Add OpenCode Go as an AI provider

.env.example

@@ -36,7 +36,7 @@ DATABASE_URL="postgresql://postgres:postgres@localhost:5432/dossier"
 
 # Defaults for Settings → AI provider; values saved in the app override them.
 # Provider: local (Ollama or llama.cpp), openai (any OpenAI-compatible
-# endpoint), anthropic, mistral or lmstudio.
+# endpoint), anthropic, mistral, lmstudio or opencode-go.
 # DOSSIER_AI_PROVIDER=local
 # DOSSIER_AI_URL=http://localhost:11434
 # DOSSIER_AI_MODEL=llama3.1:8b

DEVELOPMENT.md

@@ -135,7 +135,7 @@ Failed sign-ins are rate-limited in Postgres: after 5 failures within a minute f
 
 ## AI provider
 
-Dossier reads job postings and writes drafts with the model you choose in Settings → AI provider: a local model through Ollama or llama.cpp, LM Studio, any OpenAI-compatible server, Anthropic or Mistral. Until you save a choice there, the `DOSSIER_AI_*` variables decide (see `.env.example`). API keys are stored encrypted with a key derived from `DOSSIER_SECRET`, and never sent back to the browser.
+Dossier reads job postings and writes drafts with the model you choose in Settings → AI provider: a local model through Ollama or llama.cpp, LM Studio, any OpenAI-compatible server, Anthropic, Mistral or OpenCode Go. Until you save a choice there, the `DOSSIER_AI_*` variables decide (see `.env.example`). API keys are stored encrypted with a key derived from `DOSSIER_SECRET`, and never sent back to the browser.
 
 Long AI tasks run as background jobs (`src/lib/jobs`): the page follows their named steps as they happen and picks them up again after a reload. A job's input and result hold personal data, so they don't last: a job that works on an application names it (`startJob(…, { applicationId })`) and is deleted with it, and finished jobs are deleted once they're older than **Keep task data for** in Settings → Your data (`src/lib/jobs/retention.ts`). Keep nothing in a job that has to outlast that; Undo on a toast and a page reattaching after a reload are fine. The prompts live in `src/lib/ai/tasks`; in strict mode ("Never invent experience", on by default) a fact guard flags names and numbers that aren't in the profile or the posting.
 
@@ -156,6 +156,8 @@ docker compose exec -e AI_LIVE=ollama app yarn vitest run src/lib/ai/providers.l
 
 `AI_LIVE=mistral` checks Mistral instead, with the key saved in Settings → AI provider while Mistral is chosen there, or with `DOSSIER_AI_KEY` (and `DOSSIER_AI_MODEL`) passed with `-e`.
 
+`AI_LIVE=opencode-go` checks OpenCode Go, with the key saved in Settings while OpenCode Go is chosen there, or with `-e OPENCODE_TEST_KEY=…` (and `OPENCODE_TEST_MODEL`, by default `glm-5.3-flash`). It spends usage from that Go plan.
+
 `AI_LIVE=lmstudio` checks LM Studio, with the model saved in Settings while LM Studio is chosen there, or with `-e DOSSIER_AI_MODEL=…` (and `DOSSIER_AI_URL`, by default `http://host.docker.internal:1234`). A model that isn't loaded gets loaded and stays loaded until you eject it.
 
 ## Job search

PRIVACY.md

@@ -38,7 +38,7 @@ Exports you download and database dumps you take are copies outside Dossier's re
 
 | Who receives it | When | What |
 |---|---|---|
-| A hosted AI provider (OpenAI-compatible, Anthropic or Mistral) | Each AI task, while one is chosen in Settings → AI provider | See [below](#what-a-hosted-ai-provider-receives) |
+| A hosted AI provider (OpenAI-compatible, Anthropic, Mistral or OpenCode Go) | Each AI task, while one is chosen in Settings → AI provider | See [below](#what-a-hosted-ai-provider-receives) |
 | Google | While a Google account is connected (Settings → Google) | Signing in: your OAuth client's ID and secret, and the tokens. **Gmail:** the application emails and answers you send, with their PDFs, and a request for new mail on each check (every 10 minutes by default). **Calendar:** the next 14 days, to find your free times; each interview you add, change or remove: its time and a title with the role and company; and requests for the events in the months you look at in the dashboard's calendar, which Dossier shows and doesn't keep. Google already holds your mail and calendar; Dossier only asks for the permissions you give |
 | The job source you chose (JSearch, Adzuna or the Bundesagentur für Arbeit) | Each search in Find jobs, which is off until you choose a source, and the job agent's search, at most three times in 24 hours while it's on (never on the Bundesagentur für Arbeit) | What you search for: keywords, place, country, radius and filters. Nothing from your profile: it only fills in the form. The Bundesagentur für Arbeit has no official API: Dossier uses the access its own job search uses, without its permission, so it's searched only after you tick "I understand" under that note, and `DOSSIER_JOBS_ARBEITSAGENTUR=off` takes it out |
 | Hacker News (its official API, and Algolia's search of it) | When you press **Refresh** in Startups, then once a day while you open Startups at least every 14 days | Requests for YC startups' job posts and this month's “Who is hiring?” thread. Nothing about you: which posts suit your city and country is worked out on your server |

README.md

@@ -19,7 +19,7 @@
 
 Dossier keeps your career history in one profile and turns it into a clean resume. For every job you apply to, it reads the posting, shows how you match, and drafts a tailored resume and cover letter that only use facts from your profile. You refine them in a chat, download them as PDFs and track each application from draft to offer.
 
-It's for one person on a small server: one password, your data in your own PostgreSQL database, and an AI model of your choice. That can be a local model through Ollama or LM Studio, so your profile and documents stay on your own hardware, or a hosted one (any OpenAI-compatible API, Anthropic or Mistral), which then receives what it writes from. [Privacy](https://github.com/AienTech/dossier/wiki/Privacy) says what goes where.
+It's for one person on a small server: one password, your data in your own PostgreSQL database, and an AI model of your choice. That can be a local model through Ollama or LM Studio, so your profile and documents stay on your own hardware, or a hosted one (any OpenAI-compatible API, Anthropic, Mistral or OpenCode Go), which then receives what it writes from. [Privacy](https://github.com/AienTech/dossier/wiki/Privacy) says what goes where.
 
 **Get started:** [run it on your computer](#run-it-on-your-computer) in a few minutes, or [self-host it](#self-hosting) on a server. The [installation guide](https://github.com/AienTech/dossier/wiki) covers every way: Docker with or without Compose, Node.js, a plain Ubuntu server, CapRover, Coolify and Dokploy.
 

prisma/migrations/20261027090000_ai_provider_opencode_go/migration.sql

@@ -0,0 +1,2 @@
+-- AlterEnum
+ALTER TYPE "AiProvider" ADD VALUE 'OPENCODE_GO';

prisma/schema.prisma

@@ -306,6 +306,8 @@ enum AiProvider {
   MISTRAL
   /// LM Studio on the user's own hardware (#210).
   LMSTUDIO
+  /// OpenCode Go, through its OpenAI-compatible Go endpoint.
+  OPENCODE_GO
 }
 
 enum Creativity {

site/index.html

@@ -162,6 +162,7 @@ <h1 class="display">Your job search, <em>on your own server.</em></h1>
           <li>Ollama</li>
           <li>LM Studio</li>
           <li>OpenAI-compatible APIs</li>
+          <li>OpenCode Go</li>
           <li>Anthropic</li>
           <li>Mistral</li>
         </ul>

src/app/(app)/settings/ai-provider-section.test.tsx

@@ -32,6 +32,7 @@ describe("AiProviderSection", () => {
       "Local model",
       "LM Studio",
       "OpenAI-compatible",
+      "OpenCode Go",
       "Anthropic",
       "Mistral",
     ]);
@@ -55,6 +56,22 @@ describe("AiProviderSection", () => {
     );
   });
 
+  it("fills in OpenCode Go's address and model, with its key and privacy note", async () => {
+    render(<AiProviderSection initial={local} />);
+    await userEvent.click(screen.getByRole("radio", { name: "OpenCode Go" }));
+    expect(screen.getByLabelText("Server address")).toHaveValue("https://opencode.ai/inference/go");
+    expect(screen.getByRole("note")).toHaveTextContent(/^OpenCode receives what each task needs: .* depend on your account with OpenCode\.$/);
+    expect(screen.getByLabelText("OpenCode API key")).toBeInTheDocument();
+    expect(screen.getByText(/Go models need a Go or Go Plus plan/)).toBeInTheDocument();
+    await waitFor(() =>
+      expect(saveAiProvider).toHaveBeenCalledWith({
+        provider: "OPENCODE_GO",
+        baseUrl: "https://opencode.ai/inference/go",
+        model: "glm-5.3-flash",
+      }),
+    );
+  });
+
   it("fills in Mistral's address and model, and names Mistral in the note (#206)", async () => {
     render(<AiProviderSection initial={local} />);
     await userEvent.click(screen.getByRole("radio", { name: "Mistral" }));

src/lib/ai/config.test.ts

@@ -50,6 +50,16 @@ describe("resolveAiConfig", () => {
     });
   });
 
+  it("knows OpenCode Go's base URL and a default model", () => {
+    expect(resolveAiConfig(nothingSaved, { DOSSIER_AI_PROVIDER: "opencode-go", DOSSIER_AI_KEY: "oc_sk-env" })).toMatchObject({
+      provider: "OPENCODE_GO",
+      baseUrl: "https://opencode.ai/inference/go",
+      model: "glm-5.3-flash",
+      apiKey: "oc_sk-env",
+      timeoutMs: 120_000,
+    });
+  });
+
   it("finds LM Studio on the Docker host, as patient as a local model", () => {
     expect(resolveAiConfig(nothingSaved, { DOSSIER_AI_PROVIDER: "lmstudio" })).toMatchObject({
       provider: "LMSTUDIO",
@@ -84,6 +94,7 @@ describe("createProvider", () => {
     const base = { baseUrl: "http://x", model: "m", apiKey: "k", apiKeyProblem: null, timeoutMs: 1000 };
     expect(createProvider({ ...base, provider: "LOCAL" }).kind).toBe("LOCAL");
     expect(createProvider({ ...base, provider: "OPENAI" }).kind).toBe("OPENAI");
+    expect(createProvider({ ...base, provider: "OPENCODE_GO" }).kind).toBe("OPENCODE_GO");
     expect(createProvider({ ...base, provider: "ANTHROPIC" }).kind).toBe("ANTHROPIC");
     expect(createProvider({ ...base, provider: "MISTRAL" }).kind).toBe("MISTRAL");
     expect(createProvider({ ...base, provider: "LMSTUDIO" }).kind).toBe("LMSTUDIO");
@@ -108,5 +119,7 @@ describe("isAiUsable (#221)", () => {
     expect(await usable({ provider: "ANTHROPIC" })).toBe(false);
     expect(await usable({ provider: "ANTHROPIC", apiKeyEnc: encryptApiKey("sk-saved") })).toBe(true);
     expect(await usable({ provider: "MISTRAL", apiKeyEnc: encryptApiKey("sk", "b".repeat(32)) })).toBe(false);
+    expect(await usable({ provider: "OPENCODE_GO" })).toBe(false);
+    expect(await usable({ provider: "OPENCODE_GO", apiKeyEnc: encryptApiKey("oc_sk-saved") })).toBe(true);
   });
 });

src/lib/ai/config.ts

@@ -24,6 +24,7 @@ export const PROVIDER_DEFAULTS: Record<AiProvider, { baseUrl: string; model: str
   // Generous: a local model on a laptop can take minutes for a cover letter.
   [AiProvider.LOCAL]: { baseUrl: "http://localhost:11434", model: null, timeoutSeconds: 300 },
   [AiProvider.OPENAI]: { baseUrl: "https://api.openai.com/v1", model: null, timeoutSeconds: 120 },
+  [AiProvider.OPENCODE_GO]: { baseUrl: "https://opencode.ai/inference/go", model: "glm-5.3-flash", timeoutSeconds: 120 },
   [AiProvider.ANTHROPIC]: { baseUrl: "https://api.anthropic.com", model: "claude-sonnet-5", timeoutSeconds: 120 },
   [AiProvider.MISTRAL]: { baseUrl: "https://api.mistral.ai/v1", model: "mistral-medium-latest", timeoutSeconds: 120 },
   // As generous as a local model, and loading one can take a while too.
@@ -33,6 +34,7 @@ export const PROVIDER_DEFAULTS: Record<AiProvider, { baseUrl: string; model: str
 const FROM_ENV: Record<NonNullable<Env["DOSSIER_AI_PROVIDER"]>, AiProvider> = {
   local: AiProvider.LOCAL,
   openai: AiProvider.OPENAI,
+  "opencode-go": AiProvider.OPENCODE_GO,
   anthropic: AiProvider.ANTHROPIC,
   mistral: AiProvider.MISTRAL,
   lmstudio: AiProvider.LMSTUDIO,
@@ -126,6 +128,9 @@ export async function isAiUsable(): Promise<boolean> {
   if (!(await isAiConfigured())) return false;
   const config = await loadAiConfig();
   if (!config.model || config.apiKeyProblem) return false;
-  const needsKey = config.provider === AiProvider.ANTHROPIC || config.provider === AiProvider.MISTRAL;
+  const needsKey =
+    config.provider === AiProvider.ANTHROPIC ||
+    config.provider === AiProvider.MISTRAL ||
+    config.provider === AiProvider.OPENCODE_GO;
   return !needsKey || Boolean(config.apiKey);
 }

src/lib/ai/index.ts

@@ -10,20 +10,23 @@ import { AnthropicProvider } from "./providers/anthropic";
 import { LocalProvider } from "./providers/local";
 import { LmStudioProvider } from "./providers/lmstudio";
 import { MistralProvider } from "./providers/mistral";
+import { OpenCodeGoProvider } from "./providers/opencode-go";
 import { OpenAICompatibleProvider } from "./providers/openai-compatible";
 
 export { AiError } from "./errors";
 export type { AIProvider, ChatMessage, CompletionRequest, JsonCompletionRequest } from "./provider";
 
 /** A provider for `config`. A saved key that can't be read stops hosted providers with a clear message. */
-export function createProvider(config: AiConfig): AIProvider {
+export function createProvider(config: AiConfig, sessionId?: string): AIProvider {
   const { provider, baseUrl, model, apiKey, apiKeyProblem, timeoutMs } = config;
   if (apiKeyProblem && providerOption(provider).apiKey) throw AiError.config(apiKeyProblem);
   switch (provider) {
     case AiProvider.LOCAL:
       return new LocalProvider(baseUrl, model, timeoutMs);
     case AiProvider.OPENAI:
       return new OpenAICompatibleProvider("OPENAI", baseUrl, apiKey, model, timeoutMs);
+    case AiProvider.OPENCODE_GO:
+      return new OpenCodeGoProvider(baseUrl, apiKey, model, timeoutMs, sessionId);
     case AiProvider.ANTHROPIC:
       return new AnthropicProvider(baseUrl, apiKey, model, timeoutMs);
     case AiProvider.MISTRAL:
@@ -34,6 +37,6 @@ export function createProvider(config: AiConfig): AIProvider {
 }
 
 /** The provider configured in Settings, or by the DOSSIER_AI_* variables. */
-export async function getAiProvider(): Promise<AIProvider> {
-  return createProvider(await loadAiConfig());
+export async function getAiProvider(sessionId?: string): Promise<AIProvider> {
+  return createProvider(await loadAiConfig(), sessionId);
 }

src/lib/ai/provider-options.ts

@@ -60,6 +60,22 @@ export const PROVIDER_OPTIONS: readonly {
     defaultModel: "",
     apiKey: { label: "API key" },
   },
+  {
+    value: AiProvider.OPENCODE_GO,
+    name: "OpenCode Go",
+    vendor: "OpenCode",
+    description: "Open coding models through an OpenCode Go subscription.",
+    tag: "Hosted · Go subscription",
+    local: false,
+    defaultUrl: "https://opencode.ai/inference/go",
+    urlPlaceholder: "https://opencode.ai/inference/go",
+    urlHint: "The Go base URL from OpenCode's docs. Dossier adds the OpenAI-compatible path for chat.",
+    defaultModel: "glm-5.3-flash",
+    apiKey: {
+      label: "OpenCode API key",
+      hint: "From the OpenCode Console. The same key works with OpenCode Zen, but Go models need a Go or Go Plus plan.",
+    },
+  },
   {
     value: AiProvider.ANTHROPIC,
     name: "Anthropic",

src/lib/ai/providers.live.test.ts

@@ -5,19 +5,23 @@ import { AnthropicProvider } from "./providers/anthropic";
 import { LmStudioProvider } from "./providers/lmstudio";
 import { LocalProvider } from "./providers/local";
 import { MistralProvider } from "./providers/mistral";
+import { OpenCodeGoProvider } from "./providers/opencode-go";
 import { OpenAICompatibleProvider } from "./providers/openai-compatible";
 import type { AIProvider } from "./provider";
-import { liveLmStudio, liveMistral } from "./test-live";
+import { liveLmStudio, liveMistral, liveOpenCodeGo } from "./test-live";
 
 // Checks against real models, skipped unless asked for:
 //
 //   docker compose up -d ollama   # with qwen2.5:3b pulled (see compose.yml)
 //   AI_LIVE=ollama docker/bin/yarn vitest run src/lib/ai/providers.live.test.ts
 //   AI_LIVE=anthropic ANTHROPIC_TEST_KEY=… docker/bin/yarn vitest run src/lib/ai/providers.live.test.ts
 //   AI_LIVE=mistral docker/bin/yarn vitest run src/lib/ai/providers.live.test.ts
+//   AI_LIVE=opencode-go OPENCODE_TEST_KEY=… docker/bin/yarn vitest run src/lib/ai/providers.live.test.ts
 //
 // Mistral's key is DOSSIER_AI_KEY when that's set, or else the one saved in
-// Settings → AI provider with Mistral chosen.
+// Settings → AI provider with Mistral chosen. OpenCode Go's key is
+// OPENCODE_TEST_KEY (and OPENCODE_TEST_MODEL), or the one saved in Settings
+// with OpenCode Go chosen.
 //
 //   AI_LIVE=lmstudio docker/bin/yarn vitest run src/lib/ai/providers.live.test.ts
 //
@@ -111,6 +115,29 @@ describe.runIf(mistral)("live: Mistral", () => {
   });
 });
 
+const openCodeGo = live.has("opencode-go") ? await liveOpenCodeGo() : null;
+
+describe.runIf(openCodeGo)("live: OpenCode Go", () => {
+  const { key, model } = openCodeGo ?? { key: "", model: "" };
+  const make = (apiKey = key) => new OpenCodeGoProvider("https://opencode.ai/inference/go", apiKey, model, 120_000);
+  checks("chat completions", () => make(), model);
+
+  it("answers a reading task in plain JSON", async () => {
+    const result = await make().complete({
+      system: "Extract the job details from the posting.",
+      messages: [{ role: "user", content: posting }],
+      schema: Posting,
+      schemaName: "posting",
+      effort: "low",
+    });
+    expect(result).toMatchObject({ company: "Ledgerly", city: "Amsterdam" });
+  });
+
+  it("refuses a wrong key in plain words", async () => {
+    await expect(make("oc_sk-wrong").test()).rejects.toMatchObject({ kind: "auth" });
+  });
+});
+
 const lmStudio = live.has("lmstudio") ? await liveLmStudio() : null;
 
 describe.runIf(lmStudio)("live: LM Studio", () => {