`, `NEAR LIQUIDATION 3.7%`, `HIGH MARGIN354 122%`, `IN DRAWDOWN`, `LIQUIDATED ×1`, `CHASING`, `PAYING FUNDING`).3552. **Quant score /100** and the **six dimensions** with one line each: timing/edge, risk management,356 cost efficiency, sizing/conviction, consistency, market fit. Weights and formulas:357 `references/methodology.md`.3582b. **What you've been doing** — the strategy read: receipts (what share of trades are which class and359 side; whether longs and shorts were held at once and whether those legs actually diverge; how much of360 the P&L is just BTC; how concentrated the outcome is; sides that never paid; buys strength or weakness;361 TWAP use), a where-the-trades-went table by class × side, and the critique.3622c. **The market you're trading in — right now** — today's label (risk-on / risk-off / mixed) from the363 whole venue by class, Senpi's funding regime, where the top traders' gains sit (Hyperfeed) and whether364 you are with or against them, momentum events, and **how you trade the tape**: your own record by the365 regime of the day you entered, with today's label against your best tape.3663. **Track record** — net P&L per Hyperliquid's own ledger, return on average equity, realized on367 observed trades, win rate, max drawdown (transfer-adjusted), profit factor, trades, active days —368 and a coverage line when trade-level reads cover less than 90% of the wallet's executed volume.3694. **Where your P&L went** — gross → fees → funding → net, cost share vs the whale median.3705. **Top 3 things your agents found** — each: agent · ~$ / window · title · evidence · counterfactual · fix.3716. **Live positions — protection audit** — account value, margin used, withdrawable, net uPnL. On a372 standard account the value is perps + spot and the line names both (`account_value_perps`,373 `account_value_spot`): any balance table you show carries the perps line — never a spot-only table374 under that total; on a unified / portfolio-margin account spot already holds the perps margin. Per375 position: side, leverage, notional, uPnL, ROE, funding/day, distance to liquidation, **stop cover**376 (share of the size a resting stop covers), status (`AT RISK` / `UNPROTECTED` / `PARTLY COVERED` /377 `PROTECTED` / `UNKNOWN` — the orders could not be read: say so, never "nothing needs protecting")378 and what your quant would do. A dex whose **positions** could not be read is named instead of379 counted: never "no open positions" or "nothing here" over it.3807. **Performance** — per-coin table, long/short split, hold time winners vs losers, execution381 (taker share, fee rates, liquidations), size-vs-outcome bands.3828. **Leaks** — ranked by $ impact, each counterfactual; then the rules **tested and rejected**.3839. **You vs smart money** — two cohorts (Senpi: the proven cohort — top traders by all-time realized384 P&L with ≥ $1M — and the hot 30-day cohort; public fallback: the leaderboard's large live books). Per385 open position: your side, the cohort's side and headcount, the read (`WITH`, `WITH — BUT LATE (+6h)`,386 `AGAINST SMART MONEY`, `COHORT SPLIT`, `NO COHORT VIEW`); book-level agreement; their book by class vs387 yours; what they hold that you don't; what you hold that none of them do; your entry lag vs theirs.38810. **Market fit** — regime headline, funding across your coins, your stance and its daily funding389 cost, BTC trend; per position: trend, funding, open interest, fit.39011. **Where your edge actually is** — best setups (coin × side, hold bucket, entries before vs after the391 move) with wins/n and profit factor; the catalog families it maps to.39211b. **Live matches** — coins where the cohorts lean, the tape agrees, funding is not punitive and the393 setup fits how this trader wins, ranked and explained; "already moved today — a chase" is a demerit.39412. **What your quant would do next** — protect first · fix the biggest leak · keep the agents on.39512b. **Your quant is ready to go deeper** — three to five follow-ups from the bank of ten.396397## Reading the sources (what to say when asked "where does this come from")398399- **Public, any wallet:** fills and TWAP slices (`userFillsByTime`, `userTwapSliceFillsByTime`), funding400 payments, the wallet's own fee schedule and daily volume, the equity and P&L series, transfers, the401 live book and resting orders, hourly candles, Hyperliquid's leaderboard. No auth.402- **Coverage:** Hyperliquid keeps only the most recent TWAP slices, so a TWAP-heavy wallet's older403 executions are not returned. The desk measures the gap from the position jumps between consecutive404 fills (`startPosition` is the position before each fill) and prints the share of executed volume it405 could see; ledger figures (net P&L, equity, funding) are complete regardless.406- **With a Senpi token:** closed positions come from Senpi discovery (the complete stream, with407 leverage per trade); the proven cohort from Senpi's ALL_TIME realized-PnL ranking (≥ $1M realized, top408 100) and the hot cohort from the MONTHLY PnL ranking with open positions, both with position ages;409 Senpi's funding regime; and the Hyperfeed attention layer (where the top traders' gains sit, momentum410 events). Without one, the cohort is the largest profitable accounts on the public leaderboard, the read411 carries no entry timing, and the attention layer is absent.412- **Market makers are screened out of both cohorts** (and the public fallback) against Senpi's413 market-maker blacklist. Both cohorts rank by realized P&L — the leaderboard a quoting engine tops414 without taking a single directional view — so an unscreened cohort moves the line every book is415 measured against. Unlike the subject gate this one **fails open**: an unreadable blacklist leaves416 the cohort intact and adds a warning, because a cohort of 100 that could not be screened still417 tells the reader more than no cohort. `meta.market_maker_check.cohorts_screened` says which418 happened, and `cohorts_filtered` names what was dropped.419- **Whale median:** `references/benchmark.json`, computed by `scripts/benchmark.py` — from Senpi discovery420 with a token (whales are TWAP-heavy, so the public endpoints cannot rebuild their round trips). The table421 renders only when the benchmark holds ≥ 5 members with ≥ 10 trades; until that file ships, the smart-money422 tab shows the per-position cohort reads alone.423424## Mandatory closing (verbatim structure, after any full desk or `--section edge/next`)4254261. **Protect first** — name the AT RISK / UNPROTECTED positions, and **check first** on any UNKNOWN one427 (its orders could not be read: send the reader to Hyperliquid) or on any dex whose positions could428 not be read, and offer to help. Per rule 5, senpi429 cannot place a stop on a book the reader custodies: they set it on Hyperliquid themselves.4302. **Fix the biggest leak** — the top leak's title and its counterfactual $; the one-line fix.4313. **Keep the agents on** — "say *hire my quant* and senpi runs this desk on your book — risk guard,432 smart money, market regime, leak finder — and can code your best setup into a strategy you approve,433 deployed as **your** strategy" → `senpi-strategy-author` (it forks the closest template to the edge when434 one fits), then `senpi-strategy-ops`.435436### Handing off on *hire my quant*437438The reader has just been shown their edge AND their leaks. Open on both:439440> Good — let's code a strategy that maps to your trading style, while improving some of your leaks.441> Your edge is <the edge, in the desk's own words, with the numbers>. Let's code your best setup into442> your strategy.443444Then hand to `senpi-strategy-author` with the edge as the thesis. Two things carry across: **the edge**445(what to build) and **the leaks** (what the strategy has to fix — the exits, the sizing rule, the446maker-first entry). Name the leak each rule closes; a reader who was just told they hold losers 29.7x447longer than winners should hear which rule takes that decision away from them. Author forks the closest448template to the edge when one fits; name none before that.449450## Resilience451452The engine fails open: every optional layer (rank, cohort, candles, Senpi) degrades to a line under453`Notes:`; the trade-level analysis needs only the public fills. An address with no perp activity in the454window and no open positions returns an error document — say "nothing to read here yet" and offer the455new-trader path in plain words ("I can help you build a strategy around how you want to trade"),456routing to senpi-strategy-author yourself; never name the skill to the reader. A malformed address457returns exit 2 with the reason.458Public-API rate limits (HTTP 429) are retried with backoff; a second run inside 10 minutes is served from459the cache (`--fresh` to refetch).460461## No address given — WHOSE book, before which book462463A bare "run quant desk" with no `0x…` is ambiguous in one way that matters: they may mean **their own464book** or **someone else's**. Settle that first — it is one question and it decides everything after.465466> Your own book, or do you want me to find you someone to read?467468### If they mean their OWN book — `desk.py --my-wallets` FIRST: one list, largest first469470**Run `desk.py --my-wallets` before anything else.** It returns ONE list, `wallets`, of the reader's471saved wallets (`user_get_me`) and their senpi strategy wallets (`strategy_list`, closed ones included),472ordered by value — largest first. A trader who came from Hyperliquid and added a wallet in Your473wallets still has it the moment they have senpi strategies — **do not forget a saved wallet**, and do474not silently swap it for a senpi wallet.475476Every wallet is first-class: **keep that order and never split it by origin** — no "saved wallets"477section and "strategy wallets" section, no saved-first or senpi-first. Origin is the `kind` column and478only says what Senpi can do with the wallet. Relay `text` (the list, rendered) and ask `ask`:479480**A Senpi strategy is one row with all its wallets.** One package is one strategy: its instances are481its wallets, grouped by the package they were deployed under (`skill_name`); a wallet with no package482stamp is its own row. A strategy row lists its instances in `wallets[]` (`wallet_count` of them; a483row of several wallets has `address: null`) and is called by its package; it is closed only when every484wallet is (`closed_wallets` counts the closed ones — "some wallets closed" when it is mixed). A saved wallet is always one row of its own.485**Every row carries `run`: the exact desk arguments for it** — `--book 0x… 0x…` over a strategy's486wallets, the plain address for a one-wallet strategy or a saved wallet. When they pick a row, run487`desk.py <run>`; never pick one instance of a strategy for them.4884891. **Value.** A saved wallet's is the MCP's `state.totalValueUsd`; a strategy's is its wallets'490 Hyperliquid account values, summed; the Senpi main wallet's is its idle cash (senpi-portfolio's491 read), and its `run` is its address like any other row. `value_status: "couldnt_load"` → say its value couldn't load —492 never as $0 — it is listed last. `value_status: "partial"` → the value covers only the wallets that493 loaded: say "(1 of 2 wallets)" from `wallets_loaded` / `wallet_count`. A closed strategy stays494 labeled closed at its value (usually $0).4952. **Kind.** A saved wallet is read-only: offer it by its label, it is the one they added. Quote its496 `access` line when they ask what senpi can do with it:497 > Read-only. Senpi can analyze this wallet. It cannot place, change or cancel orders on it.498 A strategy wallet is where their senpi perp history actually is.4993. **More than one?** Then ask, because only they know which they mean today — `ask` names them500 largest first: *"Which one do you want me to run the desk on — MetaMask ($12,400), Aegis ($3,000)501 or Phalanx ($0, closed)? Or I can read them together as one book."* Offer, in plain words, to502 read them together as one book or side by side; that is often the more interesting read. **Never503 show the user a flag or a command** (`--book`, `--compare`, `desk.py …`) — those are what you run.504 `external_wallets_status: "unavailable"` → say "I couldn't load your saved wallets" and offer505 the wallets that did load; never "you have no saved wallets". `wallets_complete: false` means a506 source is missing from the list, not empty.5074. **Neither?** Ask for an address — and **offer to show them the desk on a real book in the same508 breath**. Never guess an address, but never leave a new reader with only a question either.509510 On 2026-09-23 a brand-new user's FIRST EVER prompt was the quant-desk chip. Their agent did511 everything right — read this file, checked for saved wallets (none), checked `strategy_list`512 (empty, they had no strategies yet) — and asked for an address. One turn, eight seconds, and513 they never came back. A question is the one answer that shows them nothing.514515 > I don't have a wallet for you yet — paste any Hyperliquid address and I'll read it. Or I can516 > run it on one of this week's top traders right now so you can see what it gives you.517518 `desk.py --find <band>` returns real candidates by account size. Running one on a stranger is519 analyst mode (`--other`), which is the correct voice for it.520521**A senpi user's perp history lives in their strategy wallets, not their embedded wallet.** The522embedded wallet is a FUNDING wallet: deposits land there and move out to the strategy subwallets that523actually trade. Running the desk on it returns "no PERP activity" and the reader is told they have524nothing to read, on a book that may be trading every day.525526This is not hypothetical. On launch night two of four users asked for their own book, their agents527offered the embedded wallets — reasoning correctly that "the skill says never guess an address" — and528both got a dead end. A third user pointed the desk at three strategy wallets the same night and got529three full desks, a priced leak, and a DSL fix off the back of it. Same product, same hour; the only530difference was which wallet.531532So, where they have no saved wallet: resolve their wallets with `strategy_list` and533**offer the strategy wallets first**, named by their strategy. Offer the embedded wallet second and label it — "your funding wallet, usually no534trades of its own".535536**Never promise delivery you cannot perform.** A desk that backgrounds ("Command still running")537does not come back to you on its own — nothing wakes an agent when a detached process finishes. So538"I'll relay the rest as they come in" is a promise the loop cannot keep, and the reader is left539holding a partial answer forever. Measured: 70 of 273 desk invocations in 36 hours backgrounded.540Either poll it to completion inside the turn, or tell the reader plainly what you have and what you541did not run — and let them ask for the rest.542543**Several wallets at once: `desk.py --book 0x… 0x… 0x…` or `desk.py --compare 0x… 0x… 0x…`, in ONE544call.** Not one invocation per wallet. A desk takes 30-60s (up to ~2 min on a wide book), so a separate call per wallet backgrounds545### A senpi user's book is ALL their strategy wallets — `--book`, not one wallet546547`desk.py --book 0x… 0x… 0x…` reads every wallet and **unions them into ONE desk**: one score, one548P&L, one set of leaks, plus a per-wallet table showing which strategy carried it. That is the549default for "how am I trading" / "score my trading" / "find my leaks" from a senpi user.550551Running the desk on a single strategy wallet answers a question they did not ask. A user with four552strategies has four books, and the leak that is costing them money is usually visible only in553554…(truncated)