fix(zai): declare glm-5.3-flash image input on the Chat rows

[lidge-jun]

Summary

glm-5.3-flash is a native VLM, but the two Chat-path Z.AI rows described it only by what it is not. They list the text-only siblings in noVisionModels and declared no modelInputModalities at all, which keeps Flash out of the vision sidecar but never tells the catalog what the model can read. configuredInputModalities returned undefined, the catalog fell through to the ["text"] floor, and every client export — ZCode, Pi, OMP — listed a native multimodal model as text-only, so the picker refused to attach an image.

The Responses sibling row already declared this positively, so one registry described the same model two different ways.

This adds one shared positive declaration and points both Chat rows at it. The values come from the upstream catalog, not from inference:

GET https://api.z.ai/api/v1/models
  glm-5.3        input_modalities ["text"]
  glm-5.3-flash  input_modalities ["text", "image"]

and docs.z.ai/devpack/latest-model says the same in prose: "GLM-5.3 is a text-only model, so uncheck Support Images; GLM-5.3-FLASH is a multimodal model, so Support Images can be checked".

Nothing else moves. isModelVisionSidecarConsumer checks noVisionModels first and returns early, so adding ["text"] for models already on that list changes no behavior; the advertised value stays ["text", "image"] because the sidecar appends image. Flash is on neither path now, which is the point. The bracketed aliases are looked up by their exact catalog id rather than a stripped one, so they carry their own text-only entries. glm-4.6, glm-5 and glm-5.1 stay undeclared and keep their existing fallback.

Upstream also lists video and file for Flash. Neither the internal modality vocabulary (text / image / audio) nor the export vocabulary (text / image) can express them, so image is where this stops.

The vendored snapshot behind src/generated/model-metadata.ts has no glm-5.3-flash row at all. That file is byte-synced by tests/codex-integration/model-metadata-sync.test.ts and refreshing it is a separate deliberate commit; registry declarations win over that fallback, so the defect closes here.

Closes #4296.

Verification

• bun run typecheck — clean.
• bun test tests/providers/provider-registry-parity.test.ts tests/codex-integration/catalog-vision-sidecar-modalities.test.ts — 68 pass, 0 fail.
• No GUI change.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

No structure/ or docs-site/ update is required: neither states the Chat rows' modality, so neither becomes false. Planning unit: devlog/_plan/260912_zcode_protocol_and_catalog/020_wp3_glm53_flash_modalities.md.

Summary by CodeRabbit

• New Features

  • Added clearer input modality information for Z.AI GLM models.
  • GLM-5.3-Flash is now identified as supporting both text and image inputs.
  • Other listed GLM models remain identified as text-only.

• Tests

  • Added coverage to verify modality labels across Z.AI and BigModel Coding Plan model listings.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-11T18:45:21.368319Z	38a4bde	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

✅ Deterministic PR hygiene checks passed.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 55029eca-881a-46eb-9b8d-5c5c748d2749

📥 Commits

Reviewing files that changed from the base of the PR and between 1394b34 and 38a4bde.

📒 Files selected for processing (2)

• src/providers/registry.ts
• tests/providers/provider-registry-parity.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

The registry now declares positive input modalities for GLM 5.x models. glm-5.3-flash supports text and image input, while the other listed models remain text-only. Both Coding Plan entries and parity tests use these declarations.

Changes

GLM 5.x modality declarations

Layer / File(s)	Summary
Shared modality map src/providers/registry.ts	Lines 447–470 add ZAI_GLM_5X_INPUT_MODALITIES. The map declares glm-5.3-flash as ["text", "image"] and the other listed GLM 5.x models as ["text"].
Provider wiring and parity validation src/providers/registry.ts, tests/providers/provider-registry-parity.test.ts	Lines 2631 and 2713 apply the map to both Coding Plan entries. Lines 485–499 verify the modality declarations and confirm that glm-5.3-flash is not in noVisionModels.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix · Severity of issue fixed: Medium

Merge Risk: ⚪ Minimal · up to 38a4b

The metadata correction is narrowly scoped and verified; no actionable merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly identifies the primary change: declaring image input support for glm-5.3-flash on the Z.AI Chat rows. It is concise, specific, and consistent with the pull request objectives.
Linked Issues check	✅ Passed	Issue #4296 requires positive input modality metadata for both Z.AI Chat registry rows. src/providers/registry.ts adds the shared ZAI_GLM_5X_INPUT_MODALITIES declaration and applies it to the `zai…
Out of Scope Changes check	✅ Passed	The reviewed changes are limited to src/providers/registry.ts and tests/providers/provider-registry-parity.test.ts. They implement and protect the modality metadata required by issue #4296. The ch…
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 2…

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/glm53-flash-modalities

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 38a4bded07

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Document image support for the Chat presets

This changes user-visible catalog and client-export behavior for both zai and zhipu-bigmodel-coding: their pickers now permit native image attachments to glm-5.3-flash. The provider guide currently documents this capability only inside the separate BigModel Responses section, so users of either Chat preset have no corresponding documentation; update docs-site/ to describe the newly exposed Chat-path behavior.

AGENTS.md reference: src/AGENTS.md:L29-L29

Useful? React with 👍 / 👎.

[lidge-jun]

리뷰 · 우선순위 66 / 80

설명

이 PR은 Z.AI / BigModel Coding Plan의 Chat 경로 두 줄(zai, zhipu-bigmodel-coding)에서 glm-5.3-flash가 네이티브 VLM인데도 클라이언트 내보내기에서 텍스트 전용처럼 보이던 버그를 고칩니다. 이슈 #4296을 닫습니다.

지금 dev HEAD(1394b3408, #4285 Devin 착지 직후)를 보면 Chat 두 줄은 이미 noVisionModels: ZAI_GLM_5X_SIDECAR_VISION_MODELS로 Flash만 사이드카 목록에서 빼 두었습니다. 그래서 프록시는 Flash 이미지를 설명용 사이드카로 돌리지 않습니다. 그런데 modelInputModalities는 비어 있어서 configuredInputModalities(src/codex/catalog/provider-fetch.ts)가 undefined를 돌려주고, 카탈로그가 ["text"] 바닥값으로 떨어집니다. 그 결과가 ZCode·Pi·OMP 같은 export에 그대로 실리면 피커가 이미지 첨부를 거절합니다.

같은 레지스트리 안의 Responses 형제 줄(zhipu-bigmodel-responses)은 이미 modelInputModalities: { "glm-5.3": ["text"], "glm-5.3-flash": ["text", "image"], ... }로 긍정 선언을 해 두었습니다. 즉 한 파일에서 같은 모델을 Chat은 부정 목록으로만, Responses는 긍정 모달리티로 설명하고 있었습니다. 최근 착지한 #4300(export 쪽 per-model image 광고)과 같은 계열의 카탈로그 정합성 작업입니다.

고치는 방법은 짧습니다. ZAI_GLM_5X_SIDECAR_VISION_MODELS 옆에 공유 맵 ZAI_GLM_5X_INPUT_MODALITIES를 만들고, 사이드카 대상은 ["text"], Flash만 ["text", "image"]로 넣은 뒤 Chat 두 줄에 modelInputModalities로 꽂습니다. 근거는 upstream GET https://api.z.ai/api/v1/models와 계획 유닛 evidence JSON, docs.z.ai 문장과 맞습니다. video/file은 내부·export 어휘에 자리가 없어 이번 범위 밖입니다. vendored src/generated/model-metadata.ts에 Flash 행이 없는 것도 레지스트리 선언이 fallback보다 우선하므로 이번 커밋에서 고치지 않아도 결함은 닫힙니다.

테스트는 provider-registry-parity.test.ts에 Chat 두 id의 Flash/5.3/브래킷 alias 긍정 선언과 Flash가 noVisionModels에 없다는 점을 고정합니다. 범위가 레지스트리+테스트뿐이라 types/config 분리 캠페인과도 안 겹칩니다.

src/providers/registry.ts · ZAI_GLM_5X_INPUT_MODALITIES - 실질 결함은 보이지 않는다. 사이드카 목록을 text로 펼치고 Flash만 image를 더하는 구조가 Responses 줄·이슈 제안과 같다.
src/providers/registry.ts · zai / zhipu-bigmodel-coding - modelInputModalities만 추가되고 noVisionModels·effort·context 창은 그대로라 사이드카 경로와 광고 경로가 의도대로 갈라진다.
src/generated/model-metadata.ts - 여전히 glm-5.3-flash 행이 없다. PR이 말한 대로 별도 sync 커밋 대상이며 레지스트리가 이긴다. 다만 나중에 metadata sync를 돌릴 때 Flash를 빠뜨리면 다시 헷갈릴 수 있으니 후속 이슈/커밋으로 남겨 두는 편이 좋다.
Chat models 목록의 glm-5.1 / glm-5 / glm-4.6 - 이번 맵에 안 들어간다. 지금도 text 바닥값으로 떨어지고, PR이 의도적으로 손대지 않았다. VLM이 아니면 동작은 맞지만, 나중에 그 모델이 멀티모달이 되면 이번과 같은 부정-only 구멍에 다시 빠질 수 있다.
tests/providers/provider-registry-parity.test.ts - zai와 zhipu-bigmodel-coding만 고정한다. Responses 줄은 이미 다른 선언이라 중복 assert는 없어도 된다. 회귀 방지 핀으로는 충분하다.

메인테이너의 판단이 필요한 지점

• model-metadata.ts Flash 행 sync를 이 PR에 붙일지, 아니면 말한 대로 후속 커밋으로 둘지
• glm-5.1 / glm-5 / glm-4.6에도 지금 명시적 ["text"]를 넣을지(동작 동일, 선언만 더 분명해짐)
• #4296을 이 PR merge와 함께 바로 close할지(본문에 Closes Z.AI Chat rows advertise glm-5.3-flash as text-only, so client exports block image attachments on a native VLM #4296 있음)

너의 추천
merge 권장. 현재 dev 방향(카탈로그/export 모달리티 정합, #4300 직후)과 맞고, 범위가 Chat 두 줄의공유 맵·parity 핀뿐이라 리스크가 작다. typecheck와 언급한 테스트가 로컬에서 초록이면 바로 착지하고 #4296을 닫으면 된다. metadata sync와 구형 id 명시 선언은 후속으로 분리해도 된다.

이 댓글은 grok-bot이 작성했습니다

[lidge-jun]

Merging into dev under the maintainer integration path in MAINTAINERS.md (project owner, own PR, dev only).

Head merged: 38a4bded0798b181d58ca61065963eb4cd108766

CI at that head: all 25 checks green, including test 1/4–4/4, macos 1/2–2/2, the Windows and Ubuntu matrices, gates, hygiene and enforce-target. Codex review returned no findings; CodeRabbit reported "No actionable comments were generated in the recent review".

Local gates: bun run typecheck clean, and 69 tests pass across tests/providers/provider-registry-parity.test.ts, tests/codex-integration/catalog-vision-sidecar-modalities.test.ts and tests/codex-integration/model-metadata-sync.test.ts — the last one confirming the vendored snapshot stays byte-synced while this change deliberately leaves it alone.