cloudru-qa-http-client 1.0.0__tar.gz

cloudru_qa_http_client-1.0.0/.agents/skills/http-client-usage/SKILL.md

@@ -0,0 +1,227 @@
+---
+name: http-client-usage
+description: Building HTTP clients on the http_client library — an httpx wrapper providing constructor options, event hooks, send/stream decorators, request logging with secret masking, and streaming requests. Use when creating a client for an external API, adding an endpoint to an existing client, or configuring authentication, timeouts, retries, TLS, default headers, tracing, or streaming in a project that imports http_client. Key names are HttpClient, AsyncHttpClient, ClientOption, options.with_*, send_decorator, stream_request.
+license: Apache-2.0
+metadata:
+  library-version: "1.0.0"
+  source: packages/http-client
+---
+
+# Building clients on http_client
+
+`http_client` is a thin wrapper around `httpx`. It adds exactly three things:
+
+1. **Constructor options** (`ClientOption`) — configuration is passed as positional arguments to the constructor instead of a dozen keyword parameters.
+2. **Transport decorators** — wrappers around `send` and `stream_request` for cross-cutting concerns: metrics, caching, plugging in an existing retry library.
+3. **Request/response logging** with field masking by glom paths.
+
+Everything else (`Request`, `Response`, `Auth`, `Timeout`, status codes) is plain `httpx`. Do not invent your own wrappers around what `httpx` already provides.
+
+## Hard rules
+
+1. Subclass `HttpClient` or `AsyncHttpClient`. `BaseHttpClient` is abstract and has no transport; never subclass it directly.
+2. Configure only through constructor options. Do not assign `client.timeout`, `client.auth`, etc. after construction — `httpx.Client` is already built and the change will not apply.
+3. Inside endpoint methods call `self.send(request)`, **not** `self.client.send(request)`. The direct call bypasses every decorator.
+4. Do not add dependencies to the library itself — it must stay lightweight.
+5. Never write a retry loop, backoff, or circuit breaker by hand. Take the project's existing retry library and mount it via `with_send_decorator`.
+
+Before claiming an option cannot do something, check `references/pitfalls.md` — it documents the non-obvious behaviour that trips people up.
+
+## Canonical client (sync)
+
+```python
+import httpx
+from http_client import HttpClient
+
+
+class OrdersClient(HttpClient):
+    """Client for the orders service."""
+
+    def get_order(self, order_id: str) -> httpx.Response:
+        request = self.client.build_request("GET", f"/orders/{order_id}")
+        return self.send(request)
+
+    def create_order(self, payload: dict[str, object]) -> httpx.Response:
+        request = self.client.build_request("POST", "/orders", json=payload)
+        return self.send(request)
+```
+
+Paths are relative to the `base_url` passed to the constructor — `httpx` joins them.
+
+Assembly:
+
+```python
+import logging
+
+from http_client import options
+from httpx import BasicAuth
+
+client = OrdersClient(
+    "https://api.example.com",
+    options.with_auth(BasicAuth("admin", "PazzW0rD")),
+    options.with_timeout(30),
+    options.with_logger(logging.getLogger("orders")),
+    options.with_autogenerated_request_id(),
+)
+```
+
+## Canonical client (async)
+
+Exactly two differences: the base class and `await`. The options are identical, including synchronous event hooks — those are wrapped into coroutines automatically.
+
+```python
+class AsyncOrdersClient(AsyncHttpClient):
+    async def get_order(self, order_id: str) -> httpx.Response:
+        request = self.client.build_request("GET", f"/orders/{order_id}")
+        return await self.send(request)
+```
+
+One exception: `send`/`stream_request` decorators for an async client must themselves be async — there is no auto-wrapping for them.
+
+## Choosing how to send
+
+| Task | What to call |
+|---|---|
+| Endpoint method in your own client | `self.client.build_request(...)` → `self.send(request)` |
+| One-off call without a custom client | `client.get(...)`, `client.post(...)`, `client.request(method, url, ...)` |
+| Streaming download/upload, method and URL known | `client.stream(method, url, ...)` |
+| Stream from a prepared `httpx.Request` | `client.stream_request(request)` |
+
+`send` and `stream_request` are the only points decorators pass through. The `get`/`post`/... shortcuts call `send` internally, so decorators apply to them as well.
+
+## Releasing connections
+
+The connection pool is created in the constructor, so every client must be closed. Prefer the context manager:
+
+```python
+with OrdersClient("https://api.example.com") as client:
+    client.get_order("42")
+
+async with AsyncOrdersClient("https://api.example.com") as client:
+    await client.get_order("42")
+```
+
+For a long-lived client, close it explicitly: `client.close()` / `await client.aclose()`.
+
+## Options
+
+The full catalog with exact signatures and behaviour is in `references/options.md`. Summary:
+
+| Option | Purpose |
+|---|---|
+| `with_auth(auth)` | any httpx authentication mechanism |
+| `with_timeout(timeout)` | client timeout; defaults to 60 s on every phase |
+| `with_logger(logger, sanitize_keys=..., ...)` | request and response logging |
+| `with_default_headers(headers)` | headers applied to every request |
+| `with_request_hook(hook)` / `with_response_hook(hook)` | arbitrary request/response processing |
+| `with_autogenerated_request_id()` | `X-Request-ID` header |
+| `with_autogenerated_traceparent()` | `Traceparent` header (W3C Trace Context) |
+| `with_send_decorator(dec)` / `with_stream_decorator(dec)` | transport wrappers |
+| `with_follow_redirects(max_redirects=20)` | follow redirects |
+| `with_verify_context(verify=...)` / `use_truststore()` | TLS configuration |
+
+### Writing your own option
+
+`ClientOption` is `Callable[[BaseHttpClient], None]`. Options receive the actual client inside
+`BaseHttpClient.__init__`, before the httpx transport exists. Use the public configuration properties
+(`timeout`, `auth`, `verify`, `url_suffix`) and registration methods. Accessing `client.client` inside
+an option raises `AttributeError`.
+
+```python
+import httpx
+from http_client import ClientOption
+from http_client.client import BaseHttpClient
+
+
+def with_api_key(key: str) -> ClientOption:
+    def add_key(request: httpx.Request) -> None:
+        request.headers["X-Api-Key"] = key
+
+    def option(client: BaseHttpClient) -> None:
+        client.add_request_hook(add_key, prepend=True)
+
+    return option
+```
+
+Compose options from built-ins or call `add_request_hook`, `add_response_hook`, `add_send_decorator`
+and `add_stream_decorator`. `prepend=True` puts a request hook first. Configuration properties and
+registration methods reject changes after construction with `RuntimeError`.
+
+## Wrapping the transport
+
+A decorator wraps `send`, so it sees the request on the way out and the response on the way back. The shape below is the whole pattern — this particular one just records what passed through, which is handy in tests:
+
+```python
+from collections.abc import Callable
+from functools import wraps
+
+import httpx
+from http_client.types import SyncSendMethod
+
+
+def with_call_recorder(calls: list[str]) -> Callable[[SyncSendMethod], SyncSendMethod]:
+    def decorator(func: SyncSendMethod) -> SyncSendMethod:
+        @wraps(func)
+        def wrapper(request: httpx.Request, **kwargs: object) -> httpx.Response:
+            calls.append(f"-> {request.method} {request.url}")
+            response = func(request, **kwargs)
+            calls.append(f"<- {response.status_code}")
+            return response
+
+        return wrapper
+
+    return decorator
+```
+
+**Do not hand-roll retries, backoff, or circuit breakers here.** They are solved problems with dedicated libraries; check what the project already depends on and plug it in through this hook. Writing the retry loop yourself means reimplementing jitter, budgets, and idempotency rules — and getting them subtly wrong.
+
+Annotation types live in `http_client.types`: `SyncSendMethod`, `AsyncSendMethod`, `SyncStreamRequestMethod`, `AsyncStreamRequestMethod` and the matching `*Decorator` aliases.
+
+A `stream_request` decorator must itself be a context manager — put `@contextmanager` / `@asynccontextmanager` **below** `@wraps`:
+
+```python
+@wraps(func)
+@contextmanager
+def wrapper(request, **kwargs):
+    with func(request, **kwargs) as response:
+        yield response
+```
+
+## Streaming requests
+
+```python
+with client.stream("GET", "/large-file.zip") as response:
+    response.raise_for_status()
+    with open("file.zip", "wb") as f:
+        for chunk in response.iter_bytes(chunk_size=8192):
+            f.write(chunk)
+```
+
+```python
+async with client.stream("GET", "/events") as response:
+    async for line in response.aiter_lines():
+        handle(line)
+```
+
+A streamed response body shows up in logs as `<streamed content>` — at the time the response hook fires the content has not been read yet. This is not an error.
+
+## Logging
+
+`with_logger` installs hooks that emit one record per request and one per response. The records are linked by the `id` field — an internal `trace_id` unrelated to any server-side tracing.
+
+Common secrets — `authorization`, `cookie`, `x-api-key`, `password`, `token` and friends — are masked out of the box. `sanitize_keys` adds project-specific paths on top; it does not replace the defaults.
+
+Masking is driven by glom paths applied to the body, headers, and query parameters. httpx lowercases header names, so write `"authorization"`, not `"Authorization"`.
+
+```python
+options.with_logger(
+    logger,
+    sanitize_keys=["authorization", "user.password", "items.*.token"],
+)
+```
+
+Global defaults (level, JSON indent, truncation of long bodies) come from `http_client.configure_logging(...)`. `HTTP_CLIENT_*` environment variables **override** that function's arguments. Details and the full variable list are in `references/options.md`.
+
+## Known pitfalls
+
+Before debugging "strange" behaviour, read `references/pitfalls.md`. It covers `url_suffix` join rules, decorator ordering and freezing, Content-Type body filtering, and why options cannot touch the transport.

cloudru_qa_http_client-1.0.0/.agents/skills/http-client-usage/references/options.md

@@ -0,0 +1,219 @@
+# Option catalog and configuration
+
+All options live in `http_client.options`.
+
+## Transport and connection
+
+### `with_timeout(timeout: TimeoutTypes) -> ClientOption`
+
+Client timeout. Accepts a number of seconds or an `httpx.Timeout`.
+
+The default is `httpx.Timeout(60)`, i.e. 60 seconds for each phase (`connect`, `read`, `write`, `pool`).
+
+```python
+from httpx import Timeout
+
+options.with_timeout(Timeout(connect=5.0, read=30.0, write=5.0, pool=1.0))
+```
+
+Precedence: an explicit `timeout=` on the call wins over the client timeout. Passing `timeout=None` on a call deliberately disables the timeout for that request.
+
+### `with_follow_redirects(max_redirects: int = 20) -> ClientOption`
+
+Enables redirect following and sets the limit. The default comes from `httpx._config.DEFAULT_MAX_REDIRECTS`.
+
+Without this option the client does not follow redirects. Override per request with `client.get(url, follow_redirects=True)`.
+
+### `with_verify_context(*, verify: ssl.SSLContext | str | bool) -> ClientOption`
+
+TLS verification context. Accepts an `SSLContext`, a path to a CA bundle (converted to an `SSLContext` internally), or `False` to disable verification. Applies to both `HttpClient` and `AsyncHttpClient`.
+
+⚠️ `False` disables certificate checking entirely — local debugging only, never production.
+
+### `use_truststore() -> ClientOption`
+
+Certificate verification through the system trust store. Requires the extra dependency:
+
+```shell
+pip install "cloudru-qa-http-client[truststore]"
+```
+
+Without `truststore` installed it raises `ImportError` with a hint. Internally it calls `with_verify_context`.
+
+## Authentication
+
+### `with_auth(auth: AuthTypes) -> ClientOption`
+
+Any httpx authentication mechanism: `BasicAuth`, `DigestAuth`, a callable, or your own `httpx.Auth` subclass.
+
+```python
+from httpx import BasicAuth
+
+options.with_auth(BasicAuth("user", "password"))
+options.with_auth(lambda request: request.headers.update({"Authorization": f"Bearer {token}"}))
+```
+
+Override for a single request: `client.get(url, auth=other_auth)`.
+Disable authentication entirely for a request: `client.get(url, anonymous=True)` — `auth` is then not applied at all, not even one passed explicitly. Other per-call arguments such as `follow_redirects` still apply.
+
+## Headers
+
+### `with_default_headers(headers: dict[str, str]) -> ClientOption`
+
+Adds a hook that fills in each header only when the request does not already carry it. A value passed per request wins over the default, matching `httpx` semantics.
+
+### `with_autogenerated_request_id() -> ClientOption`
+
+An `X-Request-ID` header carrying a random UUID4 per request. The hook is inserted at the **front** of the list, so the identifier is visible to later hooks, including the logging one.
+
+### `with_autogenerated_traceparent() -> ClientOption`
+
+A `Traceparent` header in W3C Trace Context format: `00-<32 hex>-<16 hex>-01`. Also inserted at the front of the list.
+
+This is standalone generation, not an OpenTelemetry integration: no context is picked up from the current span.
+
+## Event hooks
+
+### `with_request_hook(hook: EventHook) -> ClientOption`
+### `with_response_hook(hook: EventHook) -> ClientOption`
+
+Append a hook to the corresponding list. The contract matches [httpx event hooks](https://www.python-httpx.org/advanced/event-hooks/): the hook receives a `Request` (or `Response`) and its return value is ignored. Mutate the object in place.
+
+```python
+def add_tenant(request: httpx.Request) -> None:
+    request.headers["X-Tenant"] = "acme"
+```
+
+For `AsyncHttpClient`, synchronous hooks are wrapped into coroutines during `__init__`, so the same set of options works for both clients. Native async hooks are accepted too; call order is preserved regardless of mixing.
+
+## Transport decorators
+
+### `with_send_decorator(decorator: SendDecorator) -> ClientOption`
+### `with_stream_decorator(decorator: StreamRequestDecorator) -> ClientOption`
+
+Wrap `_send` and `_stream_request` respectively. Composition happens once, in `__init__`:
+
+```python
+for decorator in decorators:
+    func = decorator(func)
+```
+
+Consequences:
+
+- The last decorator added becomes the outermost.
+- Register decorators only in constructor options; late registration raises `RuntimeError`.
+- For `AsyncHttpClient` the decorator must be async — there is no auto-wrapping, unlike event hooks.
+
+Protocol types for annotations are in `http_client.types`.
+
+## Logging
+
+### `with_logger(logger, *, sanitize_keys=None, json_indent=None, use_line_breaks=None, log_level=None, use_default_sanitize_keys=True) -> ClientOption`
+
+| Parameter | Type | Meaning |
+|---|---|---|
+| `logger` | `logging.Logger` | where to write |
+| `sanitize_keys` | `Sequence[str] \| None` | glom paths to fields replaced with `****` |
+| `use_default_sanitize_keys` | `bool` | also mask the built-in secret list (default `True`) |
+| `json_indent` | `int \| None` | indent used when serializing JSON parts of the record |
+| `use_line_breaks` | `bool` | newlines between record parts instead of spaces |
+| `log_level` | `int \| None` | record level, defaults to `INFO` |
+
+Record format:
+
+```
+Request: id=<uuid> Method: [GET] https://host/path QueryParams: {...} Headers: {...} Body: {...}
+Response: id=<uuid> Status: 200 [OK] Headers: {...} Body: {...}
+```
+
+`id` is an internal identifier linking a request to its response. It is stored in `request.extensions["trace_id"]` and has nothing to do with server-side tracing.
+
+Bodies are rendered by `Content-Type`:
+
+| `Content-Type` | `Body:` shows |
+|---|---|
+| `application/json` | the parsed, masked, re-serialized body |
+| `text/plain` | the body as is |
+| `application/octet-stream` | the decoded body, or `<binary content (N bytes)>` if it is not text |
+| anything else | `<not logged, content-type: ...>` |
+| absent (no body) | nothing — the `Body:` part is omitted |
+
+So an omitted `Body:` means there was no body, while `<not logged, ...>` means there was one and the type is not rendered.
+
+A parsing failure in any part never breaks the request — the log gets `<Error parse: ExceptionType>` instead. Only the exception class is recorded: its message could carry fragments of the body or URL.
+
+### Masking
+
+A built-in list of common secrets is masked by default — `authorization`, `proxy-authorization`, `cookie`, `set-cookie`, `x-api-key`, `api-key`, `api_key`, `access_token`, `refresh_token`, `password`, `secret`, `token` (`event_hooks.DEFAULT_SANITIZE_KEYS`). Anything in `sanitize_keys` is added on top. Turn the defaults off only with `use_default_sanitize_keys=False`.
+
+Masking covers the access-log records this library writes — not `httpx` exceptions, which may carry the full URL including query. Do not put secrets in URLs.
+
+`sanitize_keys` is applied to headers, query parameters, and the parsed JSON body — independently, each path is looked up in all three. A path that does not match is silently skipped.
+
+Syntax is [glom](https://glom.readthedocs.io/en/latest/):
+
+| Path | What it masks |
+|---|---|
+| `authorization` | the `Authorization` header (httpx lowercases names) |
+| `user.password` | a nested object field |
+| `items.*.token` | a field inside every array element |
+| `*.credentials.secret` | a field when the body itself is an array of objects |
+
+### `configure_logging(...)` — global defaults
+
+```python
+from http_client import configure_logging
+from http_client.config import LoggingConfig
+
+configure_logging(
+    options=LoggingConfig(
+        log_level=logging.INFO,
+        json_indent=None,
+        use_line_breaks=False,
+        octet_stream_log_maxlength=1024,
+        plain_text_log_maxlength=None,
+        json_log_maxlength=None,
+    )
+)
+```
+
+Pass a `LoggingConfig` through `options=...`. Its fields set the defaults.
+The function writes to the module-level singleton `LOGGING_CONFIG`, from which the hooks read whenever the corresponding `with_logger` argument was not supplied.
+
+**Environment variables override the function arguments**, not the other way round:
+
+| Variable | Field |
+|---|---|
+| `HTTP_CLIENT_LOG_LEVEL` | `log_level` (integer) |
+| `HTTP_CLIENT_USE_LINE_BREAKS` | `use_line_breaks` (`true`/`y`/`yes`/`on`) |
+| `HTTP_CLIENT_JSON_INDENT` | `json_indent` |
+| `HTTP_CLIENT_OCTET_STREAM_LOG_MAXLENGTH` | `octet_stream_log_maxlength` |
+| `HTTP_CLIENT_PLAIN_TEXT_LOG_MAXLENGTH` | `plain_text_log_maxlength` |
+| `HTTP_CLIENT_JSON_LOG_MAXLENGTH` | `json_log_maxlength` |
+
+The length limits (`*_maxlength`) trim the middle of the string, keeping the head and tail and inserting `...`. They can only be set globally; `with_logger` does not accept them.
+
+## Utilities
+
+`http_client.utils.sanitizer.sanitize(value, sensitive_keys)` — masks a dict or list using the same glom paths. Returns a deep copy and leaves the original untouched.
+
+`http_client.utils.traceparent.generate()` — a traceparent string.
+
+`http_client.utils.logutils.create_request_log` / `create_response_log` — build the same log strings manually.
+
+`http_client.utils.logutils` also exports the building blocks those two are made of, so a logger built
+on a different backend renders bodies exactly the same way: `to_json`, `trim`, `decode`, `is_streamed`,
+`is_buffered`, `parse_error`. These helpers are part of the public API.
+
+## Inherited client attributes
+
+| Attribute | Where it is set | Purpose |
+|---|---|---|
+| `DEFAULT_URL_SUFFIX` | overridden in the subclass | path prefix injected by the `endpoint()` method |
+| `DEFAULT_TIMEOUT` | overridden in the subclass | default timeout, `httpx.Timeout(60)` |
+| `base_url` | from the constructor | proxies `self.client.base_url` |
+| `client` | created in `__init__` | the underlying `httpx.Client` / `httpx.AsyncClient` |
+
+Lifecycle mirrors `httpx` exactly: the connection pool is built in the constructor, `__enter__` / `__aenter__` open it, and `__exit__` / `__aexit__` close it. `HttpClient.close()` and `AsyncHttpClient.aclose()` do the same for a client not used as a context manager — the naming follows `httpx`, so the async client has `aclose()` only. A closed client cannot be reopened; doing so raises `RuntimeError`.
+
+`endpoint(path)` joins `url_suffix` with `path`. Leading slashes on `path` and a missing trailing slash on the suffix are both normalized, so `"/u-api/orders/v1"` and `"/u-api/orders/v1/"` behave identically. It accepts relative paths only: an absolute URL or a `..` segment raises `ValueError`.

cloudru_qa_http_client-1.0.0/.agents/skills/http-client-usage/references/pitfalls.md

@@ -0,0 +1,53 @@
+# Pitfalls
+
+Non-obvious behaviour that costs time when you hit it unprepared.
+
+## Options run before the httpx client exists
+
+Order inside `__init__`: `BaseHttpClient.__init__` creates typed internal configuration and runs
+the options; then `HttpClient.__init__` creates the httpx transport from that configuration.
+
+Two consequences:
+
+1. An option that touches `client.client` raises `AttributeError` — there is no transport yet.
+2. Setting `client.timeout`, `auth`, `verify` or `url_suffix` after construction raises `RuntimeError`.
+   Set them through constructor options. For transport-level changes supported by httpx, use its
+   public API directly, for example `client.client.timeout`, or rebuild the client.
+
+## Decorators are frozen once
+
+Decorators are composed in `__init__`. Register them through constructor options with
+`add_send_decorator` / `add_stream_decorator` or the built-in option factories. Late registration
+raises `RuntimeError`.
+
+Composition order is a plain loop:
+
+```python
+for decorator in decorators:
+    func = decorator(func)
+```
+
+so the last decorator added ends up outermost and runs first.
+
+## `endpoint()` only accepts relative paths
+
+An absolute URL or a `..` segment raises `ValueError` instead of silently escaping the client's base URL. Build cross-host calls with a separate client, not by passing a full URL to `endpoint()`.
+
+## A decorator that sends twice runs the hooks twice
+
+Event hooks fire inside `httpx.Client.send`, so a decorator that calls `func()` twice runs every
+request hook twice. The built-in ID hooks preserve existing `X-Request-ID` and `Traceparent` values;
+reusing one request keeps its identifiers stable across retries. A newly built request gets new
+identifiers unless the caller provided them. Other hooks, including logging, still run per attempt.
+
+## Streaming: the logged response body is always a placeholder
+
+`stream_request` marks the request with `request.extensions["is_streaming"] = True`. The logging response hook sees the marker and, while the body is unread, writes `Body: <streamed content>`. This is by design — reading the body inside the hook would defeat the purpose of streaming.
+
+## `HttpClient` and `AsyncHttpClient` are declared abstract yet instantiable
+
+Both carry `metaclass=ABCMeta`, but the only abstract member, `BaseHttpClient.client`, is implemented in each. So `HttpClient("https://...")` works — the library's own tests rely on it. Subclassing is still the preferred approach, but direct construction is not a mistake.
+
+## False alarms when reading the code
+
+- `logutils.create_request_log` writes `r.extensions["trace_id"]`, meaning the logging hook mutates the request. Keep that in mind when comparing `extensions` in tests.

cloudru_qa_http_client-1.0.0/.gitignore

@@ -0,0 +1,26 @@
+/docs/plans/
+.venv/
+.venvs/
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.hypothesis/
+.coverage
+coverage.xml
+htmlcov/
+report*.xml
+allure-results/
+allure-report/
+.allure/
+logs/
+*.log
+.env
+.env.*
+.DS_Store
+dist/
+build/
+*.egg-info/
+/packages/*/api/
+.claude/settings.local.json