fix: route Speak V2 websocket through custom transport_factory

[mayankbohradev]

Summary

A transport_factory passed to DeepgramClient / AsyncDeepgramClient is not applied to the Speak V2 websocket client. Speak V2 connects through the default websockets transport instead, with no error and no warning.

Who this affects

transport_factory exists so callers can route websocket traffic through their own transport — a proxied connection, a test double, or a custom-hosted endpoint (the .fernignore notes cite a SageMaker transport as the motivating case).

Today a caller who sets it gets custom-transport routing for Listen v1, Listen v2, Speak v1 and Agent v1, while Speak V2 opens a direct connection to api.deepgram.com. Because the bypass is silent, a caller relying on a custom transport for network-egress or auditing reasons has no signal that part of their traffic isn't going through it.

Root cause

install_transport() patches the modules listed in _TARGET_MODULES (src/deepgram/transport.py:32). The Speak V2 websocket client shipped in 7.7.0 and binds the same websockets_sync_client / websockets_client_connect symbols the patcher targets, but deepgram.speak.v2.raw_client and deepgram.speak.v2.client were not added to the list.

Reproduction on main — installing a factory and checking whether each module's symbols are shim instances:

Module	sync patched	async patched
deepgram.speak.v1.raw_client	yes	yes
deepgram.speak.v1.client	yes	yes
deepgram.speak.v2.raw_client	no	no
deepgram.speak.v2.client	no	no
deepgram.listen.v2.raw_client	yes	yes

Why the existing tests didn't catch it

The transport tests iterate _TARGET_MODULES itself:

for mod_path in _TARGET_MODULES:
    ...assert patched...

That verifies the list is internally consistent, but it can't detect a websocket client missing from the list, because the missing entry is never iterated. So the suite passes whether or not the list is complete.

This is the part worth fixing beyond the two entries: without a completeness check, the next regen that introduces a websocket client reintroduces the same silent bypass.

Changes

src/deepgram/transport.py — add deepgram.speak.v2.raw_client and deepgram.speak.v2.client to _TARGET_MODULES, positioned after the Speak V1 pair to keep the existing grouping. Also corrects the now-stale count in the comment above the list.

tests/custom/test_transport.py — add a completeness guard that discovers websocket modules by scanning the package source, independently of _TARGET_MODULES, and asserts every discovered module is registered. A second small test asserts the scan actually finds a known module, so the guard can't pass vacuously if discovery ever returns nothing.

Two deliberate choices in the guard, both easy to change if you'd prefer otherwise:

• Static source scan rather than pkgutil.walk_packages — no import side effects, no dependence on sys.modules ordering, and it still sees a module that fails to import.
• Subset rather than set equality — asserting equality would fail if _TARGET_MODULES ever intentionally lists a module that doesn't exist yet. Subset targets the case that actually causes the bug: a real websocket client that isn't registered.

I kept this to the smallest change that fixes the defect and prevents recurrence. Happy to go further and replace the hardcoded list with runtime discovery if you'd prefer that direction — I left it out to keep the diff reviewable and because it changes behaviour in a file that's deliberately hand-maintained.

Verification

• The new guard fails on the unpatched tree, naming both Speak V2 modules; passes after the fix.
• 319 passed, 1 skipped across tests/custom and tests/utils (36/36 in test_transport.py).
• mypy src/ — clean, 882 files. mypy tests/typecheck — clean.
• Re-ran the reproduction above after the fix: both Speak V2 modules now patched for sync and async.
• Warning count is unchanged from main (2 pre-existing websockets deprecation warnings).

Not run locally: the WireMock-backed tests under tests/wire need Docker, which wasn't available in my environment, so I haven't executed them. Nothing in this change touches those paths, but flagging it rather than implying a full local run.