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 at least once per delivery: durable deduplication, never a lost event, never a silent duplicate.
• Outbound — sign and deliver your own webhooks with retries, backoff, and dead-lettering, on any queue backed by Oban.

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 behaviour.
• 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: secrets are only ever resolved through your callbacks (literal secrets are rejected at compile time), endpoint URLs are checked against server-side request forgery (SSRF — a registered webhook URL can't be made to hit your internal network) at registration and again at send time, and response bodies are never stored unless you explicitly opt in for a diagnostic run. The ledgers store raw payloads, and read access to them is yours to govern — the package injects no read policies (see Security).
• Telemetry events for the whole send/receive lifecycle — structured so they can never leak secrets or payloads into your metrics backend.

Requires Elixir ~> 1.20 (OTP 28+) and Ash > 3.0. Oban (> 2.20) is needed only for outbound delivery; Phoenix or Plug only for receiving. From 1.0 the package follows semantic versioning with a named public surface — see Stability.

Installation

def deps do
  [
    {:ash_hooks, "~> 1.2"},
    # only for outbound delivery:
    {:oban, "~> 2.20"}
  ]
end

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
  # judge the row's STATUS, not just the tag: a :duplicate is re-drove
  # on redelivery, and its handler may have failed again — answer 200
  # only when the row actually finished
  {:ok, _tag, %{status: :processed}} -> 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 delivery that finished (:processed or :failed_permanent) 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 provider's event id (for action-level idempotency elsewhere in your app, our sibling package ash_onetime is an optional companion — deliberately NOT a dependency here, because it would force ash_postgres on every consumer). In short: durable deduplication, at-least-once handler invocation.

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))

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 is a small native client with every read capped, so a hostile response can't balloon worker memory; OTP's :httpc is available as an alternative, and you can inject your own adapter for tests or proxies.

Response bodies are never stored — each delivery row keeps the status and a content-type summary. When debugging a misbehaving endpoint, re-drive its row with body capture enabled and the captured body is stored only after passing the package's built-in redaction (homoglyph folding, decode-chain analysis, entropy checks — encoded secrets don't survive it), marked [captured] in the snippet:

# diagnostic: one row, one capture, floor-redacted — run/2 takes the
# same config the worker bakes (it does not recover the worker's
# settings on its own), plus the per-call capture flag:
AshHooks.Delivery.run(
  %{"endpoint_id" => row.endpoint_id, "event_uuid" => row.event_uuid},
  snippet_capture: true,
  deliveries: MyApp.Delivery,
  endpoints: MyApp.Endpoint,
  secret_resolver: {MyApp.Secrets, :webhook_secret},
  max_attempts: 10, base_backoff_seconds: 2,
  max_backoff_seconds: 3600, retry_after_cap_seconds: 86_400
)

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.

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.
• 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; ADR-0011 records the floor.

Security

The package enforces the guarantees it owns: constant-time signature compares, compile-time rejection of literal secrets, SSRF (server-side request forgery) checks at registration and send time (DNS-rebinding closed by connecting to the validated address), memory-capped HTTP reads, and redaction-gated response capture.

Read access is the one floor the package deliberately does not enforce, because it cannot know your actors: the ledger and delivery resources are YOUR resources in YOUR domain, and the package injects no policies — reads are governed entirely by the policies you write. These rows carry raw provider payloads (third-party PII), event ids, and scope keys. Mount them behind policies that deny reads 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

From 1.0.0, ash_hooks follows semantic versioning over a named public surface (the DSL, the public modules, injected attributes/actions, telemetry events, error classes) — breaking changes only in 2.0, 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 ~> 3.0, Oban ~> 2.20 (optional, outbound only). The package stays developer-portable across macOS, Linux, and Windows — a standing requirement that clone, set up, and test work on all three, kept by OS-agnostic code and tooling; CI runs the full suite on Linux. 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.

Further reading

• Get started — complete walkthrough, migrations included
• Guided tour (Livebook) — run the whole library inside one notebook
• UPGRADING.md — migration notes per release
• SECURITY.md — reporting and scope
• DSL reference
• Architecture decisions (ADRs — the reasoning behind every major design choice)
• CONTRIBUTING.md — setup, the gate suite, how to propose changes; GitHub issues are the support channel
• Full API docs at hexdocs.pm/ash_hooks

License

MIT.