Skip to content

Friction log

Publication note (2026-09-12): this is a retrospective record of September 8 failures and recovery. The native-dispatch and SDK-harness fixes belonged to the historical working branch; this PR publishes documentation only and does not establish that those fixes are released. They require separate current-code review before operational reuse. Native evidence is retained without regrading.

  • Strict controlled output: Qwen3.5-9B returned 28 instead of the requested 32 code words in all ten unique-cache capacity requests. The canaries passed, but the requests are performance-ineligible. This is retained as a negative result; no speed claim is published.
  • Exact punctuation: Qwen3.5-9B answered Tea. to the spoken-suite memory item, while the exact validator expected tea. It is semantically correct, but the native artifact remains 33/36 strict and is not regraded.
  • Realtime cancellation fence: the original SDK capture saw one inflight transcript after a client cancel request. The historical dirty-worktree harness fenced cancellation at the standard response.done cancelled acknowledgement. Current main uses the client-send boundary; the change is excluded from this publication. Subsequent baseline, Qwen3.6, and deployed-TTS captures passed their own configurations; they do not constitute a same-model Qwen3.5 repeat. The original negative observation remains part of the campaign.
  • Managed proxy and readiness: the stopped proxy was repaired through the managed surface; all four jobs were adopted into the private service inventory, dependencies were ordered, and STT readiness uses /health. No public host identity, service label, PID, or command line is retained.
  • Native dispatch (historical code change, release status unverified): the working branch modified the native-recipe dispatcher to admit the Apple MLX lane without an NVIDIA GPU role while retaining the Docker GPU gate. Its historical tests are not validation of the current release. Operational reuse remains deferred pending separate code review and the native lifecycle regression gate.
  • Kokoro candidate startup: the first 0.8.2 candidate definition resolved relative assets from the repository root and failed warmup. A pinned definition with absolute paths passed; the failure and the deployed 0.8.2 receipt remain in native evidence. The update retained model/voice asset hashes and versioned rollback definitions.
  • Download and readiness discipline: native candidate downloads remain managed recipe inputs, and only declared service dependencies/readiness checks may establish a voice-path claim. The proxy dependency receipt and STT /health correction are retained privately; public evidence records the bounded post-fix outcome rather than an operator command transcript.

Open limitations include current-code validation of those historical fixes, corpus/perceptual audio coverage, and untested reboot, physical microphone/OpenClaw, and actual rollback execution. A successful historical retry is recovery evidence, not proof that an operational defect is closed in the current product.