fix: adapt to faststream 0.7.6, and make the AsyncAPI document non-blank

[lesnik512]

FastStream 0.7.6 released today and broke this package five ways at once. uv.lock is gitignored, so CI resolves fresh on every run: lint and all six pytest jobs on #83 are red on this right now.

A sixth fix is folded in: the assembled AsyncAPI document was blank, and has been since long before 0.7.6.

The five 0.7.6 changes

1. try_it_out._get_broker_registry is gone. __init__.py monkeypatched it inside a bare try/except Exception: pass, so the failure was silent — the AsyncAPI "try it out" page just stopped knowing about TimersBroker. 0.7.6 replaces it with a declarative hook, so the 20-line block and its imports are deleted and the class statement carries broker=TimersBroker.

2. SubscriberSpec and PublisherSpec now require address. The channel key is topic:Handle / topic:Publisher, which names nothing subscribable; address is the field a reader can act on. Both get the prefixed topic, since a broker-prefixed deployment addressed by the bare topic points at a topic nothing reads.

3. SubscriberSpecification.channel_labels is abstract. TimersSubscriberSpecification could not be instantiated at all. The name override is replaced by channel_labels, and the base's _channel_key builds the key.

4. dependencies narrowed from Iterable[Dependant] to Sequence[Dependant] in broker.py, router.py and registrator.py.

5. TasksMixin.add_task no longer returns the task. The only change upstream was dropping the return task line — the supervisor and restart_on_failure behaviour are byte-identical — but subscriber/usecase.py did await consume_task on the result, so a timed-out start raised TypeError: 'NoneType' object can't be awaited instead of cancelling its poll task. Now reads self.tasks[-1], which stop() has not cleared yet at that point.

6. The AsyncAPI document was blank, and predates all of this

AsyncAPI(broker).to_specification() returned no servers, no channels and no operations for a TimersBroker. I reproduced it on main against faststream 0.7.5 before concluding anything, so it is not from the 0.7.6 work.

Upstream collects channels only for brokers that reach its broker_servers mapping, and it fills that mapping inside:

for url in specification.url:
    ...
    broker_servers.append((broker, server))

TimersBroker passed url=[], so it never entered the mapping and get_broker_channels was never called for it. Every per-endpoint get_schema() stayed correct throughout, which is exactly why specification-level tests could not see it — and why the fix in items 2 and 3 above would otherwise have shipped into a document nobody could read.

The URL is now derived from the client, and rebuilt from the pool's connection parameters rather than the caller's DSN. The document is published — FastStream serves it from the ASGI app — so echoing back a Redis.from_url("redis://user:secret@...") would publish the password. Username and password are never read, so there is nothing to mask.

Before / after, same broker:

servers:    {}                    →  {"development": {"host": "cache.example:6399", "pathname": "/3", ...}}
channels:   {}                    →  {"app:reminders:Handle": "app:reminders", "app:reminders:Publisher": "app:reminders"}
operations: []                    →  ["app:reminders:HandleSubscribe", "app:reminders:Publisher"]

built from Redis.from_url("redis://user:secret@cache.example:6399/3"), with "secret" in json.dumps(spec) returning False.

One behaviour change worth noticing

Handing the channel key to the base means title_ now names the channel, not just the operation. The old name override ignored title_ entirely. This matches the publisher and every built-in broker, and it changes the generated key for any subscriber declared with title_. Covered by test_subscriber_title_names_the_channel_not_only_the_operation.

Floor, and the cap from #84

#84 capped faststream<0.7.6 to unblock CI, with "lift once this adapts" on the line. It adapts now, so that cap is replaced by a floor of faststream>=0.7.6,<0.8. Items 1–3 use APIs that do not exist below it. The old >=0.7.1 did still import cleanly today — I checked before changing it — but it becomes a lie the moment this lands, and nothing in CI resolves at the lower bound to catch that.

Tests

Written first, red before each change:

FAILED tests/test_unit.py::test_timers_broker_registered_in_test_broker_registry
FAILED tests/test_unit.py::test_subscriber_specification_addresses_the_channel_by_its_full_topic
FAILED tests/test_unit.py::test_subscriber_title_names_the_channel_not_only_the_operation
FAILED tests/test_unit.py::test_publisher_specification_addresses_the_channel_by_its_full_topic
FAILED tests/test_unit.py::test_asyncapi_document_carries_the_declared_channels
FAILED tests/test_unit.py::test_specification_url_carries_no_credentials

The two pre-existing test_isolation.py failures cover item 5 and are green again.

Two notes on test quality, since both let real defects read as passing:

• The replaced spec tests asserted only "my-topic" in name and assert schema, which is why a specification naming an unaddressable channel looked fine.
• test_specification_url_shapes covers the unix-socket and clientless branches explicitly. The coverage gate was already satisfied for them by a MagicMock client — mock.connection_pool.connection_kwargs.get("path") is truthy, so it walked into the socket branch by accident. 100% line coverage was not evidence that branch worked.

Verification

just lint-ci clean (eof-fixer, ruff format, ruff check, ty). Full suite against Redis 8: 167 passed, 100% coverage.

Separately, not addressed here

Under TestTimersBroker, a publisher declared on a prefixed router gets a spurious extra fake subscriber: create_publisher_fake_subscriber matches on full_topic but falls back to broker.subscriber(publisher.config.topic), which re-registers the unprefixed topic. It shows up as a stray reminders:PublisherResponseSubscriber channel alongside the correct app:reminders:* ones. Test-broker path only, unrelated to 0.7.6, and out of scope here.