[unicorn1125]

Summary

Adds GLM-Image text-to-image generation through a BigModel connection, available from both the canvas and Agent commands. Includes the provider preset, supported sizes, and Chinese and English setup guides.

The adapter uses BigModel's synchronous image endpoint, validates downloaded image bytes, and saves results as project assets and candidates. Credentials stay off CDN requests. Download recovery reuses the saved result URL, and uncertain submissions are never automatically reposted.

Linked issue

• Fixes [feat] providers: add Zhipu BigModel GLM-Image generation #237

Labels

• Maintainer access needed to apply type:feat, area:provider, area:canvas, area:agent, and modality:image.

Test plan

After rebasing onto current main (e0b411f):

• .venv/bin/python -m pytest tests -q: 1,384 passed, 6 skipped, including all 38 GLM-Image cases with mocked HTTP and real image bytes.
• node --test tests/*.test.ts: 699 passed.
• Build, bilingual checks, import contracts, generated contracts, and dependency checks passed.
• Installation and local startup passed during the original feature acceptance. The watermark review follow-up below adds two live calls after the original acceptance.
• Left to CI: maintainer-triggered Release checks. Linux, Windows, other browser engines, and conversational Agent orchestration remain unverified.

The detailed record is in docs/development/glm-image-validation.md.

Evidence that it works

• Screenshots attached.

Canvas result adopted into the project:

Canvas output:

Reopened Agent project, showing the model and prompt:

Agent output:

Provider verification

• Real-provider calls described below.

Tested with a real BigModel account: a 1280×1280 canvas image and a 960×1728 image through AgentTools. Both were adopted and verified after reopening their projects. The Agent check covers business tools, not conversational model orchestration. The original acceptance submitted three requests: two succeeded and one timed out with unknown acceptance. Two further requests verified the watermark setting off and on, bringing the total to five. Billing for the timed-out request was not verified.

Bilingual UI

• Both zh-CN and en-US are included for the preset, errors, and setup documentation.

Review follow-up

OpenAI compatible is first in the image provider dropdown again, with a regression check for its position. A Chinese/English note appears above Generate when BigModel is selected and Remove watermark is on, explaining the signed-waiver requirement and its location in the BigModel account settings.

Tested the same prompt at 1280×1280 with Remove watermark off and on. The actual request sent watermark_enabled=true and false respectively, matching the saved execution settings. Both returned images still contained a visible mark. This reproduces the reported behavior despite the correct request mapping. BigModel documents that disabling watermarks requires a signed waiver; the account's waiver status and removal for a signed account remain unverified. Provider output is preserved.

Follow-up validation: 80 selected Python tests and 700 frontend tests passed, along with the production build, i18n and generated-contract checks. Browser checks covered both languages, the note appearing/disappearing when the setting changes, and provider order.

[FayeSpica]

Thanks for adding GLM-Image support, and please check whether it respects the “Remove watermark” setting.

[FayeSpica]

Please keep OpenAI compatible first in the image provider dropdown.

[FayeSpica]

The generated GLM-Image output did not follow the Remove watermark setting in local testing. Please verify the output with the setting on and off.

[FayeSpica]

please fix conflicts