Merge pull request #1810 from deepseek-harness/claude/docs-model-providers

docs(user): 新增「配置模型」入门文档，并让文档站点自带图片

.agents/notes/implemented/process/2026-07-13-documentation-site-projection.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-13-documentation-site-projection.md
-2026-07-13-documentation-site-projection.md: 2452c9dfa53e05061446df2fe650f3b4d6428c01
-2026-07-13-documentation-site-projection.zh.md: 6f1c79ac502a04714cd77f680108dbff035b048c
+2026-07-13-documentation-site-projection.md: f19d9b309aa22821a75086dc07ee302097631ba0
+2026-07-13-documentation-site-projection.zh.md: cc5e94e709f0639fd35ad81165b199cc5c9effc0

.agents/notes/implemented/process/2026-07-13-documentation-site-projection.md

@@ -18,7 +18,7 @@ Canonical Markdown remains in the repository tier that owns it. Product-facing g
 
 Locale home projections retain only the canonical YAML frontmatter. The repository-facing body can keep its H1 and bilingual source links, while the VitePress home theme owns the rendered hero and features and the site navigation owns locale switching.
 
-The projector parses Markdown links without reserializing the document. A link to another published source becomes a site-relative route; a link to an unpublished repository file becomes a GitHub source link; a repository image becomes a raw GitHub URL. Missing relative targets fail projection. Unit tests pin these transformations, and `docs:check` runs the projector tests plus a production VitePress build as part of `doc-sync` and the parallel documentation gates.
+The projector parses Markdown links without reserializing the document. A link to another published source becomes a site-relative route; a link to an unpublished repository file becomes a GitHub source link; a repository image is copied into the generated tree and referenced from there ([why](2026-08-06-doc-site-carries-its-images.md)). Missing relative targets fail projection. Unit tests pin these transformations, and `docs:check` runs the projector tests plus a production VitePress build as part of `doc-sync` and the parallel documentation gates.
 
 `website/AGENTS.md` is the only maintained Markdown file in the website subtree. The projector test enumerates tracked and unignored files and rejects any other website Markdown, so site-specific locale, route, API, or generated source copies cannot bypass the publication manifest.
 

.agents/notes/implemented/process/2026-07-13-documentation-site-projection.zh.md

@@ -18,7 +18,7 @@ Status: implemented
 
 各 locale 的首页投影只保留权威 YAML frontmatter。面向仓库的正文可以保留其 H1 和双语源文件链接，而 VitePress 首页主题负责渲染 hero 与功能区，网站导航负责切换 locale。
 
-投影器解析 Markdown 链接，但不会重新序列化文档。指向另一个已发布源文件的链接会变成站内相对路由；指向未发布仓库文件的链接会变成 GitHub 源文件链接；仓库图片会变成 GitHub raw URL。相对目标不存在时，投影会失败。单元测试会锁定这些转换行为，`docs:check` 则运行投影器测试和 VitePress 生产构建，并将二者纳入 `doc-sync` 和并行文档门禁。
+投影器解析 Markdown 链接，但不会重新序列化文档。指向另一个已发布源文件的链接会变成站内相对路由；指向未发布仓库文件的链接会变成 GitHub 源文件链接；仓库图片会被拷贝进生成树并从那里引用（[原因](2026-08-06-doc-site-carries-its-images.md)）。相对目标不存在时，投影会失败。单元测试会锁定这些转换行为，`docs:check` 则运行投影器测试和 VitePress 生产构建，并将二者纳入 `doc-sync` 和并行文档门禁。
 
 `website/AGENTS.md` 是网站子树中唯一维护的 Markdown 文件。投影器测试会枚举所有已跟踪文件和未被忽略的未跟踪文件，并拒绝网站中的任何其他 Markdown，因此网站专用的 locale、路由、API 或生成源文件副本无法绕过发布 manifest。
 

.agents/notes/implemented/process/2026-08-06-doc-site-carries-its-images.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-08-06-doc-site-carries-its-images.md
+2026-08-06-doc-site-carries-its-images.md: 9109808874579b79d85c2e22b0987110f41ddc42
+2026-08-06-doc-site-carries-its-images.zh.md: d601112e8870150c363d8533e85ef86e7f3f8ffc

.agents/notes/implemented/process/2026-08-06-doc-site-carries-its-images.md

@@ -0,0 +1,43 @@
+# Agent Note: The documentation site carries its own images
+
+Status: implemented
+
+English | [中文](2026-08-06-doc-site-carries-its-images.zh.md)
+
+## Problem
+
+`scripts/project-doc-site.ts` rewrote every repository-relative target that the publication manifest does not publish into a GitHub URL, and for an image that meant `https://raw.githubusercontent.com/<owner>/<repo>/<ref>/<path>`. Nothing in the site build copies files: `srcDir` is the disposable `.generated` tree, VitePress sets no `publicDir` (its default, `<srcDir>/public`, is inside the tree the projector deletes on every run), and only Markdown is written there.
+
+That works only for a public repository. This one is private, and `raw.githubusercontent.com` answers 404 to an unauthenticated request — a browser session on github.com does not authenticate it either, since GitHub's own UI serves private blobs through separately signed URLs. Every image on the site was therefore broken for every reader, and no gate said so: `verify-md-links` and the projector check that the target file *exists in the repository*, which is a different question from whether a site reader can fetch it.
+
+## Decision
+
+`rewriteMarkdown` takes an optional `placeImage(absPath): string`. When a page references an image the manifest does not publish as a page, the projector copies that file into the generated tree beside the page and rewrites the reference to `./<basename>`; Vite then bundles it like any other site asset. Nothing about repository visibility can reach the published page.
+
+The copy lands beside the page rather than in a shared asset directory. Each locale's route tree gets its own copy, so one relative URL is correct from both `guide/` and `en/guide/` without computing per-locale prefixes, and a page's assets are removed with the page when the manifest drops it. One map claims every projected path — pages and images alike — so a second source for one path throws, in the same spirit as the existing duplicate-route check, rather than letting whichever wrote last win.
+
+Only a regular file whose real path stays inside the repository is copied; anything else fails the projection naming the page and the target. Link rewriting needs to know a target *exists*, but publication copies its bytes onto the site, so a reference escaping the repository — through `../..` or a symlink out of the tree — would put a build-machine file on a published page. The reference's `?query` or `#fragment` rides along to the placed URL exactly as the GitHub branch has always carried it, and the file name is percent-encoded because the destination is a Markdown inline target.
+
+`docsSourceFiles()` reports the placed images alongside the Markdown, so the dev server's watcher re-projects when a screenshot is replaced instead of serving the previous copy until something touches the page.
+
+`placeImage` is optional because `rewriteMarkdown` is also called directly by its spec, where no generated tree exists. Without it the old GitHub-raw behavior stands, which keeps that seam honest: the fallback is still the correct answer for a consumer that only rewrites text.
+
+Canonical Markdown keeps writing ordinary repository-relative image paths, so the same file renders on GitHub and on the site. No document carries a site-absolute URL to satisfy VitePress.
+
+## Alternatives considered
+
+**Set `publicDir` outside `.generated` and reference site-absolute URLs.** Fewer moving parts in the projector, but every image reference would then be broken when the same Markdown is read in the repository, and canonical docs are read both ways.
+
+**Host images on the assets branch, as demo GIFs already are.** That branch exists to keep large binaries out of the main history, and its raw URLs have exactly the same visibility problem. It remains the right home for recordings; it does not solve this.
+
+**Wait for the repository to become public.** It would fix the symptom without making the site self-contained, and the site would silently depend on GitHub's availability and rate limits for every image.
+
+## Consequences
+
+Images in published documentation now work regardless of who is reading or whether the repository is public, and the site build has no runtime dependency on GitHub for them. The generated tree grows by one copy of each referenced image per locale — the four screenshots in the model-provider guide add roughly 270 KB per locale.
+
+Images referenced from *unpublished* documents are untouched: they still resolve to GitHub raw, and still fail for a private repository. Nothing consumes them today, and a document that is not on the site has no site build to carry its assets.
+
+## Testing
+
+`scripts/project-doc-site.spec.ts` covers the placer receiving the resolved absolute path and the returned URL landing in the Markdown, a placed reference keeping its fragment, a published page link still resolving to its route when a placer is present, and the unchanged GitHub-raw fallback when no placer is supplied. `publishableImage` is covered directly: a regular file inside the repository resolves, while a symlink whose target escapes it, a path outside it, and a directory are all refused. `pnpm docs:check` builds the site with the model-provider guide's screenshots and fails on a missing source; the copied files and their `./<basename>` references were verified in `website/.generated` and in a running `docs:dev` (`naturalWidth > 0` in both locales).

.agents/notes/implemented/process/2026-08-06-doc-site-carries-its-images.zh.md

@@ -0,0 +1,43 @@
+# Agent Note: 文档站点自带图片
+
+Status: implemented
+
+[English](2026-08-06-doc-site-carries-its-images.md) | 中文
+
+## Problem
+
+`scripts/project-doc-site.ts` 会把发布清单未收录的仓库相对目标一律改写成 GitHub 地址，对图片而言就是 `https://raw.githubusercontent.com/<owner>/<repo>/<ref>/<path>`。站点构建不拷贝任何文件：`srcDir` 是用完即弃的 `.generated` 树，VitePress 没有设置 `publicDir`（其默认值 `<srcDir>/public` 恰好位于投影每次运行时删除的那棵树里），而写进去的只有 Markdown。
+
+这只对公开仓库成立。本仓库是私有的，而 `raw.githubusercontent.com` 对未认证请求一律回 404——github.com 上的登录会话也不能认证它，因为 GitHub 自家界面是用另一套单独签名的地址提供私有 blob 的。于是站点上的每一张图片对每一位读者都是坏的，却没有任何门禁能说出来：`verify-md-links` 与投影校验的是目标文件**在仓库里是否存在**，那与站点读者能否取到它是两个问题。
+
+## Decision
+
+`rewriteMarkdown` 新增可选的 `placeImage(absPath): string`。当页面引用了一张清单未作为页面发布的图片时，投影把该文件复制进生成树中该页面的旁边，并把引用改写为 `./<basename>`；随后 Vite 会像处理其他站点资源一样打包它。仓库可见性再也影响不到已发布页面。
+
+副本落在页面旁边，而不是某个共享资源目录。每个 locale 的路由树各持一份副本，因此同一个相对 URL 在 `guide/` 与 `en/guide/` 下都正确，无需按 locale 计算前缀；清单撤下某页时，它的资源也随之消失。一张表登记所有被投影的路径——页面与图片一视同仁——同一路径出现第二个来源就抛错，与既有的重复路由检查同一个立场，而不是让最后写入的那个静默胜出。
+
+只有真实路径位于仓库内的普通文件才会被拷贝，其余一律让投影失败并点名页面与目标。链接改写只需要知道目标**存在**，但发布是把它的字节拷上站点，因此一个逃出仓库的引用——经由 `../..` 或指向树外的符号链接——会把构建机上的文件放到已发布页面上。引用自带的 `?query` 或 `#fragment` 会随安置后的 URL 一同保留，与 GitHub 分支一贯的做法一致；文件名做百分号编码，因为目标位于 Markdown 内联目标的位置。
+
+`docsSourceFiles()` 会连同被安置的图片一起上报，于是替换截图时开发服务器的 watcher 会重新投影，而不是一直服务旧副本直到有人碰一下页面。
+
+`placeImage` 之所以可选，是因为 `rewriteMarkdown` 也被它自己的 spec 直接调用，而那里并不存在生成树。不传它时保持原有的 GitHub raw 行为，这也让该接缝保持诚实：对只改写文本的消费方而言，这个回退仍是正确答案。
+
+正本 Markdown 照旧写普通的仓库相对图片路径，因此同一份文件在 GitHub 上和站点上都能正常显示。没有任何文档为了迁就 VitePress 而写站内绝对 URL。
+
+## Alternatives considered
+
+**把 `publicDir` 设到 `.generated` 之外，并使用站内绝对 URL。** 投影这边的活动部件更少，但同一份 Markdown 在仓库中阅读时，每一处图片引用都会是坏的，而正本文档是两种方式都要读的。
+
+**把图片放到 assets 分支，就像演示 GIF 那样。** 那个分支的存在是为了让大体积二进制不进主线历史，而它的 raw 地址有着完全相同的可见性问题。它仍然是录屏的正确归宿；但它解决不了这件事。
+
+**等仓库转为公开。** 那只是消除症状，不会让站点自给自足，而且每一张图片都会让站点隐式依赖 GitHub 的可用性与限流。
+
+## Consequences
+
+已发布文档中的图片，现在无论谁在阅读、无论仓库是否公开都能显示，站点构建也不再为图片依赖 GitHub 的运行时可达性。生成树会为每个 locale 各增加一份被引用图片的副本——配置模型指南里的四张截图，每个 locale 约 270 KB。
+
+**未发布**文档引用的图片不受影响：它们仍解析到 GitHub raw，对私有仓库仍然失败。今天没有任何消费方用到它们，而不在站点上的文档也没有站点构建可以承载其资源。
+
+## Testing
+
+`scripts/project-doc-site.spec.ts` 覆盖：placer 收到解析后的绝对路径且其返回的 URL 落进 Markdown、被安置的引用保留其 fragment、存在 placer 时已发布页面的链接仍解析到自己的路由、以及不传 placer 时不变的 GitHub raw 回退。`publishableImage` 另有直接覆盖：仓库内的普通文件被接受，而目标逃出仓库的符号链接、仓库外的路径与目录一律拒绝。`pnpm docs:check` 会带着配置模型指南的截图构建站点，并在来源缺失时失败；被拷贝的文件及其 `./<basename>` 引用已在 `website/.generated` 与运行中的 `docs:dev` 里核实（两个 locale 均 `naturalWidth > 0`）。

.agents/skills/dsh-doc-site-sync/SKILL.md

@@ -46,6 +46,7 @@ Write normal repository-relative Markdown links in canonical docs. The projector
 
 - A target present in the manifest becomes a site-relative route.
 - An existing target outside the manifest becomes a GitHub source link, including supported line suffixes.
+- An image is the exception: its file is copied into the generated tree and referenced from there, so the site serves it regardless of repository visibility. It must be a regular file inside the repository.
 - External URLs, site-absolute URLs, email links, and fragment-only links remain unchanged.
 - A missing repository-relative target fails projection instead of silently producing a broken link.
 

docs/user/guide/providers.i18n.yaml

@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write docs/user/guide/providers.md
+providers.md: 66b6cf25c61a252fbd10a85f8c79c246eeae8abe
+providers.zh.md: a2c33c90be971e09ab29e2355ca6a7ae6f947c39

docs/user/guide/providers.md

@@ -0,0 +1,120 @@
+# Configure models
+
+English | [中文](providers.zh.md)
+
+Harness ships with DeepSeek and mounts a generic multi-provider adapter alongside it, for the providers in pi-ai's installed catalog — Anthropic, OpenAI, and the rest — and for any OpenAI-compatible gateway or self-hosted server. You have two entry points: the **Models** page in the web UI, and `$DSH_HOME/settings.yaml`. Both write the same document, and a change takes effect on the next request without a restart.
+
+## Where providers come from
+
+`cordis.yml` decides which **adapters** are installed; the settings document decides which **providers** run. The shipped composition carries two LLM adapters:
+
+- `llm-deepseek` serves the `deepseek-official` route, the one available out of the box.
+- `llm-pi-ai` mounts **dormant**: zero routes and no extra entries in the model picker until an `llm-pi-ai:` settings section supplies provider profiles, at which point those routes register live and drop again when the section empties.
+
+Adding a provider therefore rarely means editing `cordis.yml` — writing settings is enough, and that is exactly what the Models page does.
+
+## Configure from the web UI
+
+Start `pnpm run dsh web` and open **Settings → Models**.
+
+![The Models page: the DeepSeek card, with Add provider and Add a custom provider below it](providers-models-page.png)
+
+**Give DeepSeek its key.** The DeepSeek card carries one API-key field; fill it in, save, and the provider is ready.
+
+**Add a provider from the installed catalog.** Choose **Add provider**, pick one of pi-ai's catalog providers (anthropic, openai, and so on), and enter that provider's API key. The endpoint, protocol, and model catalog all come from the catalog; the key is the only thing you owe.
+
+That holds for providers that authenticate with an API key. The catalog also carries Bedrock, Vertex, Azure, and Codex, which need AWS credentials and a region, an ADC project, an `api-version`, and OAuth respectively: filling in the key field alone will not make them work. Those authenticate through pi-ai's own environment discovery, with credentials prepared the way each one requires.
+
+**Add a custom provider.** Choose **Add a custom provider** for a route the catalog does not ship — a company gateway, a self-hosted server, or a provider newer than the installed catalog. It asks for a Provider ID (the lowercase identifier that names the route in requests and as its credential), a base URL, a protocol, and at least one model.
+
+![The custom provider form: Provider ID, display name, base URL, API protocol, and API key](providers-custom-form.png)
+
+**Let the endpoint report its models.** Expand **Model catalog** and choose **Fetch available models**: the interrogation asks the endpoint **the form currently shows** — including a base URL edited but not yet saved and a key typed but not yet stored — and offers what it reports as candidates to pick from. A route the installed catalog describes is answered from that catalog with no network call. Adopting a candidate only writes rows into the draft; nothing is stored until you save.
+
+Keys are write-only: the page only ever holds a redacted descriptor, never the literal secret. A key you enter is stored in `$DSH_HOME/.env`, and the profile records only the variable name that references it.
+
+## settings.yaml for advanced configuration
+
+The document lives at `$DSH_HOME/settings.yaml` (`$DSH_HOME` defaults to `~/.dsh`). The Models page writes this file, and you can edit it directly; neither source outranks the other.
+
+```yaml
+llm-deepseek:
+  reasoningEffort: high
+
+llm-pi-ai:
+  providers:
+    # Catalog route: endpoint, protocol, and models come from pi-ai; you supply
+    # the credential.
+    openai:
+      apiKeyEnv: OPENAI_API_KEY
+
+    # Also a catalog route, moved to a private proxy, with its catalog narrowed
+    # to one model and that model's capacity corrected. Every unset field still
+    # comes from the catalog.
+    anthropic:
+      apiKeyEnv: ANTHROPIC_API_KEY
+      baseURL: https://proxy.example.com:8443
+      reasoning: high
+      models:
+        - id: claude-sonnet-4-5
+          contextWindow: 200000
+
+    # Hand-declared route: pi-ai ships nothing under this key, so the profile
+    # supplies the whole provider.
+    acme-gateway:
+      displayName: Acme Gateway
+      apiKeyEnv: ACME_GATEWAY_API_KEY
+      api: openai-completions
+      baseURL: https://gateway.acme.example/v1
+      models:
+        - id: acme-large
+          name: Acme Large
+          contextWindow: 65536
+          maxTokens: 4096
+```
+
+A settings section merges over the matching `cordis.yml` configuration **per provider**, so you can override one field of one route and leave the rest as the composition set them.
+
+A profile the adapter could not serve is refused **where it is written**: a hand-declared route needs `api`, `baseURL`, and at least one model, and a profile missing any of them fails naming the offending route and model rather than being stored and quietly disabling the whole namespace. When an already-stored document is broken by an external edit, settings keeps the last good value and warns.
+
+## The model catalog
+
+A profile's `models` list *replaces* that route's installed catalog rather than extending it; omitting it or leaving it empty serves the catalog unchanged. Each entry defaults its unset fields from the installed model of the same `id`, so narrowing a route to two models, correcting one capacity, or adding a model newer than the installed catalog are each a one-line edit.
+
+Only the four fields the harness consumes are configurable: `id`, `name`, `contextWindow`, and `maxTokens`. Pricing and input modalities have no consumer, and reasoning is not per-model configurable at all — it rides the installed catalog entry.
+
+A model neither the entry nor the catalog sizes takes the route's `defaultContextWindow` (262,144) and `defaultMaxTokens` (32,768). Both are guesses by construction, which is why they are route fields: a deployment whose gateway serves smaller models corrects them once.
+
+Model ids are not lifecycle configuration. Requesting a model the route does not configure fails with `UNKNOWN_MODEL` before any provider request goes out.
+
+## Credentials
+
+Prefer `apiKeyEnv`: it is a *reference* resolved per request, so no secret enters the configuration file. A literal `apiKey` is the escape hatch. Omitting both is what leaves a route unauthenticated, which for a catalog route means pi-ai's own environment discovery. A reference that resolves to nothing fails the request with `MISSING_CREDENTIAL` rather than falling through to whatever unrelated key the environment happens to hold.
+
+References resolve from `$DSH_HOME/.env` — what the Models page's key fields write — and from the matching environment variable when no credential service is mounted. One credential serves every model on its route.
+
+## Point an agent at the new provider
+
+A configured route appears in the web model picker and can be switched at any time, which is how most people use it.
+
+A new session's default model comes from the `api-gateway` entry (`@deepseek-ai/dsh-host-apiproxy`) and its `provider` and `model`, which ship as `deepseek-official` and `deepseek-v4-flash`. To change that default, override the entry in `$DSH_HOME/config.yaml`:
+
+```yaml
+- id: api-gateway
+  config:
+    provider: acme-gateway
+    model: acme-large
+```
+
+A patch replaces that entry's whole `config`, so write out every key it needs to keep. A composition you assemble yourself — headless, for instance — sets `agent-loop`'s `agents` instead.
+
+## Troubleshooting
+
+- **`MISSING_CREDENTIAL`** — the variable the profile's `apiKeyEnv` names holds no value. Store the key once through the Models page, or export the variable.
+- **`UNKNOWN_MODEL`** — the requested model is not in the route's configured catalog. Add it to `models`, or use an id the catalog already carries.
+- **`settings-rejected`** — the written profile cannot be served, and the message names the route and model. For a hand-declared route, check that `api`, `baseURL`, and `models` are all present.
+- **Fetching available models answers 401** — the endpoint refused the interrogation. Check the key; if the base URL points at an Anthropic-style gateway, note that the interrogation reads only the OpenAI-compatible `GET /models`, so enter the models by hand instead.
+
+## Exact field reference
+
+The complete fields, types, and defaults each plugin currently supports live in the generated [plugin configuration catalog](../../config-catalog.md). Each adapter's own semantics belong to its README: [`dsh-llm-pi-ai`](../../../packages/llm/llm-pi-ai/README.md) and [`dsh-llm-deepseek`](../../../packages/llm/llm-deepseek/README.md). For `cordis.yml` itself, see [Configuration](./config.md).

docs/user/guide/providers.zh.md

@@ -0,0 +1,120 @@
+# 配置模型
+
+[English](providers.md) | 中文
+
+Harness 出厂就带 DeepSeek，同时挂着一个通用的多提供方适配器，用来接入 Anthropic、OpenAI 这类内置目录里的提供方，或任何 OpenAI 兼容的网关与自建服务。你有两个入口：Web 界面的**模型**页，以及 `$DSH_HOME/settings.yaml`。两者写的是同一份文档，改完下一次请求即生效，不用重启。
+
+## 提供方从哪里来
+
+`cordis.yml` 决定装了哪些**适配器**，settings 文档决定跑哪些**提供方**。出厂组合里有两个 LLM 适配器：
+
+- `llm-deepseek` 提供 `deepseek-official` 路由，是默认可用的那个。
+- `llm-pi-ai` 以**休眠**状态挂载：零路由，模型选择器里也不会多出条目，直到 settings 里的 `llm-pi-ai:` 段落给出 provider profile，路由才注册上来；段落清空则一并撤下。
+
+因此新增一个提供方通常不需要改 `cordis.yml`，写 settings 就够了——而模型页做的正是这件事。
+
+## 在 Web 界面里配置
+
+启动 `pnpm run dsh web`，打开**设置 → 模型**。
+
+![模型页：DeepSeek 卡片，以及添加提供方与添加自定义提供方两个入口](providers-models-page.zh.png)
+
+**填 DeepSeek 的密钥。** DeepSeek 卡片上只有一个 API 密钥输入框，填好保存即可开始用。
+
+**添加内置目录里的提供方。** 点**添加提供方**，从 pi-ai 内置目录中选一个（anthropic、openai 等），填入该提供方的 API 密钥。端点、协议和模型目录都由内置目录提供，你只需要给密钥。
+
+只对以 API 密钥认证的提供方成立。目录里也有 Bedrock、Vertex、Azure、Codex：它们分别需要 AWS 凭据与区域、ADC 项目配置、`api-version`、OAuth，只填密钥框不会让它们工作——这类提供方靠 pi-ai 自己的环境发现认证，凭据按各自的原生方式准备。
+
+**添加自定义提供方。** 点**添加自定义提供方**，用于内置目录没有的路由——公司网关、自建服务，或比内置目录更新的提供方。需要填 Provider ID（请求里点名它、也作为凭据名的小写标识）、API 地址、协议，以及至少一个模型。
+
+![自定义提供方表单：Provider ID、显示名称、API 地址、API 协议、API 密钥](providers-custom-form.zh.png)
+
+**让端点自己报模型。** 展开**模型目录**后点**获取可用模型**，会按你**当前表单里**的地址与密钥去问端点（地址改了但没保存、密钥刚输入还没存下，都算数），把它报告的模型列成候选让你勾选。内置目录里的路由直接由目录作答，不联网。采纳只是把行写进草稿，最终还是你点保存才落盘。
+
+密钥是只写的：页面拿到的永远是脱敏描述符，不是明文。写入的密钥存进 `$DSH_HOME/.env`，profile 里只记录引用它的变量名。
+
+## settings.yaml：进阶配置
+
+文档位于 `$DSH_HOME/settings.yaml`（`$DSH_HOME` 默认是 `~/.dsh`）。模型页写的就是这个文件，你也可以直接编辑它——两个来源没有主次之分。
+
+```yaml
+llm-deepseek:
+  reasoningEffort: high
+
+llm-pi-ai:
+  providers:
+    # Catalog route: endpoint, protocol, and models come from pi-ai; you supply
+    # the credential.
+    openai:
+      apiKeyEnv: OPENAI_API_KEY
+
+    # Also a catalog route, moved to a private proxy, with its catalog narrowed
+    # to one model and that model's capacity corrected. Every unset field still
+    # comes from the catalog.
+    anthropic:
+      apiKeyEnv: ANTHROPIC_API_KEY
+      baseURL: https://proxy.example.com:8443
+      reasoning: high
+      models:
+        - id: claude-sonnet-4-5
+          contextWindow: 200000
+
+    # Hand-declared route: pi-ai ships nothing under this key, so the profile
+    # supplies the whole provider.
+    acme-gateway:
+      displayName: Acme Gateway
+      apiKeyEnv: ACME_GATEWAY_API_KEY
+      api: openai-completions
+      baseURL: https://gateway.acme.example/v1
+      models:
+        - id: acme-large
+          name: Acme Large
+          contextWindow: 65536
+          maxTokens: 4096
+```
+
+settings 段落**逐个提供方**地盖在 `cordis.yml` 的同名配置之上，所以你可以只覆盖某个路由的一个字段，其余保持组合里的样子。
+
+一份服务不了的 profile 会在**写入处**被拒绝：手工声明的路由必须给出 `api`、`baseURL` 和至少一个模型，缺了会带着路由名和模型名报错，而不是存下来再让整个命名空间静默失效。已经存好的文档被外部改坏时，settings 会保留上一次的好值并告警。
+
+## 模型目录
+
+`models` 是**替换**该路由的内置目录，不是往里追加；省略或留空则原样使用内置目录。每个条目会从同 `id` 的内置模型继承自己没写的字段，所以「收窄到两个模型」「更正一个容量」「加一个比内置目录更新的模型」都是一行编辑。
+
+可配置的只有 harness 会消费的四个字段：`id`、`name`、`contextWindow`、`maxTokens`。定价与输入模态没有消费方，推理能力也不按模型配置——它随内置目录条目走。
+
+两处容量都没给出的模型，取路由级兜底 `defaultContextWindow`（262144）与 `defaultMaxTokens`（32768）。这两个数按定义就是猜测，所以它们是路由字段：网关服务的模型更小时改一次即可。
+
+模型 id 不是生命周期配置：请求一个该路由没有配置的模型，会在任何网络请求之前以 `UNKNOWN_MODEL` 失败。
+
+## 凭据
+
+优先用 `apiKeyEnv`——它是一个**引用**，每次请求时解析，密钥本身不进配置文件；`apiKey` 字面量是应急出口。两者都不给，才表示这个路由不带认证，对内置目录路由意味着交给 pi-ai 自己的环境发现。给了引用却解析不到，请求会以 `MISSING_CREDENTIAL` 失败，而不是退回去用环境里碰巧存在的某个不相干的 key 计费。
+
+引用解析自 `$DSH_HOME/.env`（模型页的密钥输入框写的就是它），没有挂载凭据服务时则直接读同名环境变量。一份凭据供该路由上的所有模型使用。
+
+## 让 agent 用上新提供方
+
+配好的路由会出现在 Web 的模型选择器里，随时可切，这也是最常用的方式。
+
+新会话的默认模型来自 `api-gateway` 那条（`@deepseek-ai/dsh-host-apiproxy`）的 `provider` 与 `model`，出厂值是 `deepseek-official` 与 `deepseek-v4-flash`。要改默认值，就在 `$DSH_HOME/config.yaml` 里覆盖该条：
+
+```yaml
+- id: api-gateway
+  config:
+    provider: acme-gateway
+    model: acme-large
+```
+
+补丁会整体替换该条的 `config`，所以要把这条需要保留的键一并写出。自行组装的 `cordis.yml`（例如 headless）改的则是 `agent-loop` 的 `agents`。
+
+## 排错
+
+- **`MISSING_CREDENTIAL`** — profile 里的 `apiKeyEnv` 指向的变量没有值。用模型页存一次密钥，或导出该环境变量。
+- **`UNKNOWN_MODEL`** — 请求的模型不在该路由配置的目录里。把它加进 `models`，或改用目录里已有的 id。
+- **`settings-rejected`** — 写入的 profile 服务不了，错误信息会点名具体的路由和模型。手工声明的路由检查 `api`、`baseURL`、`models` 是否齐全。
+- **获取可用模型返回 401** — 端点拒绝了这次探测。检查密钥；若地址指向的是 Anthropic 风格网关，注意探测只读 OpenAI 兼容的 `GET /models`，此时手工填写模型即可。
+
+## 精确字段参考
+
+每个插件当前支持的完整字段、类型与默认值见自动生成的[插件配置目录](../../config-catalog.md)。两个适配器各自的语义由它们的 README 负责：[`dsh-llm-pi-ai`](../../../packages/llm/llm-pi-ai/README.md) 与 [`dsh-llm-deepseek`](../../../packages/llm/llm-deepseek/README.md)。`cordis.yml` 本身的写法见[配置文件](./config.md)。

docs/user/guide/quickstart.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write docs/user/guide/quickstart.md
-quickstart.md: 4b43291342f80ff9c6dcb844fdc505ef797d1f6c
-quickstart.zh.md: 13294c50c2d7d4005c0fd090cc2eca3c4780d0a6
+quickstart.md: 72cc5a52c33faf81098f747799692b384fcd5f1a
+quickstart.zh.md: ebb831c1e1bec02117c78a4c2426a84af3bb9478

docs/user/guide/quickstart.md

@@ -57,5 +57,6 @@ headless-agent uses the `@deepseek-ai/dsh-cli-demo` app. `dsh web` instead boots
 
 ## Next steps
 
+- [Configure models](./providers.md) — reach providers beyond DeepSeek, and custom gateways
 - [Configuration](./config.md) — understand the `cordis.yml` format
 - [Develop a plugin](../develop/basic/) — build your own tool or backend

docs/user/guide/quickstart.zh.md

@@ -57,5 +57,6 @@ headless-agent 使用 `@deepseek-ai/dsh-cli-demo` app。`dsh web` 则启动 `web
 
 ## 下一步
 
+- [配置模型](./providers.md) — 接入 DeepSeek 之外的提供方与自定义网关
 - [配置文件](./config.md) — 了解 `cordis.yml` 的格式
 - [开发插件](../develop/basic/) — 编写自己的 tool 或后端

scripts/project-doc-site.spec.ts

@@ -1,12 +1,14 @@
 /** Tests for the documentation website projection adapter. */
 
 import { execFileSync } from 'node:child_process'
-import { existsSync, mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs'
+import { existsSync, mkdirSync, mkdtempSync, realpathSync, rmSync, symlinkSync, writeFileSync } from 'node:fs'
 import { tmpdir } from 'node:os'
 import { join, resolve } from 'node:path'
 import { afterEach, describe, expect, it } from 'vitest'
 import { docsPages, type DocsPage } from '../website/docs.ts'
-import { addProjectionFrontmatter, projectedPageContent, rewriteMarkdown } from './project-doc-site.ts'
+import {
+  addProjectionFrontmatter, projectedPageContent, publishableImage, rewriteMarkdown,
+} from './project-doc-site.ts'
 
 const roots: string[] = []
 const repositoryRoot = resolve(import.meta.dirname, '..')
@@ -63,6 +65,32 @@ describe('website source layout', () => {
   })
 })
 
+describe('publishableImage', () => {
+  it('accepts a regular file inside the repository', () => {
+    const { root } = fixture()
+    const real = realpathSync(join(root, 'packages/logo.svg'))
+    expect(publishableImage(join(root, 'packages/logo.svg'), realpathSync(root))).toBe(real)
+  })
+
+  it('refuses a target whose real path escapes the repository', () => {
+    // Publication copies the bytes onto the site, so a reference reaching a
+    // build-machine file must not be treated as an image the repository owns.
+    const { root } = fixture()
+    const outside = mkdtempSync(join(tmpdir(), 'dsh-doc-site-outside-'))
+    roots.push(outside)
+    writeFileSync(join(outside, 'secret.png'), 'not really a png\n')
+    symlinkSync(join(outside, 'secret.png'), join(root, 'packages/linked.png'))
+
+    expect(publishableImage(join(root, 'packages/linked.png'), realpathSync(root))).toBeUndefined()
+    expect(publishableImage(join(outside, 'secret.png'), realpathSync(root))).toBeUndefined()
+  })
+
+  it('refuses a directory', () => {
+    const { root } = fixture()
+    expect(publishableImage(join(root, 'packages'), realpathSync(root))).toBeUndefined()
+  })
+})
+
 describe('rewriteMarkdown', () => {
   it('maps published pages and pins unpublished source links', () => {
     const { root, pages } = fixture()
@@ -93,7 +121,7 @@ describe('rewriteMarkdown', () => {
     })).toBe('[B](./reference-root/b.md)\n')
   })
 
-  it('uses raw GitHub content for unpublished images', () => {
+  it('uses raw GitHub content for unpublished images when nothing places them', () => {
     const { root, pages } = fixture()
     expect(rewriteMarkdown('![logo](../packages/logo.svg)\n', {
       locale: 'en',
@@ -105,6 +133,57 @@ describe('rewriteMarkdown', () => {
     })).toBe('![logo](https://raw.githubusercontent.com/deepseek-harness/deepseek-harness/abc123/packages/logo.svg)\n')
   })
 
+  it('hands an image to the placer and uses the URL it returns', () => {
+    // A raw GitHub URL cannot serve a private repository, so the site build
+    // carries images itself; the placer is what puts them there. The stand-in
+    // derives its URL the way the real one does, so a placer that stopped
+    // returning the basename would fail here rather than pass on a constant.
+    const { root, pages } = fixture()
+    const placed: string[] = []
+    expect(rewriteMarkdown('![logo](../packages/logo.svg)\n', {
+      locale: 'en',
+      sourcePath: 'docs/a.md',
+      route: 'en/a.md',
+      pages,
+      repoRoot: root,
+      repositoryRef: 'abc123',
+      placeImage: (absPath) => {
+        const name = absPath.split('/').pop() ?? ''
+        placed.push(name)
+        return `./${name}`
+      },
+    })).toBe('![logo](./logo.svg)\n')
+    expect(placed).toEqual(['logo.svg'])
+  })
+
+  it('keeps a placed image\u2019s query or fragment', () => {
+    // An SVG view fragment and a Vite query both change what the reference
+    // means, and the GitHub branch has always carried them.
+    const { root, pages } = fixture()
+    expect(rewriteMarkdown('![logo](../packages/logo.svg#view)\n', {
+      locale: 'en',
+      sourcePath: 'docs/a.md',
+      route: 'en/a.md',
+      pages,
+      repoRoot: root,
+      repositoryRef: 'abc123',
+      placeImage: absPath => `./${absPath.split('/').pop() ?? ''}`,
+    })).toBe('![logo](./logo.svg#view)\n')
+  })
+
+  it('leaves a published page link to the route even when a placer exists', () => {
+    const { root, pages } = fixture()
+    expect(rewriteMarkdown('[B](b.md)\n', {
+      locale: 'en',
+      sourcePath: 'docs/a.md',
+      route: 'en/a.md',
+      pages,
+      repoRoot: root,
+      repositoryRef: 'abc123',
+      placeImage: () => { throw new Error('a page link must not be placed as an asset') },
+    })).toBe('[B](./reference/b.md)\n')
+  })
+
   it('does not rewrite Markdown-looking text inside code fences', () => {
     const { root, pages } = fixture()
     const source = '```md\n[B](b.md)\n```\n'

scripts/project-doc-site.ts

@@ -5,8 +5,10 @@
  * tier, while this adapter rewrites cross-source links for the public site.
  */
 
-import { existsSync, lstatSync, mkdirSync, readFileSync, rmSync, writeFileSync } from 'node:fs'
-import { dirname, extname, posix, relative, resolve, sep } from 'node:path'
+import {
+  copyFileSync, existsSync, lstatSync, mkdirSync, readFileSync, realpathSync, rmSync, statSync, writeFileSync,
+} from 'node:fs'
+import { basename, dirname, extname, posix, relative, resolve, sep } from 'node:path'
 import { fromMarkdown } from 'mdast-util-from-markdown'
 import { gfmFromMarkdown } from 'mdast-util-gfm'
 import { gfm } from 'micromark-extension-gfm'
@@ -38,6 +40,15 @@ export interface RewriteMarkdownOptions {
   pages: DocsPage[]
   repoRoot: string
   repositoryRef: string
+  /**
+   * Place one referenced image beside the projected page and return the URL to
+   * reach it from that page. A GitHub raw URL cannot serve this repository —
+   * `raw.githubusercontent.com` answers 404 for a private one, and no reader of
+   * the site is authenticated to it — so an image travels into the generated
+   * tree and Vite bundles it like any other site asset. Omitted by callers that
+   * only rewrite text, which then leave images pointing at the repository.
+   */
+  placeImage?: (absPath: string) => string
 }
 
 function repoPath(absPath: string, repoRoot: string): string {
@@ -222,9 +233,13 @@ export function rewriteMarkdown(source: string, options: RewriteMarkdownOptions)
       ? options.locale === 'root' ? 'en' : 'root'
       : options.locale
     const page = published.get(targetPath)?.get(targetLocale)
-    const nextUrl = page === undefined
-      ? githubTarget(absPath, line, suffix, options.repositoryRef, options.repoRoot, node.type === 'image')
-      : routeTarget(options.route, page.route, suffix)
+    const nextUrl = page !== undefined
+      ? routeTarget(options.route, page.route, suffix)
+      : node.type === 'image' && options.placeImage !== undefined
+        // The suffix rides along exactly as the GitHub branch keeps it: an SVG
+        // view fragment or a Vite query changes what the reference means.
+        ? `${options.placeImage(absPath)}${suffix}`
+        : githubTarget(absPath, line, suffix, options.repositoryRef, options.repoRoot, node.type === 'image')
 
     const start = node.position?.start.offset
     const end = node.position?.end.offset
@@ -291,17 +306,78 @@ export function projectedPageContent(markdown: string, page: DocsPage): string {
   return markdown.slice(0, closing + closingDelimiter.length)
 }
 
-/** Canonical Markdown files watched by the local VitePress dev server. */
+/**
+ * The repository file one image reference resolves to, or `undefined` when the
+ * target is not a local file this build may publish.
+ * @param absPath - resolved image target.
+ * @param repoRoot - repository root every published image must stay inside.
+ * @returns the file's real path, or `undefined` when it must not be copied.
+ *
+ * Only a regular file whose real path stays inside the repository qualifies.
+ * Publication copies the bytes into the site, so a reference escaping the
+ * repository — `../../.ssh/id_rsa`, or a symlink pointing out of the tree —
+ * would put a build-machine file on the site; `existsSync` alone, which is all
+ * link resolution needs, does not answer that.
+ */
+export function publishableImage(absPath: string, repoRoot: string): string | undefined {
+  const real = realpathSync(absPath)
+  const inside = real === repoRoot || real.startsWith(`${repoRoot}${sep}`)
+  return inside && statSync(real).isFile() ? real : undefined
+}
+
+/** Every local image a published page references, resolved to its repository file. */
+function referencedImages(): string[] {
+  const found = new Set<string>()
+  for (const page of docsPages) {
+    const sourceAbs = resolve(root, page.source)
+    if (!existsSync(sourceAbs)) continue
+    rewriteMarkdown(readFileSync(sourceAbs, 'utf8'), {
+      sourcePath: page.source,
+      locale: page.locale,
+      route: page.route,
+      pages: docsPages,
+      repoRoot: root,
+      repositoryRef: 'master',
+      placeImage: (absPath) => {
+        const real = publishableImage(absPath, root)
+        if (real !== undefined) found.add(real)
+        return ''
+      },
+    })
+  }
+  return [...found]
+}
+
+/**
+ * Files watched by the local VitePress dev server: every canonical Markdown
+ * source, plus the images they publish. Without the images, replacing a
+ * screenshot leaves the previous copy in the generated tree until something
+ * touches the Markdown beside it.
+ */
 export function docsSourceFiles(): string[] {
-  return [...new Set(docsPages.map(page => resolve(root, page.source)))]
+  return [...new Set([...docsPages.map(page => resolve(root, page.source)), ...referencedImages()])]
 }
 
 /** Rebuild the disposable VitePress source tree from the publication manifest. */
 export function projectDocs(): void {
   const routes = new Set<string>()
+  /** Projected path to the repository file that claimed it, pages and images alike. */
+  const claimed = new Map<string, string>()
   const repositoryRef = process.env.GITHUB_SHA ?? 'master'
   rmSync(generatedRoot, { recursive: true, force: true })
 
+  /** Reserve one projected path, refusing a second source for it. */
+  const claim = (target: string, sourceAbs: string): void => {
+    const holder = claimed.get(target)
+    if (holder !== undefined && holder !== sourceAbs) {
+      throw new Error(
+        `project-doc-site: ${repoPath(sourceAbs, root)} and ${repoPath(holder, root)}`
+        + ` both project to ${relative(generatedRoot, target).split(sep).join('/')}.`,
+      )
+    }
+    claimed.set(target, sourceAbs)
+  }
+
   for (const page of docsPages) {
     if (routes.has(page.route)) throw new Error(`project-doc-site: duplicate route ${JSON.stringify(page.route)}.`)
     routes.add(page.route)
@@ -310,6 +386,9 @@ export function projectDocs(): void {
       throw new Error(`project-doc-site: source ${JSON.stringify(page.source)} does not exist or is not a file.`)
     }
     const output = resolve(generatedRoot, page.route)
+    // Claimed before the images are placed: a page and an image landing on one
+    // path would otherwise overwrite each other in whichever order they ran.
+    claim(output, sourceAbs)
     mkdirSync(dirname(output), { recursive: true })
     const markdown = readFileSync(sourceAbs, 'utf8')
     const projected = rewriteMarkdown(markdown, {
@@ -319,6 +398,25 @@ export function projectDocs(): void {
       pages: docsPages,
       repoRoot: root,
       repositoryRef,
+      placeImage: (absPath) => {
+        const real = publishableImage(absPath, root)
+        if (real === undefined) {
+          throw new Error(
+            `project-doc-site: ${page.source} references image ${repoPath(absPath, root)},`
+            + ' which is not a regular file inside the repository.',
+          )
+        }
+        // Beside the page that references it, under its own basename: each
+        // locale's route tree gets its own copy, so one relative URL is correct
+        // from both.
+        const name = basename(real)
+        const target = resolve(dirname(output), name)
+        claim(target, real)
+        copyFileSync(real, target)
+        // Encoded because the destination is a Markdown inline target, where an
+        // unescaped space would end it early.
+        return `./${encodeURI(name)}`
+      },
     })
     writeFileSync(output, addProjectionFrontmatter(projectedPageContent(projected, page), page))
   }

website/docs.ts

@@ -130,13 +130,21 @@ const homeAndGuide = pairedPages([
     section: { root: '入门', en: 'Guide' },
     order: 2,
   },
+  {
+    source: 'docs/user/guide/providers.md',
+    route: 'guide/providers.md',
+    label: { root: '配置模型', en: 'Configure models' },
+    sidebar: { root: 'zh-guide', en: 'en-guide' },
+    section: { root: '入门', en: 'Guide' },
+    order: 3,
+  },
   {
     source: 'docs/user/guide/config.md',
     route: 'guide/config.md',
     label: { root: '配置文件', en: 'Configuration' },
     sidebar: { root: 'zh-guide', en: 'en-guide' },
     section: { root: '入门', en: 'Guide' },
-    order: 3,
+    order: 4,
   },
 ])