feat(kimi): consolidate Coding catalog and Responses preset

[luvs01]

Summary

Consolidate #5447 and #5429 into one Kimi Coding catalog change. Register the native k3-256k model and add an optional kimi-responses preset that shares the existing Kimi OAuth account while keeping Chat as the default. Preserve all five original contributions, including review follow-ups and author attribution.

• Use one four-model roster and capability seed across kimi, kimi-code, and kimi-responses; regenerate the metadata alias and verify cross-preset parity and tool-result adjacency.
• Add explicit K3 price namespaces using the published API reference rates and default 5-minute cache-write rate. These remain derived estimates, not Code Plan bills or quota; the 1M model and k3-256k consume different quota.
• Compatibility change: kimi-for-coding now serves K2.8 Preview, so its obsolete K2.7 price is removed from the existing presets and is not copied to Responses. Its cost remains unknown unless configured through modelCosts. Policies that exclude unknown prices or unknown cost under a cap may exclude this alias; allow policies and explicit user prices retain their existing behavior.
• Close the cost-cap fixture's SQLite history handle and clear its cached health evidence before deleting its temporary home. This resolves the reproduced Windows cleanup failure without raising timeouts or skipping cases.

Pricing/model identity was checked on 2026-09-23 against Kimi API pricing and Kimi Code model configuration. No K2.8 price or actual Code Plan billing equivalence is claimed.

Verification

• Exact local head: c86bc00b531dba6f16f15b52954d16be20b2adc1; base dev b1b60e880c61fbea12eed326ac43529cc393fb15. The last commit adds only the unchanged screenshot below; runtime, GUI source, docs and tests remain identical to 5ebd93d.
• At 7e547bf02d8215f1a195a033158cdf0558a7aa8c, the unchanged wrapper over seven explicit provider/catalog/usage/cost-cap/metadata-sync/adjacency files completed 562 passed, 0 failed, 5,616 assertions, 36.10 seconds. Earlier failed runs were retained and are not included in this total. The dev merge to 40e468c changed only nine upstream CI/layout/documentation files; the Kimi contribution files, runtime source and test harness at that point remained byte-identical to the tested head. The subsequent correction below changes documentation totals and brand aliases, not the provider behavior.
• On the preceding 40e468c head, five explicit CI/workflow/layout files completed 173 passed, 8 skipped, 0 failed, 2,070 assertions, 4.14 seconds. The existing platform/dependency-dependent shell cases were skipped, not counted as passing.
• Typecheck, structure SSOT, privacy scan, file-size ratchet and English/Chinese docs build (497 pages, existing dependencies) passed before the CI-only dev refresh. Whitespace validation passed on the final head.
• Hosted CI on 40e468c found two missed integration obligations: translated preset totals and the Responses preset brand alias. 5ebd93d updates the 17 anchored guide/quickstart/structure statements to the registry-derived 97 total / 80 key / 13 OAuth split and wires the existing Kimi brand asset and label. The two failing test files now pass locally: 24 passed, 0 failed, 61 assertions. The production browser build and 497-page docs build also passed after this correction, using existing dependencies.
• Cross-platform CI at 5ebd93d completed successfully, including all four Linux shards, gates, docs and aggregate CI. Native/macOS jobs and Windows diagnostic shards were skipped by that run's path policy; this is not a full Windows/macOS test claim.
• The existing screenshot of the real ProviderRail/ProviderIcon fixture was independently checked from its saved PNG pixels and matching DOM. All three Kimi marks and names/aliases are visible; the third mark has a complete 20x20 shape matching the first two. Source hashes match the final GUI source. The original composite viewport assertion failed, its exact failed subcondition remains unknown, and it was not rerun. This is bounded synthetic UI evidence, not live-provider or full-app interaction validation.
• Local full-suite completion, screenshot-only head hosted CI, live provider/billing verification and security review remain pending. No installation, test timeout/baseline increase or runtime setting change. This PR stays draft.

UI evidence

Actual provider rail rows for kimi, kimi-code, and selected kimi-responses, using the shared Kimi brand mark. PNG SHA256: 5977ce36680a496517cb7b50b9c734d3836d9cdb69694c79b1a708987e77e19b.

The two original contributor-owned PRs remain open for maintainer coordination.

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.

Co-authored-by: panyuanyuan panyuanyuan@hetao101.com

[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

---

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.

[lidge-jun]

리뷰 · 우선순위 58 / 80

이 PR은 Kimi Coding 쪽 두 갈래(#5447의 k3-256k 등록, #5429의 kimi-responses 프리셋)를 dev 위에 하나로 합칩니다. 피커와 시드에 네이티브 k3-256k(262K 천장)를 넣고, 선택용 kimi-responses는 같은 Kimi OAuth(oauthId: kimi)로 Responses 선을 탑니다. Chat(kimi)·API 키(kimi-code)·Responses가 같은 네 모델 명단과 능력 시드를 쓰고, 가격 네임스페이스는 프리셋마다 따로 둡니다. kimi-for-coding이 K2.8 Preview로 바뀐 뒤 옛 K2.7 추정가는 빼서, 가격을 모르는 한도/정책은 이 별칭을 막을 수 있습니다. K3 가격은 API 참고용(기본 5분 캐시 쓰기)이고 Code Plan 청구·쿼터와 같다고 말하지 않습니다. Windows에서 cost-cap 픽스처가 임시 홈을 지울 때 SQLite·헬스 캐시를 먼저 닫도록 고쳤고, 문서·구조 노트·패리티/가격 테스트도 맞춰 두었습니다. 초안이며 전체 스위트·호스트 CI는 아직입니다.

src/providers/registry/entries-core.ts (kimi-responses.promptCacheKey) - 레지스트리에 promptCacheKey: true를 켜 두었지만, EN/zh 문서는 “Responses에는 지금 안 보낸다”고 적습니다. 주석대로 Chat과 시드를 맞추려는 의도라면, 나중에 보내는 사람이 플래그만 보고 “이미 동작한다”고 착각하기 쉽습니다.

docs-site/.../providers.md vs zh-cn/.../providers.md - 영어는 구독/키 비교 표에 kimi-responses를 넣었고, 중국어는 카탈로그 표·가격 절만 고쳤습니다. 뜻은 통하지만, 표 위치가 달라서 한쪽만 고친 것처럼 보일 수 있습니다.

#5447 / #5429 - 둘 다 아직 열려 있고, 이 PR이 그 내용을 흡수한 통합본입니다. 그대로 두면 같은 카탈로그 변경이 세 갈래로 남습니다.

메인테이너의 판단이 필요한 지점

kimi-for-coding 추정가를 비워 두는 호환 변경을 지금 받아들일지(unknown-cost 정책이 별칭을 막을 수 있음). promptCacheKey를 Responses에 실제로 보내기 전까지 플래그를 둘지 말지. 머지 시점에 #5447·#5429를 닫을지(작성자도 조율이 필요하다고 적음). 초안·전체 스위트 미완을 머지 전 조건으로 둘지.

너의 추천

방향은 유지하세요. 머지 전에 #5447·#5429를 무효/중복으로 닫을 계획만 잡고, promptCacheKey는 “지금은 안 보냄”이 코드/주석/테스트에 한눈에 드러나게 맞추세요. kimi-for-coding 미가격은 문서·테스트와 같으니, 릴리스 노트에 unknown-cost 영향만 짧게 적으면 됩니다. 초안이 풀리고 합의된 게이트가 초록이면 통합본으로 가져가도 됩니다. 프리뷰 배포 이야기는 이 범위 밖입니다.

이 댓글은 grok-bot이 작성했습니다

[github-actions]

✅ READY

• all PR quality gates passed.

Hygiene

✅ Deterministic PR hygiene checks passed.

[lidge-jun]

Superseded by #5609, which is open and not yet merged. k3-256k, the kimi-responses preset, the price alignment, the brand alias and the docs (7db37ed, 63d3f19, 351f507, 63755ef, 82ff04d, 7e547bf, 5ebd93d, c86bc00) are carried there as one commit crediting panyuanyuan and you as co-authors, with preset totals recounted from the registry. The K2.8 alias and migration work is already on dev via #5403 (600e323). The source PRs #5447 and #5429 are not closed here; that is left to the maintainers. Thank you!