Walk the book at your size. The best quote never decides.

When Nasdaq is shut, Bitget's own rToken order book is the market, and its best quote prices only your first shares. Sounding answers the question that comes before you confirm: what does your whole order cost, at your own fee? It walks Bitget's book at your full size, holds it to your fee and ceiling, and gives one answer, or a named refusal with the largest size that fits. For an AI agent trading through Bitget Agent Hub 2.0, it puts that cost on the confirmation card the trader approves, and hands back the exact order only when it fits.

▶ Demo video (3:09) · Live desk ↗ · Judge it in 90 seconds ↗ · Take Bitget away ↗ · Proof ↗

▶ Demo

A trader asks to sell 178.4121 shares of rHIMS and keep it under 0.3%. The best bid says 17.48 bps; the full order walks five levels down the book to 36.87 bps and is refused, with 101.834 shares named as the size whose average fits. At a 50 bps ceiling the full order fits. An AI agent tries an order the book cannot carry through Agent Hub, whose dry run previews it as given; Sounding hands back no order, and offers the size it can send. An order that fits comes back as the exact Agent Hub order, a limit at the ceiling price, bound to a signed receipt.

Watch the demo on YouTube (3:09). Every frame is the deployed site on recorded weekend Bitget books, with Qwen 3.8 Max on Bitget's S2 endpoint; every terminal line is a production response recorded on 2026-10-08. Nothing was sent to the exchange.

Contents

• The problem I set out to solve
• What I built
• Verify it yourself in 90 seconds
• Architecture
• How I integrated Bitget
• Designed for agents: the refusal vocabulary
• Engineering decisions and the hard problems
• Numbers
• What's real: the honesty table
• Run it locally

The problem I set out to solve

On weekends and US holidays Bitget matches rToken orders on its own book, with market makers (Bitget support), and its own weekend guide tells traders to check order-book depth, use limit orders and split large orders. A quote prices the first few shares. An order is all of them. A trader, or an AI agent acting for one, looks at the best bid on a Bitget rToken, adds a fee, and sees a cost inside their limit. The order then walks down the book and fills worse. On one recorded Sunday, at Bitget's published 5 bps rToken fee, the best ask said yes and the full order said no for 28 of the 89 weekend-tradable names a 25,000 USDT buy could fill.

Bitget's own agent stack has the same gap. Agent Hub's flow is a dry run, then a confirmation card naming pair, side, quantity and account, then the send. Its pre_trade_check reads the ticker price, the balance and the positions; nothing in the flow walks the book at the order's size. Its dry run previews whatever it is given: a 5,000-share rHIMS market sell the visible book cannot fill comes back as a ready preview.

So I treated "the size does not fit" as a first-class answer, sitting right next to "the price looks fine."

What I built

1. Read. Qwen 3.8 Max reads the trader's words, English or 中文, and proposes each limit (size, fee, ceiling, deadline) with the exact words that state it.
2. Check. Code finds those words in the message and reads the number or date itself. A disagreement, an impossible date, a negative size, a corrected figure, or a limit mentioned but not read is asked back, never filled with a default.
3. Walk. The engine walks Bitget's book level by level at the full size, in exact decimals, with the taker fee charged on the traded amount. It does so on weekends and US holidays, when Bitget matches rToken orders on that book. In US sessions Bitget routes them to NASDAQ/NYSE, so Sounding refuses with ROUTED_TO_US_MARKET rather than price a book the order would not fill against.
4. Decide. Within the ceiling at a known fee, or over it with the largest size that fits. An unknown fee that changes the answer is asked.
5. Look again. Before acting, a fresh book is walked. The decision stands only within 10 bps and two minutes.
6. Hand over. For an agent, the exact Agent Hub order arguments: a limit at the trader's ceiling price, so Bitget fills no share that costs more than the ceiling, bound to a signed receipt. Or a refusal with a reason and the size that does fit.

Verify it yourself in 90 seconds

No account, no key. Nothing is sent to the exchange. These run on the recorded Sunday rHIMS book, so the answers repeat on any day; "mode":"live" reads the book now, and on a weekday it is refused with ROUTED_TO_US_MARKET.

# An order the book cannot carry: refused, with the size that fits offered as a new order
curl -s https://sounding-zeta.vercel.app/api/prepare -H 'content-type: application/json' \
  -d '{"symbol":"RHIMSUSDT","side":"sell","amount":"5000","ceilingBps":40,"userFeeBps":5,"mode":"recorded"}'
# → {"preparation":{"status":"refused","code":"INSUFFICIENT_VISIBLE_DEPTH","reason":"the visible book cannot fill this order","proposal":{"size":"66.0764","unit":"sh",…}},…}

# An order that fits: the exact Agent Hub order, bound to its receipt
curl -s https://sounding-zeta.vercel.app/api/prepare -H 'content-type: application/json' \
  -d '{"symbol":"RHIMSUSDT","side":"sell","amount":"50","ceilingBps":40,"userFeeBps":5,"mode":"recorded"}'
# → {"preparation":{"status":"prepared","order":{"action":"place","category":"SPOT","symbol":"RHIMSUSDT","side":"sell","orderType":"limit","price":"27.93","timeInForce":"ioc","qty":"50"},"command":"bgc order --action place … --dry-run","binding":{…"allInBps":"19.89","worstCaseBps":"38.88"…},"signed":…,"book":"recorded"},…}

Then open /bitget, where the engine runs one order with each Bitget input withheld, and run every number in this README again:

pnpm install && pnpm verify
# PASS  lead.best-bid                  17.48 bps
# PASS  lead.full-size                 36.87 bps, over 30
# …
# 16/16 claims verified

Architecture

flowchart LR
  W["Trader's words<br/>English · 中文"]:::you --> Q["Qwen 3.8 Max<br/>proposes each limit,<br/>quoting the words"]:::model
  Q --> C["Code<br/>finds the words,<br/>re-reads the number"]:::code
  subgraph Bitget["Bitget"]
    B1["Spot order book<br/>150 levels"]
    B2["Reality stock-info,<br/>market states, calendar"]
    B3["Instrument rules"]
    F["Your taker fee<br/>Agent Hub read-only"]
  end
  C --> E["Engine<br/>gates, exact decimals"]:::code
  B1 & B2 & B3 & F --> E
  E --> D["Decision<br/>+ signed receipt"]:::code
  D --> L["Last Look<br/>fresh book"]:::code
  D --> P["Agent Hub order<br/>only if it fits"]:::code
  classDef you fill:#0f172a,stroke:#64748b,color:#e2e8f0
  classDef model fill:#1e1b4b,stroke:#818cf8,color:#e0e7ff
  classDef code fill:#052e2b,stroke:#2dd4bf,color:#ccfbf1

Loading

The model reads and explains. It never does the arithmetic, and it cannot pass an order.

	Gate	Refuses	Runs on
1	Instrument	a symbol not on Bitget's Reality list	every answer and every prepared order
2	Session	a US session, when Bitget routes the order to NASDAQ/NYSE instead of its own book, or a session its states and calendar cannot place	every answer and every prepared order
3	Weekend eligibility	a weekend order on a name Bitget does not flag weekend-tradable	every answer and every prepared order
4	Exchange constraints	a size off Bitget's quantity precision or under its minimum, naming a size it would accept where one exists	every answer and every prepared order
5	Book validity	an empty, crossed or malformed book	every answer and every prepared order
6	Freshness	a live book older than 5 s or slower than 2 s to arrive	every live answer and prepared order
7	Stability	two soundings of the same order more than 10 bps apart	the desk, between soundings
8	Checked reading	a limit the model cannot quote from your words, or that code reads differently	orders said in words
9	Rule checks	an answer with an unpriced route, an invented figure, a promised fill, or a loosened limit	the analyst's answer
10	Ceiling at a known fee	full size over your ceiling at your fee; an unknown fee that changes the answer is asked	every answer and every prepared order
11	Last Look	a fresh book more than 10 bps worse, a flipped verdict, or a decision older than two minutes	the desk, before acting
12	Binding	a prepared order changed after checking, not signed by this server, or older than two minutes	Agent Hub orders, before sending

How I integrated Bitget

Bitget	Call	Without it
Spot order book	GET /api/v2/spot/market/orderbook?limit=150	refused: NO_EXECUTABLE_QUOTE
Reality stock-info	GET /api/v3/reality/market/stock-info	refused: INVALID_INSTRUMENT
Market states, calendar	GET /api/v3/reality/market/states · /calendar	refused: SESSION_UNKNOWN
Instrument rules	GET /api/v3/market/instruments?category=SPOT	refused: INVALID_INSTRUMENT
Agent Hub, read-only	account_overview (agent, the trader's account) · SDK 3.3.1 getAccountFeeRate (self-hosted, its operator's account)	the fee falls to the published 5 and 10 bps scenarios; an answer that depends on which is asked instead of prepared (sell 16 rHIMS under 20 bps on the recorded book: refused without a fee, prepared at 5 bps)
Agent Hub, order path	order tool · bgc order --action place	nothing receives the checked order: the answer stays advice, and an agent left to Agent Hub alone sends what it is told
Qwen on Bitget's S2 endpoint	hackathon.bitgetops.com/v1	the fallback reader cannot read "0.05%", "0.3%" or "before the 8th", and asks for limits already given

The first four rows are computed by the engine on every build (src/lib/deletion.ts); the build fails if a removal stops changing the answer. Agent Hub, recorded on 2026-10-08 with a 40 bps ceiling and a stated 5 bps fee; Sounding's side runs on the recorded Sunday rHIMS book, where Bitget's own book is the market (evidence/agenthub-20261008):

Order	Agent Hub dry run (bgc 3.0.0)	Sounding (/api/prepare, production)
sell 5000	previews it	refused, INSUFFICIENT_VISIBLE_DEPTH: the visible book cannot fill it; 66.0764 sh, what fills now inside a limit at the ceiling price, offered as a new order
side hold, qty -5	previews it	rejected as input
sell 50	previews it	prepared: 19.89 bps all-in, as the exact Agent Hub order, a limit IOC at 27.93, where one share costs 40 bps: no fill can cost more (worst case 38.88), signed

Two sizes, two questions. The desk answers on the order's average cost on this book: 101.834 sh of the lead sell average exactly 30 bps. A prepared order must hold every share inside the ceiling, because that is what a limit order lets Bitget enforce: on the same book only the first four bid levels, 66.0764 sh, sit at or above 27.96, the 30 bps ceiling price. So /api/prepare refuses 101.834 at 30 bps and offers 66.0764; at 50 bps it prepares the full 178.4121 sh at 27.90.

The fee is charged on the traded amount: buy p + f + p·f/10⁴, sell p + f − p·f/10⁴. A fee the trader states decides. An agent reads the trader's own fee with Agent Hub's account_overview and passes it in; a self-hosted Sounding with a read-only key reads its operator's fee through the SDK's getAccountFeeRate. Without either, Bitget's published 5 bps rToken rate and its 10 bps list rate are both priced, and an answer that differs between them is asked.

Designed for agents: the refusal vocabulary

Every refusal is a code an agent can branch on, never prose alone.

Where	Codes
Engine gates	INVALID_INSTRUMENT · SESSION_UNKNOWN · ROUTED_TO_US_MARKET · UNAVAILABLE_THIS_SESSION · INVALID_BOOK · NO_EXECUTABLE_QUOTE · FRESHNESS_UNKNOWN · UNSTABLE_QUOTE · INVALID_QUANTITY_PRECISION · BELOW_MIN_ORDER
Verdicts	WITHIN_CEILING_ON_THIS_SNAPSHOT · OVER_CEILING_ON_THIS_SNAPSHOT · INSUFFICIENT_VISIBLE_DEPTH
Preparation	prepared (with order, binding carrying allInBps, worstCaseBps and expected, signed, book: live | recorded) · refused with a code (an engine gate, a verdict, FEE_DECIDES, PAST_CEILING_PRICE, BELOW_MIN_ORDER), a reason and, when one exists, a proposal that itself prepares
Verification	ok, or a reason: the order differs, not issued by this server, no signing key, older than two minutes

• agent/sounding-mcp.mjs: a dependency-free MCP server with sounding_prepare_order and sounding_verify_order, to run beside @bitget-ai/bitget-agent-mcp.
• agent/skills/sounding-precheck/SKILL.md: a skill in Agent Hub's format. It runs after the dry run, adds the at-size cost to the confirmation card, and sends only the order Sounding prepared.

Engineering decisions and the hard problems

• The model never does arithmetic. Qwen proposes and explains; code re-reads every value it quotes and checks every cost figure it writes (bps, %, 基点, 千分之) against the engine's output, English or Chinese. An answer that breaks one of 16 named rules is replaced by a deterministic template.
• Ask, never default. A size in words, a corrected fee, a past or impossible date, a negative size: each is a question back to the trader, never the form's value quietly priced instead.
• The fee is charged on what is traded. Adding the fee on top misjudges orders sitting near the ceiling, so every verdict, clip, chart and census cell uses the exact formula.
• Correct in public. The research first assumed a 20 bps fee. When an independent review showed Bitget publishes 5 bps, the census was restated on the same frozen books with the original figures left beside the new ones.
• A recorded book is labelled; a prepared order expires. Recorded answers say so, and a prepared order is good for two minutes, then the book is read again.

Numbers

pnpm verify recomputes each row from the frozen books and the published runs, and fails on any difference.

claim	value	from
lead.best-bid	17.48 bps	recorded rHIMS book, 5 bps fee
lead.full-size	36.87 bps, over 30	the same book, 178.4121 sh
lead.fits	101.834 sh	the largest size whose average is within 30 bps
census.5k-50	6 / 90	best ask within 50 bps, 5,000 USDT buy over
census.25k-50	28 / 89	the same at 25,000 USDT
census.fee-decides	17 / 89	within at 5 bps, over at the 10 bps list rate
census.median-5k	23.61 bps	median all-in, 5,000 USDT buy
deletion.book	NO_EXECUTABLE_QUOTE	lead order, book withheld
deletion.stock-info	INVALID_INSTRUMENT	stock-info withheld
deletion.session	SESSION_UNKNOWN	states and calendar withheld
deletion.instruments	INVALID_INSTRUMENT	instrument rules withheld
deletion.qwen	Asked, not answered	Qwen withheld
eval.whole-task	28/30 (1 critical) vs 12/30 (3 critical)	blind set 2, Qwen vs baseline
eval.reading	142/151 vs 19/151	blind set 2, limits read
agenthub.dry-run-previews	3 of 3	bgc dry run, recorded
agenthub.sounding-prepared	1 of 3	/api/prepare, recorded

Why the model is needed: the same thirty blind tasks, with and without Qwen, run on engine 0.4.0 before the fee correction: 28 done right with 1 critical error, against 12 with 3. On 40 blind messages in English, 中文 and mixed, Qwen read 142 of 151 stated limits correctly; a regex reader read 19. Every turn is scored as rendered, and runs are published as they came out, misses included.

What's real: the honesty table

Capability	Status
Bitget books and metadata	Real. Live in live mode. Recorded mode replays frozen, hashed captures and says so on every answer.
Engine decisions and receipts	Real. Exact decimals; receipts hashed, signed, and replayable offline (pnpm replay).
Qwen reading and ruling	Real, on Bitget's S2 endpoint, checked by code on every turn.
Agent Hub fee read	Real, through Agent Hub's read-only client, recorded with a read-only key on the developer's machine (taker 5 bps, the same as Bitget's published promotional rate). The public site holds no key and never shows one account's fee as anyone else's; an agent passes the trader's own.
Which book fills	Weekends and US holidays only. Bitget's Stock 2.0 guide routes rToken orders to NASDAQ/NYSE in US sessions and matches them on its own book, with market makers, on weekends and US holidays. Sounding prices only the second case and refuses the first (ROUTED_TO_US_MARKET). On a weekday overnight (2026-10-08, 270 samples) the public ticker and the public book also disagreed: whole-share quotes on one, fractional on the other. That does not identify a venue by itself; Bitget's published routing is the reason, and the scope is the conservative reading of it (evidence/routing-20261008).
Weekend order type	Limit IOC, prepared, never sent. The prepared order is a limit at the trader's ceiling price, immediate-or-cancel: one share at that price costs exactly the ceiling at the trader's fee, so no fill can cost more, and what cannot fill at once is cancelled. An order is prepared only when its whole size fills inside that limit on the book it was checked on; otherwise the size that does is offered. Bitget's weekend FAQ lists limit orders, and Agent Hub's order tool takes price and timeInForce; whether Bitget accepts IOC on an rToken over a weekend is not tested.
Agent Hub order	Prepared, never sent. Shown as a dry run: Bitget's demo environment listed 6 of the 90 weekend-tradable rTokens, all with empty books, when checked on 2026-10-05, so a paper fill would prove nothing.
Enforcement	Advisory to send, enforced once sent. The skill tells an agent to send only what Sounding prepared; it cannot stop a client that skips it. An order it does send is a limit at the ceiling price, which Bitget's matching holds: no share fills past it.
Fills	Not promised. Every answer is conditional on the snapshot it names; hidden liquidity and the future book are out of reach.
Research	Real, corrected. The census was restated on 2026-10-07 from 57/89 to 28/89 at 25,000 USDT; the originals remain on /research. Runs made on engine 0.4.0 say so.
Sampling	Real, with a gap. Every 30 minutes on a declared schedule; a 21-hour gap on Oct 4–5 is recorded, not filled (evidence/atlas-202610).

Run it locally

pnpm install
cp .env.example .env.local
pnpm dev        # http://localhost:3000
pnpm test       # 355 tests in 16 files
pnpm verify     # every claim above, recomputed
pnpm replay fixtures/rhims-20260920T090235Z.json sell 178.4121 30 5

Variable	Purpose	Required
BITGET_QWEN_API_KEY	Qwen on Bitget's S2 endpoint; without it the deterministic reader and template answer	no
BITGET_API_KEY · BITGET_SECRET_KEY · BITGET_PASSPHRASE	a read-only Bitget key, used only to read the operator's taker fee through Agent Hub's SDK	no
SOUNDING_RECEIPT_KEY	signs receipts and prepared orders; without it nothing prepared can be verified	to verify

The suite mirrors the claims: the lead order's figures, each Bitget input withheld, the exact fee formula at the ceiling, intake asking instead of defaulting (signed, zero, worded and corrected sizes and fees, impossible dates, English and Chinese), every figure in an answer checked against the engine, and a prepared order rejected when changed, unsigned or stale.

Next.js 16 · TypeScript · decimal.js · @bitget-ai/bitget-agent-sdk · Qwen 3.8 Max · vitest

---

Built for the Bitget AI Base Camp Hackathon S2, Track 3: AI Trading Desk · Execution Assistance, by cassxbt. MIT licensed.