fix: preserve image input capabilities across integrations

[TykanN]

Summary

Closes #4667

Fix image-input capability loss across integration configurations, not only the initially reported Aside/Pi/GJC examples.

• Add shared modelInputModalities for the nine known Claude model seeds to both anthropic and anthropic-apikey. Missing source metadata previously triggered text-only/unknown export defaults; existing per-model overrides remain authoritative.
• Fix two additional exporters that dropped even declared capabilities: OpenClaw now emits schema-filtered input; Kimi CLI emits capabilities: ["image_in"] for catalog-backed image input. These exporter fixes apply to any declared image-capable model, not just Anthropic.
• Add production-catalog regressions for all 14 capability-aware registered exporters: Pi, Aside, Prime, omo, OMP, GJC, DSH, OpenCode V1/V2, ZCode, Cline, Hermes, Raycast, OpenClaw and Kimi. Also cover Codex input_modalities and Claude discovery capabilities.image_input.supported, including Desktop/readable aliases and [1m] variants, for both Anthropic auth flows.
• Preserve unknown-model defaults and existing client schema/eligibility constraints. Do not infer vision from arbitrary model names or overwrite explicit text-only settings. Update public docs and mapped architecture notes.

Scope and evidence boundary

All 15 registered exporters and the additional Codex/Claude/MMX surfaces were audited. The remaining registered exporter, MiniMax Code, exposes no image field in the current local contract, and an authoritative upstream field could not be established; no unsupported field was invented. Connection-only/cache shapes (Claude gateway cache, Desktop static config, MMX launcher) are unchanged. This is not a blanket claim that every application supports images or that every future/date-pinned Claude alias has verified metadata.

Official references: Claude models, OpenClaw schema, OpenClaw text-only default, Kimi capabilities.

Related prior work: #3474 fixed combo metadata; #3454 fixed Anthropic reasoning advertisement; #4497/#4534 fixed image wire handling. None replaced this cross-integration capability audit.

Verification

• Red before fixes: six initial export cases (three clients × two auth flows) failed with input: ["text"]; registry assertion returned undefined. OpenClaw and Kimi each separately failed for missing supported capability fields after the registry fix.
• Passed: focused provider, configuration and all-client export checks — 164 tests across five files; Codex/Claude production-catalog suites — 54 tests, 262 assertions.
• Passed: bun run typecheck, bun run privacy:scan, bun run structure:check, git diff --check; cd docs-site && bun install --frozen-lockfile && bun run build — 441 pages (nonblocking build warnings).
• Broader integration run: bun test tests/config tests/clients tests/integrations — 1178 pass, 3 skip, 2 fail. Both failures are in unchanged remote-workspace-command-runner.test.ts: the local Bun executable/ancestor path fails bubblewrap executable and parent directories must not be group or world writable. Reproduced independently; no trust check weakened.
• Changed-suite limitation: bun run test:changed selected 965 files but a Bun 1.4.2 parallel worker crashed with SIGSEGV in routing-policy-surface-parity.test.ts, aborting the remaining files. That file alone passed (6 tests). The --parallel=1 retry produced no test progress for several minutes and was stopped. Full-suite/local-CI green is not claimed.
• Actual successful image attachment in Aside after a manual config edit was reported by the issue author. This PR's tests verify generated capability declarations, not paid live model requests in every client.

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 authentication, credentials, routing or image-wire behavior changed; privacy scan passed. Submitted as draft per contributor policy and because broad validation is not fully green.

Review readiness

• All CI tests are green on my local testing.

• I pushed my PR to the latest dev commit.

• I resolved all correct Codex and CodeRabbit findings.

• My PR is ready for review.

Summary by CodeRabbit

• New Features

  • Claude models now advertise text and image input capabilities for Anthropic OAuth and API-key configurations.
  • Generated OpenClaw and Kimi Code configurations now include image-input capability metadata when supported.
  • Explicit model overrides remain respected, while unknown models receive no inferred capabilities.

• Documentation

  • Added guidance covering Anthropic image support, client configuration refresh requirements, and capability metadata across integrations.
  • Clarified that image capability metadata does not change xAI behavior or transport handling.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

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: 91d7f7c8-fcbe-4827-a4b0-b004d23a7307

📥 Commits

Reviewing files that changed from the base of the PR and between aa91958 and a9b3654.

📒 Files selected for processing (15)

• docs-site/src/content/docs/guides/providers.md
• src/clients/config-export.ts
• src/providers/registry/entries-core.ts
• src/providers/registry/model-seeds.ts
• structure/clients/integrations.md
• structure/providers/xai-grok.md
• structure/runtime.md
• structure/subagents.md
• structure/transports/inventory.md
• tests/claude-integration/claude-model-info.test.ts
• tests/codex-integration/catalog-input-modality-enum.test.ts
• tests/config/client-config-export-new-clients.test.ts
• tests/config/client-config-new-clients.test.ts
• tests/providers/provider-registry-parity.test.ts
• tests/server/management-client-config-route.test.ts

Included review availability: Your plan provides up to 10 included reviews per hour; 9 remain after this review.

---

📝 Walkthrough

Walkthrough

The change adds image-input metadata to Anthropic model seeds, propagates it through the catalog, exports it in supported client formats, and adds registry, catalog, discovery, and management-route coverage for both Anthropic authentication modes.

Changes

Anthropic capability metadata

Layer / File(s)	Summary
Registry seeds and catalog propagation src/providers/registry/model-seeds.ts, src/providers/registry/entries-core.ts, tests/providers/provider-registry-parity.test.ts, tests/codex-integration/*, tests/claude-integration/*	Known Claude seeds now declare ["text", "image"] for both Anthropic providers. Explicit overrides remain authoritative, and unknown models remain undeclared. Catalog and Claude discovery tests validate the propagated metadata.
Client capability export rules src/clients/config-export.ts, tests/config/*	OpenClaw exports supported, deduplicated catalog modalities. Kimi Code exports capabilities: ["image_in"] only for models declaring image input.
Management configuration output coverage tests/server/management-client-config-route.test.ts	The production export path now has coverage for image-capability fields across supported client configuration formats for OAuth and API-key Anthropic providers.
Capability metadata documentation docs-site/src/content/docs/guides/providers.md, structure/clients/integrations.md, structure/runtime.md, structure/subagents.md, structure/transports/inventory.md, structure/providers/xai-grok.md	Documentation describes Anthropic seed metadata, client-specific output fields, conservative handling of unknown models, and unchanged transport and eligibility behavior.

Priority: ➖ Normal

Estimated code review effort: 3 (Moderate) | ~25 minutes

Change: Bug fix · Severity of issue fixed: Medium

Sequence Diagram(s)

sequenceDiagram
  participant AnthropicProvider
  participant ModelCatalog
  participant ConfigExporters
  AnthropicProvider->>ModelCatalog: declare text and image modalities
  ModelCatalog->>ConfigExporters: provide input modalities
  ConfigExporters-->>AnthropicProvider: emit client-specific capability fields

Loading

Merge Risk: ⚪ Minimal · up to a9b36

The capability metadata and supported client exports are covered for both Anthropic authentication modes, including override and unknown-model behavior. No actionable merge risk remains.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 3 functions across 9 files. (6 skipped: 6 …	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely summarizes the main change: preserving image-input capability metadata across Anthropic provider exports and integrations.
Linked Issues check	✅ Passed	Issue #4667 coding requirements are met. src/providers/registry/model-seeds.ts adds ANTHROPIC_MODEL_INPUT_MODALITIES for all nine ANTHROPIC_MODELS entries with ["text", "image"]. `src/provider…
Out of Scope Changes check	✅ Passed	The changed source files implement the capability path required by issue #4667: shared Anthropic seed metadata, registry propagation, OpenClaw serialization, and Kimi serialization. The changed tests …

Full details: Docstring Coverage

Explanation

Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 3 functions across 9 files. (6 skipped: 6 unsupported.)

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[github-actions]

⏳ DRAFT

• review readiness checklist open (1/4 boxes ticked).

What to do

• Tick all four boxes in the PR description once you're done (currently 1/4).

Review readiness checklist

• ⬜ All CI tests are green on my local testing.
• ✅ I pushed my PR to the latest dev commit.
• ⬜ I resolved all correct Codex and CodeRabbit findings.
• ⬜ My PR is ready for review.

1/4 boxes ticked.

This PR stays in draft until every box above is ticked.

[lidge-jun]

리뷰 · 우선순위 67 / 80

이 PR은 초안(draft) 이고, 이슈 #4667을 닫으려는 카탈로그·익스포터 버그 수정이다. 지금 dev(HEAD aa91958e3, 패키지 2.56.0)에서 anthropic / anthropic-apikey 레지스트리 엔트리는 모델 목록·컨텍스트·effort 사다리는 있지만 modelInputModalities 시드가 없다. configuredInputModalities 가 비면 카탈로그가 ["text"] 바닥으로 떨어지고, Aside/Pi/GJC 같은 클라이언트가 Claude를 텍스트 전용으로 내보내 이미지 첨부가 막힌다. 이슈 작성자가 Aside 설정에 "image" 를 수동으로 넣자 첨부가 된 관찰과도 맞다.

고침은 세 층이다. 첫째, model-seeds.ts 에 ANTHROPIC_MODEL_INPUT_MODALITIES 를 두고 아홉 Claude 시드에 ["text","image"] 를 심은 뒤, entries-core.ts 의 OAuth·API-key 양쪽 엔트리에 붙인다. enrichProviderFromRegistry 는 fill-only 라서 운영자가 이미 claude-sonnet-5: ["text"] 처럼 적어 둔 값은 덮어쓰지 않고, 없는 키만 채운다. 둘째, OpenClaw 익스포터가 선언된 inputModalities 를 스키마가 아는 text|image|video|audio 만 골라 input 배열로 내보낸다(없으면 필드 생략 — OpenClaw 기본이 text-only로 떨어지는 구멍). 셋째, Kimi 익스포터가 카탈로그에 image가 있을 때만 capabilities: ["image_in"] 를 낸다. 예전에 “capabilities를 절대 안 낸다”던 주석·테스트는 이 계약에 맞게 바뀌었다.

회귀는 넓다. 레지스트리 패리티, OpenClaw/Kimi 단위, management 경로에서 capability-aware 익스포터 14종(Aside/Pi/Prime/omo/OMP/GJC/DSH/OpenCode/ZCode/Cline/Hermes/Raycast/OpenClaw/Kimi), Codex input_modalities, Claude discovery capabilities.image_input.supported(readable·desktop 별칭과 [1m] 포함)까지 본다. MiniMax Code는 로컬 계약에 image 필드가 없어 가짜 필드를 만들지 않았다고 문서에 적어 두었다. docs-site providers 가이드와 structure/clients/integrations.md / runtime.md 도 같은 이야기를 한다. xai-grok·subagents·transports 쪽 한두 줄 링크는 runtime 앵커를 가리키는 structure 교차 참조로 보인다(구조 SSOT 패턴).

types.ts / config.ts 모놀리스 본문을 고치지 않는다. close-don't-rebase 대상이 아니다. #3474·#3454·#4497/#4534 와 겹치는 “다른 층”을 PR 본문이 잘 구분했다. 다만 작성자가 스스로 full-suite/local-CI 초록을 주장하지 않았다. bubblewrap 권한 실패 2건은 이 PR과 무관해 보이고, test:changed 병렬 SIGSEGV는 환경 이슈로 적혀 있다. draft 체크리스트도 아직 비어 있다(최신 dev push만 체크됨).

라인 / 경로 수준 주의 포인트:

src/providers/registry/model-seeds.ts - ANTHROPIC_MODEL_INPUT_MODALITIES 가 ANTHROPIC_MODELS 전 원소에 vision을 준다. 공식 overview와 맞추려는 선택이다. 시드에 새 모델이 들어오면 자동으로 따라가므로, 나중에 text-only Claude가 생기면 예외 목록이 필요해진다
src/providers/registry/entries-core.ts (anthropic / anthropic-apikey) - 시드 spread가 양쪽 auth에 동일하게 들어갔다. OAuth·키 비대칭 구멍은 없다
src/clients/config-export.ts buildOpenclawClientConfig - pdf 같은 비지원 modality는 버리고, 지원 목록만 dedupe한다. 빈 배열이면 input 키 자체를 생략한다(OpenClaw 기본 text-only를 피하려는 의도)
같은 파일 buildKimiClientConfig - image가 있을 때만 image_in. 다른 capability는 이름에서 추론하지 않는다. 주석·테스트가 예전 “never emit”에서 이 계약으로 바뀌었는지 리뷰어가 한 번만 보면 된다
structure/runtime.md - “modelInputModalities: ["text", "image"] per model” 표현은 실제 타입이 모델 id → 배열 맵인데 문장이 조금 단순하다. 의미는 통한다
tests/server/management-client-config-route.test.ts - production loadExportModels 로 14 익스포터를 한 번에 밀어 버그 재발을 잘 막는다. 초안 상태에서 CI 전체 초록은 아직 작성자 주장 밖이다

메인테이너의 판단이 필요한 지점

• draft를 ready로 올리기 전에 작성자가 말한 bubblewrap 2실패·SIGSEGV와 무관함을 메인테이너가 어느 선까지 재확인할지
• OpenClaw/Kimi 익스포터 변경이 Anthropic만이 아니라 “카탈로그에 image가 있는 모든 모델”에 적용되는데, 그 확장을 이번 PR에 같이 넣을지(이슈 본문은 의도적으로 같이 넣으라고 함)
• Claude discovery 테스트가 생성된 모든 alias(포함 [1m])에 image_input.supported === true 를 요구하는 것이 장기적으로 맞는지
• structure 교차 링크(xai-grok/subagents/transports)를 최소 runtime+integrations만 남길지, 지금처럼 넓게 둘지(structure:check가 통과했다면 유지해도 됨)

너의 추천
초안이므로 지금 머지하지 말고, 체크리스트를 채운 뒤 ready로 전환하는 쪽이 맞다. 뿌리 원인(Anthropic 시드에 modelInputModalities 부재 + OpenClaw/Kimi가 선언조차 버리던 구멍)은 dev 코드와 이슈 재현이 일치하고, fill-only·unknown 미추론·명시 오버라이드 보존은 기존 카탈로그 불변식과 맞다. ready 전에 (1) 포커스 스위트만이라도 다시 초록 확인하고 (2) runtime.md 한 줄 표현만 다듬으면 충분하다. types/config 스플릿 때문에 닫을 PR이 아니다. 충돌 없으면 유지·rebase 후 ready 전환을 권한다.

이 댓글은 grok-bot이 작성했습니다