fix(runway): make an initial session request when max_retry is zero

[coygeek]

Summary

APIConnectOptions(max_retry=0) is valid and means zero retries, but the Runway avatar plugin performs zero initial HTTP attempts and fails session startup immediately. In livekit-plugins-runway 1.8.4, _create_session iterates over range(max_retry). With zero retries, it skips the request and raises APIConnectionError("Failed to start Runway Avatar Session after all retries") even when a local transport is ready to return success.

Steps to reproduce

1. Create a disposable Python 3.12 environment and install livekit-agents==1.8.4 and livekit-plugins-runway==1.8.4.
2. Run the following minimized local reproduction. The private method and HTTP stub isolate the session-creation retry loop; all strings are unused synthetic placeholders, and the stub performs no network request.

import asyncio
from livekit.agents import APIConnectOptions, APIConnectionError
from livekit.plugins.runway import AvatarSession

class Response:
    ok = True
    async def __aenter__(self): return self
    async def __aexit__(self, *args): pass
    async def json(self): return {"id": "local-session"}

class FakeHTTP:
    def __init__(self): self.attempts = 0
    def post(self, *args, **kwargs):
        self.attempts += 1
        return Response()

async def main():
    for retries in (0, 1):
        http = FakeHTTP()
        avatar = AvatarSession(
            preset_id="cat-character", api_key="unused-local-placeholder",
            conn_options=APIConnectOptions(max_retry=retries),
        )
        avatar._http_session = http
        avatar._local_participant_identity = "local-agent"
        error = None
        try:
            await avatar._create_session("wss://unused.invalid", "unused-token", "local-room")
        except APIConnectionError as exc:
            error = str(exc)
        print({"max_retry": retries, "http_attempts": http.attempts, "error": error})

asyncio.run(main())

3. Observe max_retry=0 print http_attempts: 0 and the startup error. The max_retry=1 control prints http_attempts: 1 with no error.
4. The same zero-attempt behavior also occurs through public AvatarSession.start() with a local fake room, fake agent session, fake job context, explicit synthetic LiveKit arguments, and the same HTTP stub. The public-start path was exercised separately so the result is not limited to direct private-method invocation.

Expected behavior

Zero retries still permits one initial HTTP attempt. A successful first response starts the session. A failed initial request is surfaced without another attempt when max_retry=0.

Actual behavior

No POST is attempted with max_retry=0, and startup fails immediately. Positive values also count total attempts rather than retries, so the loop is inconsistent with the documented retry count.

Affected area

livekit-plugins/livekit-plugins-runway/livekit/plugins/runway/avatar.py, specifically _create_session, and its use of the shared APIConnectOptions contract.

Runtime or environment

Operating system: macOS arm64. Python 3.12.13; livekit-plugins-runway 1.8.4; livekit-agents 1.8.4; livekit 1.1.20; livekit-api 1.2.1; aiohttp 3.14.3. Models used: none; the reproduction needs no STT, LLM, TTS, provider account, or live session. Both LiveKit packages remain the current PyPI releases, and the current repository revision is 96341b0db2d0224403a36612cb04a1e38d60d405.

Evidence

• APIConnectOptions documentation and validation describe max_retry as the maximum number of retries and permit zero.
• The Runway creation loop iterates range(self._conn_options.max_retry) before its final error.
• Fresh executions of both the minimized creation loop and the public start path reproduced zero HTTP attempts for max_retry=0. The positive-count control reached the fake HTTP transport and succeeded.

Impact

Applications that disable retries cannot start a Runway avatar session, regardless of provider availability. The failure occurs before the initial request.

Additional context

A regression check should assert that zero retries makes exactly one attempt, that a successful initial response returns successfully, and that a retryable failure makes no additional attempt with zero retries. For a configured retry count N, the maximum attempt count should match the shared contract of one initial attempt plus at most N retries. Session, room, and call IDs are unnecessary because all reproduction objects are local.