[diyuyi-agora]

No description provided.

[github-actions]

Review: qwen_asr_python

I reviewed this against docs/ai/L1/L2/asr_plugin_design_review.md. I found no MUST-level code violations in the extension. The items below need attention before merge.

1. This PR rewrites the review baseline it is judged against (process blocker)
docs/ai/L1/L2/asr_plugin_design_review.md and docs/ai/L1/L2/testing.md drop "A bounded retry strategy MUST escalate to fatal" and the "Retry ceiling reached → FATAL_ERROR" row, and make unlimited non-fatal retry the default. Under the current guide, _reconnect_loop (no ceiling) is arguably already allowed. Even so, changing the framework-wide ASR merge contract inside a vendor PR with an empty description skips the "framework contract change with migration" path the guide requires. Please split the doc change into its own docs: PR and get maintainer sign-off there. That PR should also explain how existing bounded-retry plugins relate to the new default.

2. No guarder or standalone test evidence (§13.2)
The PR description is empty. Please attach results from task test-extension EXTENSION=agents/ten_packages/extension/qwen_asr_python and task asr-guarder-test EXTENSION=qwen_asr_python CONFIG_DIR=tests/configs. The guarder two-cycle, reconnect, invalid-credential, and dump-equality checks matter most here because finalize ends the vendor task and starts a new one.

3. The audio consumer blocks on vendor round-trips under _swap_lock (performance, §6.2/§9.3)

• send_audio() takes _swap_lock for every frame (extension.py:306). On a session change or when no task is ready, it awaits client.finish_task() and client.start_task() inside that lock (_prepare_task_for_session_locked, extension.py:835-860). Each of those calls can wait up to connect_timeout_s (10 s).
• _finalize_impl holds _swap_lock while it awaits finish_task() for up to finalize_timeout_ms (extension.py:381-394). Every send_audio call queues behind it for that whole time.
• If the socket drops while one of these waits is in flight, _handle_close runs on the reader task and also waits for _swap_lock. Disconnect reporting and reconnect are then delayed until the 5–10 s timeout fires, instead of starting immediately.

This does not deadlock, but the pipeline stalls for several seconds at each turn boundary and on each failure. Two possible fixes: set a "task transition" flag under the lock and do the vendor wait outside it (frames are already rebuffered through _keep_buffer_frame), or have the client close path set _task_failed so waiters fail fast.

4. Minor

• _send_audio_locked computes duration as int(len / (sr/1000*2)) (extension.py:827-829). It ignores input_audio_channels()/input_audio_sample_width() (unlike _keep_buffer_frame), and truncating per frame lets the timeline drift for frame sizes that are not a whole number of milliseconds. Use the §6.4 formula and accumulate in float, or in bytes.
• In start_connection, buffered frames flushed to the vendor are dumped only when ready_to_report_connected is true (extension.py:237-248). If the generation check fails after a successful flush, that audio was sent but never dumped, which can break dump byte equality.
• manifest.json has no trailing newline.

ASR checklist

Item	Result
Lifecycle (super calls, init-failure latch, stop cancels tasks)	Pass
Connection state (connected only after task-started, generation guard, on_disconnected on failures)	Pass
Buffering (Keep mode, 10 MB bound, per-frame session metadata on flush, flush after recovery)	Pass
Finalize (finish-task/task-finished, timeout, always one asr_finalize_end with saved id/metadata, serialized)	Pass (see 3 for latency)
Reconnect (single owner, capped backoff, fatal latch, stop cancel, ConnectionClosedOK emits no error)	Pass under current guide; policy change needs separate review (1)
Result shape (metadata.session_id, asr_info, BCP-47 language, empty/heartbeat results dropped)	Pass
Metrics (connect delay after handshake, base TTFW/TTLW, cumulative timeline across tasks)	Pass (minor duration rounding, 4)
Tests (standalone tests cover finalize/reconnect/stale/redaction/timeline)	Standalone tests present; guarder evidence missing (2)