diff --git a/packages/ai/package.json b/packages/ai/package.json
new file mode 100644
index 0000000..42d474a
--- /dev/null
+++ b/packages/ai/package.json
@@ -0,0 +1,25 @@
+{
+  "name": "@pubscript/ai",
+  "version": "0.1.0",
+  "type": "module",
+  "main": "./src/index.ts",
+  "types": "./src/index.ts",
+  "exports": {
+    ".": "./src/index.ts"
+  },
+  "scripts": {
+    "dev": "echo \"@pubscript/ai has no dev server\" (exit 0)",
+    "build": "tsc --noEmit",
+    "test": "vitest run",
+    "lint": "eslint .",
+    "typecheck": "tsc --noEmit"
+  },
+  "dependencies": {
+    "@pubscript/core": "workspace:*"
+  },
+  "devDependencies": {
+    "@types/node": "^22.10.2",
+    "typescript": "^5.7.2",
+    "vitest": "^2.1.8"
+  }
+}
diff --git a/packages/ai/src/budget.test.ts b/packages/ai/src/budget.test.ts
new file mode 100644
index 0000000..1dd4a50
--- /dev/null
+++ b/packages/ai/src/budget.test.ts
@@ -0,0 +1,162 @@
+import { describe, expect, it } from "vitest";
+
+import {
+  allocateBudget,
+  countTokens,
+  DEFAULT_BUDGET,
+  enforceBudget,
+  measureBudget,
+} from "./index.js";
+import type { Message } from "./index.js";
+
+describe("countTokens", () => {
+  it("counts tokens for a known string", () => {
+    // "hello world" is 2 words ≈ 3-4 tokens; assert a stable lower bound.
+    const n = countTokens("hello world");
+    expect(n).toBeGreaterThan(0);
+    expect(Number.isInteger(n)).toBe(true);
+  });
+
+  it("returns 0 for an empty string", () => {
+    expect(countTokens("")).toBe(0);
+  });
+
+  it("scales with input length", () => {
+    const short = countTokens("a");
+    const long = countTokens("a".repeat(1000));
+    expect(long).toBeGreaterThan(short);
+  });
+});
+
+describe("allocateBudget", () => {
+  it("splits a 128k context window into per-slot caps", () => {
+    const budget = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 128_000,
+    });
+    expect(budget.system).toBe(12_800);
+    expect(budget.context).toBe(51_200);
+    expect(budget.history).toBe(38_400);
+    expect(budget.output).toBe(25_600);
+    expect(budget.total).toBe(128_000);
+  });
+
+  it("scales the output budget with context window size", () => {
+    const small = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 8_000,
+    });
+    const large = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 128_000,
+    });
+    expect(large.output).toBeGreaterThan(small.output);
+  });
+
+  it("never allocates more than half the context window to output", () => {
+    const budget = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 16_000,
+    });
+    expect(budget.output).toBeLessThanOrEqual(16_000 / 2);
+  });
+});
+
+describe("measureBudget", () => {
+  it("breaks down tokens by dimension", () => {
+    const messages: Message[] = [
+      { role: "system", content: "You are a helpful editor." },
+      { role: "user", content: "Summarize this chapter." },
+    ];
+    const measured = measureBudget(
+      messages.find((m) => m.role === "system")!.content,
+      "retrieved context",
+      messages.filter((m) => m.role !== "system"),
+      DEFAULT_BUDGET.output,
+    );
+    expect(measured.system).toBeGreaterThan(0);
+    expect(measured.context).toBeGreaterThan(0);
+    expect(measured.history).toBeGreaterThan(0);
+    expect(measured.output).toBe(DEFAULT_BUDGET.output);
+  });
+});
+
+describe("enforceBudget", () => {
+  it("passes when all dimensions are within budget", () => {
+    const budget = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 128_000,
+    });
+    const measured = {
+      system: 100,
+      context: 200,
+      history: 300,
+      output: 500,
+    };
+    expect(() => enforceBudget(measured, budget)).not.toThrow();
+  });
+
+  it("throws when the context dimension overflows", () => {
+    const budget = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 8_000,
+    });
+    const measured = {
+      system: 100,
+      context: budget.context + 1,
+      history: 100,
+      output: 100,
+    };
+    expect(() => enforceBudget(measured, budget)).toThrow(/context/);
+  });
+
+  it("throws when the total overflows even if each dimension fits", () => {
+    // Caps sum to more than the total, so each dimension can fit individually
+    // while the aggregate still exceeds the context window.
+    const budget = allocateBudget({
+      system: 30,
+      context: 30,
+      history: 30,
+      output: 30,
+      total: 1_000,
+    });
+    const measured = { system: 300, context: 300, history: 300, output: 300 };
+    expect(() => enforceBudget(measured, budget)).toThrow(/total/);
+  });
+
+  it("reports all overflowing dimensions", () => {
+    const budget = allocateBudget({
+      system: 10,
+      context: 40,
+      history: 30,
+      output: 20,
+      total: 8_000,
+    });
+    const measured = {
+      system: budget.system + 1,
+      context: budget.context + 1,
+      history: 100,
+      output: 100,
+    };
+    expect(() => enforceBudget(measured, budget)).toThrow(/system/);
+    expect(() => enforceBudget(measured, budget)).toThrow(/context/);
+  });
+});
diff --git a/packages/ai/src/budget.ts b/packages/ai/src/budget.ts
new file mode 100644
index 0000000..5ce5528
--- /dev/null
+++ b/packages/ai/src/budget.ts
@@ -0,0 +1,15 @@
+/**
+ * Token budget re-exports.
+ *
+ * The budget engine lives in `packages/core` (the single source of truth for
+ * token accounting). This module re-exports it so the AI layer has a single
+ * import surface, per CHUNK#8.md §3.
+ */
+
+export {
+  countTokens,
+  allocateBudget,
+  DEFAULT_BUDGET,
+} from "@pubscript/core";
+
+export type { Budget } from "@pubscript/core";
diff --git a/packages/ai/src/client.test.ts b/packages/ai/src/client.test.ts
new file mode 100644
index 0000000..880a165
--- /dev/null
+++ b/packages/ai/src/client.test.ts
@@ -0,0 +1,201 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
+
+import {
+  BudgetExceededError,
+  collectStream,
+  complete,
+  streamCompletion,
+  UnknownProviderError,
+} from "./index.js";
+import type { ResolvedProvider } from "./index.js";
+
+/** Build a `text/event-stream` Response from a list of SSE event payloads. */
+function sseResponse(events: string[], status = 200): Response {
+  const encoder = new TextEncoder();
+  const body = new ReadableStream<Uint8Array<ArrayBuffer>>({
+    start(controller) {
+      for (const event of events) {
+        controller.enqueue(encoder.encode(`data: ${event}\n\n`));
+      }
+      controller.enqueue(encoder.encode("data: [DONE]\n\n"));
+      controller.close();
+    },
+  });
+  return new Response(body, {
+    status,
+    headers: { "content-type": "text/event-stream" },
+  });
+}
+
+/** Build a JSON Response (for non-streaming completions). */
+function jsonResponse(body: unknown, status = 200): Response {
+  return new Response(JSON.stringify(body), {
+    status,
+    headers: { "content-type": "application/json" },
+  });
+}
+
+const config: ResolvedProvider = {
+  id: "test-openai",
+  kind: "openai-compatible",
+  baseUrl: "https://api.openai.com/v1",
+  apiKey: "sk-test",
+  model: "gpt-4o",
+  contextWindow: 128_000,
+};
+
+let savedFetch: typeof fetch;
+
+beforeEach(() => {
+  savedFetch = globalThis.fetch;
+});
+
+afterEach(() => {
+  globalThis.fetch = savedFetch;
+  vi.restoreAllMocks();
+});
+
+describe("streamCompletion", () => {
+  it("yields text deltas in order", async () => {
+    globalThis.fetch = (() =>
+      sseResponse([
+        JSON.stringify({ choices: [{ delta: { content: "Hello" } }] }),
+        JSON.stringify({ choices: [{ delta: { content: " world" } }] }),
+      ])) as unknown as typeof fetch;
+
+    const chunks: Array<{ type: string; text?: string }> = [];
+    for await (const chunk of streamCompletion(
+      config,
+      [{ role: "user", content: "hi" }],
+    )) {
+      chunks.push(chunk);
+    }
+
+    const textChunks = chunks.filter((c) => c.type === "text");
+    expect(textChunks.map((c) => c.text)).toEqual(["Hello", " world"]);
+  });
+
+  it("parses reasoning deltas separately from text", async () => {
+    globalThis.fetch = (() =>
+      sseResponse([
+        JSON.stringify({ choices: [{ delta: { reasoning_content: "let me think" } }] }),
+        JSON.stringify({ choices: [{ delta: { content: "the answer" } }] }),
+      ])) as unknown as typeof fetch;
+
+    const reasoning: string[] = [];
+    const text: string[] = [];
+    for await (const chunk of streamCompletion(
+      config,
+      [{ role: "user", content: "hi" }],
+    )) {
+      if (chunk.type === "reasoning" && chunk.reasoning) {
+        reasoning.push(chunk.reasoning);
+      } else if (chunk.type === "text" && chunk.text) {
+        text.push(chunk.text);
+      }
+    }
+
+    expect(reasoning).toEqual(["let me think"]);
+    expect(text).toEqual(["the answer"]);
+  });
+
+  it("tracks usage on the final done chunk", async () => {
+    globalThis.fetch = (() =>
+      sseResponse([
+        JSON.stringify({ choices: [{ delta: { content: "ok" } }] }),
+        JSON.stringify({
+          choices: [{ delta: {} }],
+          usage: { prompt_tokens: 12, completion_tokens: 3, total_tokens: 15 },
+        }),
+      ])) as unknown as typeof fetch;
+
+    const result = await collectStream(
+      config,
+      [{ role: "user", content: "hi" }],
+    );
+    expect(result.text).toBe("ok");
+    expect(result.usage).toEqual({ promptTokens: 12, completionTokens: 3, total: 15 });
+  });
+});
+
+describe("collectStream", () => {
+  it("assembles full text, reasoning, and usage", async () => {
+    globalThis.fetch = (() =>
+      sseResponse([
+        JSON.stringify({ choices: [{ delta: { reasoning_content: "hmm" } }] }),
+        JSON.stringify({ choices: [{ delta: { content: "Final" } }] }),
+        JSON.stringify({ choices: [{ delta: { content: " answer" } }] }),
+        JSON.stringify({
+          choices: [{ delta: {} }],
+          usage: { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7 },
+        }),
+      ])) as unknown as typeof fetch;
+
+    const result = await collectStream(
+      config,
+      [{ role: "user", content: "q" }],
+    );
+    expect(result.text).toBe("Final answer");
+    expect(result.reasoning).toBe("hmm");
+    expect(result.usage).toEqual({ promptTokens: 5, completionTokens: 2, total: 7 });
+  });
+});
+
+describe("complete (non-streaming)", () => {
+  it("returns the full text and usage", async () => {
+    globalThis.fetch = (() =>
+      jsonResponse({
+        choices: [{ message: { content: "full text" } }],
+        usage: { prompt_tokens: 1, completion_tokens: 2, total_tokens: 3 },
+      })) as unknown as typeof fetch;
+
+    const result = await complete(
+      config,
+      [{ role: "user", content: "hi" }],
+    );
+    expect(result.text).toBe("full text");
+    expect(result.usage).toEqual({ promptTokens: 1, completionTokens: 2, total: 3 });
+  });
+});
+
+describe("retry on 429", () => {
+  it("retries a rate-limited request and succeeds", async () => {
+    let calls = 0;
+    globalThis.fetch = ((() => {
+      calls += 1;
+      if (calls === 1) {
+        return jsonResponse({ error: { message: "rate limited" } }, 429);
+      }
+      return sseResponse([JSON.stringify({ choices: [{ delta: { content: "recovered" } }] })]);
+    })) as unknown as typeof fetch;
+
+    const result = await collectStream(
+      config,
+      [{ role: "user", content: "hi" }],
+    );
+    expect(result.text).toBe("recovered");
+    expect(calls).toBe(2);
+  });
+});
+
+describe("budget enforcement", () => {
+  it("throws BudgetExceededError before any network call", async () => {
+    const tinyConfig: ResolvedProvider = { ...config, contextWindow: 8 };
+    const fetchSpy = vi.fn();
+    globalThis.fetch = fetchSpy as unknown as typeof fetch;
+
+    await expect(
+      collectStream(tinyConfig, [{ role: "system", content: "x".repeat(200) }]),
+    ).rejects.toBeInstanceOf(BudgetExceededError);
+    expect(fetchSpy).not.toHaveBeenCalled();
+  });
+});
+
+describe("provider resolution", () => {
+  it("throws UnknownProviderError for an unrecognized kind", async () => {
+    const badConfig = { ...config, kind: "nope" } as unknown as ResolvedProvider;
+    await expect(
+      collectStream(badConfig, [{ role: "user", content: "hi" }]),
+    ).rejects.toBeInstanceOf(UnknownProviderError);
+  });
+});
diff --git a/packages/ai/src/client.ts b/packages/ai/src/client.ts
new file mode 100644
index 0000000..13ca6eb
--- /dev/null
+++ b/packages/ai/src/client.ts
@@ -0,0 +1,171 @@
+/**
+ * The completion client: the public entry point for talking to a provider.
+ *
+ * It enforces the token budget (via `packages/core`), resolves the adapter for
+ * a provider, and exposes:
+ * - `streamCompletion`: an async iterator over `StreamChunk`s.
+ * - `collectStream`: collect all chunks + usage into a single result.
+ * - `complete`: a non-streaming completion (used for tool use).
+ *
+ * See CHUNK#8.md §3.
+ */
+
+import {
+  allocateBudget,
+  countTokens,
+  DEFAULT_BUDGET,
+} from "@pubscript/core";
+
+import { BudgetExceededError } from "./errors.js";
+import { getProvider } from "./providers/adapters.js";
+import type {
+  CallOptions,
+  CompleteResult,
+  CompletionUsage,
+  Message,
+  ProviderAdapter,
+  ResolvedProvider,
+  StreamChunk,
+} from "./providers/types.js";
+
+/** The measured token breakdown used by the budget engine. */
+export interface MeasuredBudget {
+  system: number;
+  context: number;
+  history: number;
+  output: number;
+}
+
+/** The assembled result of a collected stream. */
+export interface StreamResult {
+  text: string;
+  reasoning?: string;
+  usage: CompletionUsage;
+}
+
+/** Measure the token breakdown of a prompt against the budget engine. */
+export function measureBudget(
+  system: string,
+  context: string,
+  history: Message[],
+  output: number,
+): MeasuredBudget {
+  return {
+    system: countTokens(system),
+    context: countTokens(context),
+    history: countTokens(history.map((m) => `${m.role}: ${m.content}`).join("\n")),
+    output,
+  };
+}
+
+/**
+ * Enforce a measured budget against the configured token budget.
+ *
+ * @throws {BudgetExceededError} if any dimension overflows.
+ */
+export function enforceBudget(
+  measured: MeasuredBudget,
+  budget: ReturnType<typeof allocateBudget>,
+): void {
+  const overflows: string[] = [];
+  if (measured.system > budget.system) {
+    overflows.push(`system (${measured.system}/${budget.system})`);
+  }
+  if (measured.context > budget.context) {
+    overflows.push(`context (${measured.context}/${budget.context})`);
+  }
+  if (measured.history > budget.history) {
+    overflows.push(`history (${measured.history}/${budget.history})`);
+  }
+  if (measured.output > budget.output) {
+    overflows.push(`output (${measured.output}/${budget.output})`);
+  }
+  const total =
+    measured.system + measured.context + measured.history + measured.output;
+  if (total > budget.total) {
+    overflows.push(`total (${total}/${budget.total})`);
+  }
+  if (overflows.length > 0) {
+    throw new BudgetExceededError(
+      `Prompt exceeds context budget; overflowed: ${overflows.join(", ")}`,
+    );
+  }
+}
+
+/** The output budget to reserve, based on the thinking config. */
+function outputBudget(opts: CallOptions): number {
+  const thinking = opts.thinking;
+  return typeof thinking?.budget === "number" ? thinking.budget : DEFAULT_BUDGET.output;
+}
+
+/**
+ * Enforce the token budget for a set of messages + context, then resolve the
+ * provider adapter.
+ *
+ * @throws {BudgetExceededError} if the prompt exceeds the budget.
+ */
+function getAdapter(
+  config: ResolvedProvider,
+  messages: Message[],
+  opts: CallOptions,
+): ProviderAdapter {
+  const output = outputBudget(opts);
+  const system = messages
+    .filter((m) => m.role === "system")
+    .map((m) => m.content)
+    .join("\n");
+  const history = messages.filter((m) => m.role !== "system");
+  const measured = measureBudget(system, opts.contextText ?? "", history, output);
+  enforceBudget(measured, allocateBudget({ ...DEFAULT_BUDGET, total: config.contextWindow }));
+  return getProvider(config);
+}
+
+/**
+ * Stream a completion, enforcing the budget first. Yields `StreamChunk`s.
+ */
+export async function* streamCompletion(
+  config: ResolvedProvider,
+  messages: Message[],
+  opts: CallOptions = {},
+): AsyncIterableIterator<StreamChunk> {
+  const adapter = getAdapter(config, messages, opts);
+  yield* adapter.stream(messages, opts);
+}
+
+/**
+ * Collect a full completion stream into a single result.
+ */
+export async function collectStream(
+  config: ResolvedProvider,
+  messages: Message[],
+  opts: CallOptions = {},
+): Promise<StreamResult> {
+  const adapter = getAdapter(config, messages, opts);
+  let text = "";
+  let reasoning: string | undefined;
+  let usage: CompletionUsage = { promptTokens: 0, completionTokens: 0, total: 0 };
+
+  for await (const chunk of adapter.stream(messages, opts)) {
+    if (chunk.type === "text" && chunk.text) {
+      text += chunk.text;
+    } else if (chunk.type === "reasoning" && chunk.reasoning) {
+      reasoning = (reasoning ?? "") + chunk.reasoning;
+    } else if (chunk.type === "done" && chunk.usage) {
+      usage = chunk.usage;
+    }
+  }
+
+  return { text, ...(reasoning ? { reasoning } : {}), usage };
+}
+
+/**
+ * Run a non-streaming completion (used for tool use and offline calls).
+ */
+export async function complete(
+  config: ResolvedProvider,
+  messages: Message[],
+  opts: CallOptions = {},
+): Promise<CompleteResult> {
+  const adapter = getAdapter(config, messages, opts);
+  return adapter.complete(messages, opts);
+}
diff --git a/packages/ai/src/discover.test.ts b/packages/ai/src/discover.test.ts
new file mode 100644
index 0000000..30ecc8a
--- /dev/null
+++ b/packages/ai/src/discover.test.ts
@@ -0,0 +1,81 @@
+import { describe, expect, it } from "vitest";
+
+import { discoverLocalProviders } from "./index.js";
+
+/** Build a mock fetch that returns a models catalog for a given URL. */
+function mockFetch(
+  catalog: Record<string, unknown>,
+): typeof fetch {
+  return ((url: string) => {
+    if (catalog[url]) {
+      return Promise.resolve(
+        new Response(JSON.stringify(catalog[url]), {
+          status: 200,
+          headers: { "content-type": "application/json" },
+        }),
+      );
+    }
+    return Promise.reject(new Error("not found"));
+  }) as unknown as typeof fetch;
+}
+
+describe("discoverLocalProviders", () => {
+  it("detects a running Ollama server", async () => {
+    const results = await discoverLocalProviders({
+      timeoutMs: 200,
+      fetchImpl: mockFetch({
+        "http://localhost:11434/v1/models": {
+          object: "list",
+          data: [
+            { id: "llama3.1", owned_by: "ollama" },
+            { id: "mistral", owned_by: "ollama" },
+          ],
+        },
+      }),
+    });
+
+    expect(results).toHaveLength(1);
+    expect(results[0]!.endpoint).toBe("http://localhost:11434/v1/models");
+    expect(results[0]!.kind).toBe("openai-compatible");
+    expect(results[0]!.models.map((m) => m.id)).toEqual(["llama3.1", "mistral"]);
+  });
+
+  it("returns an empty array when no endpoints respond", async () => {
+    const results = await discoverLocalProviders({
+      timeoutMs: 100,
+      fetchImpl: mockFetch({}),
+    });
+    expect(results).toEqual([]);
+  });
+
+  it("ignores endpoints that return no models", async () => {
+    const results = await discoverLocalProviders({
+      timeoutMs: 200,
+      fetchImpl: mockFetch({
+        "http://localhost:1234/v1/models": { object: "list", data: [] },
+      }),
+    });
+    expect(results).toEqual([]);
+  });
+
+  it("probes multiple endpoints and collects all responders", async () => {
+    const results = await discoverLocalProviders({
+      timeoutMs: 200,
+      fetchImpl: mockFetch({
+        "http://localhost:11434/v1/models": {
+          object: "list",
+          data: [{ id: "llama3.1" }],
+        },
+        "http://localhost:1234/v1/models": {
+          object: "list",
+          data: [{ id: "qwen2.5" }],
+        },
+      }),
+    });
+
+    expect(results).toHaveLength(2);
+    const ids = results.flatMap((r) => r.models.map((m) => m.id));
+    expect(ids).toContain("llama3.1");
+    expect(ids).toContain("qwen2.5");
+  });
+});
diff --git a/packages/ai/src/discover.ts b/packages/ai/src/discover.ts
new file mode 100644
index 0000000..9651183
--- /dev/null
+++ b/packages/ai/src/discover.ts
@@ -0,0 +1,97 @@
+/**
+ * Local provider discovery.
+ *
+ * Scans a set of well-known local endpoints for OpenAI-compatible model
+ * catalogs, so a user can point PubScript at a running Ollama / llama.cpp /
+ * LM Studio server without hand-entering config. See CHUNK#8.md §3.
+ */
+
+import type { ProviderKind } from "./providers/types.js";
+
+/** A model as reported by a provider's `/v1/models` endpoint. */
+export interface DetectedModel {
+  id: string;
+  owned_by?: string;
+}
+
+/** A detected local provider and its available models. */
+export interface DetectedProvider {
+  endpoint: string;
+  kind: ProviderKind;
+  models: DetectedModel[];
+}
+
+/** A well-known local endpoint to probe. */
+interface LocalEndpoint {
+  url: string;
+  kind: ProviderKind;
+}
+
+/**
+ * The local endpoints probed during discovery. Each returns an OpenAI-style
+ * `{ data: [{ id, owned_by }] }` catalog.
+ */
+export const LOCAL_ENDPOINTS: LocalEndpoint[] = [
+  { url: "http://localhost:11434/v1/models", kind: "openai-compatible" }, // Ollama
+  { url: "http://localhost:1234/v1/models", kind: "openai-compatible" }, // llama.cpp / LM Studio
+  { url: "http://localhost:8080/v1/models", kind: "openai-compatible" }, // generic
+];
+
+/** The shape of a provider's `/v1/models` response. */
+interface ModelsResponse {
+  data?: Array<{ id?: string; owned_by?: string }>;
+}
+
+/**
+ * Discover local OpenAI-compatible providers.
+ *
+ * @param options.timeoutMs  Per-endpoint timeout in milliseconds.
+ * @param options.fetchImpl  Injectable fetch (for testing).
+ */
+export async function discoverLocalProviders(
+  options: { timeoutMs?: number; fetchImpl?: typeof fetch } = {},
+): Promise<DetectedProvider[]> {
+  const timeoutMs = options.timeoutMs ?? 800;
+  const fetchImpl = options.fetchImpl ?? fetch;
+
+  const results: DetectedProvider[] = [];
+
+  await Promise.all(
+    LOCAL_ENDPOINTS.map(async (endpoint) => {
+      try {
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(), timeoutMs);
+        try {
+          const response = await fetchImpl(endpoint.url, {
+            signal: controller.signal,
+          });
+          if (!response.ok) {
+            return;
+          }
+          const json = (await response.json()) as ModelsResponse;
+          const models = (json.data ?? [])
+            .filter((model) => typeof model.id === "string")
+            .map((model) => ({
+              id: model.id as string,
+              ...(typeof model.owned_by === "string"
+                ? { owned_by: model.owned_by }
+                : {}),
+            }));
+          if (models.length > 0) {
+            results.push({
+              endpoint: endpoint.url,
+              kind: endpoint.kind,
+              models,
+            });
+          }
+        } finally {
+          clearTimeout(timer);
+        }
+      } catch {
+        // Endpoint unreachable or timed out; skip it.
+      }
+    }),
+  );
+
+  return results;
+}
diff --git a/packages/ai/src/errors.ts b/packages/ai/src/errors.ts
new file mode 100644
index 0000000..b1a2322
--- /dev/null
+++ b/packages/ai/src/errors.ts
@@ -0,0 +1,49 @@
+/**
+ * Error types for the AI provider layer.
+ *
+ * Errors are classified so callers can react appropriately: auth failures are
+ * terminal (the key is wrong), rate-limit/overloaded errors are retryable, and
+ * budget violations are a client-side guard.
+ *
+ * See CHUNK#8.md §3.
+ */
+
+/** Base class for all provider-layer errors. */
+export class ProviderError extends Error {
+  constructor(message: string) {
+    super(message);
+    this.name = "ProviderError";
+  }
+}
+
+/** The provider rejected the request due to missing/invalid credentials. */
+export class ProviderAuthError extends ProviderError {
+  constructor(message = "Provider authentication failed (invalid API key).") {
+    super(message);
+    this.name = "ProviderAuthError";
+  }
+}
+
+/** The provider rate-limited or overloaded the request; safe to retry. */
+export class ProviderRateLimitError extends ProviderError {
+  constructor(message = "Provider rate limit exceeded; retrying.") {
+    super(message);
+    this.name = "ProviderRateLimitError";
+  }
+}
+
+/** The request exceeded the configured token budget. */
+export class BudgetExceededError extends ProviderError {
+  constructor(message: string) {
+    super(message);
+    this.name = "BudgetExceededError";
+  }
+}
+
+/** No provider adapter is registered for the given kind. */
+export class UnknownProviderError extends ProviderError {
+  constructor(kind: string) {
+    super(`Unknown provider kind: ${kind}`);
+    this.name = "UnknownProviderError";
+  }
+}
diff --git a/packages/ai/src/http.ts b/packages/ai/src/http.ts
new file mode 100644
index 0000000..43c0798
--- /dev/null
+++ b/packages/ai/src/http.ts
@@ -0,0 +1,170 @@
+/**
+ * Low-level HTTP + SSE helpers for the provider layer.
+ *
+ * This module provides:
+ * - `readSSEEvents`: parse a `text/event-stream` `ReadableStream` into
+ *   concatenated `data:` payloads (one per event).
+ * - `fetchWithRetry`: a `fetch` wrapper that retries on 429/5xx with
+ *   exponential backoff and surfaces auth errors immediately.
+ *
+ * It is intentionally dependency-free and browser-safe (uses the global
+ * `fetch`/`TextDecoder`/`ReadableStream`). See CHUNK#8.md §3.
+ */
+
+import {
+  ProviderAuthError,
+  ProviderError,
+  ProviderRateLimitError,
+} from "./errors.js";
+
+/** Read a `text/event-stream` body and yield each event's `data:` payload. */
+export async function* readSSEEvents(
+  stream: ReadableStream<Uint8Array<ArrayBuffer>>,
+): AsyncGenerator<string> {
+  const reader = stream.getReader();
+  const decoder = new TextDecoder();
+  let buffer = "";
+  const dataLines: string[] = [];
+
+  const flush = (): string | null => {
+    if (dataLines.length === 0) {
+      return null;
+    }
+    const value = dataLines.join("\n");
+    dataLines.length = 0;
+    return value;
+  };
+
+  for (;;) {
+    const { done, value } = await reader.read();
+    if (value) {
+      buffer += decoder.decode(value, { stream: true });
+    }
+
+    let newlineIndex: number;
+    while ((newlineIndex = buffer.indexOf("\n")) >= 0) {
+      let line = buffer.slice(0, newlineIndex);
+      buffer = buffer.slice(newlineIndex + 1);
+      if (line.endsWith("\r")) {
+        line = line.slice(0, -1);
+      }
+
+      if (line === "") {
+        // Blank line terminates an SSE event.
+        const payload = flush();
+        if (payload !== null) {
+          yield payload;
+        }
+      } else if (line.startsWith("data:")) {
+        dataLines.push(line.slice(5).replace(/^ /, ""));
+      }
+      // Ignore `event:`, `id:`, `retry:` lines.
+    }
+
+    if (done) {
+      break;
+    }
+  }
+
+  // Flush any trailing data without a terminating blank line.
+  if (buffer.length > 0) {
+    if (buffer.startsWith("data:")) {
+      dataLines.push(buffer.slice(5).replace(/^ /, ""));
+    }
+    const payload = flush();
+    if (payload !== null) {
+      yield payload;
+    }
+  }
+}
+
+/** Parse a JSON SSE event, returning `null` for malformed input. */
+export function parseSSEEvent<T = unknown>(event: string): T | null {
+  try {
+    return JSON.parse(event) as T;
+  } catch {
+    return null;
+  }
+}
+
+/** Sleep for `ms` milliseconds. */
+export function sleep(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+/** Build the retry backoff delay for a given attempt (0-based). */
+function backoffDelay(attempt: number, maxMs: number): number {
+  const base = 250 * 2 ** attempt;
+  return Math.min(base, maxMs);
+}
+
+/**
+ * `fetch` wrapper that retries on 429/5xx with exponential backoff. Auth
+ * errors (401/403) are thrown immediately; other non-retryable errors are
+ * surfaced as-is.
+ */
+export async function fetchWithRetry(
+  url: string,
+  init: RequestInit,
+  options: { retries?: number; maxDelayMs?: number; signal?: AbortSignal } = {},
+): Promise<Response> {
+  const retries = options.retries ?? 2;
+  const maxDelayMs = options.maxDelayMs ?? 5000;
+  let lastError: unknown;
+
+  for (let attempt = 0; attempt <= retries; attempt++) {
+    if (attempt > 0) {
+      await sleep(backoffDelay(attempt - 1, maxDelayMs));
+    }
+
+    try {
+      const response = await fetch(url, init);
+
+      if (response.ok) {
+        return response;
+      }
+
+      const detail = await safeText(response);
+      if (response.status === 401 || response.status === 403) {
+        throw new ProviderAuthError(
+          `Provider rejected credentials (${response.status}).`,
+        );
+      }
+      if (response.status === 429 || response.status >= 500) {
+        lastError = new ProviderRateLimitError(
+          `Provider rate-limited or overloaded (${response.status}).`,
+        );
+        if (attempt < retries) {
+          continue;
+        }
+        throw lastError;
+      }
+
+      throw new ProviderError(
+        `Provider request failed (${response.status}): ${detail}`,
+      );
+    } catch (error) {
+      if (error instanceof Error && error.name === "AbortError") {
+        throw error;
+      }
+      lastError = error;
+      if (attempt >= retries) {
+        throw error;
+      }
+    }
+  }
+
+  // Unreachable, but satisfies the type checker.
+  throw lastError instanceof Error
+    ? lastError
+    : new ProviderError("Request failed after retries.");
+}
+
+/** Read a response body as text, tolerating errors. */
+async function safeText(response: Response): Promise<string> {
+  try {
+    return await response.text();
+  } catch {
+    return "(no body)";
+  }
+}
diff --git a/packages/ai/src/index.ts b/packages/ai/src/index.ts
new file mode 100644
index 0000000..fd817fd
--- /dev/null
+++ b/packages/ai/src/index.ts
@@ -0,0 +1,72 @@
+/**
+ * `@pubscript/ai` — the AI provider layer for PubScript.
+ *
+ * A self-contained, SSE-based provider abstraction that talks to any
+ * OpenAI-compatible endpoint with streaming, thinking/reasoning params, and
+ * the token-budget engine from `packages/core`. See CHUNK#8.md.
+ *
+ * Public API:
+ * - `streamCompletion`, `collectStream`, `complete` — the completion client.
+ * - `measureBudget`, `enforceBudget` — budget helpers.
+ * - `discoverLocalProviders` — local provider discovery.
+ * - `getProvider` — resolve a provider config into an adapter.
+ * - `normalizeThinking` — thinking-param mapping.
+ * - Re-exports of the token budget engine from `packages/core`.
+ * - All types and error classes.
+ */
+
+// Completion client
+export {
+  streamCompletion,
+  collectStream,
+  complete,
+  measureBudget,
+  enforceBudget,
+} from "./client.js";
+export type { MeasuredBudget, StreamResult } from "./client.js";
+
+// Provider adapters + factory
+export { getProvider } from "./providers/adapters.js";
+
+// Thinking-param mapping
+export { normalizeThinking, reasoningBudgetTokens } from "./thinking.js";
+
+// Local provider discovery
+export {
+  discoverLocalProviders,
+  LOCAL_ENDPOINTS,
+} from "./discover.js";
+export type { DetectedModel, DetectedProvider } from "./discover.js";
+
+// Token budget engine (re-exported from core)
+export {
+  countTokens,
+  allocateBudget,
+  DEFAULT_BUDGET,
+} from "./budget.js";
+export type { Budget } from "./budget.js";
+
+// Types
+export type {
+  ProviderKind,
+  ThinkingConfig,
+  ThinkingEffort,
+  ProviderConfig,
+  ResolvedProvider,
+  Message,
+  CallOptions,
+  CompletionUsage,
+  StreamEventType,
+  StreamChunk,
+  CompleteResult,
+  ProviderAdapter,
+} from "./providers/types.js";
+
+// Errors
+export {
+  ProviderError,
+  ProviderAuthError,
+  ProviderRateLimitError,
+  BudgetExceededError,
+  UnknownProviderError,
+} from "./errors.js";
diff --git a/packages/ai/src/providers/adapters.ts b/packages/ai/src/providers/adapters.ts
new file mode 100644
index 0000000..8e8a414
--- /dev/null
+++ b/packages/ai/src/providers/adapters.ts
@@ -0,0 +1,424 @@
+/**
+ * Provider adapters: translate the provider-agnostic contract into concrete
+ * wire formats for OpenAI-compatible, Anthropic, Google (Gemini), and
+ * OpenRouter endpoints.
+ *
+ * Each adapter implements `ProviderAdapter` (see `providers/types.ts`): a
+ * `stream()` method that yields `StreamChunk`s and a `complete()` method for
+ * non-streaming calls. See CHUNK#8.md §3.
+ */
+
+import { ProviderAuthError, UnknownProviderError } from "../errors.js";
+import { fetchWithRetry, readSSEEvents } from "../http.js";
+import { normalizeThinking } from "../thinking.js";
+import type {
+  CallOptions,
+  CompletionUsage,
+  Message,
+  ProviderAdapter,
+  ResolvedProvider,
+} from "./types.js";
+
+/** Normalize a raw provider usage object into our canonical shape. */
+function normalizeUsage(usage: Record<string, unknown>): CompletionUsage {
+  const promptTokens = (usage.prompt_tokens as number) ?? 0;
+  const completionTokens =
+    (usage.completion_tokens as number) ??
+    (usage.completion_tokens_details as number) ??
+    0;
+  const total =
+    (usage.total_tokens as number) ??
+    promptTokens + completionTokens;
+  return { promptTokens, completionTokens, total };
+}
+
+/** Build the auth + content-type headers for a provider. */
+function buildHeaders(config: ResolvedProvider): Record<string, string> {
+  const headers: Record<string, string> = {
+    "content-type": "application/json",
+  };
+
+  if (config.kind === "anthropic") {
+    if (config.apiKey) {
+      headers["x-api-key"] = config.apiKey;
+    }
+    if (config.apiVersion) {
+      headers["anthropic-api-version"] = config.apiVersion;
+    }
+  } else if (config.apiKey) {
+    headers["authorization"] = `Bearer ${config.apiKey}`;
+  }
+
+  if (config.kind === "openrouter") {
+    headers["HTTP-Referer"] = "https://pubscript.dev";
+    headers["X-Title"] = "PubScript";
+  }
+
+  if (config.apiVersion && config.kind !== "anthropic") {
+    headers["openai-api-version"] = config.apiVersion;
+  }
+
+  Object.assign(headers, config.headers);
+  return headers;
+}
+
+/** Join a base URL and a path, avoiding double slashes. */
+function joinUrl(baseUrl: string, path: string): string {
+  return `${baseUrl.replace(/\/+$/, "")}/${path.replace(/^\/+/, "")}`;
+}
+
+/** POST a JSON body and return the raw Response. */
+async function post(
+  config: ResolvedProvider,
+  path: string,
+  body: Record<string, unknown>,
+  opts: CallOptions,
+): Promise<Response> {
+  const url = joinUrl(config.baseUrl, path);
+  return fetchWithRetry(url, {
+    method: "POST",
+    headers: buildHeaders(config),
+    body: JSON.stringify(body),
+    signal: opts.signal ?? null,
+  });
+}
+
+// ---------------------------------------------------------------------------
+// OpenAI-compatible adapter
+// ---------------------------------------------------------------------------
+
+function createOpenAICompatibleAdapter(
+  config: ResolvedProvider,
+): ProviderAdapter {
+  const path = "/chat/completions";
+
+  return {
+    kind: "openai-compatible",
+    async* stream(messages: Message[], opts: CallOptions) {
+      const body = {
+        model: config.model,
+        messages: messages.map((m) => ({ role: m.role, content: m.content })),
+        stream: true,
+        temperature: opts.temperature ?? config.temperature ?? 0.7,
+        max_tokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+        ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      for await (const event of readSSEEvents(response.body!)) {
+        if (event === "[DONE]") {
+          yield { type: "done" };
+          return;
+        }
+        const json = event ? JSON.parse(event) : null;
+        if (!json) {
+          continue;
+        }
+        const choice = json.choices?.[0] as
+          | { delta?: { content?: string; reasoning_content?: string } }
+          | undefined;
+        const delta = choice?.delta ?? {};
+        if (delta.reasoning_content) {
+          yield { type: "reasoning", reasoning: delta.reasoning_content };
+        }
+        if (delta.content) {
+          yield { type: "text", text: delta.content };
+        }
+        if (json.usage) {
+          yield { type: "done", usage: normalizeUsage(json.usage) };
+        }
+      }
+    },
+    async complete(messages: Message[], opts: CallOptions) {
+      const body = {
+        model: config.model,
+        messages: messages.map((m) => ({ role: m.role, content: m.content })),
+        stream: false,
+        temperature: opts.temperature ?? config.temperature ?? 0.7,
+        max_tokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+        ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      const json = (await response.json()) as {
+        choices?: Array<{ message?: { content?: string } }>;
+        usage?: Record<string, unknown>;
+      };
+      const text = json.choices?.[0]?.message?.content ?? "";
+      return {
+        text,
+        usage: json.usage ? normalizeUsage(json.usage) : { promptTokens: 0, completionTokens: 0, total: 0 },
+      };
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Anthropic adapter
+// ---------------------------------------------------------------------------
+
+function createAnthropicAdapter(config: ResolvedProvider): ProviderAdapter {
+  const path = "/v1/messages";
+
+  return {
+    kind: "anthropic",
+    async* stream(messages: Message[], opts: CallOptions) {
+      const system = messages
+        .filter((m) => m.role === "system")
+        .map((m) => m.content)
+        .join("\n");
+      const nonSystem = messages.filter((m) => m.role !== "system");
+
+      const body = {
+        model: config.model,
+        max_tokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+        temperature: opts.temperature ?? config.temperature ?? 0.7,
+        stream: true,
+        anthropic_version: config.apiVersion ?? "2023-06-01",
+        ...(system ? { system } : {}),
+        messages: nonSystem.map((m) => ({
+          role: m.role as "user" | "assistant",
+          content: m.content,
+        })),
+        ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      for await (const event of readSSEEvents(response.body!)) {
+        if (event === "[DONE]") {
+          yield { type: "done" };
+          return;
+        }
+        const json = event ? JSON.parse(event) : null;
+        if (!json) {
+          continue;
+        }
+        if (json.type === "content_block_delta") {
+          const delta = json.delta as { type?: string; text?: string; thinking?: string };
+          if (delta.type === "text_delta" && delta.text) {
+            yield { type: "text", text: delta.text };
+          } else if (delta.type === "thinking_delta" && delta.thinking) {
+            yield { type: "reasoning", reasoning: delta.thinking };
+          } else if (delta.type === "input_json_delta" && delta.thinking) {
+            yield { type: "reasoning", reasoning: delta.thinking };
+          }
+        } else if (json.type === "message_delta") {
+          const usage = json.delta?.usage as { output_tokens?: number } | undefined;
+          if (usage?.output_tokens) {
+            yield {
+              type: "done",
+              usage: {
+                promptTokens: json.usage?.input_tokens ?? 0,
+                completionTokens: usage.output_tokens,
+                total: (json.usage?.input_tokens ?? 0) + usage.output_tokens,
+              },
+            };
+          }
+        }
+      }
+    },
+    async complete(messages: Message[], opts: CallOptions) {
+      const system = messages
+        .filter((m) => m.role === "system")
+        .map((m) => m.content)
+        .join("\n");
+      const nonSystem = messages.filter((m) => m.role !== "system");
+
+      const body = {
+        model: config.model,
+        max_tokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+        temperature: opts.temperature ?? config.temperature ?? 0.7,
+        stream: false,
+        anthropic_version: config.apiVersion ?? "2023-06-01",
+        ...(system ? { system } : {}),
+        messages: nonSystem.map((m) => ({
+          role: m.role as "user" | "assistant",
+          content: m.content,
+        })),
+        ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      const json = (await response.json()) as {
+        content?: Array<{ type?: string; text?: string; thinking?: string }>;
+        usage?: { input_tokens?: number; output_tokens?: number };
+      };
+      const text = (json.content ?? [])
+        .filter((block) => block.type === "text")
+        .map((block) => block.text ?? "")
+        .join("");
+      const reasoning = (json.content ?? [])
+        .filter((block) => block.type === "thinking")
+        .map((block) => block.thinking ?? "")
+        .join("");
+      const usage = json.usage
+        ? {
+            promptTokens: json.usage.input_tokens ?? 0,
+            completionTokens: json.usage.output_tokens ?? 0,
+            total: (json.usage.input_tokens ?? 0) + (json.usage.output_tokens ?? 0),
+          }
+        : { promptTokens: 0, completionTokens: 0, total: 0 };
+      return { text, ...(reasoning ? { reasoning } : {}), usage };
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Google (Gemini) adapter
+// ---------------------------------------------------------------------------
+
+/** Shape of a Gemini `streamGenerateContent` response. */
+interface GeminiResponse {
+  candidates?: Array<{
+    content?: {
+      parts?: Array<{
+        text?: string;
+        thinking?: { reasoning_text?: string };
+      }>;
+    };
+  }>;
+  usageMetadata?: {
+    promptTokenCount?: number;
+    candidatesTokenCount?: number;
+    totalTokenCount?: number;
+  };
+}
+
+function createGoogleAdapter(config: ResolvedProvider): ProviderAdapter {
+  const path = "/v1beta/models/chat:streamGenerateContent";
+
+  return {
+    kind: "google",
+    async* stream(messages: Message[], opts: CallOptions) {
+      const body = {
+        contents: messages.map((m) => ({
+          role: m.role === "system" ? "user" : m.role,
+          parts: [{ text: m.content }],
+        })),
+        generationConfig: {
+          temperature: opts.temperature ?? config.temperature ?? 0.7,
+          maxOutputTokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+          ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        },
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      for await (const event of readSSEEvents(response.body!)) {
+        if (event === "[DONE]") {
+          yield { type: "done" };
+          return;
+        }
+        const json = event ? JSON.parse(event) : null;
+        if (!json) {
+          continue;
+        }
+        const candidates = json.candidates ?? [];
+        for (const candidate of candidates) {
+          const parts = candidate.content?.parts ?? [];
+          for (const part of parts) {
+            if (part.text) {
+              yield { type: "text", text: part.text };
+            }
+            if (part.thinking?.reasoning_text) {
+              yield { type: "reasoning", reasoning: part.thinking.reasoning_text };
+            }
+          }
+        }
+        if (json.usageMetadata) {
+          const meta = json.usageMetadata as {
+            promptTokenCount?: number;
+            candidatesTokenCount?: number;
+            totalTokenCount?: number;
+          };
+          yield {
+            type: "done",
+            usage: {
+              promptTokens: meta.promptTokenCount ?? 0,
+              completionTokens: meta.candidatesTokenCount ?? 0,
+              total: meta.totalTokenCount ?? 0,
+            },
+          };
+        }
+      }
+    },
+    async complete(messages: Message[], opts: CallOptions) {
+      const body = {
+        contents: messages.map((m) => ({
+          role: m.role === "system" ? "user" : m.role,
+          parts: [{ text: m.content }],
+        })),
+        generationConfig: {
+          temperature: opts.temperature ?? config.temperature ?? 0.7,
+          maxOutputTokens: opts.maxTokens ?? config.maxTokens ?? 4096,
+          ...normalizeThinking(config.kind, opts.thinking ?? config.thinking),
+        },
+        ...opts.extraBody,
+      };
+
+      const response = await post(config, path, body, opts);
+      const json = (await response.json()) as GeminiResponse;
+      const text = (json.candidates ?? [])
+        .flatMap((c) => c.content?.parts ?? [])
+        .filter((p) => p.text)
+        .map((p) => p.text)
+        .join("");
+      const reasoning = (json.candidates ?? [])
+        .flatMap((c) => c.content?.parts ?? [])
+        .filter((p) => p.thinking?.reasoning_text)
+        .map((p) => p.thinking?.reasoning_text)
+        .join("");
+      const meta = json.usageMetadata ?? {};
+      const usage = {
+        promptTokens: meta.promptTokenCount ?? 0,
+        completionTokens: meta.candidatesTokenCount ?? 0,
+        total: meta.totalTokenCount ?? 0,
+      };
+      return { text, ...(reasoning ? { reasoning } : {}), usage };
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// OpenRouter adapter (OpenAI-compatible with extra headers)
+// ---------------------------------------------------------------------------
+
+function createOpenRouterAdapter(config: ResolvedProvider): ProviderAdapter {
+  // OpenRouter speaks the OpenAI chat-completions wire format; the only
+  // differences are the extra referer/title headers, handled in buildHeaders.
+  return createOpenAICompatibleAdapter({
+    ...config,
+    baseUrl: config.baseUrl ?? "https://openrouter.ai/api/v1",
+  });
+}
+
+// ---------------------------------------------------------------------------
+// Factory
+// ---------------------------------------------------------------------------
+
+/**
+ * Resolve a provider config into a concrete adapter.
+ *
+ * @throws {UnknownProviderError} if the kind is not recognized.
+ */
+export function getProvider(config: ResolvedProvider): ProviderAdapter {
+  switch (config.kind) {
+    case "openai-compatible":
+      return createOpenAICompatibleAdapter(config);
+    case "anthropic":
+      return createAnthropicAdapter(config);
+    case "google":
+      return createGoogleAdapter(config);
+    case "openrouter":
+      return createOpenRouterAdapter(config);
+    default:
+      throw new UnknownProviderError((config as { kind: string }).kind);
+  }
+}
+
+/** Re-export for parity with the auth-error surface used by adapters. */
+export { ProviderAuthError };
diff --git a/packages/ai/src/providers/types.ts b/packages/ai/src/providers/types.ts
new file mode 100644
index 0000000..949e77f
--- /dev/null
+++ b/packages/ai/src/providers/types.ts
@@ -0,0 +1,117 @@
+/**
+ * Provider-agnostic types for the AI provider layer.
+ *
+ * These interfaces describe the *contract* between PubScript and any
+ * OpenAI-compatible endpoint, independent of the concrete provider. The
+ * adapters in `providers/adapters.ts` translate these into provider-specific
+ * wire formats; the client in `client.ts` orchestrates them.
+ *
+ * See CHUNK#8.md §3.
+ */
+
+/** The family of a provider, which drives the wire-format translation. */
+export type ProviderKind =
+  | "openai-compatible"
+  | "anthropic"
+  | "google"
+  | "openrouter";
+
+/** Reasoning budget expressed either in tokens or at a coarse effort level. */
+export type ThinkingEffort = "low" | "medium" | "high";
+
+/**
+ * Thinking/reasoning configuration. When enabled, the client maps `effort` or
+ * `budget` onto the provider-specific wire format (see `thinking.ts`).
+ */
+export interface ThinkingConfig {
+  enabled: boolean;
+  /** Reasoning budget in tokens. Maps to `reasoning_effort` when omitted. */
+  budget?: number;
+  /** Coarse effort level. Maps to `reasoning_effort` on OpenAI-style APIs. */
+  effort?: ThinkingEffort;
+}
+
+/** A provider configuration as stored in the project store. */
+export interface ProviderConfig {
+  id: string;
+  kind: ProviderKind;
+  /** Base URL of the endpoint (e.g. `https://api.openai.com/v1`). */
+  baseUrl?: string;
+  /** API key. Never logged or serialized to disk in plaintext. */
+  apiKey?: string;
+  /** Model id to request. */
+  model: string;
+  temperature?: number;
+  maxTokens?: number;
+  thinking?: ThinkingConfig;
+  /** Context window in tokens, used by the budget engine. */
+  contextWindow: number;
+  /** Extra headers forwarded to the provider (e.g. OpenRouter referer). */
+  headers?: Record<string, string>;
+  /** `anthropic-api-version` or `openai` API version header. */
+  apiVersion?: string;
+}
+
+/** A provider config with a resolved, non-optional `baseUrl`. */
+export interface ResolvedProvider extends ProviderConfig {
+  baseUrl: string;
+}
+
+/** A chat message. Only text content is carried in v1. */
+export interface Message {
+  role: "system" | "user" | "assistant";
+  content: string;
+}
+
+/** Options for a single completion call. */
+export interface CallOptions {
+  temperature?: number;
+  maxTokens?: number;
+  thinking?: ThinkingConfig;
+  signal?: AbortSignal;
+  /** Extra body params merged into the request (e.g. tools, metadata). */
+  extraBody?: Record<string, unknown>;
+  /** Retrieved context (world state) counted separately by the budget engine. */
+  contextText?: string;
+}
+
+/** Token usage for a single completion. */
+export interface CompletionUsage {
+  promptTokens: number;
+  completionTokens: number;
+  total: number;
+}
+
+/** The kind of event emitted on a completion stream. */
+export type StreamEventType = "text" | "reasoning" | "done";
+
+/** A single chunk on a completion stream. */
+export interface StreamChunk {
+  type: StreamEventType;
+  /** Text delta for `"text"`. */
+  text?: string;
+  /** Reasoning delta for `"reasoning"`. */
+  reasoning?: string;
+  /** Final usage, carried on the `"done"` chunk. */
+  usage?: CompletionUsage;
+}
+
+/** The fully-assembled result of a non-streaming completion. */
+export interface CompleteResult {
+  text: string;
+  reasoning?: string;
+  usage: CompletionUsage;
+}
+
+/**
+ * The interface every provider adapter implements. Adapters translate the
+ * provider-agnostic `Message[]`/`CallOptions` into the provider's wire format
+ * and yield `StreamChunk`s.
+ */
+export interface ProviderAdapter {
+  readonly kind: ProviderKind;
+  /** Stream a completion, yielding text/reasoning deltas then a final `done`. */
+  stream(messages: Message[], opts: CallOptions): AsyncIterableIterator<StreamChunk>;
+  /** Run a non-streaming completion (used for tool use and offline calls). */
+  complete(messages: Message[], opts: CallOptions): Promise<CompleteResult>;
+}
diff --git a/packages/ai/src/thinking.test.ts b/packages/ai/src/thinking.test.ts
new file mode 100644
index 0000000..94c65be
--- /dev/null
+++ b/packages/ai/src/thinking.test.ts
@@ -0,0 +1,97 @@
+import { describe, expect, it } from "vitest";
+
+import { normalizeThinking, reasoningBudgetTokens } from "./index.js";
+import type { ThinkingConfig } from "./index.js";
+
+describe("normalizeThinking", () => {
+  it("returns an empty object when thinking is disabled", () => {
+    expect(normalizeThinking("openai-compatible", { enabled: false })).toEqual({});
+    expect(normalizeThinking("openai-compatible", undefined)).toEqual({});
+  });
+
+  describe("openai-compatible", () => {
+    it("maps effort to reasoning_effort", () => {
+      const cfg: ThinkingConfig = { enabled: true, effort: "high" };
+      expect(normalizeThinking("openai-compatible", cfg)).toEqual({
+        reasoning_effort: "high",
+      });
+    });
+
+    it("defaults effort to medium", () => {
+      const cfg: ThinkingConfig = { enabled: true };
+      expect(normalizeThinking("openai-compatible", cfg)).toEqual({
+        reasoning_effort: "medium",
+      });
+    });
+
+    it("adds max_tokens when a budget is set", () => {
+      const cfg: ThinkingConfig = { enabled: true, budget: 4096 };
+      expect(normalizeThinking("openai-compatible", cfg)).toEqual({
+        reasoning_effort: "medium",
+        max_tokens: 4096,
+      });
+    });
+  });
+
+  describe("openrouter", () => {
+    it("maps effort to reasoning_effort", () => {
+      const cfg: ThinkingConfig = { enabled: true, effort: "low" };
+      expect(normalizeThinking("openrouter", cfg)).toEqual({
+        reasoning_effort: "low",
+      });
+    });
+  });
+
+  describe("anthropic", () => {
+    it("emits a thinking object and thinking_budget", () => {
+      const cfg: ThinkingConfig = { enabled: true, budget: 4096 };
+      const result = normalizeThinking("anthropic", cfg);
+      expect(result).toEqual({
+        thinking: { enabled: true, budget_tokens: 4096 },
+        thinking_budget: 4096,
+      });
+    });
+
+    it("caps the budget at the Anthropic maximum", () => {
+      const cfg: ThinkingConfig = { enabled: true, budget: 100_000 };
+      const result = normalizeThinking("anthropic", cfg) as {
+        thinking: { budget_tokens: number };
+      };
+      expect(result.thinking.budget_tokens).toBe(32768);
+    });
+
+    it("defaults the budget when unset", () => {
+      const cfg: ThinkingConfig = { enabled: true };
+      const result = normalizeThinking("anthropic", cfg) as {
+        thinking: { budget_tokens: number };
+      };
+      expect(result.thinking.budget_tokens).toBe(2048);
+    });
+  });
+
+  describe("google", () => {
+    it("emits reasoning_effort and thinking_budget", () => {
+      const cfg: ThinkingConfig = { enabled: true, effort: "high", budget: 1024 };
+      const result = normalizeThinking("google", cfg);
+      expect(result).toEqual({
+        reasoning_effort: "high",
+        thinking_budget: 1024,
+      });
+    });
+  });
+});
+
+describe("reasoningBudgetTokens", () => {
+  it("returns 0 when thinking is disabled", () => {
+    expect(reasoningBudgetTokens({ enabled: false })).toBe(0);
+    expect(reasoningBudgetTokens(undefined)).toBe(0);
+  });
+
+  it("returns the configured budget", () => {
+    expect(reasoningBudgetTokens({ enabled: true, budget: 4096 })).toBe(4096);
+  });
+
+  it("defaults to 2048 when unset", () => {
+    expect(reasoningBudgetTokens({ enabled: true })).toBe(2048);
+  });
+});
diff --git a/packages/ai/src/thinking.ts b/packages/ai/src/thinking.ts
new file mode 100644
index 0000000..6df207e
--- /dev/null
+++ b/packages/ai/src/thinking.ts
@@ -0,0 +1,90 @@
+/**
+ * Thinking/reasoning parameter mapping.
+ *
+ * Different providers express the reasoning budget differently:
+ *
+ * - OpenAI-style (`openai-compatible`, `openrouter`): `reasoning_effort`
+ *   (`"low" | "medium" | "high"`), optionally `max_tokens` for the reasoning
+ *   budget.
+ * - Anthropic: a `thinking` object `{ enabled: boolean, budget_tokens: number }`
+ *   plus `thinking_budget` (must be <= 32768 and > 0).
+ * - Google (Gemini): `reasoning_effort` + `thinking_budget`.
+ *
+ * This module normalizes the provider-agnostic `ThinkingConfig` into the
+ * provider-specific wire format. See CHUNK#8.md §3.
+ */
+
+import type { ProviderKind, ThinkingConfig } from "./providers/types.js";
+
+/** The maximum reasoning budget Anthropic accepts. */
+const ANTHROPIC_MAX_THINKING_TOKENS = 32768;
+
+/** Normalize a thinking config into the wire-format body fragment for a kind. */
+export function normalizeThinking(
+  kind: ProviderKind,
+  thinking?: ThinkingConfig,
+): Record<string, unknown> {
+  if (!thinking?.enabled) {
+    return {};
+  }
+
+  switch (kind) {
+    case "anthropic":
+      return normalizeAnthropicThinking(thinking);
+    case "google":
+      return normalizeGoogleThinking(thinking);
+    case "openai-compatible":
+    case "openrouter":
+    default:
+      return normalizeOpenAIThinking(thinking);
+  }
+}
+
+/** OpenAI-style: `reasoning_effort` + optional `max_tokens`. */
+function normalizeOpenAIThinking(thinking: ThinkingConfig): Record<string, unknown> {
+  const effort = thinking.effort ?? "medium";
+  const body: Record<string, unknown> = {
+    reasoning_effort: effort,
+  };
+  if (typeof thinking.budget === "number") {
+    body.max_tokens = thinking.budget;
+  }
+  return body;
+}
+
+/** Anthropic: `thinking` object + `thinking_budget`. */
+function normalizeAnthropicThinking(thinking: ThinkingConfig): Record<string, unknown> {
+  const budget =
+    typeof thinking.budget === "number"
+      ? Math.min(thinking.budget, ANTHROPIC_MAX_THINKING_TOKENS)
+      : 2048;
+  return {
+    thinking: {
+      enabled: true,
+      budget_tokens: budget,
+    },
+    thinking_budget: budget,
+  };
+}
+
+/** Google: `reasoning_effort` + `thinking_budget`. */
+function normalizeGoogleThinking(thinking: ThinkingConfig): Record<string, unknown> {
+  const effort = thinking.effort ?? "medium";
+  const budget = typeof thinking.budget === "number" ? thinking.budget : 2048;
+  return {
+    reasoning_effort: effort,
+    thinking_budget: budget,
+  };
+}
+
+/**
+ * Compute the reasoning budget (in tokens) to advertise to the budget engine,
+ * so the output budget accounts for the reasoning spend. Returns 0 when
+ * thinking is disabled.
+ */
+export function reasoningBudgetTokens(thinking?: ThinkingConfig): number {
+  if (!thinking?.enabled) {
+    return 0;
+  }
+  return typeof thinking.budget === "number" ? thinking.budget : 2048;
+}
diff --git a/packages/ai/tsconfig.json b/packages/ai/tsconfig.json
new file mode 100644
index 0000000..98f36d9
--- /dev/null
+++ b/packages/ai/tsconfig.json
@@ -0,0 +1,8 @@
+{
+  "extends": "../../tsconfig.base.json",
+  "compilerOptions": {
+    "outDir": "./dist",
+    "rootDir": "./src"
+  },
+  "include": ["src/**/*.ts"]
+}
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index bb7ba2c..bb231fc 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -85,6 +85,22 @@ importers:
         specifier: ^2.1.8
         version: 2.1.9(@types/node@22.20.5)(jsdom@25.0.1)
 
+  packages/ai:
+    dependencies:
+      '@pubscript/core':
+        specifier: workspace:*
+        version: link:../core
+    devDependencies:
+      '@types/node':
+        specifier: ^22.10.2
+        version: 22.20.5
+      typescript:
+        specifier: ^5.7.2
+        version: 5.9.3
+      vitest:
+        specifier: ^2.1.8
+        version: 2.1.9(@types/node@22.20.5)(jsdom@25.0.1)
+
   packages/core:
     dependencies:
       dexie: