bug: Anthropic image input capabilities missing from integration configurations #4667

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):

• 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.