AshHooks

Webhooks for Ash Framework, in both directions:

Inbound — receive provider webhooks, verify their signatures, deduplicate them on a ledger (a table recording every delivery — the dedup record itself), and run your handler with durable deduplication and at-least-once processing.
Outbound — sign and deliver your own webhooks with retries, backoff, and dead-lettering through Oban or your own scheduler.

The two halves work independently: if you only receive webhooks, you need no queue infrastructure at all.

Verify signatures for ComplyCube and HubSpot v3 out of the box; bring your own scheme with a one-module provider behavior.
Duplicate and replayed deliveries are deduplicated on the ledger — exactly one row per delivery; crashes mid-flight resume on redelivery instead of losing events (handlers run at-least-once).
Outbound webhooks follow the Standard Webhooks spec, so receivers verify with any conformant library. Key rotation and a legacy-envelope mode for receivers mid-migration are built in.
Retries honor Retry-After, back off with jitter, dead-letter at a ceiling, and durably disable endpoints that return 410.
Safe defaults: inbound declarations use secret resolvers rather than literal binaries, and endpoint rows store secret references. Endpoint URLs are checked against server-side request forgery (SSRF) at registration and again at send time, and response bodies are never stored unless you explicitly opt in for a diagnostic run. Inbound ledgers retain decoded payloads and a digest of the signed bytes; outbound ledgers retain the exact bytes to send. Read access is yours to govern — the package injects no read policies (see Security).
Telemetry events for the whole send/receive lifecycle, with classified reasons and identifiers rather than bodies or signing secrets.

Requires Elixir _{> 1.20 (OTP 28+) and Ash >= 3.34.3, < 4.0. Add Oban (}> 2.20) for the generated delivery worker, and Phoenix or Plug for the HTTP ingress. Direct delivery and signature verification work without either dependency. From 1.0 the package follows semantic versioning with a named public surface — see Stability.

Installation

def deps do
  [
    {:ash_hooks, "~> 2.0"},
    # only for outbound delivery:
    {:oban, "~> 2.20"}
  ]
end

Upgrading from 1.x? Follow UPGRADING.md before starting 2.0 workers. The delivery schema and HubSpot's default dedup identity have changed.

Or mix igniter.install ash_hooks, which also tries to patch your endpoint's Plug.Parsers with a raw-body reader. Signature schemes sign the exact wire bytes, and a router plug cannot recover what the parser already consumed — so if the automatic patch didn't apply, add it yourself:

plug Plug.Parsers,
  parsers: [:json],
  pass: ["*/*"],
  body_reader: {AshHooks.BodyReader, :read_body, []},
  json_decoder: Phoenix.json_library()

By default every parsed request carries a cached copy of its raw body; pass [only: ["/webhooks"]] as the reader's third element to limit that memory cost to your webhook routes.

You also create your own tables — ash_hooks injects fields and identities onto your resources, and your migrations carry them, including the unique indexes the deduplication guarantee rests on. Complete, runnable migrations, resources, and setup are in the get-started tutorial.

Receiving webhooks

Declare an inbound source on a ledger resource:

use Ash.Resource,
  data_layer: AshPostgres.DataLayer,
  extensions: [AshHooks, AshHooks.InboundDelivery]

inbound_delivery do
  # provider event ids aren't unique across accounts — your scope slots
  # extend the dedup identity (each must be a non-nullable attribute)
  scope_identity([:account_id])
end

attributes do
  attribute(:account_id, :string, allow_nil?: false)
end

webhooks do
  inbound :comply_cube do
    secret {:app_env, [:my_app, :complycube_secret]}
  end
end

Secrets are always sources — an {m, f, a} callback, an {:app_env, path}, a zero-arity function, or (multi-tenant apps) a one-arity function that receives the tenant — never literal values.

Your provider module also defines the handler — the handle_event/2 callback that receives the verified payload (the guided-tour Livebook builds one from scratch in a few lines).

From your controller, one call runs the whole pipeline: verify the signature over the raw bytes, persist the payload, deduplicate, claim under a lease (a time-limited ownership claim — rows whose worker died are reclaimed when it expires), run your handler, and record the outcome.

case AshHooks.Ingress.ingest(Ledger, :comply_cube, conn.private[:ash_hooks_raw_body], %{
       signature: List.first(get_req_header(conn, "complycube-signature")),
       headers: Map.new(conn.req_headers),
       scope: %{"account_id" => conn.params["account_id"]}
     }) do
  # Acknowledge terminal rows, including permanent failures.
  # A duplicate may have retried its handler: inspect its current status.
  {:ok, _tag, %{status: status}} when status in [:processed, :failed_permanent, :superseded] ->
    send_resp(conn, 200, "")
  {:ok, _tag, _row} -> send_resp(conn, 500, "")    # handler failed
  {:error, _} -> send_resp(conn, 400, "")          # bad signature/payload
end

Delivery semantics. A terminal delivery (:processed, :failed_permanent, or :superseded) is never processed again. A crash after your handler ran but before the ledger recorded it will re-run the handler on redelivery — so write handlers idempotent, keyed on the logical event or batch identity (for action-level idempotency elsewhere in your app, our sibling package ash_onetime is an optional companion). Permanent failures belong on your operator surface; the successful acknowledgment prevents a provider from retrying a row that cannot run again.

HubSpot's v3 scheme also signs the HTTP method and the full request URI, so its controller passes both — build the public URI from a base URL you configure, not from conn, which behind a TLS proxy carries the internal host and port.

claim_delivery/3, mark_processed/4, and friends are public if you want to drive the lease machine from your own async pipeline; AshHooks.Ingress.reap/2 re-drives deliveries whose claims died with an expired lease.

Sending webhooks

Declare the event on the emitting resource, point it at your subscription and delivery resources, and define one worker module:

defmodule MyApp.WebhookDeliveryWorker do
  use AshHooks.Worker,
    deliveries: MyApp.OutboundDelivery,
    endpoints: MyApp.WebhookEndpoint,
    secret_resolver: {MyApp.Secrets, :webhook_secret},
    queue: :webhooks
end

The secret resolver maps an endpoint's secret reference to its value — generate values with AshHooks.Signing.generate_secret/0, store them whole in your secret store, and return them unchanged.

Dispatch, wiring the worker's generated enqueue function:

{:ok, event} =
  AshHooks.Event.new(
    type: :order_paid,
    payload: Jason.encode!(%{order_id: order.id}),
    # Reuse this ID when retrying the same logical event.
    id: "msg_order-" <> to_string(order.id)
  )

AshHooks.dispatch(Order, :order_paid, event,
  enqueue: {MyApp.WebhookDeliveryWorker, :enqueue}
)

Every matching enabled endpoint gets a durable delivery row carrying the exact bytes to sign. The worker signs per Standard Webhooks (the same webhook-id on every retry), succeeds only on 2xx, never follows redirects, honors Retry-After (bounded), backs off with jitter on 5xx and transport errors, dead-letters other client errors, and durably disables the endpoint on 410. An endpoint's failure never blocks delivery to its siblings.

Without Oban, dispatch still works — every matching endpoint gets a durable :pending row — but nothing drives those rows until you define the worker (or call AshHooks.Delivery.run/2 yourself): the delivery row owns the retry policy, and the queue is only its trigger (ADR-0008).

The default HTTP adapter bounds response headers, retained body bytes, and the complete operation time. OTP's :httpc is an alternative with additional buffering limits described in its module docs. You can also provide an adapter for your application's transport requirements.

Schedule AshHooks.reconcile_pending/3 for each declaration and tenant using the same named worker callback. It repairs enqueue gaps, due retries, expired send leases, and pending endpoint disables while preserving retry times and attempt counts. Delivery rows retain their declaration and route, so one declaration cannot recover another's work. Anonymous enqueue callbacks can use a stable enqueue_key for recovery. Application nodes must keep UTC clocks synchronized for lease decisions.

Each result belongs to a live, fenced attempt. Network delivery remains at-least-once: a receiver may accept a request before its sender dies. Receivers deduplicate by the stable webhook-id.

Response bodies are never stored by default — each delivery row keeps the status and a content-type summary. For a diagnostic run, enable capture on one row. The captured body is bounded, passes built-in redaction, and is marked [captured] in the snippet:

# Diagnostic capture for one row. Supply the resource and resolver config;
# retry overrides are independent of the worker's configured settings.
AshHooks.Delivery.run(
  %{"endpoint_id" => row.endpoint_id, "event_uuid" => row.event_uuid},
  snippet_capture: true,
  deliveries: MyApp.OutboundDelivery,
  endpoints: MyApp.WebhookEndpoint,
  secret_resolver: {MyApp.Secrets, :webhook_secret},
  max_attempts: 10, base_backoff_seconds: 2,
  max_backoff_seconds: 3600, retry_after_cap_seconds: 86_400
)

The driver defaults to the retry values shown. Its max_attempts option controls the delivery row; the worker calls that option delivery_max_attempts because the worker's max_attempts controls Oban jobs. Copy the worker's retry policy for a diagnostic run: a lower delivery ceiling can dead-letter a row the worker would still retry.

Only non-terminal rows are driven — a row that already finished will not re-send; re-drive a failed one, or wait for its retry.

For a domain-specific denylist, also pass a snippet_redactor ({module, function} or fn body -> body | nil — return the redacted binary, or nil to leave the body uncaptured). It runs before the built-in redaction; a crashing redactor leaves the body uncaptured. See AshHooks.Worker for the worker-macro form.

Signing modes. :standard (default) needs only the endpoint's secret_ref. :dual and :legacy additionally require a legacy_secret_ref — :dual emits both envelopes so receivers can migrate, :legacy emits only the old one.

Fitting the extensions to your domain

The injected fields and actions carry opinionated names by default; your domain may reserve those names or own the lifecycle itself. Every knob is a per-resource DSL option, fail-closed at compile:

# your domain reserves `payload` for its own sole payload store —
# rename the ledger's exact-bytes column instead
outbound_delivery do
  payload_attribute :event_bytes
end

# append-only audit ledger: no destroy action is injected, and
# the retention hook fails loud (deletion is your own surface)
outbound_delivery do
  payload_attribute :event_bytes
  prune_action :none
end

# your register already has an enable switch — map the durable
# enable/disable onto it (the 410 breaker and operators flip your
# attribute; the package injects no `status` of its own)
endpoint do
  status_attribute :active
  enabled_values [true]
  disabled_value false
end

Ledger operations use the complete resource primary key, including custom and composite keys. Every component needs a default, a data-layer generator, or a supported create input; inbound scope fields can supply key components. Endpoint and subscription references require a single UUID-compatible key, which may be renamed or generated with uuid_v7_primary_key. A subscription register may be a closed {:array, :atom} enum (your constraints, your default — matching handles atoms and strings identically, wildcard included), and injected-PK ledgers still classify :created/:duplicate exactly.

Policies. The package injects write actions, not the authorization around them. Cover the generated machine actions with an action-specific policy of your own (the runtime's internal calls bypass policies by design), or keep the resources off every actor-facing surface — the full action set and obligation are stated at each extension site in the module docs.

Observability

Attach one handler to see the whole lifecycle — inbound verify/dedup/claim, enqueue failures, delivery attempt/result/backoff/dead-letter/endpoint-disable. Events carry ids, integers, fixed atoms, and classified reasons — never secrets, bodies, or payloads. The exact event list and a copy-paste attach_many block are in the AshHooks.Telemetry docs and the get-started tutorial.

Retention

Ledger and delivery rows accumulate by default (they are the dedup and audit record). When you want them bounded, drive the retention hooks on a schedule of your choosing (an Oban cron job, a mix task, a nightly job):

AshHooks.Ingress.prune/2 and AshHooks.Delivery.prune/2 delete terminal rows older than a cutoff — retryable and in-flight rows are never touched. They key off the resource's inserted_at, so add Ash's timestamps() to the resource and its migration. Append-only ledgers may omit the destroy action entirely (prune_action :none) — prune/2 then fails loud with a named error, and deletion is your own surface.
AshHooks.Ingress.redact_payload/5 rewrites a claimed row's payload under the claim fence (scrub sensitive fields while keeping the dedup identity; the original-bytes digest is preserved for audit).

Deleting a terminal row re-opens its dedup identity — a replayed webhook re-processes, a re-emitted outbound event re-sends — so set the TTL beyond any replay or re-emission horizon.

Multi-tenancy

If one app serves multiple organizations, ash_hooks works the way Ash does: declare attribute multitenancy on the four resources — your Subscription, Endpoint, inbound ledger, and outbound delivery ledger — and pass a :tenant to every call:

# the four resources declare the same contract
multitenancy do
  strategy :attribute
  attribute :org_id
end

# outbound: the tenant scopes the fanout, endpoint resolution, and rows
AshHooks.dispatch(Order, :order_paid, event, tenant: org.id)

# inbound: the tenant rides the request context
AshHooks.Ingress.ingest(WebhookLedger, :stripe, raw_body, %{
  signature: sig,
  tenant: org.id
})

From there the isolation is structural, not advisory: a dispatch for tenant A cannot create a delivery row for tenant B's endpoint, a subscription pointing at another tenant's endpoint id resolves to not-found and is skipped, the inbound dedup identity is per-tenant, and the claim fence and every mark are per-tenant. Touch a multitenant resource without a tenant and you get {:error, :tenant_required} before any data access; mix tenancy declarations across the resources one operation touches (say, a tenant-scoped delivery ledger beside an undeclared endpoint table) and you get {:error, :tenancy_mismatch} — both named errors, both fail-closed. Each operation checks the set it actually reads: an inbound-only app declares tenancy on its ledger alone and needs none of the outbound resources.

The async path carries its own weight: the Oban job args include the row's tenant (the worker recovers full context after any restart), the retention sweeps take a :tenant (with AshHooks.reap_all/2 / AshHooks.prune_all/2 sugar to sweep a tenant list), and AshHooks.reconcile_pending/3 repairs rows stranded between the row write and the enqueue — per tenant, with single-winner semantics. If your signing secrets are per-tenant, resolution can be too: the worker macro's tenant_aware_secrets: true (the resolver becomes f(ref, tenant)), an inbound secret fn tenant -> {:ok, secret} end, and an optional provider webhook_signing_secret(connection, tenant) callback.

Single-tenant app? Do none of this. No tenant passed means no tenant threaded — behavior is identical.

Adopting tenancy on tables that already have rows is an ordered transition (backfill, then regenerate indexes, then enable) — the adoption checklist walks it.

Security

Signatures are compared in constant time. Secret values come from your configured sources. Endpoint URLs are checked at registration and send time; the supplied adapters connect to a validated address to prevent DNS rebinding. The default adapter bounds HTTP reads, and diagnostic response capture runs through redaction before storage.

Your resource policies govern access to the ledgers. Inbound rows contain decoded provider payloads, signed-body digests, event IDs, and scope keys; outbound rows contain exact event bytes. Those payloads can contain personal information. Define read policies that deny access by default:

policies do
  # deny by default; open exactly what your app needs
  policy action_type(:read) do
    authorize_if actor_attribute_equals(:admin, true)
  end
end

This assumes Ash.Policy.Authorizer in the resource's authorizers; match the snippet to your actual actors. The policies block lives inside the ledger/delivery resource module (with authorizers: [Ash.Policy.Authorizer] in the use Ash.Resource options — Ash's policies guide covers the full model). Vulnerability reports: SECURITY.md — never a public issue.

Stability

ash_hooks follows semantic versioning over a named public surface (the DSL, the public modules, injected attributes/actions, telemetry events, error classes) — the current major is 2; breaking changes next ship in 3.0, deprecations run two minors minimum, safety corrections ship as fixes (ADR-0010).

Minimum supported versions: Elixir ~> 1.20 (OTP 28+; CI-tested on Erlang/OTP 28 and 29), Ash >= 3.34.3 and < 4.0, Oban ~> 2.20 (optional, outbound only). Development is supported on macOS and Linux; Windows developers use WSL2. CI runs on Linux, including an AshPostgres leg exercising a uuid_v7-keyed consumer shape. On Ash 3.33+ your application must also set Ash's required default_string_length_count config — an Ash requirement for every app compiling resources, not an ash_hooks one (UPGRADING.md). One nuance: a fix that closes a safety hole can change behavior in a patch release (a delivery that wrongly succeeded may now retry, for example) — such corrections are always called out under "Fixed" in the CHANGELOG. Moving off a supported version is a minor release with an UPGRADING.md note.

License

MIT.