Implementing a venue package

The shape

defmodule DpExchange.YourVenue do
  @behaviour DpExchange.Core.Venue
end

That module is the entire public API of your package. Everything else — transport, signing, session handling, supervision — is internal, and the conformance suite asserts it.

Declare the behaviour. The compiler's missing-callback check is the cheapest assertion in the whole contract and it runs before any test does.

Do not add functions to the facade

A venue does not add functions; it declares which ones it answers. A public function only your venue has is a function a caller must know which venue it is holding to call — which is the coupling the family exists to remove.

This is not hypothetical. One venue shipped get_staking_balances/2 as a public venue-specific function. It does not cross the facade, and the consumer loses that call at migration.

If a capability is genuinely missing from the facade, that is a Core change with a deliberate release behind it, not a local addition.

Your job is to get the data and keep the connections up — not to process it

Market data passes through. What a package may hold is state describing what it is doing, and whether it is fulfilling what the host asked for.

That is the whole boundary, and it is worth stating as its own rule rather than leaving it to be inferred from the shapes the contract happens to offer, because a contract with only a snapshot type and no incremental one reads — wrongly — as an invitation to accumulate. It is not one. Decoding a venue's frames into Core.Types.* and delivering them to your subscribers is the job. Reconstructing the venue's own state from a stream of updates is a second job nobody asked your package to do, and the host is doing it already — this data is already flowing into the host's own store.

Hold — bookkeeping about your own job:

• Which pairs are subscribed to which shard or connection. This is the one piece of market topology you are allowed to keep, because it is not market data — it is a record of what the host asked for and where you put it.
• The delivery and coverage tracking backing coverage/1 and coverage_by_kind/1 — whether a symbol is actually arriving, by which route, and by which kind. This is a fact about whether your package is doing its job, not a fact about the market.
• A venue catalogue or alias map needed to attribute an inbound frame to the symbol the host subscribed to under. Necessary to route the frame at all; still not the market's own state.

Do not hold — reconstructions of what the venue itself is doing:

• An order book. Types.OrderBookDelta exists so a book stream passes through as the venue's own deltas instead of being folded into one.
• A running last price, an accumulating candle, or any other rebuild of venue state from a stream of updates. If a caller needs a maintained view, it builds one on its own side of the facade, from the values you already hand it.

Why the line is drawn exactly here. The host is already streaming this same data into a time-series store of its own, so a package-side copy is duplicated state sitting in the one place that can least afford to hold it — and holding it costs the job the package actually has. dp_exchange_coinbase's Socket was the one package in this family that got this wrong: it held a full order book per symbol and rebuilt the whole thing on every delta, which cost 65–110 ms per frame. That work ran inside the same process responsible for WebSockex.send_frame/2, so a socket that was never idle rebuilding a book it was never asked to keep could not service its own sends — which is the :send_timeout behind issue #22. Maintaining state we were not supposed to hold is what broke the connections we were supposed to keep. The other four venue packages in this family already decode a frame and pass it on without holding a copy; this section states the convention they already follow, for the one package that did not.

capabilities/0 is a claim about a real venue

def capabilities do
  DpExchange.Core.Capabilities.new(
    endpoints: %{
      {:get_price, 2} => :experimental,
      {:get_transfers, 2} => :unsupported
    },
    supported_quotes: ~w(USD USDC),
    historical_timeframes: ~w(1m 1h 1d),
    credential_benefit: :higher_ceiling,
    public_ceiling: %{limit: 10, per_ms: 1_000},
    authenticated_ceiling: %{limit: 100, per_ms: 1_000},
    measured_at: ~D[2026-08-27],
    measured_against: "GET /api/v3/exchangeInfo"
  )
end

Build it with new/1. Assembling %Capabilities{} directly skips every validation, and the validations are the point.

Declare what you measured, not what you assume

If a value was measured, say when and against what. If it was read from documentation and never probed, say that instead. An unlabelled number is worse than a missing one.

This is not fussiness. A state table drafted by people who knew the system was wrong in 7 of 21 rows; the measured version replaced it.

The declaration and the behaviour may not disagree

• :proven or :experimental → the function works. It may not answer {:error, :not_supported}.
• :unsupported → the function exists and returns {:error, :not_supported}. Not a raise, not undefined, not degraded data.

Over-declaring fails in your caller's hands at runtime. Under-declaring hides working functionality. The suite checks both directions because checking one leaves the other open.

Never declare transport

There is no has_websocket and there must never be one. Both endpoints exist on every venue. What a caller legitimately needs is which kinds of data stream — streamable: [:quotes, :order_book] — not which channels carry them. "level2" is your venue's word; :order_book is everyone's.

The domain vocabularies are closed lists, and new/1 checks them

supported_order_types and supported_time_in_force are validated against a fixed list at new/1 time — a value outside it raises rather than being carried through silently. Read the current lists from DpExchange.Core.Capabilities's source when in doubt; they grow only when a venue proves it needs a word this contract does not yet have.

Current supported_time_in_force: :gtc, :ioc, :fok, :gtd, :day, :gfw, :gfm. The last two — "good for week" and "good for month" — are real Robinhood values, added because the vendor's own OpenAPI schema names them in both the order request and response schemas and this contract had no slot for them before. Purely additive: a venue declaring a subset of this list is unaffected by the addition.

Current supported_order_types: :market, :limit, :stop, :stop_limit, :post_only, :ioc, :fok, :trailing_stop, :trailing_stop_limit, :market_on_close, :limit_on_close. The last four exist because Schwab accepts them and Core had no word for them; declaring one says the venue accepts the type, not that Core can express every parameter it takes — place_order/3's request map is for that.

Prefer Types.*.new/1 over a struct literal in your decoder

Every Core.Types.* module exposes a validating new/1, built on DpExchange.Core.Types.Validate: struct!/2, plus a check that every field the type's own @enforce_keys names is present and non-nil, raising ArgumentError naming the offending field when it is not.

@enforce_keys alone guards presence, not nil — %Candle{open: nil, high: ..., low: ..., close: ..., ...} builds without complaint even though Candle's typespec calls open a Decimal.t(), never a Decimal.t() | nil. That gap is not academic: a nil in a field the typespec forbids is exactly what a decode bug on a venue key that got renamed produces, and without new/1 the failure surfaces several calls downstream — inside Decimal or similar — with nothing pointing at which field was actually the problem.

%Candle{...} and every other struct literal still work; nothing here removes defstruct or @enforce_keys, and internal code or a test building a known-good value by hand is unaffected. new/1 is the path your own decoder should prefer, because it turns a decode bug into an ArgumentError at the boundary instead of a crash three calls downstream with no indication which venue field caused it.

Types.Order is the one type where this needs a caveat: it enforces the presence of seven keys, but its own moduledoc documents that all but :provider legitimately admit nil — "the venue's word, or nothing," since a venue can acknowledge a cancel with an id and nothing else. So Order.new/1 narrows its check to :provider alone. Check a type's own moduledoc rather than assuming every enforced key must come out non-nil.

Fail closed; never substitute

The recurring failure in this family is a nearby substitute where there should be an error. A missing granularity becoming the closest one. A missing endpoint becoming synthetic data. Every value stays plausible and only the meaning is wrong, which is why it does not surface as a failure.

If asked for a timeframe you do not serve, return an error. Do not serve the nearest width. AdapterContract's assertion 21 checks this against your Fake: it picks a width from Timeframe.nameable/0 your capabilities().historical_timeframes does not name and asks get_historical_prices/4 for it, which must not answer {:ok, _}.

Timestamps are the venue's own

Use what the venue gave you, unchanged. Where it gave nothing, nil is the honest answer — a substituted local clock is a plausible value with the wrong meaning.

Balance is the one exception, and it is stated: its timestamp is when you asked, because a balance has no venue event time and its freshness is the only thing a caller can reason about.

Carry the incident, not just the code

Where a moduledoc explains why a guard exists, that explanation is the most valuable thing in the file. Carry it when the code moves or is copied. A guard without its reason reads as defensive padding, and the next person tidying up deletes it.

The surface is 88 callbacks, and almost all of them are optional

Venue.required_callbacks/0 is the list the compiler enforces; everything else is declared :unsupported and answers {:error, :not_supported}. A new package does not implement 88 functions. It implements what its venue serves and declares the rest — which is exactly the work, because the declaring is where the thinking is.

One of the optional ones is optional for a different reason than the rest: coverage_by_kind/1 is not ceremony to skip, it is a callback Core ships ahead of any venue adopting it, so that publishing it never breaks a venue package mid-release. Adopt it when you can — see usage-rules/feeds.md for the incident it exists to make visible — but a package that has not yet is not a package doing anything wrong.

Venue.peripheral_endpoints/0 names the ones a consumer can live without, with the reason for each. It is what tells a package author which absences are survivable and which will cost a consumer the migration.

Options: opts is the venue's own vocabulary, and that is deliberate

The facade takes keyword() on nearly every callback, and packages read venue-specific keys out of it — category: on Webull, portfolio: on Coinbase, account_number: on Robinhood. That looks like the coupling this family exists to remove, and it is not, because of one rule:

A caller that passes no options must get a correct answer. Options select among things the venue offers; they never carry something the call cannot work without. Where a venue genuinely requires a parameter this contract has no word for — Robinhood v2's account number — the package refuses locally, by name, rather than sending a request the venue will reject with something less specific.

Two things follow that are worth stating because they are easy to get backwards:

• Never route on an option the caller did not pass. Webull's five categories are five separate endpoints with five different parameter sets; guessing which one a caller meant produces a plausible answer from the wrong market.
• An option this package does not recognise is ignored, not an error. A consumer moving between venues carries options that only one of them reads, and refusing them would make the uniform facade unusable for the thing it is for.

A forwarded opts turns "never configured" into key: nil, not into absence

Every venue package in this family forwards its own opts unchanged, by convention, through several layers — a Feed passes its opts straight to PollingFeed.start_link/1, which never itself set interval_ms. When nothing upstream ever configured a key, it does not vanish from the list; it arrives as key: nil, explicit and present, because something upstream read it with a bare Keyword.get/2 and passed the nil straight through.

Keyword.get(opts, key, default) only substitutes default for an ABSENT key, never for one that is present and nil. Against interval_ms: nil it returns nil, not a sane default — and a nil reaching Process.send_after/3, or arithmetic further downstream, crashes the calling process, which this library does not supervise. DpExchange.Core.Config.opt/3 is Keyword.get/3 with exactly that one difference: a present-and-nil value is treated the same as an absent one. Reach for it, not Keyword.get/3 or ||, at every default-bearing option your decoder or Feed reads out of forwarded opts — deliberately not ||, because || is falsy on false too and would silently turn an explicit log_requests: false back into its default.

A call with nothing waiting on it must set rate_limit_blocking: true

Core.HttpClient reads :rate_limit_blocking to decide between acquire/3 — wait for capacity — and check/3 — fail immediately if there is none. It defaults to false, because a caller a human is waiting on should get a fast, honest refusal rather than an unexplained pause.

That default is wrong for every call made from a timer, a Process.send_after/3, or a background task, and getting it wrong there is not a slow call — it is a permanent silent degradation, because the work that was supposed to happen simply does not, and nothing is blocked to notice.

This has now been the same defect three separate times, in three packages:

Every one is a background call, with nothing waiting on the result, failing rather than waiting a second — while Core's own throttle message names the fix in the text it returns. Reading that message requires already having shipped the bug, which is why it is written down here instead.

So, when you add a call:

• Ask who is waiting on the result. Nobody? Set rate_limit_blocking: true. A one-second wait on a 60-second timer is free; a failure is total.
• Set it at the layer that knows. A Feed knows its resubscribe is unattended and should default it to true; a Rest module does not know whether its caller is a background replay or a user-facing one-off, so it should forward the option and never default it. dp_exchange_robinhood's and dp_exchange_webull's feeds are the worked examples.
• Check every allowlist between you and HttpClient. Webull's took three: Feed built the options, Subscription filtered them, Rest filtered them again. A fix that stops at the first layer passes every test that asserts "the keyword list contains the option" and changes nothing on the wire.

An internal function nothing in lib/ calls is a defect, not a stub

"Mechanism built, documented, and never wired." Six instances in one week across this family, every one shipped green because a test called the function directly:

Every one of those functions was reachable from a test. Coverage stayed green and the suite stayed silent, because a test is not a caller.

Core.AdapterContract's assertion 16 — "internal wiring" — checks the thing a human reviewer checks by eye and a test suite cannot: does anything in this package's own lib/ actually reach this function. It reads DpExchange.Core.UnwiredCheck's call graph from :xref, the same OTP tool assertion 7 already reads for the purity check, so a captured &Mod.fun/1 or a literal apply(Mod, :fun, args) counts as real usage the way a grep never would. What it excludes — the facade, the fake, every behaviour's own callbacks, child_spec/1 and 2, start_link/1, and every compiler-injected export — is read from the same venue:/fake: bindings and behaviour declarations your contract test already supplies, never a hand-maintained list. See Core.UnwiredCheck's moduledoc for the full account of what it does and does not catch, including the one real gap: a function reachable only through your own Fake and dead on every real path is not flagged, because Fake is part of lib/ too.

When assertion 16 fails, the finding names a real function at a real line. Read it as one of two things, never a third: either the function is genuinely dead and should go, or it is exactly the shape above — reachable, tested, and never actually wired to the thing that was supposed to call it. Wire it, or delete it. Do not add it to an exclusion list; there is no per-function list to add it to.

Assertion 17 — your Fake must not be more capable than your real venue

Found independently in two venue packages the same week this assertion was added: six credentialed functions on a venue where every request is signed and there is no anonymous endpoint answered {:ok, _} for %{} or nil credentials, because nothing in the fake checked the argument at all. Tier 1 (in-process fakes) is the only tier that runs on every CI run and the only one most consumers ever exercise — a fake that succeeds where the real venue would refuse silently certifies consumer code that forgot to supply credentials.

If your capabilities/0 declares credential_benefit: :required, assertion 17 calls every active endpoint on your fake: with every credential-bearing argument stripped and refuses {:ok, _} back. It is Fake-only — it never dials the real venue — so it carries none of the risk a live-network assertion would.

"Every active endpoint", not "every endpoint that takes a credentials() argument". :required means every active endpoint on your venue needs a credential — that is what declaring it says — so this checks them all: Venue.behaviour_info(:callbacks) minus child_spec/1/start_link/1 (never called here; starting a real process from inside this suite is a risk no assertion should take) minus exactly two exemptions, named below. It used to check only a fixed list of eleven callback names that take credentials as their own first positional argument — a callback that reads a credential out of opts instead (get_option_chain/2, get_news/1, get_corporate_events/1 and quantization/1 are the shape a venue that signs every request actually uses this for) was invisible to that list no matter how it answered with no credential. If your venue declares :required and any active endpoint takes its credential through opts rather than as an argument, this now reaches it: stripped_credential_args/2 passes [credentials: %{}] for every opts position, not merely an opts list that never carried the key at all.

Two things it deliberately does not do:

• It does not run at all unless you declare :required. A venue where credentials are :no_difference or :higher_ceiling may legitimately serve some of these endpoints without one, and this assertion would otherwise be inventing a rule your venue never claimed.

• It excludes exactly two callbacks from the gate, by name, even on a :required venue: test_connection/2 and get_rate_limit_status/2. Both document credentials() | nil on purpose — test_connection/2's whole job is answering "the credential, IF GIVEN, is accepted" — so both are expected to answer plain reachability with none at all. This is a two-name exemption list stated in the assertion's own code, not a judgement call your venue's test file makes — if your fake has another endpoint that is genuinely credential-free by design (a purely local computation that never calls the venue at all, for instance), this assertion will fail on it, and that failure is the finding: either your credential_benefit: :required declaration overstates your venue, or that endpoint's exemption belongs argued in your own package's review, not assumed silently.

  dp_exchange_webull and dp_exchange_robinhood's market_status/1 were exactly this case as of the widening above, and both are now resolved — differently, because the two venues turned out not to share a reason:

  • market_status/1 is not exempt by name, the way the two callbacks above are. Its own callback doc makes an unconditional claim — "crypto venues answer :open" — but dp_exchange_schwab serves equities and its real market_status/1 calls an authenticated /markets endpoint; its fake correctly refuses without a credential, and a name-based exemption would have silenced that protection on the one venue where this assertion is doing real work, purely to accommodate two venues where it currently is not.
  • Instead there is a second, narrower exemption, scoped by what the doc's claim is actually about: skipped only when your venue's own asset_classes/0 is exactly [:crypto]. Crypto has no exchange-mandated trading session for a credential to gate, so dp_exchange_robinhood (crypto-only) is exempt on that ground — its {:ok, :open} with no credential is the documented behaviour for a crypto-only venue, argued in its own review and recorded in AdapterContract's "17. credential gate" comment rather than assumed silently.
  • dp_exchange_webull is not crypto-only (asset_classes/0 is [:crypto, :equity, :option, :future, :event_contract]), so it does not qualify and stays fully gated. It satisfies the gate by declaring market_status/1 :unsupported ({:error, :not_supported}) instead: its OpenAPI documents no market-status or trading-calendar endpoint at all — the one such endpoint Webull publishes anywhere belongs to a separate Broker API product on a different host, reachable only with a broker-tier credential this contract's credentials() does not model.

  See DpExchange.Core.Venue's own market_status/1 doc and AdapterContract's "17. credential gate" comment for the full argument.

  dp_exchange_webull's get_fees/2 is a second, later case, and it needed a third mechanism rather than reusing either of the first two: it is not exempt by name (a :credentialed-style list, correctly retired for rotting), and it is not exempt by asset class (get_fees/2 answers a flat crypto spread even though this venue is not crypto-only, so market_status_crypto_exempt?/2's ground does not apply). What actually distinguishes it is a fact about the ENDPOINT, not the venue: it answers a rate captured from Webull's own published pricing and never builds a request, so no credential could change what it returns. Capabilities.no_venue_contact names that fact directly — a list of {name, arity} your capabilities/0 declares, the same per-endpoint shape endpoints already uses, so it cannot rot into one more hand-maintained name list. Declare an endpoint there only when you can point at its real implementation and show the absence of any request-building call; assertion 17 trusts the declaration, so a wrong one defeats the same protection a wrong credential_benefit would. See Capabilities's own moduledoc for the full argument.

• It does not assert that your fake's refusal has the same shape as your real venue's ({:error, {:missing_credentials, :your_venue}} vs whatever your fake returns) — only that it is not {:ok, _}. Matching shapes would need to call the real venue with stripped credentials too, and that is a live network call this suite will not make on your behalf.

If this fails, the fix is in your fake: make it check credentials — wherever your real facade actually reads them, positional argument or opts[:credentials] — the way your real Rest/Auth module does, not in this package.

Assertion 18 — a process that links a child it starts must trap exits

Found live in four of five venue packages on 2026-09-07: Feed.init/1 never called Process.flag(:trap_exit, true), and Socket.start_link/1 ran from inside a Feed callback — which links the socket to Feed itself, not to a supervisor. An abnormal socket exit was therefore untrappable and killed Feed, and the Supervisor restarted it from its static start opts — every subscribe/2 a consumer had made since boot, gone in the same instant. One socket dying anywhere took the whole feed's subscription state with it.

Assertion 18 is static, not a behavioural "start the tree and kill a linked child" test — that was the first design and it was rejected, because starting a venue's real (non-fake) tree is not reliably network-free. dp_exchange_schwab's Feed dials its Streamer unconditionally from init/1's own {:continue, :connect}, regardless of whether anything has ever been subscribed, and the only way around that from Core is a venue-specific injection option name this suite is expressly forbidden from knowing (see this file's own "never declare transport"). A static check never starts a process at all, so it cannot dial out for any venue, present or future — see DpExchange.Core.LinkSafetyCheck's moduledoc for the full reasoning.

It scans every module in your lib/ that declares GenServer, :gen_statem, GenStateMachine or WebSockex as a @behaviour, and for each one asks two questions of its own compiled code, taken as a whole rather than function by function: does it create a link — a remote call named start_link (any target module, any arity, since X.start_link always links by OTP convention), Process.link/1, or spawn_link — and does it also call Process.flag(:trap_exit, true) somewhere. The first without the second is the violation: a module that manufactures a link to a process it started and has no way to survive that process dying abnormally.

start_link/1, start_link/2, child_spec/1 and child_spec/2 are never scanned as the source of a link — every process-behaviour module's own start_link/N delegates to that behaviour's own start_link (WebSockex.start_link/4, GenServer.start_link/3) to bootstrap itself, and that link belongs to whoever calls it, not to the module being bootstrapped. This is the same exclusion assertion 16 already makes for the same functions, reused rather than reinvented.

What it does not check: whether you handle the resulting {:EXIT, pid, reason} correctly — clearing coverage, firing a :link_down notice, reopening the child. A process that traps exits but defines no matching handle_info/2 still survives (use GenServer injects a default that logs and continues), which is the literal claim this assertion makes and no more. Whether your feed recovers usefully is genuinely different per venue and is exactly the kind of mechanism this contract's own rule ("never declare transport") forbids an assertion from encoding — that part is on you, and the five real fixes (dp_exchange_coinbase e77b542, dp_exchange_gemini 66acd3b, dp_exchange_webull d0c54a8, dp_exchange_schwab 90dddc6, dp_exchange_robinhood 51ad189) are the reference for what a good recovery looks like.

If this fails, add Process.flag(:trap_exit, true) as the first thing your process's init/1 does, and a handle_info({:EXIT, pid, reason}, state) clause that isolates the crash to whatever it actually broke.

Assertion 19 — a struct holding a secret must redact it under inspect/1

Found live in four of five venue packages on 2026-09-07: Feed/Socket held :credentials as a bare map for their whole lifetime, and OTP's default crash report prints a process's state in full on termination — a plain map prints every key it holds, secrets included. Proven by crashing an equivalent process holding %{api_key: "...", api_secret: "..."} as a bare state field and reading the log back. A second leak was found the same way: a FunctionClauseError's stacktrace prints the actual arguments a failed clause was called with, so a bad call handed the same raw map to a function whose every clause failed to match printed it too. Process.flag(:sensitive, true) does not help — it changes what :sys.get_state/1 and :dbg can see, not how a crash report or a stacktrace is formatted.

If your package holds a credential anywhere for longer than one function call, wrap it in a dedicated struct — %YourVenue.Credentials{} — the moment it enters that long-lived process, and derive Inspect with except: naming every secret field:

defmodule DpExchange.YourVenue.Credentials do
  @derive {Inspect, except: [:api_key, :api_secret]}
  defstruct [:api_key, :api_secret]
end

Four venues shipped exactly this fix independently: dp_exchange_coinbase 4d00669, dp_exchange_webull 80eaf02, dp_exchange_schwab 336cbd8, dp_exchange_robinhood cfc4861. dp_exchange_gemini needed none — it signs and discards inside stateless pipelines and never holds a credential in process state.

Assertion 19 checks this behaviourally, not by looking for @derive in your source. For every struct your lib/ defines, if any field is named one of fifteen known secret names (api_key, api_secret, app_key, app_secret, secret, password, passphrase, token, access_token, refresh_token, client_secret, private_key, signature, authorization, bearer — see DpExchange.Core.CredentialRedactionCheck's moduledoc for why each name is on the list, and why client_id deliberately is not), it builds a real instance of your struct with a distinctive value in that field and searches the actual inspect/1 output for it. A hand-written defimpl Inspect, for: YourStruct that never mentions @derive at all passes exactly as validly as the derived form above — this checks what your struct prints, never how you made it print that way.

Read this as a floor, not a ceiling. A struct field named something this list does not cover but that genuinely carries a secret — your own venue may have a field this family has not seen yet — should still be wrapped and redacted; assertion 19 not flagging it is not permission to leave it in cleartext, only a statement of what this particular check happens to look for today.

Be honest about what this does not catch. The original defect was a raw map, not a struct — Feed/Socket, pre-fix, held Keyword.get(opts, :credentials) directly in state, an opaque value never constructed as a literal anywhere in either module's own compiled code. This assertion locks the struct-based fix in; it cannot and does not reach back to catch the shape of the bug before that fix existed. If your package holds a credential as a bare map or keyword list anywhere longer-lived than one function call, that is a defect this assertion will not find for you. Wrap it in a struct regardless of whether assertion 19 currently has an opinion about your field names.

If this fails, add @derive {Inspect, except: [...]} (or an equivalent hand-written defimpl Inspect) naming every secret field your struct holds, following the four commits above as the reference for the mechanism.

Assertion 20 — subscribe/2 must push a Core.Types.* struct, tagged with runtime_id/0

DpExchange.Core.Venue.subscribe/2's own doc makes an unconditional claim: events arrive "tagged so a process subscribed to several venues can tell them apart," and "the payload is a DpExchange.Core.Types.* struct — the same value the pull endpoints return." Nothing checked either half before this assertion existed. Assertion 16 (internal wiring) catches a decoder with no caller, but a decoder that is wired and simply never gets called before the raw response reaches the sink passes every other assertion — the function that forwards it has a caller, and subscribe/2 still answers :ok.

Assertion 20 calls your Fake's subscribe/2 with @sample_pairs and to: self(), and checks the first message that arrives: it must be {:dp_exchange, runtime_id, payload} where runtime_id == your_venue.runtime_id() and payload is a struct under DpExchange.Core.Types. Fake-only, like assertion 17 — every fake in this family pushes synchronously inside the call that returns :ok, so nothing here ever dials out.

If this fails, make sure your subscribe/2 builds the same Types.* struct your pull endpoint for the same data returns — never the venue's raw decoded JSON — and tags every send with runtime_id(), not a literal atom.

Assertion 21 — a timeframe outside historical_timeframes must be refused

The behavioural half of "Fail closed; never substitute" above, run against your Fake. Assertion 21 picks a width from Timeframe.nameable/0 that your own capabilities().historical_timeframes does not name, and calls get_historical_prices/4 with it. The answer must not be {:ok, _} — your fake either errors on the unrecognised width or (if you gate on credentials first) refuses for that reason instead, either of which proves the width was never silently served at the nearest one you do have.

A venue that declares the entire Timeframe.nameable/0 vocabulary has nothing left for this assertion to pick, and the check is a no-op for it — it is not a way to avoid the check by declaring narrowly, since narrowing what you declare only gives the assertion more widths to try.

Asset classes are a statement about today

asset_classes says what the package serves now — :crypto, :equity, :option, :future, :event_contract. It is never a permanent scope boundary, and it must never be used to justify not implementing something: "this venue's options endpoints are out of scope because we declared crypto" is the argument in its wrong form, and it has been made in this family and was wrong.

The only test of scope is does the venue provide it.

Two lists, not one: absence has two causes

Split your :unsupported endpoints into what the venue does not serve and what this package has not ported, and expose the first through venue_does_not_serve/0.

Both answer a caller identically. Only one of them can ever change, and a host planning around a gap needs to know which it is looking at.

The mislabel goes both ways, and both are defects. A venue's absence filed as a backlog item invents work that cannot be done and quietly implies an endpoint the vendor does not publish. A backlog item filed as the venue's absence hides a capability a consumer could have had. Robinhood shipped four of the first kind and they were found by auditing, not by tests — nothing fails when a comment is wrong.

Every negative gets an audit

Write docs/reference/<venue>/negative-claims.md, tabulating every place your package says the venue does not do something, with the source and the date you consulted it.

An unverified negative is a substitution exactly like an invented value. "The venue has no order book" and "we never looked" produce the same {:error, :not_supported}, and only one of them is true. Across five venues this audit found nine false negatives — every one of them a working endpoint a consumer was being refused.

Two patterns produced most of them:

• A true statement about one endpoint restated as a claim about the venue. "The stock snapshot does not serve options" is correct; "this venue does not serve options" is not.
• A derived artefact read instead of the vendor. A claim originating in the host application's own adapter, or in a third-party wrapper's README, carried forward until somebody read the vendor's pages.

Dependency floors are a claim exactly like a capability

A ~> X.Y requirement is a claim, in public, that your package runs against every version from X.Y.0 up. mix.lock never tests that claim — it always resolves the newest version the requirement allows, so a floor that is actually too low compiles and passes every ordinary CI run. Only a consumer who resolves an older, still-permitted version finds out, and finds out as a crash in their application, not a failure in yours.

Four real instances, found in one week, all the same shape:

• dp_exchange_webull declared {:websockex, "~> 0.4"} while calling WebSockex.send_frame/3, which exists only from 0.5.1. A consumer who resolved 0.4.x got :undef.
• dp_exchange_webull declared {:dp_exchange_core, "~> 0.1.48"} while capabilities/0 declares no_venue_contact, a Capabilities field added in Core 0.1.68. Capabilities.new/1 builds with struct!/2, so an older Core raised KeyError on every capabilities/0 call — not on the one path that changed, on the one every consumer calls first, at boot.
• dp_exchange_gemini declared ~> 0.1.48 while calling Types.OrderBookDelta.new/1, added in Core 0.1.53.
• The first fix for the webull/websockex instance was itself wrong. It raised the floor to ~> 0.5, reasoning that the third send_frame argument "only exists from 0.5". ~> 0.5 still permits 0.5.0, and send_frame/3 is new in 0.5.1 — one patch later. The corrected floor was reasoned about, not resolved, and reasoning about a floor is exactly the failure mode this whole section is about. It was only caught because script/check_dependency_floor.sh (below) resolved it for real the same day.

A per-API pinning test — function_exported?/3 after Code.ensure_loaded!/1, or a struct-field check — guards a floor against being lowered later. It cannot catch a floor that was wrong when it was written, because a test like that has no way to know what "new" means; it only knows what the author already thought to check. Every instance above shipped with its code passing whatever tests already existed.

The fix is to resolve the floor, not read it. script/check_dependency_floor.sh rewrites a scratch copy's mix.exs, pinning every dependency declared without only: (the ones a consumer actually resolves — never credo, ex_doc, or anything else that ships only to :dev/:test) to == the exact floor version its own requirement string names, then runs mix compile --warnings-as-errors plus the package's own AdapterContract conformance test against that pinned set. Compiling catches an undefined remote call (the websockex shape); the conformance test — which calls capabilities/0 under a fully offline Fake — catches a struct or module the floor does not ship yet (the other two shapes). It runs weekly and on demand, deliberately not on every push: it freshly resolves the transitive dependencies of whatever it pins (Core's own req, for instance), which float to whatever is newest today and are not this package's claim to keep correct, so an unrelated package's release can turn this red for a reason that has nothing to do with your floor. See the script's own header for the full reasoning, including why a full mix test is deliberately not what it runs.

Raising mix.lock is never a substitute for raising the floor in mix.exs. Every instance above passed CI because CI always resolves the newest allowed version — the bug was invisible from inside this repository by construction. If a change starts calling a Core function, uses a new Types.* module, or relies on any behaviour a dependency added after its floor's release, the floor in mix.exs goes up in the same commit, stated as a version and the reason, the way every floor comment in every venue's mix.exs already does. Do not wait for the weekly check to say so.

Never do blocking work in a process that owes a reply

This is the failure this family has paid for most often, and every instance looked different until they were lined up:

• #16, #23 — work on the reply path inside a venue's own feed.
• #28 — Core.PollingFeed ran its fetch in a task and then blocked on that task inside handle_info. The task bounded a hang; it did nothing for the mailbox. A read-only coverage/1 timed out, the exit propagated out of the venue Feed's handle_call/3, and the feed died. It restarted without its subscription state, and a live venue went from 61 pairs to 0 and stayed there — alive, idle, passing every liveness probe.

A sweep for the class then found three more, none of which had failed in production yet: a Streamer bootstrap (a signed HTTP round trip plus a WebSocket connect) inline in handle_call/3; a whole-catalogue HTTP fetch inline in handle_info/2; and a socket connect inline in handle_call/3.

The rule

If a callback can block for longer than a caller's timeout, it must not block the process. Run the slow part in a task, let the result arrive as a message, and defer the reply with {:noreply, state} + GenServer.reply/2. dp_exchange_webull's spawn_reconcile/3 is the reference shape; do not invent a second one.

Three things that are easy to get wrong:

• A task that bounds a hang does not unblock a mailbox. Wrapping work in Task.async/1 and then calling Task.yield/2 is still blocking — it only bounds how long. That is precisely what #28 was.
• start_link links to its caller. A socket opened inside a task dies when the task exits, moments later. Fetch in the task; connect in the GenServer. Only the fetch is slow enough to matter.
• Convert exceptions inside the task. Task.async/1 links, so an unconverted raise arrives as an {:EXIT, …} the GenServer has no clause for — and anything parked waiting on that task is never answered at all.

Reads must carry an explicit timeout too

coverage/1 and status/1 are what a consumer's health check calls. Left on GenServer.call/2's implicit five seconds while writes name a generous one, any moment the feed is legitimately busy turns a health check into an exit — and into a dead consumer process, when the read happens inside the consumer's own handle_call/3.

Every venue in this family now passes @call_timeout on reads as well as writes. It is the second line of defence, never the fix: a read that has to queue behind something should wait for it, not die of it.

And the answer must not lie

Where a read cannot reach what it is reporting on, say the least that is true. An empty coverage map plus a :link_down notice is honest; a remembered coverage map asserts arrivals nobody confirmed, which is the "nearby substitute where an error belongs" failure this family keeps finding. "We could not ask" is not "nothing arrived", and only the notice tells a consumer which one they are looking at.

If your vendor publishes an index, diff it

Across five vendors, a changelog diff caught nothing and an index diff was the only mechanism that ever fired. Three instances now, all real:

• A rate-limit table on developer.webull.com that had been published for weeks while this family declared a ceiling five times too permissive, on a venue whose documented penalty is a temporary IP block.
• A WebSocket channel on developer.gemini.com withdrawn with no changelog entry — the same venue that once withdrew an entire market-data API the same way.
• Coinbase's rate-limit pages, previously recorded as "could not be located", sitting in the vendor's own sitemap.xml the whole time.

So every venue package carries script/check_endpoint_inventory.sh, weekly and non-blocking, comparing what the vendor says it serves against a committed record. What it compares depends on what the vendor offers, in this order of preference:

1. A machine-readable specification — OpenAPI, AsyncAPI. Diff the operation and channel lists. This is the strongest form; only Gemini publishes one today.
2. A sitemap whose pages are one-per-endpoint — diff the set of those paths. Coinbase and Webull are this shape.
3. Nothing fetchable — Schwab answers 403 to an anonymous reader. There is no automated form; its specification is committed to the repository and re-capture is a human signing in. That is a distinct class, not a degraded one.

Two rules that make the difference between a check and a rubber stamp:

• Fix the claim before you update the record. A difference means a claim this package makes may now be false. Updating the committed inventory first, so the check goes green, destroys the only evidence that anything changed.
• Diff the specification, never the rendered page. An operation list is structured and every entry means something. The pages around it carry build hashes and rotating banners, and a content diff on those is red every week for reasons that are never the reason you care about.

Absent from the documentation is not absent from the venue

When something vanishes, what you have established is that the vendor stopped publishing it — not that the venue stopped serving it. Gemini has diverged from its own documentation in both directions: a socket URL it still published and no longer served, and candle widths it served before it documented them.

So a withdrawal is a reason to label a claim, not automatically to delete it. Deleting asserts a new negative ("this venue does not have X"), and an unverified negative is a substitution exactly like an invented value. Where the thing is private and probing it needs a credential the repository must never hold, labelled is the most honest state available.