bug: Anthropic image input capabilities missing from integration configurations

[TykanN]

Client or integration

Other

Area

Catalog / models

Summary

This is a cross-integration capability-metadata defect, not an Aside/Pi/GJC-only bug. The built-in anthropic and anthropic-apikey providers omit modelInputModalities for their static Claude model catalog. All integration configuration paths consuming that metadata need auditing for omitted image-input declarations or their schema-equivalent capability fields.

Aside, Pi and GJC (internal client id gajae) were the initial reproduction examples, not the scope boundary. The audit covers all 15 registered config exporters plus Codex and Claude integration surfaces.

Two independent exporter gaps also need fixing: OpenClaw drops even declared input modalities (its omitted input defaults to ["text"]); Kimi CLI drops the documented capabilities: ["image_in"] flag. These two fixes must apply to all catalog-backed image models, not only Anthropic.

I observed that adding "image" to the input arrays for Fable, Opus, Sonnet and Haiku in my Aside configuration enabled image attachments in Aside chat. This manual observation is separate from the deterministic configuration-generation reproduction below; Pi/GJC live image requests have not been tested here.

Expected: every integration that declares model input capabilities must advertise image support for the known image-capable Claude models, using "input": ["text", "image"] or the integration's supported equivalent (such as modalities.input or a vision boolean), without manual client-file edits. Configurations with no per-model capability field must not receive invented unsupported fields. Unknown models should retain the existing conservative fallback, and explicit operator overrides should remain authoritative.

Reproduction

1. Use a bare built-in Anthropic provider config, with liveModels: false to reproduce offline:

   const config = {
     port: 10100, hostname: "127.0.0.1", defaultProvider: "anthropic",
     providers: {
       anthropic: {
         adapter: "anthropic", baseUrl: "https://api.anthropic.com",
         authMode: "oauth", liveModels: false,
       },
     },
   };

2. Call the production loadExportModels(config) from src/server/management/model-rows.ts, then buildClientConfig(client, { config, models, baseUrl: "http://127.0.0.1:10100/v1" }) for each of aside, pi, gajae.
3. Inspect document.providers.opencodex.models: Claude entries have input: ["text"] instead of ["text", "image"].

A regression added beside the existing bare-Anthropic reasoning test in tests/server/management-client-config-route.test.ts fails for all three clients. A registry-level test also fails because providerConfigSeed(entry).modelInputModalities?.[model] is undefined.

Version

Reproduced on upstream dev commit aa91958e3b050084e1edc07dcd66b05ef6eac604. Release main examined at 1cc89cf88c39160e4bb1ee21f0fd37404dd6930b (2.55.0). The installed version used for my manual Aside observation was not recorded.

Operating system

macOS / Darwin 25.6.0; Bun 1.4.2 (automated reproduction).

Provider and model

Built-in anthropic (OAuth) and anthropic-apikey (API key); Claude Fable 5.1/5, Sonnet 5/4.6, Opus 5/4.8/4.7/4.6, Haiku 4.5 static catalog.

Logs or error output

bun test tests/server/management-client-config-route.test.ts --test-name-pattern 'native Anthropic image input'
3 fail: aside, pi, gajae
Expected input: ["text", "image"]
Received input: ["text"]

bun test tests/providers/provider-registry-parity.test.ts --test-name-pattern 'Anthropic providers seed image'
1 fail
Expected: ["text", "image"]
Received: undefined

Screenshots and supporting files

Audit matrix (capability declarations, not a claim of live image-request testing in every application):

Integration	Image-input declaration / finding
Pi, Aside, Prime, omo, OMP, GJC, DSH	input arrays; consume shared catalog metadata
OpenCode V1 and V2, ZCode	modalities.input; OpenCode also attachment
Cline	modalities.input and supportsVision
Hermes	supports_vision
Raycast	abilities.vision.supported
OpenClaw	Additional missing field: model input; defaults to text-only when omitted
Kimi CLI	Additional missing field: model capabilities: ["image_in"]
Codex injected model catalog	input_modalities
Claude model discovery	capabilities.image_input.supported
Claude gateway cache / Desktop static config / MMX launcher	No image capability field in these connection/cache shapes; do not invent one
MiniMax Code	Current exporter has no image declaration; upstream support for such a field could not be established, so this remains an explicit evidence gap

• OpenClaw input schema, attachment gate, text-only default.

• Kimi CLI capability schema and image input gate.

• Official model documentation: https://platform.claude.com/docs/en/models/overview — "All current models support text and image input".

• Vision API documentation: https://platform.claude.com/docs/en/build-with-claude/vision

• Current registry: src/providers/registry/entries-core.ts has context windows and reasoning efforts for both Anthropic providers, but no input modalities.

• src/providers/derive.ts already fills modality declarations per model while preserving explicit overrides.

• src/clients/config-export/model-metadata.ts already allows both text and image; the text-only fallback for missing metadata is intentional, not an Anthropic-specific restriction.

• Related but different: fix(catalog,anthropic): keep Claude combo image/effort capabilities and honor provider output budget #3332 / feat(catalog,codex): carry Claude combo capabilities and give reset-credit redeems a stable identity #3474 addressed combo metadata fallback; fix(providers): advertise the reasoning-effort ladder for native Anthropic models #3454 added Anthropic reasoning metadata; fix(chat): accept Pi and Anthropic-shaped image parts on the chat wire #4497 / [agent] fix: normalize inbound Chat images before route selection and preserve an explicit reasoning disable #4534 addressed image wire conversion rather than exported capabilities. I found no exact existing report for this direct-provider export gap in all-state issue/PR searches.

Redacted configuration

{"id":"anthropic/claude-fable-5-1","input":["text"]}

Proposed fix: declare image input for the existing shared Anthropic model seed on both auth variants, audit every integration configuration writer, fix any remaining dropped image declaration, and cover every supported output shape through the registry-to-client-document chain. No blanket image claim for unknown models, and no unsupported fields added to client schemas.

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.

[lidge-jun]

리뷰 · 우선순위 60 / 80

이 이슈는 내장 anthropic / anthropic-apikey 정적 카탈로그에 modelInputModalities 시드가 없어서, 클라이언트로 내보내는 설정이 Claude 를 텍스트 전용(input: ["text"])으로 만든다는 메타데이터 버그다. 지금 dev(HEAD 9b711073a, 패키지 2.56.0)의 src/providers/registry/entries-core.ts 에서 두 Anthropic 엔트리는 models / modelContextWindows / modelReasoningEfforts 는 있지만 modelInputModalities 키가 없다. model-seeds.ts 에도 ANTHROPIC_MODEL_INPUT_MODALITIES 심볼이 없다. configuredInputModalities 가 비면 카탈로그·익스포터가 text-only/unknown 바닥으로 떨어지는 기존 계약과 맞다. Aside 설정에 "image" 를 수동으로 넣자 첨부가 됐다는 관찰과도 일치한다.

이슈 범위는 Aside/Pi/GJC 만이 아니다. capability 를 선언하는 익스포터 전반과, 선언이 있어도 버리는 OpenClaw·Kimi 구멍을 같이 본다. 현재 dev 의 buildOpenclawClientConfig 는 모델에 id/name/contextWindow 만 넣고 input 배열을 내지 않는다(OpenClaw 기본이 text-only). buildKimiClientConfig 는 capabilities: ["image_in"] 를 내지 않는다. Codex input_modalities·Claude discovery capabilities.image_input.supported 도 같은 시드에 의존하므로 뿌리는 레지스트리 한곳이다.

이미 고치는 초안 PR #4668 이 열려 있다(작성자 TykanN, Closes #4667). Anthropic 시드에 image 를 심고 OpenClaw/Kimi 익스포터를 고치며 capability-aware 익스포터 회귀를 넓힌다. grok-bot 은 #4668 에 별도 리뷰를 이미 남겼다(우선순위 67). 이 이슈는 재현 절차·감사 표·관련 이슈(#3332/#3474/#3454/#4497/#4534) 구분이 잘 되어 있고, 와이어 변환(#4497 계열)과 “내보내기 메타데이터”를 섞지 않았다. types.ts/config.ts 스플릿과 무관하다.

라인 / 경로 수준:

경로 src/providers/registry/entries-core.ts (anthropic / anthropic-apikey) - 현재 dev 에 modelInputModalities 부재. 이슈 재현의 뿌리이다
경로 src/providers/registry/model-seeds.ts - Anthropic 전용 input modality 시드 심볼 없음
경로 src/clients/config-export.ts buildOpenclawClientConfig / buildKimiClientConfig - 선언된 image 도 클라이언트 문서로 안 나간다(이슈의 “추가 구멍”)
경로 src/clients/config-export/model-metadata.ts - text|image 허용·누락 시 text-only 폴백은 의도적. Anthropic 시드만 채우면 Aside/Pi/GJC 쪽은 따라온다
경로 PR #4668 - draft. ready·포커스 스위트 초록 후 머지하면 이 이슈를 닫을 수 있다

메인테이너의 판단이 필요한 지점

• fix: preserve image input capabilities across integrations #4668 의 OpenClaw/Kimi 확장을 이번 이슈 범위에 같이 넣을지(이슈 본문은 같이 넣으라고 함)
• unknown 모델에 image 를 추론하지 않는 보수 폴백을 유지할지(권장: 유지)
• draft CI/환경 이슈(bubblewrap·SIGSEGV)를 메인테이너가 어느 선까지 재확인할지

너의 추천
새 PR을 또 열지 말고 #4668 초안을 ready 로 밀어서 이 이슈를 닫는 쪽을 권한다. 레지스트리 시드 + 익스포터 두 구멍이 이슈·PR·현재 dev 코드에서 같은 이야기다. 중복 이슈로 닫을 대상이 아니라 PR 대기 상태로 두면 된다. types/config 스플릿 때문에 닫을 이슈가 아니다.

이 댓글은 grok-bot이 작성했습니다