modeldispatcher-browser-agent 0.6.0

package/dist/providers/ollama.d.ts

@@ -0,0 +1,10 @@
+/**
+ * Ollama — a local (or self-hosted) model server, no API key. Always uses
+ * `/api/chat` (message-list based) rather than switching between `/generate`
+ * and `/chat` per call shape — a single-turn prompt is just a one-message
+ * chat, so one code path covers both without the caller needing to know
+ * which Ollama endpoint that implies.
+ */
+import type { ChatMessage, ProviderCallConfig, RequestOptions } from "../types.js";
+export declare function completeOllama(cfg: ProviderCallConfig, messages: readonly ChatMessage[], options?: RequestOptions): Promise<string>;
+export declare function streamOllama(cfg: ProviderCallConfig, messages: readonly ChatMessage[], onChunk: (fullTextSoFar: string) => void, options?: RequestOptions): Promise<string>;

package/dist/providers/ollama.js

@@ -0,0 +1,61 @@
+/**
+ * Ollama — a local (or self-hosted) model server, no API key. Always uses
+ * `/api/chat` (message-list based) rather than switching between `/generate`
+ * and `/chat` per call shape — a single-turn prompt is just a one-message
+ * chat, so one code path covers both without the caller needing to know
+ * which Ollama endpoint that implies.
+ */
+import { consumeLines, fetchWithRetry, validateOllamaUrl } from "../http.js";
+const DEFAULT_MAX_TOKENS = 1024;
+function toApiMessages(messages, systemInstruction) {
+    const hasSystemTurn = messages.some((m) => m.role === "system");
+    const system = !hasSystemTurn && systemInstruction ? [{ role: "system", content: systemInstruction }] : [];
+    return [...system, ...messages.map((m) => ({ role: m.role, content: m.content }))];
+}
+function buildBody(cfg, messages, options, stream) {
+    return {
+        model: cfg.model,
+        messages: toApiMessages(messages, options.systemInstruction),
+        stream,
+        options: { num_predict: options.maxTokens ?? DEFAULT_MAX_TOKENS },
+    };
+}
+export async function completeOllama(cfg, messages, options = {}) {
+    const baseUrl = validateOllamaUrl(cfg.ollamaUrl);
+    const res = await fetchWithRetry(`${baseUrl}/api/chat`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify(buildBody(cfg, messages, options, false)),
+    }, options.signal);
+    if (!res.ok)
+        throw new Error(`Ollama error: HTTP ${res.status}. Is Ollama running?`);
+    const data = await res.json();
+    return data.message?.content ?? "";
+}
+export async function streamOllama(cfg, messages, onChunk, options = {}) {
+    const baseUrl = validateOllamaUrl(cfg.ollamaUrl);
+    const res = await fetchWithRetry(`${baseUrl}/api/chat`, {
+        method: "POST",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify(buildBody(cfg, messages, options, true)),
+    }, options.signal);
+    if (!res.ok)
+        throw new Error(`Ollama error: HTTP ${res.status}. Is Ollama running?`);
+    let full = "";
+    await consumeLines(res, (line) => {
+        if (!line.trim())
+            return;
+        try {
+            const text = JSON.parse(line).message?.content;
+            if (text) {
+                full += text;
+                onChunk(full);
+            }
+        }
+        catch {
+            // Skip a malformed/partial NDJSON line.
+        }
+    });
+    return full;
+}
+//# sourceMappingURL=ollama.js.map

package/dist/providers/ollama.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ollama.js","sourceRoot":"","sources":["../../src/providers/ollama.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,YAAY,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,YAAY,CAAC;AAG7E,MAAM,kBAAkB,GAAG,IAAI,CAAC;AAEhC,SAAS,aAAa,CAAC,QAAgC,EAAE,iBAA0B;IACjF,MAAM,aAAa,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IAChE,MAAM,MAAM,GAAG,CAAC,aAAa,IAAI,iBAAiB,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,iBAAiB,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC3G,OAAO,CAAC,GAAG,MAAM,EAAE,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC;AACrF,CAAC;AAED,SAAS,SAAS,CAAC,GAAuB,EAAE,QAAgC,EAAE,OAAuB,EAAE,MAAe;IACpH,OAAO;QACL,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,QAAQ,EAAE,aAAa,CAAC,QAAQ,EAAE,OAAO,CAAC,iBAAiB,CAAC;QAC5D,MAAM;QACN,OAAO,EAAE,EAAE,WAAW,EAAE,OAAO,CAAC,SAAS,IAAI,kBAAkB,EAAE;KAClE,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,cAAc,CAClC,GAAuB,EACvB,QAAgC,EAChC,UAA0B,EAAE;IAE5B,MAAM,OAAO,GAAG,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IACjD,MAAM,GAAG,GAAG,MAAM,cAAc,CAC9B,GAAG,OAAO,WAAW,EACrB;QACE,MAAM,EAAE,MAAM;QACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE;QAC/C,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC;KAC/D,EACD,OAAO,CAAC,MAAM,CACf,CAAC;IACF,IAAI,CAAC,GAAG,CAAC,EAAE;QAAE,MAAM,IAAI,KAAK,CAAC,sBAAsB,GAAG,CAAC,MAAM,sBAAsB,CAAC,CAAC;IACrF,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,IAAI,EAAE,CAAC;IAC9B,OAAO,IAAI,CAAC,OAAO,EAAE,OAAO,IAAI,EAAE,CAAC;AACrC,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,YAAY,CAChC,GAAuB,EACvB,QAAgC,EAChC,OAAwC,EACxC,UAA0B,EAAE;IAE5B,MAAM,OAAO,GAAG,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IACjD,MAAM,GAAG,GAAG,MAAM,cAAc,CAC9B,GAAG,OAAO,WAAW,EACrB;QACE,MAAM,EAAE,MAAM;QACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE;QAC/C,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC;KAC9D,EACD,OAAO,CAAC,MAAM,CACf,CAAC;IACF,IAAI,CAAC,GAAG,CAAC,EAAE;QAAE,MAAM,IAAI,KAAK,CAAC,sBAAsB,GAAG,CAAC,MAAM,sBAAsB,CAAC,CAAC;IAErF,IAAI,IAAI,GAAG,EAAE,CAAC;IACd,MAAM,YAAY,CAAC,GAAG,EAAE,CAAC,IAAI,EAAE,EAAE;QAC/B,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE;YAAE,OAAO;QACzB,IAAI,CAAC;YACH,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,OAAO,CAAC;YAC/C,IAAI,IAAI,EAAE,CAAC;gBACT,IAAI,IAAI,IAAI,CAAC;gBACb,OAAO,CAAC,IAAI,CAAC,CAAC;YAChB,CAAC;QACH,CAAC;QAAC,MAAM,CAAC;YACP,wCAAwC;QAC1C,CAAC;IACH,CAAC,CAAC,CAAC;IACH,OAAO,IAAI,CAAC;AACd,CAAC"}

package/dist/providers/openaiCompatible.d.ts

@@ -0,0 +1,9 @@
+/**
+ * OpenAI and Groq — both speak the same chat-completions REST shape, direct
+ * from the browser, no SDK. Groq is just the same code pointed at its own
+ * base URL (matches how the Python core's GroqProvider relates to
+ * OpenAIProvider — same idea, browser side).
+ */
+import type { ChatMessage, ProviderCallConfig, RequestOptions } from "../types.js";
+export declare function completeOpenAICompatible(provider: "openai" | "groq", cfg: ProviderCallConfig, messages: readonly ChatMessage[], options?: RequestOptions): Promise<string>;
+export declare function streamOpenAICompatible(provider: "openai" | "groq", cfg: ProviderCallConfig, messages: readonly ChatMessage[], onChunk: (fullTextSoFar: string) => void, options?: RequestOptions): Promise<string>;

package/dist/providers/openaiCompatible.js

@@ -0,0 +1,64 @@
+/**
+ * OpenAI and Groq — both speak the same chat-completions REST shape, direct
+ * from the browser, no SDK. Groq is just the same code pointed at its own
+ * base URL (matches how the Python core's GroqProvider relates to
+ * OpenAIProvider — same idea, browser side).
+ */
+import { consumeLines, describeHttpError, fetchWithRetry } from "../http.js";
+const DEFAULT_MAX_TOKENS = 1024;
+const BASE_URLS = {
+    openai: "https://api.openai.com/v1/chat/completions",
+    groq: "https://api.groq.com/openai/v1/chat/completions",
+};
+function toApiMessages(messages, systemInstruction) {
+    const hasSystemTurn = messages.some((m) => m.role === "system");
+    const system = !hasSystemTurn && systemInstruction ? [{ role: "system", content: systemInstruction }] : [];
+    return [...system, ...messages.map((m) => ({ role: m.role, content: m.content }))];
+}
+function buildBody(cfg, messages, options, stream) {
+    return {
+        model: cfg.model,
+        max_tokens: options.maxTokens ?? DEFAULT_MAX_TOKENS,
+        messages: toApiMessages(messages, options.systemInstruction),
+        ...(options.jsonMode ? { response_format: { type: "json_object" } } : {}),
+        ...(stream ? { stream: true } : {}),
+    };
+}
+async function request(provider, cfg, messages, options, stream) {
+    const res = await fetchWithRetry(BASE_URLS[provider], {
+        method: "POST",
+        headers: { "Content-Type": "application/json", Authorization: `Bearer ${cfg.apiKey}` },
+        body: JSON.stringify(buildBody(cfg, messages, options, stream)),
+    }, options.signal);
+    if (!res.ok)
+        throw new Error(await describeHttpError(res));
+    return res;
+}
+export async function completeOpenAICompatible(provider, cfg, messages, options = {}) {
+    const res = await request(provider, cfg, messages, options, false);
+    const data = await res.json();
+    return data.choices?.[0]?.message?.content ?? "";
+}
+export async function streamOpenAICompatible(provider, cfg, messages, onChunk, options = {}) {
+    const res = await request(provider, cfg, messages, options, true);
+    let full = "";
+    await consumeLines(res, (line) => {
+        if (!line.startsWith("data: "))
+            return;
+        const raw = line.slice(6).trim();
+        if (raw === "[DONE]")
+            return;
+        try {
+            const delta = JSON.parse(raw).choices?.[0]?.delta?.content;
+            if (delta) {
+                full += delta;
+                onChunk(full);
+            }
+        }
+        catch {
+            // Skip a malformed/partial SSE frame.
+        }
+    });
+    return full;
+}
+//# sourceMappingURL=openaiCompatible.js.map

package/dist/providers/openaiCompatible.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openaiCompatible.js","sourceRoot":"","sources":["../../src/providers/openaiCompatible.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,YAAY,EAAE,iBAAiB,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAG7E,MAAM,kBAAkB,GAAG,IAAI,CAAC;AAEhC,MAAM,SAAS,GAAsC;IACnD,MAAM,EAAE,4CAA4C;IACpD,IAAI,EAAE,iDAAiD;CACxD,CAAC;AAEF,SAAS,aAAa,CAAC,QAAgC,EAAE,iBAA0B;IACjF,MAAM,aAAa,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC;IAChE,MAAM,MAAM,GAAG,CAAC,aAAa,IAAI,iBAAiB,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,iBAAiB,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IAC3G,OAAO,CAAC,GAAG,MAAM,EAAE,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC;AACrF,CAAC;AAED,SAAS,SAAS,CAAC,GAAuB,EAAE,QAAgC,EAAE,OAAuB,EAAE,MAAe;IACpH,OAAO;QACL,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,UAAU,EAAE,OAAO,CAAC,SAAS,IAAI,kBAAkB;QACnD,QAAQ,EAAE,aAAa,CAAC,QAAQ,EAAE,OAAO,CAAC,iBAAiB,CAAC;QAC5D,GAAG,CAAC,OAAO,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,EAAE,IAAI,EAAE,aAAa,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QACzE,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACpC,CAAC;AACJ,CAAC;AAED,KAAK,UAAU,OAAO,CACpB,QAA2B,EAC3B,GAAuB,EACvB,QAAgC,EAChC,OAAuB,EACvB,MAAe;IAEf,MAAM,GAAG,GAAG,MAAM,cAAc,CAC9B,SAAS,CAAC,QAAQ,CAAC,EACnB;QACE,MAAM,EAAE,MAAM;QACd,OAAO,EAAE,EAAE,cAAc,EAAE,kBAAkB,EAAE,aAAa,EAAE,UAAU,GAAG,CAAC,MAAM,EAAE,EAAE;QACtF,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,SAAS,CAAC,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;KAChE,EACD,OAAO,CAAC,MAAM,CACf,CAAC;IACF,IAAI,CAAC,GAAG,CAAC,EAAE;QAAE,MAAM,IAAI,KAAK,CAAC,MAAM,iBAAiB,CAAC,GAAG,CAAC,CAAC,CAAC;IAC3D,OAAO,GAAG,CAAC;AACb,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,wBAAwB,CAC5C,QAA2B,EAC3B,GAAuB,EACvB,QAAgC,EAChC,UAA0B,EAAE;IAE5B,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,EAAE,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,KAAK,CAAC,CAAC;IACnE,MAAM,IAAI,GAAG,MAAM,GAAG,CAAC,IAAI,EAAE,CAAC;IAC9B,OAAO,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,IAAI,EAAE,CAAC;AACnD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,sBAAsB,CAC1C,QAA2B,EAC3B,GAAuB,EACvB,QAAgC,EAChC,OAAwC,EACxC,UAA0B,EAAE;IAE5B,MAAM,GAAG,GAAG,MAAM,OAAO,CAAC,QAAQ,EAAE,GAAG,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC;IAClE,IAAI,IAAI,GAAG,EAAE,CAAC;IACd,MAAM,YAAY,CAAC,GAAG,EAAE,CAAC,IAAI,EAAE,EAAE;QAC/B,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAC;YAAE,OAAO;QACvC,MAAM,GAAG,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;QACjC,IAAI,GAAG,KAAK,QAAQ;YAAE,OAAO;QAC7B,IAAI,CAAC;YACH,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC;YAC3D,IAAI,KAAK,EAAE,CAAC;gBACV,IAAI,IAAI,KAAK,CAAC;gBACd,OAAO,CAAC,IAAI,CAAC,CAAC;YAChB,CAAC;QACH,CAAC;QAAC,MAAM,CAAC;YACP,sCAAsC;QACxC,CAAC;IACH,CAAC,CAAC,CAAC;IACH,OAAO,IAAI,CAAC;AACd,CAAC"}

package/dist/registry.d.ts

@@ -0,0 +1,18 @@
+/**
+ * The provider registry — one canonical source for the metadata every app's
+ * settings screen was independently duplicating (name, default model, key
+ * placeholder, where to get a key).
+ */
+import type { ProviderId, ProviderInfo } from "./types.js";
+export declare const PROVIDERS: Record<ProviderId, ProviderInfo>;
+export declare function isKnownProvider(id: string): id is ProviderId;
+/**
+ * A curated, hand-maintained shortlist of current models per vendor, for
+ * rendering a model `<select>` instead of a free-text field. Necessarily
+ * goes stale as vendors ship new models — update this list by hand when
+ * that happens; it isn't fetched from anywhere. `PROVIDERS[id].defaultModel`
+ * is always included (first) even if it's ever edited out of this list.
+ */
+export declare const MODEL_OPTIONS: Record<ProviderId, readonly string[]>;
+/** `MODEL_OPTIONS[provider]` with `defaultModel` guaranteed present and first. */
+export declare function modelOptionsFor(provider: ProviderId): readonly string[];

package/dist/registry.js

@@ -0,0 +1,77 @@
+/**
+ * The provider registry — one canonical source for the metadata every app's
+ * settings screen was independently duplicating (name, default model, key
+ * placeholder, where to get a key).
+ */
+export const PROVIDERS = {
+    gemini: {
+        id: "gemini",
+        name: "Google Gemini",
+        free: false,
+        defaultModel: "gemini-2.0-flash",
+        placeholder: "AIza...",
+        infoUrl: "https://aistudio.google.com/app/apikey",
+        infoText: "Get a free key from Google AI Studio",
+    },
+    groq: {
+        id: "groq",
+        name: "Groq",
+        free: true,
+        defaultModel: "llama-3.1-8b-instant",
+        placeholder: "gsk_...",
+        infoUrl: "https://console.groq.com/keys",
+        infoText: "Get a free key from Groq Console",
+    },
+    ollama: {
+        id: "ollama",
+        name: "Ollama (local)",
+        free: true,
+        noKey: true,
+        defaultModel: "llama3.2",
+        placeholder: "http://localhost:11434",
+        infoUrl: "https://ollama.ai",
+        infoText: "Install Ollama on your machine",
+    },
+    anthropic: {
+        id: "anthropic",
+        name: "Anthropic Claude",
+        free: false,
+        defaultModel: "claude-haiku-4-5-20251001",
+        placeholder: "sk-ant-...",
+        infoUrl: "https://console.anthropic.com/settings/keys",
+        infoText: "Get a key from the Anthropic Console",
+    },
+    openai: {
+        id: "openai",
+        name: "OpenAI",
+        free: false,
+        defaultModel: "gpt-4o-mini",
+        placeholder: "sk-...",
+        infoUrl: "https://platform.openai.com/api-keys",
+        infoText: "Get a key from the OpenAI Platform",
+    },
+};
+export function isKnownProvider(id) {
+    return Object.hasOwn(PROVIDERS, id);
+}
+/**
+ * A curated, hand-maintained shortlist of current models per vendor, for
+ * rendering a model `<select>` instead of a free-text field. Necessarily
+ * goes stale as vendors ship new models — update this list by hand when
+ * that happens; it isn't fetched from anywhere. `PROVIDERS[id].defaultModel`
+ * is always included (first) even if it's ever edited out of this list.
+ */
+export const MODEL_OPTIONS = {
+    gemini: ["gemini-2.0-flash", "gemini-2.5-flash", "gemini-2.5-pro"],
+    groq: ["llama-3.1-8b-instant", "llama-3.3-70b-versatile", "openai/gpt-oss-120b"],
+    ollama: ["llama3.2", "llama3.1", "mistral", "qwen2.5"],
+    anthropic: ["claude-haiku-4-5-20251001", "claude-sonnet-4-5", "claude-opus-4-8"],
+    openai: ["gpt-4o-mini", "gpt-4o", "gpt-4.1-mini"],
+};
+/** `MODEL_OPTIONS[provider]` with `defaultModel` guaranteed present and first. */
+export function modelOptionsFor(provider) {
+    const options = MODEL_OPTIONS[provider];
+    const defaultModel = PROVIDERS[provider].defaultModel;
+    return options.includes(defaultModel) ? options : [defaultModel, ...options];
+}
+//# sourceMappingURL=registry.js.map

package/dist/registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"registry.js","sourceRoot":"","sources":["../src/registry.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAIH,MAAM,CAAC,MAAM,SAAS,GAAqC;IACzD,MAAM,EAAE;QACN,EAAE,EAAE,QAAQ;QACZ,IAAI,EAAE,eAAe;QACrB,IAAI,EAAE,KAAK;QACX,YAAY,EAAE,kBAAkB;QAChC,WAAW,EAAE,SAAS;QACtB,OAAO,EAAE,wCAAwC;QACjD,QAAQ,EAAE,sCAAsC;KACjD;IACD,IAAI,EAAE;QACJ,EAAE,EAAE,MAAM;QACV,IAAI,EAAE,MAAM;QACZ,IAAI,EAAE,IAAI;QACV,YAAY,EAAE,sBAAsB;QACpC,WAAW,EAAE,SAAS;QACtB,OAAO,EAAE,+BAA+B;QACxC,QAAQ,EAAE,kCAAkC;KAC7C;IACD,MAAM,EAAE;QACN,EAAE,EAAE,QAAQ;QACZ,IAAI,EAAE,gBAAgB;QACtB,IAAI,EAAE,IAAI;QACV,KAAK,EAAE,IAAI;QACX,YAAY,EAAE,UAAU;QACxB,WAAW,EAAE,wBAAwB;QACrC,OAAO,EAAE,mBAAmB;QAC5B,QAAQ,EAAE,gCAAgC;KAC3C;IACD,SAAS,EAAE;QACT,EAAE,EAAE,WAAW;QACf,IAAI,EAAE,kBAAkB;QACxB,IAAI,EAAE,KAAK;QACX,YAAY,EAAE,2BAA2B;QACzC,WAAW,EAAE,YAAY;QACzB,OAAO,EAAE,6CAA6C;QACtD,QAAQ,EAAE,sCAAsC;KACjD;IACD,MAAM,EAAE;QACN,EAAE,EAAE,QAAQ;QACZ,IAAI,EAAE,QAAQ;QACd,IAAI,EAAE,KAAK;QACX,YAAY,EAAE,aAAa;QAC3B,WAAW,EAAE,QAAQ;QACrB,OAAO,EAAE,sCAAsC;QAC/C,QAAQ,EAAE,oCAAoC;KAC/C;CACF,CAAC;AAEF,MAAM,UAAU,eAAe,CAAC,EAAU;IACxC,OAAO,MAAM,CAAC,MAAM,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;AACtC,CAAC;AAED;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,aAAa,GAA0C;IAClE,MAAM,EAAE,CAAC,kBAAkB,EAAE,kBAAkB,EAAE,gBAAgB,CAAC;IAClE,IAAI,EAAE,CAAC,sBAAsB,EAAE,yBAAyB,EAAE,qBAAqB,CAAC;IAChF,MAAM,EAAE,CAAC,UAAU,EAAE,UAAU,EAAE,SAAS,EAAE,SAAS,CAAC;IACtD,SAAS,EAAE,CAAC,2BAA2B,EAAE,mBAAmB,EAAE,iBAAiB,CAAC;IAChF,MAAM,EAAE,CAAC,aAAa,EAAE,QAAQ,EAAE,cAAc,CAAC;CAClD,CAAC;AAEF,kFAAkF;AAClF,MAAM,UAAU,eAAe,CAAC,QAAoB;IAClD,MAAM,OAAO,GAAG,aAAa,CAAC,QAAQ,CAAC,CAAC;IACxC,MAAM,YAAY,GAAG,SAAS,CAAC,QAAQ,CAAC,CAAC,YAAY,CAAC;IACtD,OAAO,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,YAAY,EAAE,GAAG,OAAO,CAAC,CAAC;AAC/E,CAAC"}

package/dist/types.d.ts

@@ -0,0 +1,97 @@
+/**
+ * Shared vocabulary for the browser agent.
+ *
+ * Mirrors the split model-dispatcher's Python core uses: plain, dependency-free
+ * value types here, behaviour in `providers/*` and `agent.ts`.
+ */
+export type ProviderId = "gemini" | "openai" | "anthropic" | "groq" | "ollama";
+/** Static metadata about a provider — enough to render a settings picker. */
+export interface ProviderInfo {
+    readonly id: ProviderId;
+    readonly name: string;
+    /** Whether this provider has a genuinely free tier (not just "has a trial"). */
+    readonly free: boolean;
+    /** True for providers that need no API key at all (e.g. a local Ollama). */
+    readonly noKey?: boolean;
+    readonly defaultModel: string;
+    /** UI placeholder text for the key input (e.g. "sk-ant-..."). */
+    readonly placeholder: string;
+    readonly infoUrl: string;
+    readonly infoText: string;
+}
+/**
+ * One configured vendor: its model, and the key(s) pooled for it. More than
+ * one key is tried in order (e.g. a personal key and a team key) before
+ * this vendor is considered exhausted and the dispatcher moves to the next
+ * configured provider. Ignored for `provider: "ollama"`, which authenticates
+ * via `AgentConfig.ollamaUrl` instead of a key.
+ */
+export interface ProviderCredential {
+    readonly provider: ProviderId;
+    readonly model: string;
+    readonly apiKeys: readonly string[];
+}
+/**
+ * The full BYOK configuration: an ordered fallback list of vendors (each
+ * with its own pooled keys) to try in turn, plus Ollama's local server URL
+ * (shared across the whole config — there's only one local Ollama, not one
+ * per fallback slot). `dispatch`/`complete`/`streamComplete` walk this list
+ * candidate by candidate, mirroring the Python core's fallback chain.
+ */
+export interface AgentConfig {
+    readonly providers: readonly ProviderCredential[];
+    readonly ollamaUrl: string;
+}
+/**
+ * One concrete attempt: a specific vendor, model, and (for keyed vendors) a
+ * single already-resolved key — what an individual `providers/*` module
+ * needs to make one call. Distinct from `AgentConfig`, which holds the whole
+ * configured fallback list before the dispatch loop has picked a candidate.
+ */
+export interface ProviderCallConfig {
+    readonly provider: ProviderId;
+    readonly model: string;
+    readonly apiKey: string;
+    readonly ollamaUrl: string;
+}
+export type ChatRole = "system" | "user" | "assistant";
+export interface ChatMessage {
+    readonly role: ChatRole;
+    readonly content: string;
+}
+/** Options shared by both complete() and streamComplete()/streamChat(). */
+export interface RequestOptions {
+    /** System prompt / instruction, applied via each provider's native mechanism. */
+    systemInstruction?: string;
+    /** Request the provider's structured-JSON response mode where it has one
+     * (OpenAI-compatible providers only — Gemini's `responseMimeType` is
+     * always requested by callers that need it via a raw prompt instruction,
+     * Anthropic and Ollama have no such mode). */
+    jsonMode?: boolean;
+    maxTokens?: number;
+    /** Cancel the request (e.g. component unmount). Independent of, and
+     * merged with, the library's own request timeout. */
+    signal?: AbortSignal;
+}
+export interface StreamOptions extends RequestOptions {
+    /** Called with the *cumulative* text so far after every chunk, matching
+     * every existing app's streaming callback shape. */
+    onChunk: (fullTextSoFar: string) => void;
+}
+/** A free, public AI chat web product `openExternalChat` can hand a question
+ * off to — the no-API-key escape hatch, distinct from the `ProviderId`s above
+ * (which all require a BYOK credential). */
+export type ExternalChatProviderId = "chatgpt" | "claude" | "gemini" | "groq";
+/** Static metadata about one external chat provider. */
+export interface ExternalChatProviderInfo {
+    readonly id: ExternalChatProviderId;
+    readonly name: string;
+    /** Where `openExternalChat` sends the user when there's no known
+     * query-prefill parameter for this provider (or as the base URL the
+     * parameter is appended to). */
+    readonly homeUrl: string;
+    /** Build the deep-link URL for a given question, when this provider has a
+     * known (unofficial, reverse-engineered) prefill parameter. `null` when it
+     * doesn't — `openExternalChat` falls back to `homeUrl` plus the clipboard. */
+    readonly buildUrl: ((question: string) => string) | null;
+}

package/dist/types.js

@@ -0,0 +1,8 @@
+/**
+ * Shared vocabulary for the browser agent.
+ *
+ * Mirrors the split model-dispatcher's Python core uses: plain, dependency-free
+ * value types here, behaviour in `providers/*` and `agent.ts`.
+ */
+export {};
+//# sourceMappingURL=types.js.map

package/dist/types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA;;;;;GAKG"}

package/package.json

@@ -0,0 +1,40 @@
+{
+  "name": "modeldispatcher-browser-agent",
+  "version": "0.6.0",
+  "description": "Browser-native, bring-your-own-key AI agent core with cross-provider/cross-key fallback dispatch: multi-provider (Gemini/OpenAI/Anthropic/Groq/Ollama), streaming and non-streaming, no backend and no vendor SDK required. The client-side sibling to model-dispatcher for apps with no server to run it on.",
+  "license": "MIT",
+  "type": "module",
+  "types": "./dist/index.d.ts",
+  "module": "./dist/index.js",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    }
+  },
+  "files": [
+    "dist",
+    "src"
+  ],
+  "sideEffects": false,
+  "scripts": {
+    "build": "tsc -p tsconfig.build.json",
+    "typecheck": "tsc --noEmit",
+    "test": "vitest run"
+  },
+  "keywords": [
+    "llm",
+    "ai-agent",
+    "browser",
+    "gemini",
+    "openai",
+    "anthropic",
+    "groq",
+    "ollama",
+    "streaming"
+  ],
+  "devDependencies": {
+    "typescript": "^5.5.0",
+    "vitest": "^2.0.0"
+  }
+}

package/src/agent.ts

@@ -0,0 +1,206 @@
+/**
+ * The public entry points — now multi-provider: `cfg.providers` is an
+ * ordered fallback list (each vendor with one or more pooled keys) instead
+ * of a single vendor+key, so a rate limit, bad key, or outage on the first
+ * configured provider falls through to the next instead of failing the
+ * whole request. Mirrors the Python core's fallback chain, simplified for a
+ * browser client with no shared quota/routing state to manage.
+ *
+ *   - complete()       — one-shot prompt in, full text back (HighFive, StepByLearn)
+ *   - streamComplete()  — one-shot prompt in, incremental text via onChunk (JobFlowTracker's runStream)
+ *   - streamChat()      — a UI chat history in (auto-normalised via buildMessages), incremental text back (JobFlowTracker's/KanDOne's streamChat)
+ *
+ * All three dispatch on each candidate's `provider` to the matching
+ * provider module; callers never need their own provider switch statement.
+ */
+
+import { buildMessages, type BuildMessagesOptions } from "./messages.js";
+import { completeAnthropic, streamAnthropic } from "./providers/anthropic.js";
+import { completeGemini, streamGemini } from "./providers/gemini.js";
+import { completeOllama, streamOllama } from "./providers/ollama.js";
+import { completeOpenAICompatible, streamOpenAICompatible } from "./providers/openaiCompatible.js";
+import type {
+  AgentConfig,
+  ChatMessage,
+  ProviderCallConfig,
+  RequestOptions,
+  StreamOptions,
+} from "./types.js";
+
+/** One candidate that failed, kept only to compose
+ * {@link AllProvidersExhaustedError}'s message. */
+interface FailedAttempt {
+  readonly provider: string;
+  readonly error: string;
+}
+
+/** `cfg.providers` has nothing usable at all — no vendor has a pooled key,
+ * and Ollama isn't configured either. Raised before any network call, the
+ * same way the Python core's `NoProviderAvailableError` is. */
+export class NoProviderConfiguredError extends Error {
+  constructor() {
+    super("No AI provider is configured — add a provider and key (or Ollama) in AI settings.");
+    this.name = "NoProviderConfiguredError";
+  }
+}
+
+/** Every configured provider/key was tried and every one failed. */
+export class AllProvidersExhaustedError extends Error {
+  readonly attempts: readonly FailedAttempt[];
+
+  constructor(attempts: readonly FailedAttempt[]) {
+    super(`All configured AI providers failed: ${attempts.map((a) => `${a.provider} (${a.error})`).join("; ")}`);
+    this.name = "AllProvidersExhaustedError";
+    this.attempts = attempts;
+  }
+}
+
+/**
+ * Flatten the configured provider list into concrete, single-key attempts,
+ * in fallback order: each provider's own pooled keys are tried in the order
+ * they were added before moving to the next configured provider. Ollama
+ * contributes at most one candidate (it authenticates via `cfg.ollamaUrl`,
+ * not a key) and only when its own entry is present in the list — a
+ * provider with no usable credential is never offered as a candidate at
+ * all, the same rule the Python core's registry-building applies.
+ */
+function candidatesOf(cfg: AgentConfig): ProviderCallConfig[] {
+  const candidates: ProviderCallConfig[] = [];
+  for (const cred of cfg.providers) {
+    if (cred.provider === "ollama") {
+      candidates.push({ provider: "ollama", model: cred.model, apiKey: "", ollamaUrl: cfg.ollamaUrl });
+      continue;
+    }
+    for (const apiKey of cred.apiKeys) {
+      if (apiKey) candidates.push({ provider: cred.provider, model: cred.model, apiKey, ollamaUrl: cfg.ollamaUrl });
+    }
+  }
+  return candidates;
+}
+
+function messagesOf(promptOrMessages: string | readonly ChatMessage[]): readonly ChatMessage[] {
+  return typeof promptOrMessages === "string" ? [{ role: "user", content: promptOrMessages }] : promptOrMessages;
+}
+
+function completeOne(
+  candidate: ProviderCallConfig,
+  messages: readonly ChatMessage[],
+  options: RequestOptions,
+): Promise<string> {
+  switch (candidate.provider) {
+    case "gemini":
+      return completeGemini(candidate, messages, options);
+    case "anthropic":
+      return completeAnthropic(candidate, messages, options);
+    case "ollama":
+      return completeOllama(candidate, messages, options);
+    case "openai":
+    case "groq":
+      return completeOpenAICompatible(candidate.provider, candidate, messages, options);
+  }
+}
+
+/** Non-streaming completion, tried across every configured provider/key in
+ * order until one succeeds. `prompt` is wrapped as a single user turn; pass
+ * `messages` instead for an already-built multi-turn history.
+ *
+ * Raises:
+ *   NoProviderConfiguredError: `cfg.providers` has no usable credential at all.
+ *   AllProvidersExhaustedError: every candidate was tried and failed;
+ *     `.attempts` lists each provider and its error.
+ */
+export async function complete(
+  cfg: AgentConfig,
+  promptOrMessages: string | readonly ChatMessage[],
+  options: RequestOptions = {},
+): Promise<string> {
+  const candidates = candidatesOf(cfg);
+  if (candidates.length === 0) throw new NoProviderConfiguredError();
+  const messages = messagesOf(promptOrMessages);
+
+  const attempts: FailedAttempt[] = [];
+  for (const candidate of candidates) {
+    try {
+      return await completeOne(candidate, messages, options);
+    } catch (err) {
+      attempts.push({ provider: candidate.provider, error: (err as Error).message });
+    }
+  }
+  throw new AllProvidersExhaustedError(attempts);
+}
+
+function streamOne(
+  candidate: ProviderCallConfig,
+  messages: readonly ChatMessage[],
+  onChunk: (fullTextSoFar: string) => void,
+  options: RequestOptions,
+): Promise<string> {
+  switch (candidate.provider) {
+    case "gemini":
+      return streamGemini(candidate, messages, onChunk, options);
+    case "anthropic":
+      return streamAnthropic(candidate, messages, onChunk, options);
+    case "ollama":
+      return streamOllama(candidate, messages, onChunk, options);
+    case "openai":
+    case "groq":
+      return streamOpenAICompatible(candidate.provider, candidate, messages, onChunk, options);
+  }
+}
+
+/**
+ * Streaming completion, tried across every configured provider/key in
+ * order — but only up to the first chunk actually streamed. Once any text
+ * has reached `onChunk` for a candidate, a later failure on that same
+ * candidate is surfaced immediately rather than silently retried on the
+ * next provider: the caller has already rendered a partial answer, and
+ * switching providers mid-stream would duplicate or contradict it rather
+ * than complete it.
+ *
+ * Raises the same errors as {@link complete}.
+ */
+export async function streamComplete(
+  cfg: AgentConfig,
+  promptOrMessages: string | readonly ChatMessage[],
+  options: StreamOptions,
+): Promise<string> {
+  const candidates = candidatesOf(cfg);
+  if (candidates.length === 0) throw new NoProviderConfiguredError();
+  const messages = messagesOf(promptOrMessages);
+  const { onChunk, ...rest } = options;
+
+  const attempts: FailedAttempt[] = [];
+  for (const candidate of candidates) {
+    let startedStreaming = false;
+    try {
+      return await streamOne(
+        candidate,
+        messages,
+        (text) => {
+          startedStreaming = true;
+          onChunk(text);
+        },
+        rest,
+      );
+    } catch (err) {
+      if (startedStreaming) throw err;
+      attempts.push({ provider: candidate.provider, error: (err as Error).message });
+    }
+  }
+  throw new AllProvidersExhaustedError(attempts);
+}
+
+/**
+ * Streaming multi-turn chat: normalises a raw UI message list via
+ * {@link buildMessages} (role validation, alternation, sentinel-dropping)
+ * before dispatching — the one thing `streamComplete()` deliberately
+ * doesn't do, since its `messages` input is assumed already well-formed.
+ */
+export async function streamChat(
+  cfg: AgentConfig,
+  uiMessages: readonly ChatMessage[],
+  options: StreamOptions & BuildMessagesOptions,
+): Promise<string> {
+  const normalised = buildMessages(uiMessages, options);
+  return streamComplete(cfg, normalised, options);
+}