feat(web-search): add Kimi search with shared chat credentials (opt-in)

[wyuc]

Summary

Add Kimi web search using the existing Kimi chat account's API key and base URL, with explicit opt-in. This AI-assisted PR implements the web-search portion of #1909.

Moonshot Open Platform accounts use /tools/search_pro and combine page chunks, falling back to snippets. Coding Plan accounts use /search, enable page crawling, and prefer page content over snippets. The adapter uses Bearer authentication, a 30-second server timeout, the shared 400-character query limit, at most 20 results enforced locally for both endpoints, and proxy-aware fetch with caller cancellation.

Configuration and review corrections

• Opt-in: A Kimi chat key alone creates no legacy search entry, search default, or operator-preferred backend. Existing SearXNG/Tavily defaults and user Tavily requests remain intact. Explicitly assign webSearch in openmaic.yml or workspace settings, add web-search.kimi in server-providers.yml, or set WEB_SEARCH_KIMI_ENABLED=true to opt into legacy search with shared credentials.
• Disable switch and model pins: WEB_SEARCH_KIMI_ENABLED=false force-disables search. WEB_SEARCH_KIMI_ENABLED is the only Kimi search switch. KIMI_MODELS remains chat-only.
• Shared accounts: Legacy search entries cannot overwrite chat credentials, endpoint, or model pins. Conflicting accounts produce a warning and are skipped during translation/default assignment. Separate accounts must use separate provider ids in openmaic.yml.
• Wizard: It checks the same official account-base list used by server search validation and does not recommend search for a custom workspace chat proxy, even if a view offers search capability.
• Documentation: Restore the existing localized Coding Plan names, append localized search capability names, translate the new supported-provider sections in five non-English locales, and restore Arabic list separators. Configuration additions remain English to match their surrounding sections.
• File-backed configuration: When openmaic.yml exists without a webSearch assignment, a loaded legacy Kimi search opt-in produces a specific one-time startup warning telling the operator to declare the provider and assign slots.webSearch. The docs include this requirement in every maintained locale.
• Proxy paths: After removing trailing slashes, a base path ending in /coding/v1 selects /search; otherwise the adapter selects /tools/search_pro. Deployment-configured proxies must preserve this path layout.

Opting in on a deployment with no other search provider makes Kimi the default search provider for all workspaces.

The existing switch dispatch remains. Existing search defaults retain precedence after Kimi opt-in; webFetch and published packages are unchanged. Existing Kimi labels and logo are reused across settings; the 12 settings locales already contain the label.

Verification

All commands run with Node 22.22.0 and the required environment cleanup. Dependencies were installed with pnpm install; standalone render-service test dependencies remain installed from CI's PUPPETEER_SKIP_DOWNLOAD=true npm ci step.

Command	Result
pnpm check	Passed
pnpm lint	Passed; 0 errors, 18 warnings in unrelated files
pnpm exec tsc --noEmit	Passed
pnpm check:i18n-keys	Passed; 12 locales
pnpm exec vitest run tests/web-search tests/config tests/model-settings tests/server tests/providers	2,116 tests passed, 141 skipped; 155 files passed, 7 skipped
pnpm test	11,234 tests passed, 243 skipped; 929 files passed, 29 skipped
pnpm build	Passed; Edge Runtime and middleware warnings

Regression coverage includes both API contracts and result mapping; query/result limits, including 19 returned results for a request limited to 3 on both endpoints; malformed/HTTP errors and cancellation; chat-only Kimi; Kimi plus SearXNG/Tavily; user Tavily requests under operator preference; explicit env/YAML opt-in; no activation without credentials; the dedicated search switch; opted-in Kimi preserving existing SearXNG/Brave/Doubao defaults while remaining assignable; chat pins; conflicting legacy accounts; wizard official/proxy bases; and the ignored file-backed opt-in warning, its once-only logging, and absence for assigned/disabled slots or inactive opt-ins.

Self-review checked opt-in activation and both default-selection paths, account collision handling, credential/base-URL precedence, endpoint validation, locale consistency, and scope. All local automated Kimi requests are mocked; this implementation run made no live authenticated calls. No open design questions remain for this scope.

Maintainer live verification

Open Platform search was verified end-to-end by a maintainer with a real key against api.moonshot.cn: two full course generations used current search results in their outlines, and page chunks provided substantial body content. The maintainer also verified the legacy regressions: a chat key alone leaves search disabled; explicit opt-in with no other search provider makes Kimi the default; an existing Tavily provider keeps precedence. With openmaic.yml, declaring the Kimi provider and assigning slots.webSearch enabled search.

Coding Plan /search response fields were verified with a real key: content, date, icon, mime, site_name, snippet, title, and url. The endpoint ignored limit: 3 and returned 19 results; the adapter therefore enforces the requested result cap locally. This verifies the Coding Plan response contract, rather than a full Coding Plan course-generation run.

International Open Platform search Pro availability remains unverified.

Related #1909

[CXuPercy]

Live-tested with a real Moonshot Open Platform key (api.moonshot.cn).

API contract: /tools/search_pro responses match the adapter's mapping exactly; search_results[].chunks[].text joins into substantial body content (6k+ chars for a single result).

End-to-end (two runs, both full course generations):

• Topic "2026 年 Kimi K3 大模型的最新进展与能力评测" (agent-driven test): search returned 10 highly relevant sources (Xinhua, TMTPost, Tencent Cloud, CSDN, …); the outline contained facts that only exist in that week's search results (release date, WAIC context, 2.8T parameters), and one scene title came verbatim from the TMTPost review's headline.
• Topic "我想通过一节课学习2026的诺贝尔奖的相关情况" (manual test): all 10 returned sources were published within the past few days (People's Daily Online / Xinhua: the three Physiology-or-Medicine laureates, Physics laureate Francis Halzen, …); the generated outline accurately stated the 2026 winners and dates (manually cross-checked).

Legacy regressions: KIMI_API_KEY alone (switch off) → 0 search providers loaded, search stays disabled; WEB_SEARCH_KIMI_ENABLED=true with no other search provider → Kimi becomes the default; with a Tavily key already present, the default remains Tavily.

One configuration-interplay observation (repro included), as a supplement: with an openmaic.yml present that assigns no webSearch slot, setting KIMI_API_KEY + WEB_SEARCH_KIMI_ENABLED=true does load 1 WebSearch provider server-side, yet GET /api/generate-classroom/capabilities still returns webSearch: false and course generation skips search; the only log line is the generic "openmaic.yml is present, so slot resolution uses it and not the legacy provider variables" warning — nothing Kimi-specific. Adding providers.kimi (preset: kimi) + slots.webSearch: kimi fixes it. Per the PR description this is by design; sharing it here as a supplement.

[wyuc]

Thanks for the thorough live test.

• The openmaic.yml interplay is now explicit: when openmaic.yml is present and no webSearch slot is assigned, setting WEB_SEARCH_KIMI_ENABLED=true logs a one-time warning that names the fix (declare a kimi provider and assign slots.webSearch). The docs say the same in all locales (bdd8d90).
• Coding Plan: I checked POST https://api.kimi.com/coding/v1/search with a real Coding Plan key. The response fields match the adapter (search_results[].content, snippet, title, url, …), but the endpoint ignores limit (it returned 19 results when asked for 3). Both endpoints now truncate to the requested count, and there are tests for it.

[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 2026-10-16T06:21:57.596039Z	522d59d	Draft marked ready

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.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 522d59da48

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]

Preserve search for official Kimi alternate URLs

When a workspace user selects an official non-default Kimi endpoint such as https://api.moonshot.ai/v1, modelSettingsView classifies every stored chat baseUrl as custom and exposes that provider as chat-only, so offered is already undefined before this allowlist check runs. Consequently the wizard does not assign web search; if an existing assignment survives, resolve-slot.ts marks the URL as customBaseUrl and media.ts rejects the search at runtime. The official Kimi URLs need to be recognized in the model-settings capability and resolution path, not only here.

Useful? React with 👍 / 👎.