ML-21: Classify API errors as transient vs permanent

backlog/tasks/ml-21 - Classify-API-errors-as-transient-vs-permanent-in-workers.md

@@ -1,15 +1,16 @@
 ---
 id: ML-21
 title: Classify API errors as transient vs permanent in workers
-status: To Do
+status: Done
 assignee: []
 created_date: '2026-04-20 08:50'
-updated_date: '2026-04-24 11:12'
+updated_date: '2026-04-24 11:59'
 labels: []
 dependencies: []
 references:
   - 'https://github.com/cloud8421/music_library/issues/158'
 priority: medium
+ordinal: 1000
 ---
 
 ## Description
@@ -69,10 +70,61 @@ Out of scope (tracked separately): honouring `Retry-After` / `X-*-Reset` headers
 
 ## Acceptance Criteria
 <!-- AC:BEGIN -->
-- [ ] #1 MusicBrainz, Discogs, Wikipedia REST v1, Brave Search, and OpenAI API modules expose a classifier that maps {status, body} to :retry or :cancel
+- [x] #1 MusicBrainz, Discogs, Wikipedia REST v1, Brave Search, and OpenAI API modules expose a classifier that maps {status, body} to :retry or :cancel
-- [ ] #2 Wikipedia Action API responses (wbgetentities, prop=extracts) decode HTTP 200 bodies containing {"error": ...} into {:error, reason} instead of {:ok, body}
+- [x] #2 Wikipedia Action API responses (wbgetentities, prop=extracts) decode HTTP 200 bodies containing {"error": ...} into {:error, reason} instead of {:ok, body}
-- [ ] #3 OpenAI classifier distinguishes rate_limit_exceeded (retry) from insufficient_quota (cancel) despite both being HTTP 429
+- [x] #3 OpenAI classifier distinguishes rate_limit_exceeded (retry) from insufficient_quota (cancel) despite both being HTTP 429
-- [ ] #4 MusicBrainz classifier treats 503 as the rate-limit signal (not 429) and is documented as such
+- [x] #4 MusicBrainz classifier treats 503 as the rate-limit signal (not 429) and is documented as such
-- [ ] #5 Workers using these APIs use the classifier to return {:snooze, n} / {:error, reason} / {:cancel, reason} instead of bubbling raw bodies
+- [x] #5 Workers using these APIs use the classifier to return {:snooze, n} / {:error, reason} / {:cancel, reason} instead of bubbling raw bodies
-- [ ] #6 Existing LastFm.API.ErrorResponse behaviour is preserved
+- [x] #6 Existing LastFm.API.ErrorResponse behaviour is preserved
 <!-- AC:END -->
+
+## Final Summary
+
+<!-- SECTION:FINAL_SUMMARY:BEGIN -->
+## Summary
+
+Introduced structured per-API `ErrorResponse` modules for MusicBrainz, Discogs, Wikipedia, Brave Search, and OpenAI, so workers can now distinguish transient failures (rate limit, 5xx, timeout) from permanent ones (4xx, not found, auth, quota). Workers emit `{:snooze, seconds}` / `{:cancel, reason}` / `{:error, reason}` instead of bubbling raw bodies. Preserved the existing `LastFm.API.ErrorResponse` behaviour and added struct-based `retryable?/1` / `retry_delay_seconds/1` helpers so Last.fm plugs into the same handler.
+
+## What changed
+
+**New shared modules**
+- `MusicLibrary.HttpError` — default HTTP status → kind mapping
+- `MusicLibrary.Worker.ErrorHandler.to_oban_result/1` — maps any known `ErrorResponse` struct to the correct Oban tuple; passes through `{:ok, _}`, `{:cancel, _}`, and atom-reason errors unchanged
+
+**New per-API `ErrorResponse` modules**
+- `MusicBrainz.API.ErrorResponse` — maps 503 to `:rate_limit` (MusicBrainz-specific; 429 is not used upstream)
+- `Discogs.API.ErrorResponse`
+- `Wikipedia.API.ErrorResponse` — has a dedicated `from_action_api_body/1` for HTTP 200 Action API error envelopes (AC2 silent-bug fix)
+- `BraveSearch.API.ErrorResponse`
+- `OpenAI.API.ErrorResponse` — disambiguates HTTP 429 between `rate_limit_exceeded` (retry) and `insufficient_quota` (cancel) via body `code` (AC3)
+
+**Updated API modules** — each now attaches a `parse_error/1` Req response step that halts with the appropriate struct on failure:
+- `MusicBrainz.API`, `Discogs.API`, `Wikipedia.API`, `BraveSearch.API`
+- `OpenAI.API` — `gpt/2`, `get_embeddings/2`, and `chat_stream/6` now return `{:error, %OpenAI.API.ErrorResponse{}}` instead of `{:error, "OpenAI API error: " <> inspect(body)}` strings
+
+**Updated workers** to route through `ErrorHandler.to_oban_result/1`, preserving existing atom-cancel branches (`:no_english_wikipedia`, `:cover_not_available`, `:image_not_found`, `:no_discogs_data`):
+- `ArtistRefreshMusicBrainzData`, `ArtistRefreshDiscogsData`, `ArtistRefreshWikipediaData`
+- `FetchArtistInfo`, `FetchArtistImage`, `FetchArtistLastFmData`
+- `RefreshCover`, `RecordRefreshMusicBrainzData`
+- `ImportFromMusicbrainzRelease`, `ImportFromMusicbrainzReleaseGroup`
+- `PopulateGenres`, `GenerateRecordEmbedding`
+
+`RefreshScrobbles` unchanged (already had its own `ErrorResponse` routing).
+
+## Notable fixes along the way
+
+- Wikipedia silent bug (AC2): Action API returned `{:ok, %{"error" => ...}}` to callers when the body contained an error envelope. Now properly surfaced as `{:error, %Wikipedia.API.ErrorResponse{}}`.
+- `lib/music_library_web/components/chat.ex`: Logger interpolation of the error reason (`"Chat streaming error: #{reason}"`) broke on struct errors. Switched to `inspect(reason)`.
+
+## Tests
+
+- New: `test/music_library/worker/error_handler_test.exs`, `test/music_brainz/api_test.exs`, `test/last_fm/api/error_response_test.exs`
+- Added per-API error-classification assertions to existing `test/open_ai/api_test.exs`, `test/brave_search/api_test.exs`, `test/discogs_test.exs`, `test/wikipedia_test.exs`
+- Updated callers that previously asserted on raw error bodies / string errors (`records_test.exs`, `similarity_test.exs`, worker tests)
+
+Full verification: `mise run dev:precommit` — shellcheck, credo --strict, sobelow, gettext, format, unused deps, 871 tests pass (43 doctests). Dialyzer shows 2 pre-existing warnings unrelated to this change.
+
+## Out of scope (ML-146)
+
+Honouring `Retry-After` / `X-*-Reset` headers to derive precise snooze durations — each ErrorResponse module currently returns fixed per-kind defaults (30–60 s).
+<!-- SECTION:FINAL_SUMMARY:END -->

lib/brave_search/api.ex

@@ -3,12 +3,13 @@ defmodule BraveSearch.API do
   Interface to the Brave Search API.
   """
 
+  alias BraveSearch.API.ErrorResponse
   alias Req.Request
 
   require Logger
 
   @spec search_images(String.t(), keyword(), BraveSearch.Config.t()) ::
-          {:ok, [map()]} | {:error, term()}
+          {:ok, [map()]} | {:error, ErrorResponse.t() | Exception.t()}
   def search_images(query, opts, config) do
     params = [q: query, count: Keyword.get(opts, :count, 20)]
 
@@ -61,16 +62,20 @@ defmodule BraveSearch.API do
     |> Request.merge_options(config.req_options)
     |> Req.RateLimiter.attach(name: :brave_search, cooldown: config.api_cooldown)
     |> Request.append_request_steps(log_attempt: &log_attempt/1)
-    |> Request.append_response_steps(log_error: &log_error/1)
+    |> Request.append_response_steps(parse_error: &parse_error/1)
   end
 
   defp get_request(request) do
     case Req.get(request) do
-      {:ok, response} when response.status == 200 ->
+      {:ok, %{status: status, body: body}} when status in 200..299 ->
-        {:ok, response.body}
+        {:ok, body}
 
-      {:ok, response} ->
+      {:ok, %{body: %ErrorResponse{} = error}} ->
-        {:error, response.body}
+        {:error, error}
+
+      # Image download path does not attach parse_error; fall back to raw body.
+      {:ok, %{body: body}} ->
+        {:error, body}
 
       error ->
         error
@@ -93,4 +98,17 @@ defmodule BraveSearch.API do
 
     {request, response}
   end
+
+  defp parse_error({request, %{status: status} = response}) when status not in 200..299 do
+    error = ErrorResponse.from_response(response)
+
+    Logger.error(fn ->
+      url = URI.to_string(request.url)
+      "Failed to fetch data from #{url}, status: #{status}, reason: #{inspect(response.body)}"
+    end)
+
+    Request.halt(request, %{response | body: error})
+  end
+
+  defp parse_error(tuple), do: tuple
 end

lib/brave_search/api/error_response.ex

@@ -0,0 +1,60 @@
+defmodule BraveSearch.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for Brave Search API calls.
+
+  Brave Search returns a consistent JSON envelope on errors:
+
+      %{"type" => "ErrorResponse",
+        "error" => %{"status" => 422, "code" => "SUBSCRIPTION_TOKEN_INVALID",
+                     "detail" => "...", "meta" => %{...}}}
+
+  HTTP status codes are the primary classifier (429 for rate limit, 5xx for
+  server errors). Brave uses 422 for most validation failures including
+  authentication errors that other APIs would return as 401.
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
+  alias MusicLibrary.HttpError
+
+  @type t :: %__MODULE__{
+          status: integer() | nil,
+          code: String.t() | nil,
+          message: String.t() | nil,
+          kind: HttpError.kind(),
+          body: term()
+        }
+
+  defstruct [:status, :code, :message, :kind, :body]
+
+  @spec from_response(Req.Response.t() | map()) :: t()
+  def from_response(%{status: status, body: body} = _response) do
+    %__MODULE__{
+      status: status,
+      code: extract_code(body),
+      message: extract_message(body),
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{kind: kind}) when kind in [:rate_limit, :server_error, :timeout],
+    do: true
+
+  def retryable?(%__MODULE__{}), do: false
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{kind: :rate_limit}), do: 60
+  def retry_delay_seconds(%__MODULE__{kind: :server_error}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :timeout}), do: 10
+  def retry_delay_seconds(%__MODULE__{}), do: 30
+
+  defp extract_code(%{"error" => %{"code" => code}}) when is_binary(code), do: code
+  defp extract_code(_), do: nil
+
+  defp extract_message(%{"error" => %{"detail" => msg}}) when is_binary(msg), do: msg
+  defp extract_message(_), do: nil
+end

lib/discogs/api.ex

@@ -3,11 +3,13 @@ defmodule Discogs.API do
   Interface to the Discogs API.
   """
 
+  alias Discogs.API.ErrorResponse
   alias Req.Request
 
   require Logger
 
-  @spec get_artist(integer() | String.t(), Discogs.Config.t()) :: {:ok, map()} | {:error, term()}
+  @spec get_artist(integer() | String.t(), Discogs.Config.t()) ::
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_artist(id, config) do
     config
     |> new_request()
@@ -41,17 +43,20 @@ defmodule Discogs.API do
     |> Request.merge_options(config.req_options)
     |> Req.RateLimiter.attach(name: :discogs, cooldown: config.api_cooldown)
     |> Request.append_request_steps(log_attempt: &log_attempt/1)
-    |> Request.append_response_steps(log_error: &log_error/1)
+    |> Request.append_response_steps(parse_error: &parse_error/1)
   end
 
   defp get_request(request) do
     case Req.get(request) do
-      {:ok, response} when response.status == 200 ->
+      {:ok, %{status: status, body: body}} when status in 200..299 ->
-        {:ok, response.body}
+        {:ok, body}
 
-      # all non-success responses can be treated as errors
+      {:ok, %{body: %ErrorResponse{} = error}} ->
-      {:ok, response} ->
+        {:error, error}
-        {:error, response.body}
+
+      # Image download path does not attach parse_error; fall back to raw body.
+      {:ok, %{body: body}} ->
+        {:error, body}
 
       error ->
         error
@@ -74,4 +79,17 @@ defmodule Discogs.API do
 
     {request, response}
   end
+
+  defp parse_error({request, %{status: status} = response}) when status not in 200..299 do
+    error = ErrorResponse.from_response(response)
+
+    Logger.error(fn ->
+      url = URI.to_string(request.url)
+      "Failed to fetch data from #{url}, status: #{status}, reason: #{inspect(response.body)}"
+    end)
+
+    Request.halt(request, %{response | body: error})
+  end
+
+  defp parse_error(tuple), do: tuple
 end

lib/discogs/api/error_response.ex

@@ -0,0 +1,48 @@
+defmodule Discogs.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for Discogs API calls.
+
+  Discogs uses HTTP status codes as the error channel. Error bodies are JSON with
+  a single `"message"` field and no application-level error codes.
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
+  alias MusicLibrary.HttpError
+
+  @type t :: %__MODULE__{
+          status: integer() | nil,
+          message: String.t() | nil,
+          kind: HttpError.kind(),
+          body: term()
+        }
+
+  defstruct [:status, :message, :kind, :body]
+
+  @spec from_response(Req.Response.t() | map()) :: t()
+  def from_response(%{status: status, body: body} = _response) do
+    %__MODULE__{
+      status: status,
+      message: extract_message(body),
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{kind: kind}) when kind in [:rate_limit, :server_error, :timeout],
+    do: true
+
+  def retryable?(%__MODULE__{}), do: false
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{kind: :rate_limit}), do: 60
+  def retry_delay_seconds(%__MODULE__{kind: :server_error}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :timeout}), do: 10
+  def retry_delay_seconds(%__MODULE__{}), do: 30
+
+  defp extract_message(%{"message" => message}) when is_binary(message), do: message
+  defp extract_message(_), do: nil
+end

lib/last_fm/api/error_response.ex

@@ -1,4 +1,19 @@
 defmodule LastFm.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for Last.fm API calls.
+
+  Last.fm is unusual among the APIs this project consumes: it returns HTTP 200
+  with a `%{"error" => N, "message" => "..."}` JSON body on failure, carrying a
+  numeric application-level error code in the range 2–29. Those codes map to
+  the atoms declared in `t:error_atom/0`.
+
+  The struct-based helpers (`retryable?/1`, `retry_delay_seconds/1`) exist to
+  share a uniform protocol with the per-HTTP-status `ErrorResponse` modules
+  used by the other APIs — see `MusicLibrary.Worker.ErrorHandler`.
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
   defstruct [:error, :message]
 
   @type error_atom ::
@@ -66,4 +81,28 @@ defmodule LastFm.API.ErrorResponse do
   # 5 seconds
   def retry_delay(:operation_failed), do: 5_000
   def retry_delay(_), do: nil
+
+  @doc """
+  Struct-based retryability predicate shared with other API `ErrorResponse` modules.
+
+  Enables `MusicLibrary.Worker.ErrorHandler.to_oban_result/1` to treat Last.fm
+  uniformly alongside HTTP-status-based errors without requiring callers to
+  unwrap the atom first.
+  """
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{error: error}), do: retryable_error?(error)
+
+  @doc """
+  Struct-based retry delay in seconds. Falls back to 30 s when the underlying
+  atom has no specific delay (see `retry_delay/1`).
+  """
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{error: error}) do
+    case retry_delay(error) do
+      ms when is_integer(ms) -> div(ms, 1000)
+      nil -> 30
+    end
+  end
 end

lib/music_brainz/api.ex

@@ -6,6 +6,7 @@ defmodule MusicBrainz.API do
   - Extend the metadata associated with existing records
   """
 
+  alias MusicBrainz.API.ErrorResponse
   alias MusicBrainz.{Artist, ReleaseGroupSearchResult, ReleaseSearchResult}
   alias Req.Request
 
@@ -207,7 +208,8 @@ defmodule MusicBrainz.API do
         }
   """
 
-  @spec get_release_group(String.t(), MusicBrainz.Config.t()) :: {:ok, map()} | {:error, term()}
+  @spec get_release_group(String.t(), MusicBrainz.Config.t()) ::
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_release_group(id, config) do
     config
     |> new_request()
@@ -284,7 +286,8 @@ defmodule MusicBrainz.API do
         "title": "Clark (Soundtrack From the Netflix Series)"
       }
   """
-  @spec get_release(String.t(), MusicBrainz.Config.t()) :: {:ok, map()} | {:error, term()}
+  @spec get_release(String.t(), MusicBrainz.Config.t()) ::
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_release(id, config) do
     config
     |> new_request()
@@ -299,7 +302,7 @@ defmodule MusicBrainz.API do
   end
 
   @spec get_releases(String.t(), keyword(), MusicBrainz.Config.t()) ::
-          {:ok, map()} | {:error, term()}
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_releases(release_group_id, opts, config) do
     Keyword.validate!(opts, [:limit, :offset])
 
@@ -320,7 +323,7 @@ defmodule MusicBrainz.API do
   end
 
   @spec search_release_by_barcode(String.t(), MusicBrainz.Config.t()) ::
-          {:ok, [ReleaseSearchResult.t()]} | {:error, term()}
+          {:ok, [ReleaseSearchResult.t()]} | {:error, ErrorResponse.t() | Exception.t()}
   def search_release_by_barcode(barcode, config) do
     config
     |> new_request()
@@ -439,7 +442,7 @@ defmodule MusicBrainz.API do
   """
   @spec search_release_group(String.t(), keyword(), MusicBrainz.Config.t()) ::
           {:ok, %{total_count: non_neg_integer(), release_groups: [ReleaseGroupSearchResult.t()]}}
-          | {:error, term()}
+          | {:error, ErrorResponse.t() | Exception.t()}
   def search_release_group(query, opts, config) do
     Keyword.validate!(opts, [:limit, :offset])
 
@@ -461,7 +464,8 @@ defmodule MusicBrainz.API do
     |> get_request()
   end
 
-  @spec get_artist(String.t(), MusicBrainz.Config.t()) :: {:ok, Artist.t()} | {:error, term()}
+  @spec get_artist(String.t(), MusicBrainz.Config.t()) ::
+          {:ok, Artist.t()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_artist(musicbrainz_id, config) do
     config
     |> new_request()
@@ -507,22 +511,41 @@ defmodule MusicBrainz.API do
     |> Request.merge_options(config.req_options)
     |> Req.RateLimiter.attach(name: :music_brainz, cooldown: config.api_cooldown)
     |> Request.append_request_steps(log_attempt: &log_attempt/1)
+    |> Request.append_response_steps(parse_error: &parse_error/1)
   end
 
   defp get_request(request) do
     case Req.get(request) do
-      {:ok, response} when response.status == 200 ->
+      {:ok, %{status: status, body: body}} when status in 200..299 ->
-        {:ok, response.body}
+        {:ok, body}
 
-      # all non-success responses can be treated as errors
+      {:ok, %{body: %ErrorResponse{} = error}} ->
-      {:ok, response} ->
+        {:error, error}
-        {:error, response.body}
+
+      # Fallback for pipelines without parse_error (e.g. cover-art binary path)
+      # or Req exhausting retries before the step runs. Callers like
+      # get_cover_art/2 normalise this further.
+      {:ok, %{body: body}} ->
+        {:error, body}
 
       error ->
         error
     end
   end
 
+  defp parse_error({request, %{status: status} = response}) when status not in 200..299 do
+    error = ErrorResponse.from_response(response)
+
+    Logger.error(fn ->
+      url = URI.to_string(request.url)
+      "Failed to fetch data from #{url}, status: #{status}, reason: #{inspect(response.body)}"
+    end)
+
+    Request.halt(request, %{response | body: error})
+  end
+
+  defp parse_error(tuple), do: tuple
+
   defp log_attempt(request) do
     url = URI.to_string(request.url)
     Logger.debug("Fetching data from #{url}")

lib/music_brainz/api/error_response.ex

@@ -0,0 +1,64 @@
+defmodule MusicBrainz.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for MusicBrainz API calls.
+
+  MusicBrainz uses classic HTTP status codes as the error channel. The body is a
+  flat JSON `{"error": "message"}` with no numeric application codes.
+
+  ## Rate limiting
+
+  MusicBrainz signals rate limiting with **HTTP 503**, not 429. The service does
+  not use 429 at all — a 503 response with a `Retry-After` header is the rate
+  limit signal. `from_response/1` therefore maps 503 to `:rate_limit` (not the
+  generic `:server_error` classification from `MusicLibrary.HttpError`).
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
+  alias MusicLibrary.HttpError
+
+  @type t :: %__MODULE__{
+          status: integer() | nil,
+          message: String.t() | nil,
+          kind: HttpError.kind(),
+          body: term()
+        }
+
+  defstruct [:status, :message, :kind, :body]
+
+  @spec from_response(Req.Response.t() | map()) :: t()
+  def from_response(%{status: 503, body: body} = _response) do
+    %__MODULE__{
+      status: 503,
+      message: extract_message(body),
+      kind: :rate_limit,
+      body: body
+    }
+  end
+
+  def from_response(%{status: status, body: body} = _response) do
+    %__MODULE__{
+      status: status,
+      message: extract_message(body),
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{kind: kind}) when kind in [:rate_limit, :server_error, :timeout],
+    do: true
+
+  def retryable?(%__MODULE__{}), do: false
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{kind: :rate_limit}), do: 60
+  def retry_delay_seconds(%__MODULE__{kind: :server_error}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :timeout}), do: 10
+  def retry_delay_seconds(%__MODULE__{}), do: 30
+
+  defp extract_message(%{"error" => message}) when is_binary(message), do: message
+  defp extract_message(_), do: nil
+end

lib/music_library/error_response.ex

@@ -0,0 +1,23 @@
+defmodule MusicLibrary.ErrorResponse do
+  @moduledoc """
+  Behaviour shared by all per-API `ErrorResponse` modules.
+
+  Each API (MusicBrainz, Discogs, Wikipedia, Brave Search, OpenAI, Last.fm)
+  returns a struct implementing this behaviour on HTTP failure, so
+  `MusicLibrary.Worker.ErrorHandler.to_oban_result/1` can dispatch uniformly
+  to produce the right Oban tuple (`{:snooze, n}` / `{:cancel, reason}`)
+  without needing to know which API raised the error.
+  """
+
+  @doc """
+  Returns `true` if the error is transient and the worker should retry after
+  a delay; `false` if the error is permanent and the worker should cancel.
+  """
+  @callback retryable?(struct()) :: boolean()
+
+  @doc """
+  Returns the snooze delay in seconds for retryable errors. Implementations
+  may return any positive integer; common defaults are 30–60 s.
+  """
+  @callback retry_delay_seconds(struct()) :: pos_integer()
+end

lib/music_library/http_error.ex

@@ -0,0 +1,37 @@
+defmodule MusicLibrary.HttpError do
+  @moduledoc """
+  Default HTTP status → error kind mapping shared by per-API `ErrorResponse` modules.
+
+  Each API's `ErrorResponse.from_response/1` uses this as a baseline before applying
+  API-specific overrides (e.g. MusicBrainz treats 503 as a rate limit, not a server
+  error; OpenAI splits HTTP 429 into `:rate_limit` vs `:auth_error` based on the
+  body `code`).
+
+  ## Kinds
+
+    * `:rate_limit`   — back off and retry (transient)
+    * `:server_error` — retry with backoff (transient)
+    * `:timeout`      — retry with shorter backoff (transient)
+    * `:auth_error`   — permanent until credentials change
+    * `:not_found`    — permanent
+    * `:client_error` — permanent (malformed request)
+    * `:unknown`      — unclassified; treated as permanent
+  """
+
+  @type kind ::
+          :rate_limit
+          | :server_error
+          | :timeout
+          | :auth_error
+          | :not_found
+          | :client_error
+          | :unknown
+
+  @spec default_kind(integer()) :: kind()
+  def default_kind(429), do: :rate_limit
+  def default_kind(status) when status in 500..599, do: :server_error
+  def default_kind(status) when status in [401, 403], do: :auth_error
+  def default_kind(404), do: :not_found
+  def default_kind(status) when status in 400..499, do: :client_error
+  def default_kind(_), do: :unknown
+end

lib/music_library/worker/artist_refresh_discogs_data.ex

@@ -1,8 +1,14 @@
 defmodule MusicLibrary.Worker.ArtistRefreshDiscogsData do
+  @moduledoc false
+
   use Oban.Worker, queue: :discogs, max_attempts: 3
 
+  alias MusicLibrary.Worker.ErrorHandler
+
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_info_id}}) do
-    MusicLibrary.Artists.refresh_discogs_data(artist_info_id)
+    artist_info_id
+    |> MusicLibrary.Artists.refresh_discogs_data()
+    |> ErrorHandler.to_oban_result()
   end
 end

lib/music_library/worker/artist_refresh_music_brainz_data.ex

@@ -1,8 +1,14 @@
 defmodule MusicLibrary.Worker.ArtistRefreshMusicBrainzData do
+  @moduledoc false
+
   use Oban.Worker, queue: :music_brainz, max_attempts: 3
 
+  alias MusicLibrary.Worker.ErrorHandler
+
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_info_id}}) do
-    MusicLibrary.Artists.refresh_musicbrainz_data(artist_info_id)
+    artist_info_id
+    |> MusicLibrary.Artists.refresh_musicbrainz_data()
+    |> ErrorHandler.to_oban_result()
   end
 end

lib/music_library/worker/artist_refresh_wikipedia_data.ex

@@ -1,11 +1,15 @@
 defmodule MusicLibrary.Worker.ArtistRefreshWikipediaData do
+  @moduledoc false
+
   use Oban.Worker, queue: :wikipedia, max_attempts: 3
 
+  alias MusicLibrary.Worker.ErrorHandler
+
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_info_id}}) do
-    with {:error, :no_english_wikipedia} <-
+    case MusicLibrary.Artists.refresh_wikipedia_data(artist_info_id) do
-           MusicLibrary.Artists.refresh_wikipedia_data(artist_info_id) do
+      {:error, :no_english_wikipedia} -> {:cancel, :no_english_wikipedia}
-      {:cancel, :no_english_wikipedia}
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/error_handler.ex

@@ -0,0 +1,49 @@
+defmodule MusicLibrary.Worker.ErrorHandler do
+  @moduledoc """
+  Converts context-layer results into Oban worker return values.
+
+  Workers call external APIs that return `{:error, %ErrorResponse{}}` (per-API
+  structs) on HTTP failures. This helper recognises any of the known error
+  response structs and translates them into the correct Oban tuple:
+
+    * retryable errors → `{:snooze, seconds}` so the attempt isn't consumed
+    * non-retryable errors → `{:cancel, reason}` so Oban stops retrying
+    * unknown `{:error, reason}` → passed through for Oban's default backoff
+
+  Workers that have app-layer atom-cancel reasons (e.g. `:no_english_wikipedia`,
+  `:cover_not_available`) must match those **before** calling this helper, since
+  atoms fall through to the generic `{:error, reason}` branch here.
+  """
+
+  @error_structs [
+    LastFm.API.ErrorResponse,
+    MusicBrainz.API.ErrorResponse,
+    Discogs.API.ErrorResponse,
+    Wikipedia.API.ErrorResponse,
+    BraveSearch.API.ErrorResponse,
+    OpenAI.API.ErrorResponse
+  ]
+
+  @type oban_result ::
+          :ok
+          | {:ok, term()}
+          | {:error, term()}
+          | {:cancel, term()}
+          | {:snooze, pos_integer()}
+
+  @spec to_oban_result(term()) :: oban_result()
+  def to_oban_result(:ok), do: :ok
+  def to_oban_result({:ok, _} = result), do: result
+
+  def to_oban_result({:error, %mod{} = response}) when mod in @error_structs do
+    if mod.retryable?(response) do
+      {:snooze, mod.retry_delay_seconds(response)}
+    else
+      {:cancel, response}
+    end
+  end
+
+  def to_oban_result({:error, reason}), do: {:error, reason}
+  def to_oban_result({:cancel, _} = result), do: result
+  def to_oban_result({:snooze, _} = result), do: result
+end

lib/music_library/worker/fetch_artist_image.ex

@@ -1,20 +1,17 @@
 defmodule MusicLibrary.Worker.FetchArtistImage do
+  @moduledoc false
+
   use Oban.Worker, queue: :heavy_writes, max_attempts: 3
 
+  alias MusicLibrary.Worker.ErrorHandler
+
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_id}}) do
     case MusicLibrary.Artists.refresh_image(artist_id) do
-      {:ok, _artist_info} ->
+      {:ok, _artist_info} -> :ok
-        :ok
+      {:error, :image_not_found} -> {:cancel, :image_not_found}
-
+      {:error, :no_discogs_data} -> {:cancel, :no_discogs_data}
-      {:error, :image_not_found} ->
+      other -> ErrorHandler.to_oban_result(other)
-        {:cancel, :image_not_found}
-
-      {:error, :no_discogs_data} ->
-        {:cancel, :no_discogs_data}
-
-      error ->
-        error
     end
   end
 end

lib/music_library/worker/fetch_artist_info.ex

@@ -1,8 +1,11 @@
 defmodule MusicLibrary.Worker.FetchArtistInfo do
+  @moduledoc false
+
   use Oban.Worker, queue: :default, max_attempts: 3
 
   alias MusicLibrary.Artists
   alias MusicLibrary.Records.Similarity
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_id}}) do
@@ -13,7 +16,7 @@ defmodule MusicLibrary.Worker.FetchArtistInfo do
       Similarity.regenerate_artist_embeddings(artist_id)
     else
       {:error, :no_english_wikipedia} -> {:cancel, :no_english_wikipedia}
-      error -> error
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/fetch_artist_last_fm_data.ex

@@ -1,11 +1,15 @@
 defmodule MusicLibrary.Worker.FetchArtistLastFmData do
+  @moduledoc false
+
   use Oban.Worker, queue: :last_fm, max_attempts: 3
 
+  alias MusicLibrary.Worker.ErrorHandler
+
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => artist_id}}) do
     case MusicLibrary.Artists.refresh_lastfm_data(artist_id) do
       {:ok, _artist_info} -> :ok
-      error -> error
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/generate_record_embedding.ex

@@ -1,8 +1,11 @@
 defmodule MusicLibrary.Worker.GenerateRecordEmbedding do
+  @moduledoc false
+
   use Oban.Worker, queue: :heavy_writes, max_attempts: 3
 
   alias MusicLibrary.Records
   alias MusicLibrary.Records.Similarity
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"record_id" => record_id}}) do
@@ -11,7 +14,7 @@ defmodule MusicLibrary.Worker.GenerateRecordEmbedding do
     case Similarity.generate_embedding(record) do
       :noop -> :ok
       {:ok, _} -> Records.notify_update(record)
-      {:error, _} = error -> error
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/import_from_musicbrainz_release.ex

@@ -8,6 +8,7 @@ defmodule MusicLibrary.Worker.ImportFromMusicbrainzRelease do
   use Oban.Worker, queue: :music_brainz, max_attempts: 3
 
   alias MusicLibrary.Records.Record
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"release_id" => release_id} = args}) do
@@ -19,7 +20,7 @@ defmodule MusicLibrary.Worker.ImportFromMusicbrainzRelease do
 
     case MusicLibrary.Records.import_from_musicbrainz_release(release_id, opts) do
       {:ok, _record} -> :ok
-      {:error, reason} -> {:error, reason}
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/import_from_musicbrainz_release_group.ex

@@ -13,6 +13,7 @@ defmodule MusicLibrary.Worker.ImportFromMusicbrainzReleaseGroup do
 
   alias MusicLibrary.Records
   alias MusicLibrary.Records.Record
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"release_group_id" => release_group_id} = args}) do
@@ -23,7 +24,7 @@ defmodule MusicLibrary.Worker.ImportFromMusicbrainzReleaseGroup do
 
     case Records.import_from_musicbrainz_release_group(release_group_id, opts) do
       {:ok, _record} -> :ok
-      {:error, reason} -> {:error, reason}
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/populate_genres.ex

@@ -1,16 +1,20 @@
 defmodule MusicLibrary.Worker.PopulateGenres do
+  @moduledoc false
+
   use Oban.Worker, queue: :heavy_writes, max_attempts: 10
 
   alias MusicLibrary.Records
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => record_id}}) do
     record = Records.get_record!(record_id)
 
     with {:ok, updated_record} <- Records.populate_genres(record),
-         {:ok, _worker} <-
+         {:ok, _worker} <- Records.Similarity.generate_embedding_async(updated_record) do
-           Records.Similarity.generate_embedding_async(updated_record) do
       Records.notify_update(updated_record)
+    else
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/record_refresh_music_brainz_data.ex

@@ -1,14 +1,18 @@
 defmodule MusicLibrary.Worker.RecordRefreshMusicBrainzData do
+  @moduledoc false
+
   use Oban.Worker, queue: :music_brainz, max_attempts: 3
 
   alias MusicLibrary.Records
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => record_id}}) do
     record = Records.get_record!(record_id)
 
-    with {:ok, updated_record} <- Records.refresh_musicbrainz_data(record) do
+    case Records.refresh_musicbrainz_data(record) do
-      Records.notify_update(updated_record)
+      {:ok, updated_record} -> Records.notify_update(updated_record)
+      other -> ErrorHandler.to_oban_result(other)
     end
   end
 end

lib/music_library/worker/refresh_cover.ex

@@ -1,21 +1,19 @@
 defmodule MusicLibrary.Worker.RefreshCover do
+  @moduledoc false
+
   use Oban.Worker, queue: :heavy_writes, max_attempts: 3
 
   alias MusicLibrary.Records
+  alias MusicLibrary.Worker.ErrorHandler
 
   @impl Oban.Worker
   def perform(%Oban.Job{args: %{"id" => record_id}}) do
     record = Records.get_record!(record_id)
 
     case Records.refresh_cover(record) do
-      {:ok, updated_record} ->
+      {:ok, updated_record} -> Records.notify_update(updated_record)
-        Records.notify_update(updated_record)
+      {:error, :cover_not_available} -> {:cancel, :cover_not_available}
-
+      other -> ErrorHandler.to_oban_result(other)
-      {:error, :cover_not_available} ->
-        {:cancel, :cover_not_available}
-
-      error ->
-        error
     end
   end
 end

lib/music_library_web/components/chat.ex

@@ -399,7 +399,7 @@ defmodule MusicLibraryWeb.Components.Chat do
           )
 
         {:error, reason} ->
-          Logger.error("Chat streaming error: #{reason}")
+          Logger.error("Chat streaming error: #{inspect(reason)}")
 
           LiveView.send_update(parent_pid, __MODULE__,
             id: component_id,

lib/open_ai.ex

@@ -17,7 +17,7 @@ defmodule OpenAI do
     API.gpt(completion, config())
   end
 
-  @spec chat_stream([map()], chat_stream_opts()) :: :ok | {:error, String.t()}
+  @spec chat_stream([map()], chat_stream_opts()) :: :ok | {:error, term()}
   def chat_stream(messages, opts \\ []) when is_list(messages) do
     model = Keyword.get(opts, :model, "gpt-4.1")
     temperature = Keyword.get(opts, :temperature, 0.7)

lib/open_ai/api.ex

@@ -1,13 +1,23 @@
 defmodule OpenAI.API do
   @moduledoc """
   Low-level HTTP client for the OpenAI API (chat completions, streaming responses, embeddings).
+
+  HTTP errors are returned as `OpenAI.API.ErrorResponse` structs, which classify
+  rate-limit failures (retryable) separately from billing-quota failures
+  (`insufficient_quota` — non-retryable) despite both using HTTP 429.
+
+  Mid-stream failures in `chat_stream/6` come from parsed SSE event payloads
+  rather than HTTP responses and remain as message strings passed through the
+  stream callback plumbing.
   """
 
+  alias OpenAI.API.ErrorResponse
   alias Req.Request
 
   require Logger
 
-  @spec gpt(OpenAI.Completion.t(), OpenAI.Config.t()) :: {:ok, map()} | {:error, term()}
+  @spec gpt(OpenAI.Completion.t(), OpenAI.Config.t()) ::
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def gpt(completion, config) do
     case config
          |> new_request()
@@ -27,17 +37,17 @@ defmodule OpenAI.API do
         content = get_in(body, ["choices", Access.at(0), "message", "content"])
         JSON.decode(content)
 
-      {:ok, %{body: body}} ->
+      {:ok, response} ->
-        {:error, "OpenAI API error: #{inspect(body)}"}
+        {:error, ErrorResponse.from_response(response)}
 
       {:error, exception} ->
-        {:error, "Connection error: #{Exception.message(exception)}"}
+        {:error, exception}
     end
   end
 
   @spec chat_stream([map()], String.t(), String.t(), float(), OpenAI.Config.t(), (String.t() ->
                                                                                     any())) ::
-          :ok | {:error, String.t()}
+          :ok | {:error, ErrorResponse.t() | Exception.t() | String.t()}
   def chat_stream(messages, instructions, model, temperature, config, cb) do
     config
     |> new_request()
@@ -86,18 +96,18 @@ defmodule OpenAI.API do
           :ok
         end
 
-      {:ok, %{body: body}} ->
+      {:ok, response} ->
-        {:error, "OpenAI API error: #{inspect(body)}"}
+        {:error, ErrorResponse.from_response(response)}
 
       {:error, exception} ->
-        {:error, "Connection error: #{Exception.message(exception)}"}
+        {:error, exception}
     end
   end
 
-  @spec get_embeddings(String.t(), OpenAI.Config.t()) :: {:ok, [float()]} | {:error, term()}
+  @spec get_embeddings(String.t(), OpenAI.Config.t()) ::
+          {:ok, [float()]} | {:error, ErrorResponse.t() | Exception.t()}
   def get_embeddings(text, config) do
-    resp =
+    case config
-      config
          |> new_request()
          |> Req.merge(
            url: "/v1/embeddings",
@@ -106,13 +116,15 @@ defmodule OpenAI.API do
              model: "text-embedding-3-small"
            }
          )
-      |> Req.post!()
+         |> Req.post() do
+      {:ok, %{status: 200, body: body}} ->
+        {:ok, get_in(body, ["data", Access.at(0), "embedding"])}
 
-    if resp.status == 200 do
+      {:ok, response} ->
-      embeddings = get_in(resp.body, ["data", Access.at(0), "embedding"])
+        {:error, ErrorResponse.from_response(response)}
-      {:ok, embeddings}
+
-    else
+      {:error, exception} ->
-      {:error, resp.body}
+        {:error, exception}
     end
   end
 

lib/open_ai/api/error_response.ex

@@ -0,0 +1,83 @@
+defmodule OpenAI.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for OpenAI API calls.
+
+  OpenAI uses HTTP status codes as the primary classifier, with one critical
+  body-peek exception:
+
+    * HTTP **429** covers both `code: "rate_limit_exceeded"` (transient — retry)
+      and `code: "insufficient_quota"` (permanent — billing failure, cancel).
+      `from_response/1` reads `error.code` to disambiguate.
+
+  Error body shape (non-streaming):
+
+      %{"error" => %{"message" => ..., "type" => ..., "code" => ..., "param" => ...}}
+
+  Mid-stream SSE `error` / `response.failed` events are handled separately in
+  `OpenAI.API.chat_stream/6` and do not flow through this module.
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
+  alias MusicLibrary.HttpError
+
+  @type t :: %__MODULE__{
+          status: integer() | nil,
+          code: String.t() | nil,
+          type: String.t() | nil,
+          message: String.t() | nil,
+          kind: HttpError.kind(),
+          body: term()
+        }
+
+  defstruct [:status, :code, :type, :message, :kind, :body]
+
+  @spec from_response(Req.Response.t() | map()) :: t()
+  def from_response(%{status: 429, body: %{"error" => %{"code" => "insufficient_quota"} = e}} = r) do
+    %__MODULE__{
+      status: 429,
+      code: "insufficient_quota",
+      type: e["type"],
+      message: e["message"],
+      kind: :auth_error,
+      body: r.body
+    }
+  end
+
+  def from_response(%{status: status, body: %{"error" => err} = body} = _response)
+      when is_map(err) do
+    %__MODULE__{
+      status: status,
+      code: err["code"],
+      type: err["type"],
+      message: err["message"],
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  def from_response(%{status: status, body: body} = _response) do
+    %__MODULE__{
+      status: status,
+      code: nil,
+      type: nil,
+      message: nil,
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{kind: kind}) when kind in [:rate_limit, :server_error, :timeout],
+    do: true
+
+  def retryable?(%__MODULE__{}), do: false
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{kind: :rate_limit}), do: 60
+  def retry_delay_seconds(%__MODULE__{kind: :server_error}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :timeout}), do: 10
+  def retry_delay_seconds(%__MODULE__{}), do: 30
+end

lib/wikipedia/api.ex

@@ -1,14 +1,23 @@
 defmodule Wikipedia.API do
   @moduledoc """
   Interface to the Wikidata and Wikipedia APIs.
+
+  Two API surfaces are used, with different error handling:
+
+    * **Action API** (`/w/api.php`) returns HTTP 200 with an `{"error": {...}}`
+      payload on failure. `parse_error/1` promotes those bodies into
+      `Wikipedia.API.ErrorResponse` errors.
+    * **REST v1 API** (`/api/rest_v1/page/summary/:title`) uses classic HTTP
+      status codes. `parse_error/1` handles non-2xx statuses the same way.
   """
 
   alias Req.Request
+  alias Wikipedia.API.ErrorResponse
 
   require Logger
 
   @spec get_wikipedia_title(String.t(), Wikipedia.Config.t()) ::
-          {:ok, String.t() | nil} | {:error, term()}
+          {:ok, String.t() | nil} | {:error, ErrorResponse.t() | Exception.t()}
   def get_wikipedia_title(wikidata_id, config) do
     case config
          |> new_request("https://www.wikidata.org")
@@ -32,7 +41,8 @@ defmodule Wikipedia.API do
     end
   end
 
-  @spec get_article_summary(String.t(), Wikipedia.Config.t()) :: {:ok, map()} | {:error, term()}
+  @spec get_article_summary(String.t(), Wikipedia.Config.t()) ::
+          {:ok, map()} | {:error, ErrorResponse.t() | Exception.t()}
   def get_article_summary(title, config) do
     config
     |> new_request("https://en.wikipedia.org")
@@ -41,7 +51,7 @@ defmodule Wikipedia.API do
   end
 
   @spec get_article_extract(String.t(), Wikipedia.Config.t()) ::
-          {:ok, String.t() | nil} | {:error, term()}
+          {:ok, String.t() | nil} | {:error, ErrorResponse.t() | Exception.t()}
   def get_article_extract(title, config) do
     case config
          |> new_request("https://en.wikipedia.org")
@@ -80,16 +90,22 @@ defmodule Wikipedia.API do
     |> Request.merge_options(config.req_options)
     |> Req.RateLimiter.attach(name: :wikipedia, cooldown: config.api_cooldown)
     |> Request.append_request_steps(log_attempt: &log_attempt/1)
-    |> Request.append_response_steps(log_error: &log_error/1)
+    |> Request.append_response_steps(parse_error: &parse_error/1)
   end
 
   defp get_request(request) do
     case Req.get(request) do
-      {:ok, response} when response.status == 200 ->
+      {:ok, %{status: 200, body: %ErrorResponse{} = error}} ->
-        {:ok, response.body}
+        {:error, error}
 
-      {:ok, response} ->
+      {:ok, %{status: status, body: body}} when status in 200..299 ->
-        {:error, response.body}
+        {:ok, body}
+
+      {:ok, %{body: %ErrorResponse{} = error}} ->
+        {:error, error}
+
+      {:ok, %{body: body}} ->
+        {:error, body}
 
       error ->
         error
@@ -102,14 +118,32 @@ defmodule Wikipedia.API do
     request
   end
 
-  defp log_error({request, response}) do
+  # Action API: HTTP 200 with an error envelope in the body.
-    if response.status in 400..499 or response.status in 500..599 do
+  defp parse_error(
+         {request,
+          %{status: 200, body: %{"error" => %{"code" => _, "info" => _}} = body} = response}
+       ) do
+    error = ErrorResponse.from_action_api_body(body)
+
     Logger.error(fn ->
       url = URI.to_string(request.url)
-        "Failed to fetch data from #{url}, reason: #{inspect(response.body)}"
+      "Failed to fetch data from #{url}, code: #{error.code}, info: #{error.message}"
     end)
+
+    Request.halt(request, %{response | body: error})
   end
 
-    {request, response}
+  # REST v1 API and other non-2xx responses.
+  defp parse_error({request, %{status: status} = response}) when status not in 200..299 do
+    error = ErrorResponse.from_response(response)
+
+    Logger.error(fn ->
+      url = URI.to_string(request.url)
+      "Failed to fetch data from #{url}, status: #{status}, reason: #{inspect(response.body)}"
+    end)
+
+    Request.halt(request, %{response | body: error})
   end
+
+  defp parse_error(tuple), do: tuple
 end

lib/wikipedia/api/error_response.ex

@@ -0,0 +1,94 @@
+defmodule Wikipedia.API.ErrorResponse do
+  @moduledoc """
+  Structured error response for Wikipedia/Wikidata API calls.
+
+  Wikipedia exposes two API surfaces with different error paradigms:
+
+    * **Action API** (`/w/api.php`, used by `get_wikipedia_title/2` and
+      `get_article_extract/2`) returns **HTTP 200 with an error in the body**:
+      `%{"error" => %{"code" => ..., "info" => ...}}`. Without body inspection
+      the error would be silently returned as `{:ok, body}`.
+
+    * **REST v1 API** (`/api/rest_v1/page/summary/:title`, used by
+      `get_article_summary/2`) uses classic HTTP status codes.
+
+  `from_response/1` handles both paths. `from_action_api_body/1` is a dedicated
+  entry point for the HTTP 200 + body-error case.
+
+  ## Non-error body shapes
+
+  These are NOT errors and should continue to return `{:ok, body}`:
+
+    * `%{"entities" => %{_id => %{"missing" => ""}}}` from `wbgetentities` —
+      the facade converts this to `{:error, :no_english_wikipedia}` at a
+      higher level.
+    * `get_in(body, ["query", "pages", _, "extract"])` returning `nil` — a
+      title exists but has no extract.
+
+  The discriminator for an actual error is the literal top-level `"error"` key
+  whose value is a map with both `"code"` and `"info"`.
+  """
+
+  @behaviour MusicLibrary.ErrorResponse
+
+  alias MusicLibrary.HttpError
+
+  @type t :: %__MODULE__{
+          status: integer() | nil,
+          code: String.t() | nil,
+          message: String.t() | nil,
+          kind: HttpError.kind(),
+          body: term()
+        }
+
+  defstruct [:status, :code, :message, :kind, :body]
+
+  @spec from_response(Req.Response.t() | map()) :: t()
+  def from_response(%{status: status, body: body} = _response) do
+    %__MODULE__{
+      status: status,
+      code: extract_rest_code(body),
+      message: extract_rest_message(body),
+      kind: HttpError.default_kind(status),
+      body: body
+    }
+  end
+
+  @spec from_action_api_body(map()) :: t()
+  def from_action_api_body(%{"error" => %{"code" => code, "info" => info}} = body) do
+    %__MODULE__{
+      status: 200,
+      code: code,
+      message: info,
+      kind: action_api_kind(code),
+      body: body
+    }
+  end
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retryable?(t()) :: boolean()
+  def retryable?(%__MODULE__{kind: kind}) when kind in [:rate_limit, :server_error, :timeout],
+    do: true
+
+  def retryable?(%__MODULE__{}), do: false
+
+  @impl MusicLibrary.ErrorResponse
+  @spec retry_delay_seconds(t()) :: pos_integer()
+  def retry_delay_seconds(%__MODULE__{kind: :rate_limit}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :server_error}), do: 30
+  def retry_delay_seconds(%__MODULE__{kind: :timeout}), do: 10
+  def retry_delay_seconds(%__MODULE__{}), do: 30
+
+  defp action_api_kind("ratelimited"), do: :rate_limit
+  defp action_api_kind("maxlag"), do: :rate_limit
+  defp action_api_kind("readonly"), do: :server_error
+  defp action_api_kind("internal_api_error_" <> _), do: :server_error
+  defp action_api_kind(_), do: :client_error
+
+  defp extract_rest_code(%{"httpCode" => code}), do: to_string(code)
+  defp extract_rest_code(_), do: nil
+
+  defp extract_rest_message(%{"messageTranslations" => %{"en" => msg}}), do: msg
+  defp extract_rest_message(%{"detail" => msg}) when is_binary(msg), do: msg
+  defp extract_rest_message(_), do: nil
+end

test/brave_search/api_test.exs

@@ -2,6 +2,7 @@ defmodule BraveSearch.APITest do
   use ExUnit.Case, async: true
 
   alias BraveSearch.API
+  alias BraveSearch.API.ErrorResponse
 
   @config %BraveSearch.Config{
     api_key: "test_key",
@@ -52,14 +53,44 @@ defmodule BraveSearch.APITest do
     end
 
     @tag :capture_log
-    test "returns error on non-200 response" do
+    test "returns a rate-limit ErrorResponse on 429" do
       Req.Test.stub(__MODULE__, fn conn ->
         conn
         |> Plug.Conn.put_status(429)
-        |> Req.Test.json(%{"error" => "rate limited"})
+        |> Req.Test.json(%{
+          "type" => "ErrorResponse",
+          "error" => %{"status" => 429, "code" => "RATE_LIMITED", "detail" => "Too many requests"}
+        })
       end)
 
-      assert {:error, %{"error" => "rate limited"}} = API.search_images("test", [], @config)
+      assert {:error, %ErrorResponse{} = err} =
+               API.search_images("test", [], @config)
+
+      assert err.status == 429
+      assert err.code == "RATE_LIMITED"
+      assert err.kind == :rate_limit
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "returns a client-error ErrorResponse on 422" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(422)
+        |> Req.Test.json(%{
+          "type" => "ErrorResponse",
+          "error" => %{
+            "status" => 422,
+            "code" => "SUBSCRIPTION_TOKEN_INVALID",
+            "detail" => "The provided subscription token is invalid."
+          }
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :client_error} = err} =
+               API.search_images("test", [], @config)
+
+      refute ErrorResponse.retryable?(err)
     end
 
     test "passes count option as query param" do

test/brave_search_test.exs

@@ -76,14 +76,14 @@ defmodule BraveSearchTest do
     end
 
     @tag :capture_log
-    test "returns the decoded error body on non-200 responses" do
+    test "returns an ErrorResponse struct on non-200 responses" do
       Req.Test.stub(BraveSearch.API, fn conn ->
         conn
         |> Plug.Conn.put_resp_content_type("application/json")
         |> Plug.Conn.send_resp(429, ~s({"error":"rate_limited"}))
       end)
 
-      assert {:error, %{"error" => "rate_limited"}} =
+      assert {:error, %BraveSearch.API.ErrorResponse{status: 429, kind: :rate_limit}} =
                BraveSearch.search_images("too many requests")
     end
   end

test/discogs_test.exs

@@ -1,6 +1,7 @@
 defmodule DiscogsTest do
   use ExUnit.Case, async: true
 
+  alias Discogs.API.ErrorResponse
   alias Discogs.Fixtures
 
   describe "get_artist/1" do
@@ -16,5 +17,47 @@ defmodule DiscogsTest do
 
       assert {:ok, expected_info} == Discogs.get_artist(discogs_id)
     end
+
+    @tag :capture_log
+    test "returns a rate-limit ErrorResponse on 429" do
+      Req.Test.stub(Discogs.API, fn conn ->
+        conn
+        |> Plug.Conn.put_status(429)
+        |> Req.Test.json(%{"message" => "You are making requests too quickly."})
+      end)
+
+      assert {:error, %ErrorResponse{} = err} = Discogs.get_artist("1")
+      assert err.status == 429
+      assert err.kind == :rate_limit
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "returns a not-found ErrorResponse on 404" do
+      Req.Test.stub(Discogs.API, fn conn ->
+        conn
+        |> Plug.Conn.put_status(404)
+        |> Req.Test.json(%{"message" => "The requested resource was not found."})
+      end)
+
+      assert {:error, %ErrorResponse{kind: :not_found} = err} =
+               Discogs.get_artist("1")
+
+      refute ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "returns a server-error ErrorResponse on 500" do
+      Req.Test.stub(Discogs.API, fn conn ->
+        conn
+        |> Plug.Conn.put_status(500)
+        |> Req.Test.json(%{"message" => "Internal server error"})
+      end)
+
+      assert {:error, %ErrorResponse{kind: :server_error} = err} =
+               Discogs.get_artist("1")
+
+      assert ErrorResponse.retryable?(err)
+    end
   end
 end

test/last_fm/api/error_response_test.exs

@@ -0,0 +1,62 @@
+defmodule LastFm.API.ErrorResponseTest do
+  use ExUnit.Case, async: true
+
+  alias LastFm.API.ErrorResponse
+
+  describe "new/2" do
+    test "maps known error codes to atoms" do
+      assert %ErrorResponse{error: :rate_limit_exceeded, message: "limit"} =
+               ErrorResponse.new(29, "limit")
+
+      assert %ErrorResponse{error: :invalid_session_key} = ErrorResponse.new(9, "bad")
+      assert %ErrorResponse{error: :service_offline} = ErrorResponse.new(11, "down")
+    end
+  end
+
+  describe "retryable_error?/1" do
+    test "returns true for transient error atoms" do
+      for atom <- [:transient_error, :service_offline, :rate_limit_exceeded, :operation_failed] do
+        assert ErrorResponse.retryable_error?(atom), "expected #{atom} to be retryable"
+      end
+    end
+
+    test "returns false for permanent error atoms" do
+      for atom <- [:invalid_session_key, :invalid_api_key, :authentication_failed] do
+        refute ErrorResponse.retryable_error?(atom), "expected #{atom} to be non-retryable"
+      end
+    end
+  end
+
+  describe "retry_delay/1 and retry_delay_seconds/1" do
+    test "returns a delay in milliseconds for retryable atoms" do
+      assert ErrorResponse.retry_delay(:rate_limit_exceeded) == 60_000
+      assert ErrorResponse.retry_delay(:service_offline) == 30_000
+      assert ErrorResponse.retry_delay(:transient_error) == 5_000
+      assert ErrorResponse.retry_delay(:operation_failed) == 5_000
+    end
+
+    test "returns nil for non-retryable atoms" do
+      for atom <- [:invalid_session_key, :invalid_api_key, :authentication_failed] do
+        assert ErrorResponse.retry_delay(atom) == nil,
+               "expected #{atom} to have no retry delay"
+      end
+    end
+
+    test "struct-based retry_delay_seconds/1 returns the delay in seconds" do
+      assert ErrorResponse.retry_delay_seconds(ErrorResponse.new(29, "")) == 60
+      assert ErrorResponse.retry_delay_seconds(ErrorResponse.new(11, "")) == 30
+      assert ErrorResponse.retry_delay_seconds(ErrorResponse.new(16, "")) == 5
+    end
+
+    test "struct-based retry_delay_seconds/1 falls back to 30 s for non-retryable atoms" do
+      assert ErrorResponse.retry_delay_seconds(ErrorResponse.new(9, "")) == 30
+    end
+  end
+
+  describe "retryable?/1 (struct-based)" do
+    test "agrees with retryable_error?/1" do
+      assert ErrorResponse.retryable?(ErrorResponse.new(29, ""))
+      refute ErrorResponse.retryable?(ErrorResponse.new(9, ""))
+    end
+  end
+end

test/music_brainz/api_test.exs

@@ -0,0 +1,85 @@
+defmodule MusicBrainz.APITest do
+  use ExUnit.Case, async: true
+
+  alias MusicBrainz.API
+
+  @config %MusicBrainz.Config{
+    user_agent: "test_agent",
+    req_options: [plug: {Req.Test, __MODULE__}, max_retries: 0],
+    api_cooldown: 0
+  }
+
+  describe "error classification" do
+    @tag :capture_log
+    test "503 is treated as :rate_limit (MusicBrainz's rate-limit signal, not 429)" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(503)
+        |> Req.Test.json(%{"error" => "Your requests are exceeding the allowable rate limit."})
+      end)
+
+      assert {:error, %API.ErrorResponse{} = err} =
+               API.get_artist("mbid", @config)
+
+      assert err.status == 503
+      assert err.kind == :rate_limit
+      assert API.ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "429 falls under :rate_limit via default mapping (even though MusicBrainz doesn't use it)" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(429)
+        |> Req.Test.json(%{"error" => "Too many requests"})
+      end)
+
+      assert {:error, %API.ErrorResponse{kind: :rate_limit} = err} =
+               API.get_artist("mbid", @config)
+
+      assert API.ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "404 is treated as :not_found and is not retryable" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(404)
+        |> Req.Test.json(%{"error" => "Not Found"})
+      end)
+
+      assert {:error, %API.ErrorResponse{kind: :not_found} = err} =
+               API.get_artist("mbid", @config)
+
+      refute API.ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "400 is treated as :client_error and is not retryable" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(400)
+        |> Req.Test.json(%{"error" => "Invalid MBID"})
+      end)
+
+      assert {:error, %API.ErrorResponse{kind: :client_error} = err} =
+               API.get_artist("mbid", @config)
+
+      refute API.ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "500 is treated as :server_error and is retryable" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(500)
+        |> Req.Test.json(%{"error" => "Internal Server Error"})
+      end)
+
+      assert {:error, %API.ErrorResponse{kind: :server_error} = err} =
+               API.get_artist("mbid", @config)
+
+      assert API.ErrorResponse.retryable?(err)
+    end
+  end
+end

test/music_brainz_test.exs

@@ -121,6 +121,7 @@ defmodule MusicBrainzTest do
       assert releases == []
     end
 
+    @tag :capture_log
     test "error on a later page is returned immediately" do
       full_page = release_page(100)
 
@@ -138,7 +139,7 @@ defmodule MusicBrainzTest do
         end
       end)
 
-      assert {:error, %{"error" => "service unavailable"}} =
+      assert {:error, %MusicBrainz.API.ErrorResponse{status: 503, kind: :rate_limit}} =
                MusicBrainz.get_all_releases(@release_group_id)
     end
   end

test/music_library/records/similarity_test.exs

@@ -298,10 +298,10 @@ defmodule MusicLibrary.Records.SimilarityTest do
       Req.Test.stub(OpenAI.API, fn conn ->
         conn
         |> Plug.Conn.put_status(500)
-        |> Req.Test.json(%{"error" => "internal server error"})
+        |> Req.Test.json(%{"error" => %{"message" => "internal server error"}})
       end)
 
-      assert {:error, %{"error" => "internal server error"}} =
+      assert {:error, %OpenAI.API.ErrorResponse{status: 500, kind: :server_error}} =
                Similarity.generate_embedding(record)
     end
   end

test/music_library/records_test.exs

@@ -383,10 +383,15 @@ defmodule MusicLibrary.RecordsTest do
       record = record(%{genres: []})
 
       Req.Test.stub(OpenAI.API, fn conn ->
-        Plug.Conn.send_resp(conn, 500, JSON.encode!(%{"error" => "internal server error"}))
+        Plug.Conn.send_resp(
+          conn,
+          500,
+          JSON.encode!(%{"error" => %{"message" => "internal server error"}})
+        )
       end)
 
-      assert {:error, "OpenAI API error:" <> _} = Records.populate_genres(record)
+      assert {:error, %OpenAI.API.ErrorResponse{status: 500, kind: :server_error}} =
+               Records.populate_genres(record)
     end
   end
 end

test/music_library/worker/error_handler_test.exs

@@ -0,0 +1,97 @@
+defmodule MusicLibrary.Worker.ErrorHandlerTest do
+  use ExUnit.Case, async: true
+
+  alias MusicLibrary.Worker.ErrorHandler
+
+  describe "to_oban_result/1 — successes" do
+    test ":ok passes through" do
+      assert ErrorHandler.to_oban_result(:ok) == :ok
+    end
+
+    test "{:ok, result} passes through unchanged" do
+      payload = %{anything: true}
+      assert ErrorHandler.to_oban_result({:ok, payload}) == {:ok, payload}
+    end
+  end
+
+  describe "to_oban_result/1 — retryable ErrorResponses snooze" do
+    test "MusicBrainz 503 → {:snooze, 60}" do
+      err =
+        MusicBrainz.API.ErrorResponse.from_response(%{status: 503, body: %{"error" => "rate"}})
+
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 60}
+    end
+
+    test "Discogs 429 → {:snooze, 60}" do
+      err = Discogs.API.ErrorResponse.from_response(%{status: 429, body: %{"message" => "rate"}})
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 60}
+    end
+
+    test "Wikipedia Action API ratelimited → {:snooze, 30}" do
+      err =
+        Wikipedia.API.ErrorResponse.from_action_api_body(%{
+          "error" => %{"code" => "ratelimited", "info" => "rate"}
+        })
+
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 30}
+    end
+
+    test "BraveSearch 500 → {:snooze, 30}" do
+      err =
+        BraveSearch.API.ErrorResponse.from_response(%{status: 500, body: %{"error" => "boom"}})
+
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 30}
+    end
+
+    test "OpenAI 429 rate_limit_exceeded → {:snooze, 60}" do
+      err =
+        OpenAI.API.ErrorResponse.from_response(%{
+          status: 429,
+          body: %{"error" => %{"code" => "rate_limit_exceeded", "message" => "rate"}}
+        })
+
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 60}
+    end
+
+    test "Last.fm rate_limit_exceeded → {:snooze, 60}" do
+      err = LastFm.API.ErrorResponse.new(29, "Rate limit exceeded")
+      assert ErrorHandler.to_oban_result({:error, err}) == {:snooze, 60}
+    end
+  end
+
+  describe "to_oban_result/1 — non-retryable ErrorResponses cancel" do
+    test "MusicBrainz 404 → {:cancel, err}" do
+      err = MusicBrainz.API.ErrorResponse.from_response(%{status: 404, body: %{"error" => "nf"}})
+      assert {:cancel, ^err} = ErrorHandler.to_oban_result({:error, err})
+    end
+
+    test "OpenAI 429 insufficient_quota → {:cancel, err}" do
+      err =
+        OpenAI.API.ErrorResponse.from_response(%{
+          status: 429,
+          body: %{"error" => %{"code" => "insufficient_quota", "message" => "quota"}}
+        })
+
+      assert {:cancel, ^err} = ErrorHandler.to_oban_result({:error, err})
+    end
+
+    test "Last.fm invalid_session_key → {:cancel, err}" do
+      err = LastFm.API.ErrorResponse.new(9, "Invalid session key")
+      assert {:cancel, ^err} = ErrorHandler.to_oban_result({:error, err})
+    end
+  end
+
+  describe "to_oban_result/1 — passthrough for unstructured errors and atom reasons" do
+    test "unknown {:error, reason} passes through unchanged" do
+      assert ErrorHandler.to_oban_result({:error, :something_else}) == {:error, :something_else}
+    end
+
+    test "{:cancel, _} passes through unchanged" do
+      assert ErrorHandler.to_oban_result({:cancel, :manual}) == {:cancel, :manual}
+    end
+
+    test "{:snooze, _} passes through unchanged" do
+      assert ErrorHandler.to_oban_result({:snooze, 120}) == {:snooze, 120}
+    end
+  end
+end

test/music_library/worker/generate_record_embedding_test.exs

@@ -44,16 +44,16 @@ defmodule MusicLibrary.Worker.GenerateRecordEmbeddingTest do
     end
 
     @tag :capture_log
-    test "returns {:error, reason} when OpenAI API fails" do
+    test "snoozes when OpenAI API fails with a 5xx (transient)" do
       record = record()
 
       Req.Test.stub(OpenAI.API, fn conn ->
         conn
         |> Plug.Conn.put_status(500)
-        |> Req.Test.json(%{"error" => "internal server error"})
+        |> Req.Test.json(%{"error" => %{"message" => "internal server error"}})
       end)
 
-      assert {:error, %{"error" => "internal server error"}} =
+      assert {:snooze, 30} =
                perform_job(GenerateRecordEmbedding, %{"record_id" => record.id})
     end
 

test/music_library/worker/populate_genres_test.exs

@@ -31,14 +31,38 @@ defmodule MusicLibrary.Worker.PopulateGenresTest do
     end
 
     @tag :capture_log
-    test "returns error when OpenAI API fails" do
+    test "snoozes when OpenAI API returns a 5xx (transient)" do
       record = record(%{genres: []})
 
       Req.Test.stub(OpenAI.API, fn conn ->
-        Plug.Conn.send_resp(conn, 500, JSON.encode!(%{"error" => "internal server error"}))
+        conn
+        |> Plug.Conn.put_status(500)
+        |> Req.Test.json(%{"error" => %{"message" => "internal server error"}})
       end)
 
-      assert {:error, "OpenAI API error:" <> _} =
+      assert {:snooze, 30} = perform_job(PopulateGenres, %{"id" => record.id})
+    end
+
+    @tag :capture_log
+    test "cancels when OpenAI API returns 429 insufficient_quota (permanent)" do
+      record = record(%{genres: []})
+
+      Req.Test.stub(OpenAI.API, fn conn ->
+        conn
+        |> Plug.Conn.put_resp_content_type("application/json")
+        |> Plug.Conn.send_resp(
+          429,
+          JSON.encode!(%{
+            "error" => %{
+              "code" => "insufficient_quota",
+              "type" => "insufficient_quota",
+              "message" => "quota exceeded"
+            }
+          })
+        )
+      end)
+
+      assert {:cancel, %OpenAI.API.ErrorResponse{code: "insufficient_quota"}} =
                perform_job(PopulateGenres, %{"id" => record.id})
     end
   end

test/open_ai/api_test.exs

@@ -2,6 +2,7 @@ defmodule OpenAI.APITest do
   use ExUnit.Case, async: true
 
   alias OpenAI.API
+  alias OpenAI.API.ErrorResponse
 
   @config %OpenAI.Config{
     api_key: "test_key",
@@ -26,15 +27,62 @@ defmodule OpenAI.APITest do
     end
 
     @tag :capture_log
-    test "returns error on non-2xx response" do
+    test "returns a rate-limit error response on 429 rate_limit_exceeded" do
       Req.Test.stub(__MODULE__, fn conn ->
         conn
         |> Plug.Conn.put_status(429)
-        |> Req.Test.json(%{"error" => "rate limited"})
+        |> Req.Test.json(%{
+          "error" => %{
+            "code" => "rate_limit_exceeded",
+            "type" => "rate_limit_error",
+            "message" => "Rate limit reached"
+          }
+        })
       end)
 
       completion = %{model: "gpt-4.1-mini", content: "test", role: "user", temperature: 0.5}
-      assert {:error, "OpenAI API error:" <> _} = API.gpt(completion, @config)
+
+      assert {:error, %ErrorResponse{} = err} = API.gpt(completion, @config)
+      assert err.status == 429
+      assert err.code == "rate_limit_exceeded"
+      assert err.kind == :rate_limit
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "returns an auth-error response on 429 insufficient_quota (non-retryable)" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(429)
+        |> Req.Test.json(%{
+          "error" => %{
+            "code" => "insufficient_quota",
+            "type" => "insufficient_quota",
+            "message" => "You exceeded your current quota"
+          }
+        })
+      end)
+
+      completion = %{model: "gpt-4.1-mini", content: "test", role: "user", temperature: 0.5}
+
+      assert {:error, %ErrorResponse{} = err} = API.gpt(completion, @config)
+      assert err.code == "insufficient_quota"
+      assert err.kind == :auth_error
+      refute ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "returns a server-error response on 500" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        conn
+        |> Plug.Conn.put_status(500)
+        |> Req.Test.json(%{"error" => %{"message" => "boom"}})
+      end)
+
+      completion = %{model: "gpt-4.1-mini", content: "test", role: "user", temperature: 0.5}
+
+      assert {:error, %ErrorResponse{kind: :server_error}} =
+               API.gpt(completion, @config)
     end
   end
 
@@ -54,14 +102,26 @@ defmodule OpenAI.APITest do
     end
 
     @tag :capture_log
-    test "returns error on non-200 response" do
+    test "returns a retryable server-error ErrorResponse on 500" do
       Req.Test.stub(__MODULE__, fn conn ->
         conn
         |> Plug.Conn.put_status(500)
-        |> Req.Test.json(%{"error" => "internal error"})
+        |> Req.Test.json(%{"error" => %{"message" => "internal error"}})
       end)
 
-      assert {:error, %{"error" => "internal error"}} = API.get_embeddings("test text", @config)
+      assert {:error, %ErrorResponse{status: 500, kind: :server_error} = err} =
+               API.get_embeddings("test text", @config)
+
+      assert ErrorResponse.retryable?(err)
+    end
+
+    test "returns the transport exception on connection failure" do
+      Req.Test.stub(__MODULE__, fn conn ->
+        Req.Test.transport_error(conn, :timeout)
+      end)
+
+      assert {:error, %Req.TransportError{reason: :timeout}} =
+               API.get_embeddings("test text", @config)
     end
   end
 
@@ -114,27 +174,27 @@ defmodule OpenAI.APITest do
     end
 
     @tag :capture_log
-    test "returns error on non-2xx response" do
+    test "returns an ErrorResponse struct on non-2xx response" do
       Req.Test.stub(__MODULE__, fn conn ->
         conn
         |> Plug.Conn.put_status(500)
-        |> Req.Test.json(%{"error" => "server error"})
+        |> Req.Test.json(%{"error" => %{"message" => "server error"}})
       end)
 
       cb = fn _chunk -> :ok end
 
-      assert {:error, "OpenAI API error:" <> _} =
+      assert {:error, %ErrorResponse{status: 500, kind: :server_error}} =
                API.chat_stream([], "instructions", "gpt-4.1", 0.7, @config, cb)
     end
 
-    test "returns error on connection failure" do
+    test "returns the transport exception on connection failure" do
       Req.Test.stub(__MODULE__, fn conn ->
         Req.Test.transport_error(conn, :timeout)
       end)
 
       cb = fn _chunk -> :ok end
 
-      assert {:error, "Connection error:" <> _} =
+      assert {:error, %Req.TransportError{reason: :timeout}} =
                API.chat_stream([], "instructions", "gpt-4.1", 0.7, @config, cb)
     end
 

test/wikipedia/api_test.exs

@@ -0,0 +1,164 @@
+defmodule Wikipedia.APITest do
+  use ExUnit.Case, async: true
+
+  alias Wikipedia.API
+  alias Wikipedia.API.ErrorResponse
+
+  @config %Wikipedia.Config{
+    user_agent: "test_agent",
+    req_options: [plug: {Req.Test, Wikipedia.API}, max_retries: 0],
+    api_cooldown: 0
+  }
+
+  describe "Action API silent-bug fix (HTTP 200 + body error envelope)" do
+    @tag :capture_log
+    test "promotes ratelimited Action API error into a retryable ErrorResponse" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{"code" => "ratelimited", "info" => "You've exceeded your rate limit."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{} = err} = API.get_wikipedia_title("Q352766", @config)
+      assert err.status == 200
+      assert err.code == "ratelimited"
+      assert err.kind == :rate_limit
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "promotes maxlag Action API error into a retryable ErrorResponse" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{
+            "code" => "maxlag",
+            "info" => "Waiting for a database server: 7 seconds lagged."
+          }
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :rate_limit} = err} =
+               API.get_wikipedia_title("Q352766", @config)
+
+      assert err.code == "maxlag"
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "promotes readonly Action API error into a retryable server error" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{"code" => "readonly", "info" => "The wiki is currently in read-only mode."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :server_error} = err} =
+               API.get_wikipedia_title("Q352766", @config)
+
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "promotes internal_api_error_* codes into retryable server errors" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{
+            "code" => "internal_api_error_DBQueryError",
+            "info" => "[abc] Exception Caught"
+          }
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :server_error} = err} =
+               API.get_wikipedia_title("Q352766", @config)
+
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "classifies unknown error codes as non-retryable client errors" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{"code" => "badtoken", "info" => "Invalid CSRF token."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :client_error} = err} =
+               API.get_wikipedia_title("Q352766", @config)
+
+      refute ErrorResponse.retryable?(err)
+    end
+  end
+
+  describe "Action API non-error body shapes still return {:ok, _}" do
+    test "wbgetentities with a missing entity returns {:ok, body} — not an error" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "entities" => %{"Q99999999999" => %{"id" => "Q99999999999", "missing" => ""}}
+        })
+      end)
+
+      assert {:ok, nil} = API.get_wikipedia_title("Q99999999999", @config)
+    end
+
+    test "wbgetentities with no enwiki sitelink returns {:ok, nil}" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "entities" => %{"Q352766" => %{"id" => "Q352766", "sitelinks" => %{}}}
+        })
+      end)
+
+      assert {:ok, nil} = API.get_wikipedia_title("Q352766", @config)
+    end
+
+    test "query with empty extract returns {:ok, nil}" do
+      Req.Test.stub(API, fn conn ->
+        Req.Test.json(conn, %{
+          "query" => %{
+            "pages" => %{"123" => %{"pageid" => 123, "title" => "X", "extract" => nil}}
+          }
+        })
+      end)
+
+      assert {:ok, nil} = API.get_article_extract("X", @config)
+    end
+  end
+
+  describe "REST v1 API error classification" do
+    @tag :capture_log
+    test "404 on article summary surfaces as a not-found ErrorResponse" do
+      Req.Test.stub(API, fn conn ->
+        conn
+        |> Plug.Conn.put_status(404)
+        |> Req.Test.json(%{
+          "httpCode" => 404,
+          "httpReason" => "Not Found",
+          "messageTranslations" => %{"en" => "The specified title does not exist."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{status: 404, kind: :not_found} = err} =
+               API.get_article_summary("Missing%20Title", @config)
+
+      refute ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "429 on article summary surfaces as a retryable rate-limit ErrorResponse" do
+      Req.Test.stub(API, fn conn ->
+        conn
+        |> Plug.Conn.put_status(429)
+        |> Req.Test.json(%{
+          "httpCode" => 429,
+          "httpReason" => "Too Many Requests",
+          "messageTranslations" => %{"en" => "Slow down."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{status: 429, kind: :rate_limit} = err} =
+               API.get_article_summary("Steven%20Wilson", @config)
+
+      assert ErrorResponse.retryable?(err)
+    end
+  end
+end

test/wikipedia_test.exs

@@ -1,6 +1,8 @@
 defmodule WikipediaTest do
   use ExUnit.Case, async: true
 
+  alias Wikipedia.API.ErrorResponse
+
   describe "get_artist_summary/1" do
     test "resolves wikidata ID to Wikipedia summary with full intro" do
       wikidata_id = "Q352766"
@@ -38,5 +40,40 @@ defmodule WikipediaTest do
 
       assert {:error, :no_english_wikipedia} = Wikipedia.get_artist_summary(wikidata_id)
     end
+
+    @tag :capture_log
+    test "promotes HTTP 200 Action API error envelopes into ErrorResponse structs" do
+      wikidata_id = "Q352766"
+
+      Req.Test.stub(Wikipedia.API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{"code" => "ratelimited", "info" => "You've exceeded your rate limit."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{} = err} =
+               Wikipedia.get_artist_summary(wikidata_id)
+
+      assert err.status == 200
+      assert err.code == "ratelimited"
+      assert err.kind == :rate_limit
+      assert ErrorResponse.retryable?(err)
+    end
+
+    @tag :capture_log
+    test "classifies non-retryable Action API error codes as client errors" do
+      wikidata_id = "Q352766"
+
+      Req.Test.stub(Wikipedia.API, fn conn ->
+        Req.Test.json(conn, %{
+          "error" => %{"code" => "badtoken", "info" => "Invalid CSRF token."}
+        })
+      end)
+
+      assert {:error, %ErrorResponse{kind: :client_error} = err} =
+               Wikipedia.get_artist_summary(wikidata_id)
+
+      refute ErrorResponse.retryable?(err)
+    end
   end
 end