fix(moa): route MoA provider through in-process facade in streaming path

[wangzengzhi]

The non-streaming code path (chat_completion_helpers.py line ~423) already
handles provider='moa' correctly by calling agent.client.chat.completions.create()
through the MoAClient facade. The streaming path in interruptible_streaming_api_call
was missing this check, causing it to build an HTTP client targeting the
virtual 'moa://local' URL, resulting in 404 errors on platforms with stream
consumers (Feishu, WeChat).

This fix adds the same provider=='moa' guard to the streaming path, making
MoA work consistently across both streaming and non-streaming scenarios.

[alt-glitch]

This was generated by AI during triage.

Related to closed streaming predecessors #57314/#57328 and the merged non-streaming facade fix #53826. This is an active reimplementation, not a duplicate of closed work. The unrelated lark-oapi version bump should be split or justified.

[teknium1]

Thanks for tracing the virtual-provider path.

Problems

• Current main already provides the reported streaming behavior: agent/chat_completion_helpers.py:2334 calls _create_request_openai_client(), and run_agent.py:4201-4202 returns the existing MoA facade for provider == "moa". This was added by 66aeda35501d4998fddbc87fde92b3c25d1b4486 (fix(moa): keep virtual provider on MoA client). The new direct streaming branch therefore duplicates the current shared routing behavior.
• The lark-oapi pin change is independent of this MoA fix. As noted in the existing review discussion, please split it into a focused Feishu dependency change or document its validation separately.

Suggested changes

• Drop the redundant MoA branch and retain the shared helper route.
• Split or independently justify the Feishu dependency update.

Automated hermes-sweeper review.

[teknium1]

Current main already reaches the MoA facade through the shared helper: _create_request_openai_client() returns the primary client for provider == "moa" (run_agent.py:4201-4202). This branch duplicates that behavior; keeping the common route preserves one request-client lifecycle path.

[teknium1]

This Feishu SDK update is unrelated to the MoA streaming change. Please split it into a focused dependency fix or add dedicated justification and validation in that scoped change.

[teknium1]

Closing — the fix this PR proposes is already on current main: _create_request_openai_client (run_agent.py) short-circuits provider == "moa" and returns the in-process facade (landed via #53826, commit 66aeda3), and the streaming path routes through it. Verified live on today's main. Note for future PRs: the branch also bundles an unrelated lark-oapi 1.6.8→1.6.9 bump in pyproject.toml/uv.lock — dependency changes should never ride inside a fix PR. Thanks for the report-with-repro regardless!