[Zhangbo-cn]

Summary

• Add Zhipu BigModel glm-tts as an audio provider on the existing connection-scoped voice library ([feat] audio: add a connection-scoped voice library with on-demand MOSS pagination #218): 7 documented preset voices plus the account's cloned voices. UI and agent tools share the same service for clone (upload → glm-tts-clone → delete sample → read back), select, synthesize and delete. Voice IDs are scoped to the provider connection.
• Only non-streaming 24 kHz WAV is requested, and only a RIFF body is accepted. JSON errors that arrive with HTTP 200 fail the job. BigModel 1113 maps to quota and 1214 "音色不存在" maps to voice_not_found. Format and sample rate are recorded in the execution snapshot.
• Text length (1024), recording type (mp3/wav), recording duration (3–30 s) and duplicate names are checked before any job or remote call is created.
• Zhipu is also added as a conversation-agent preset (models from /models). When no conversation model is configured, the assistant footnote now explains why Send is disabled and links to settings.
• The OFFICIAL voice list is not used: entries such as jam1 and streamer fail synthesis on the live API.

Linked issue

• Fixes [feat] providers: add Zhipu BigModel GLM-TTS and voice management #238 (child of [RFC] providers: add Zhipu BigModel for text, image, video, and speech generation #234). The lease problem found during manual testing is not caused by this change; it is tracked in [bug] 工程：租约未释放时刷新被当作其他窗口拒绝，英文提示误报存储繁忙 #277 and fixed in [bugfix] projects: reopen a reloaded tab's own lease and name the other-window conflict (#277) #279.

Labels

• Suggested: type:feat, area:provider, area:canvas, modality:audio. I can't apply labels from a fork.

Test plan

• Local:
  • pytest: 1611 passed, 98 skipped (includes 31 new tests in tests/test_zhipu.py).
  • npm test: 757 passed.
  • npm run check:i18n, npm run check:browser-manifest, npm run build and tsc --noEmit are clean.
  • tests/browser/zhipu_real.mjs passed against the real service in both locales. This script is a paid manifest helper and is not run in any lane.
• Left to CI: Release checks.

Evidence that it works

• Media attached

	zh-CN	en-US
Preset library
Preset narration result
Clone form
Cloned voice selected and synthesized
Delete confirmation
After delete
No conversation model: hint
Zhipu agent models fetched

Provider verification

• Real-provider calls are described honestly
• Real BigModel API (open.bigmodel.cn/api/paas/v4) with a funded key, run on 2026-11-10:
  • Preset narration (tongtong) and the full clone lifecycle (upload, create, select, synthesize, preview, delete) passed in both locales. Afterwards the account had no PRIVATE voices and no files.
  • /models discovery listed the GLM models. Agent connection tests succeeded with glm-4.5-flash and glm-5.3-flash.
  • In a manual agent run, glm-5.3-flash produced xiaochen narration through generation submit.
• Covered by mocked protocol tests only: error-code mapping (1000/1113/1214), HTTP-200 JSON errors, the non-RIFF body guard, and the preview-URL HTTPS guard.
• Not verified:
  • Similarity of the cloned voice. This needs a human listening check.
  • The AIGC WAV metadata chunk. BigModel always embeds it unless de-watermarking is enabled in the console, and it cannot be controlled from code.
• Known limits:
  • Two model configs that share one key count as separate connections. This is unchanged shared behavior.
  • BigModel keeps the clone-output preview file after a voice is deleted.

Bilingual UI

• zh-CN and en-US updated together, and both user guides are updated.