Python SDK retries, exceptions, constants

TL;DR Every client retries by default: RetryPolicy() = 2 retries, exponential backoff 0.5s → 5.0s with 25% jitter, retryable statuses {408, 429, 500–599}, Retry-After honored, 30s total retry budget per call. HTTP timeouts are separate and default to 10.0s per operation. Catch TypeSafeError for everything, TypeSafeAPIError for HTTP failures (.status, .body, .headers, .endpoint, .request_id), and TypeSafeAPIConnectionError / TypeSafeAPITimeoutError for transport failures. Disable retries with RetryPolicy(max_retries=0).

RetryPolicy

A frozen dataclass, importable as from typesafe_sdk import RetryPolicy. It may be passed on the client (retry=) or per call (system_one(..., retry=...) / models.list(retry=...)); a per-call policy fully replaces the client-level one for that call.

RetryPolicy(
    max_retries: int = 2,
    backoff_initial: float = 0.5,
    backoff_max: float = 5.0,
    backoff_jitter: float = 0.25,
    http_statuses: set[int] = {408, 429, *range(500, 600)},
    respect_retry_after: bool = True,
    api_connection_error: bool = True,
    api_timeout_error: bool = True,
    exceptions: set[type[BaseException]] = set(),
    predicate: Callable[[BaseException], bool] | None = None,
    timeout: float | None = 30.0,
)

Validation (__post_init__)

This validation is the 0.6.0 bug fix "handle invalid values in RetryPolicy".

Which failures are retried

RetryPolicy._retryable(error) returns True when any of these hold:

1. isinstance(error, TypeSafeAPITimeoutError) and api_timeout_error is True; else
2. isinstance(error, TypeSafeAPIConnectionError) and api_connection_error is True; else
3. isinstance(error, TypeSafeAPIError) and error.status in http_statuses; else not retryable by the built-in rules — then
4. isinstance(error, tuple(self.exceptions)), or
5. predicate is not None and predicate(error).

The three built-in branches are mutually exclusive (TypeSafeAPITimeoutError is checked first because it subclasses TypeSafeAPIConnectionError). Note that TypeSafeAPIResponseValidationError subclasses TypeSafeAPIError and carries the successful status code, so it is retried only if that status happens to be in http_statuses — with the defaults, a 200 is not.

Default retryable statuses: 408 (Request Timeout), 429 (Too Many Requests), and every 5xx from 500 through 599. Non-retryable by default: 400, 401, 403, 404, 422. See HTTP status codes, rate limits, retry semantics.

Backoff and Retry-After

For attempt n (1-based, n = 1 is the wait before the first retry), _backoff computes:

• exponential = min(backoff_initial * 2**(n-1), backoff_max)
• delay = exponential * (1 - random() * backoff_jitter), rounded to 3 decimals, capped at exponential
• if backoff_initial == 0 or backoff_max == 0, the delay is 0.0

With defaults that is ~0.5s then ~1.0s (each minus up to 25%).

When respect_retry_after is True and the error is a TypeSafeAPIError carrying a retry header, that header overrides the computed backoff entirely. parse_retry_after checks headers in this order:

Rules from _core/errors.py:parse_retry_after:

• A numeric value must be finite; negative values on retry-after-ms are skipped, a negative retry-after returns None.
• An empty/whitespace value is treated as 0.
• A non-numeric retry-after is parsed as an HTTP-date; the wait is max(0, date - now) in milliseconds.
• The function returns milliseconds; the retry layer divides by 1000 to get seconds.

The total budget is enforced with tenacity's stop_after_attempt(max_retries + 1) | stop_before_delay(timeout) and reraise=True, so the original exception (not a RetryError) surfaces.

Each retried request carries X-TypeSafe-Retry-Count: <attempt> and logs "<METHOD> <url> retry <n>" at INFO.

Examples

from typesafe_sdk import RetryPolicy, TypeSafeClient

client = TypeSafeClient(
    retry=RetryPolicy(max_retries=3, timeout=10.0, http_statuses={429, 500, 502, 503, 504})
)

from typesafe_sdk import Noul, RetryPolicy, TypeSafeClient

with TypeSafeClient() as client:
    # client-level
    client_policy = RetryPolicy(max_retries=3, backoff_max=0.2, timeout=1.0)

    # per-call override
    result = client.system_one(
        "I was charged twice.",
        {"billing": Noul(instructions="Is this about billing?")},
        retry=RetryPolicy(max_retries=0),  # no retries for this call
    )
    print(result.nouls["billing"].noul)

Retry on an extra exception type, or with a custom predicate:

import httpx2

from typesafe_sdk import RetryPolicy, TypeSafeAPIError, TypeSafeClient

policy = RetryPolicy(
    exceptions={httpx2.RemoteProtocolError},
    predicate=lambda error: isinstance(error, TypeSafeAPIError) and error.status == 409,
)
client = TypeSafeClient(retry=policy)

Timeouts

Two independent budgets:

Precedence for the HTTP timeout: per-call timeout → client timeout → http_client.timeout (when you supplied an http_client and passed no timeout) → DEFAULT_TIMEOUT (10.0).

Pass an httpx2.Timeout for fine-grained control:

import httpx2

from typesafe_sdk import TypeSafeClient

client = TypeSafeClient(timeout=httpx2.Timeout(connect=2.0, read=20.0, write=10.0, pool=5.0))

An expired HTTP timeout raises TypeSafeAPITimeoutError, which is retryable by default; when the retry budget is exhausted it propagates.

Exception hierarchy

Exception
└── TypeSafeError
    ├── TypeSafeAPIError
    │   ├── TypeSafeBadRequestError            (400)
    │   ├── TypeSafeAuthenticationError        (401)
    │   ├── TypeSafePermissionDeniedError      (403)
    │   ├── TypeSafeNotFoundError              (404)
    │   ├── TypeSafeUnprocessableEntityError   (422)
    │   ├── TypeSafeRateLimitError             (429)
    │   ├── TypeSafeInternalServerError        (5xx)
    │   └── TypeSafeAPIResponseValidationError (2xx with an unusable body)
    └── TypeSafeAPIConnectionError   (also subclasses ConnectionError)
        └── TypeSafeAPITimeoutError  (also subclasses TimeoutError)

Status→class mapping (STATUS_ERROR_TYPES in _core/errors.py): 400, 401, 403, 404, 422, 429 map to the classes above; any other status ≥ 500 becomes TypeSafeInternalServerError; anything else becomes a plain TypeSafeAPIError.

Because TypeSafeAPIConnectionError also subclasses ConnectionError and TypeSafeAPITimeoutError also subclasses TimeoutError, generic except ConnectionError / except TimeoutError handlers in surrounding code will catch them.

Error attributes in detail

• status — HTTP response status code.
• body — the server's JSON error body, plain response text, or None for an empty body.
• headers — httpx2.Headers of the response.
• endpoint — the request method and URL, without credentials, query parameters, or fragment, when available (e.g. POST https://api.typesafe.ai/v1/systemone). This is the 0.6.0 "error messages include http details and metadata" improvement.
• request_id — the x-typesafe-request-id response header, or None if absent.
• retry_after_ms (429 only) — the server's requested wait in milliseconds, or None.
• field_path (validation only) — dotted path to the first offending field.

__str__ renders "{endpoint}: {status} {message} (request_id={id})", omitting the parts that are unavailable. The message is derived from the body by extract_message, which tries, in order: body as a string → body["error"] (string) → body["error"]["message"] → body["message"] → body["detail"] (string) → body["detail"]["message"] → a FastAPI-style body["detail"] list joined as "loc.path: msg; …". If nothing matches, the serialized body is truncated to MAX_ERROR_BODY_LENGTH (200) characters with an ellipsis; an empty body yields "status code (no body)". __repr__ deliberately omits the body and headers. Exceptions are picklable as of 0.6.0.

Handling examples

from typesafe_sdk import TypeSafeAPIError, TypeSafeClient

try:
    client.system_one(state, questions)
except TypeSafeAPIError as error:
    print(error.status, error.request_id)

Full triage:

import logging
import time

from typesafe_sdk import (
    Noul,
    TypeSafeAPIConnectionError,
    TypeSafeAPIError,
    TypeSafeAPIResponseValidationError,
    TypeSafeAPITimeoutError,
    TypeSafeAuthenticationError,
    TypeSafeClient,
    TypeSafeError,
    TypeSafeRateLimitError,
)

with TypeSafeClient() as client:
    try:
        result = client.system_one(
            "I was charged twice.",
            {"billing": Noul(instructions="Is this about billing?")},
        )
    except TypeSafeAuthenticationError:
        raise SystemExit("Check TYPESAFE_API_KEY")
    except TypeSafeRateLimitError as error:
        wait = (error.retry_after_ms or 1000) / 1000
        logging.warning("rate limited, waiting %.1fs (request %s)", wait, error.request_id)
        time.sleep(wait)
    except TypeSafeAPIResponseValidationError as error:
        logging.error("bad response body at %s: %s", error.field_path, error.endpoint)
    except TypeSafeAPITimeoutError as error:
        logging.error("timed out after %s", error.timeout)
    except TypeSafeAPIConnectionError as error:
        logging.error("connection failure: %s", error)
    except TypeSafeAPIError as error:
        logging.error("%s %s", error.status, error)
    except TypeSafeError as error:  # config / validation problems, no HTTP made
        logging.error("client-side error: %s", error)
    else:
        print(result.nouls["billing"].noul)

Catch order matters: TypeSafeAPITimeoutError before TypeSafeAPIConnectionError, the specific TypeSafeAPIError subclasses before TypeSafeAPIError, and TypeSafeError last.

Constants

Public — typesafe_sdk.constants

from typesafe_sdk import constants

print(constants.DEFAULT_BASE_URL, constants.DEFAULT_MODEL, constants.DEFAULT_TIMEOUT)

Internal — typesafe_sdk._core.constants

Not part of the public API (absent from the docs and from __all__), but they define the wire behavior you will see on the network:

Treat these as version-specific implementation detail; depend on the public constants module instead.

Version notes

0.6.0 changed retry/error behavior in two ways recorded in the changelog: RetryPolicy now rejects invalid values (see the validation table), and errors carry HTTP details and metadata (endpoint, richer __str__). Exceptions and responses became picklable. See Python SDK changelog.

Related

• Python SDK: install, clients, system_one() — clients, system_one(), headers, logging
• Python SDK responses, answers, usage, models — what you get when no exception is raised
• Python SDK changelog — release history
• HTTP status codes, rate limits, retry semantics — HTTP status codes and retry semantics at the API level
• TYPESAFE_* environment variables across SDKs — TYPESAFE_* across SDKs
• JavaScript SDK error classes, RetryPolicy, RequestOptions — the JS/TS equivalents
• HTTP API: POST /v1/systemone and GET /v1/models — endpoints and error bodies

Sources

• raw/docs/sdk__python__api__retries.md (https://docs.typesafe.ai/sdk/python/api/retries.md)
• raw/docs/sdk__python__api__exceptions.md (https://docs.typesafe.ai/sdk/python/api/exceptions.md)
• raw/docs/sdk__python__api__constants.md (https://docs.typesafe.ai/sdk/python/api/constants.md)
• raw/docs/sdk__python__usage.md (https://docs.typesafe.ai/sdk/python/usage.md)
• raw/github/typesafe-sdk-python/src/typesafe_sdk/_core/retry.py, _core/errors.py, _core/constants.py, _core/transport.py, _core/config.py, constants.py (https://github.com/typesafe-ai/typesafe-sdk-python @ 420ef4ffb612d5a539a1e0f0fe883ff6770340af)