feat(agent): AgentHarness streams through a required Models instance (phase 6)

AgentHarnessOptions.models is required; the harness stream path,
compaction, and branch summarization go through models.streamSimple()/
completeSimple() instead of the compat globals. getApiKeyAndHeaders
stays and wins per-field over provider-resolved auth, but is no longer
required: without it, requests resolve through provider auth.

compact()/generateSummary()/generateBranchSummary() take a Models
parameter; explicit apiKey becomes optional. StreamFn is redefined
structurally (Models.streamSimple satisfies it), dropping the compat
type dependency from agent types.

Harness tests build per-file Models collections with fauxProvider()
and unique provider ids instead of mutating the global api-registry.

packages/agent/docs/models.md

@@ -820,9 +820,9 @@ Check items off as they land. Keep this list current; it is the working state fo
 
 ### Phase 6 — AgentHarness
 
-- [ ] `AgentHarnessOptions.models` required; harness stream path uses `models.streamSimple()`.
-- [ ] Compaction/branch-summarization paths use the harness `Models` instance.
-- [ ] Harness tests use `createModels()` + faux provider.
+- [x] `AgentHarnessOptions.models` required (`readonly models` on the harness); the harness stream path uses `models.streamSimple()`. `StreamFn` redefined structurally (no compat type dependency); `Models.streamSimple` satisfies it.
+- [x] Compaction/branch-summarization take the harness `Models` instance; explicit `getApiKeyAndHeaders` auth stays and wins per-field, but is no longer required — requests resolve through provider auth otherwise (the hard "No auth available" throws are gone).
+- [x] Harness tests use `createModels()` + `fauxProvider()` with unique per-fake provider ids; no global api-registry state, no unregister bookkeeping.
 
 ### Phase 7 — coding-agent bridge (minimal)
 
@@ -843,8 +843,9 @@ The full AuthStorage deletion (`FileCredentialStore` + decorators, see "Replacin
 
 ### Deferred / follow-ups
 
-- [ ] Web OAuth implementations (sitegeist-style) behind `oauth: "web"`.
+- [ ] Web OAuth implementations (sitegeist-style) as an alternative `OAuthAuth`.
 - [ ] coding-agent `ModelRegistry` -> session `ModelManager` migration; delete `/compat`.
+- [ ] Move ALL internal `/compat` imports to the new API before compat is deleted: every package's src, all tests, and the example extensions (examples then demonstrate the new API; the extension-loader root-to-compat alias dies with compat). Nothing inside the repo may import `/compat` at that point.
 - [ ] Images API registry redesign (untouched in this pass).
 
 ## Error behavior

packages/agent/src/harness/agent-harness.ts

@@ -1,10 +1,4 @@
-import {
-	type AssistantMessage,
-	type ImageContent,
-	type Model,
-	streamSimple,
-	type UserMessage,
-} from "@earendil-works/pi-ai/compat";
+import type { AssistantMessage, ImageContent, Model, Models, UserMessage } from "@earendil-works/pi-ai";
 import { runAgentLoop } from "../agent-loop.ts";
 import type {
 	AgentContext,
@@ -178,6 +172,7 @@ export class AgentHarness<
 > {
 	readonly env: ExecutionEnv;
 	private session: Session;
+	readonly models: Models;
 	private phase: AgentHarnessPhase = "idle";
 	private runAbortController?: AbortController;
 	private runPromise?: Promise<void>;
@@ -200,6 +195,7 @@ export class AgentHarness<
 	constructor(options: AgentHarnessOptions<TSkill, TPromptTemplate, TTool>) {
 		this.env = options.env;
 		this.session = options.session;
+		this.models = options.models;
 		this.resources = options.resources ?? {};
 		this.streamOptions = cloneStreamOptions(options.streamOptions);
 		this.systemPrompt = options.systemPrompt;
@@ -382,7 +378,7 @@ export class AgentHarness<
 				headers: mergeHeaders(turnState.streamOptions.headers, auth?.headers),
 			};
 			const requestOptions = await this.emitBeforeProviderRequest(model, turnState.sessionId, snapshotOptions);
-			return streamSimple(model, context, {
+			return this.models.streamSimple(model, context, {
 				cacheRetention: requestOptions.cacheRetention,
 				headers: requestOptions.headers,
 				maxRetries: requestOptions.maxRetries,
@@ -713,8 +709,8 @@ export class AgentHarness<
 		try {
 			const model = this.model;
 			if (!model) throw new AgentHarnessError("invalid_state", "No model set for compaction");
+			// Explicit auth wins; otherwise the request resolves through provider auth.
 			const auth = await this.getApiKeyAndHeaders?.(model);
-			if (!auth) throw new AgentHarnessError("auth", "No auth available for compaction");
 			const branchEntries = await this.session.getBranch();
 			const preparationResult = prepareCompaction(branchEntries, DEFAULT_COMPACTION_SETTINGS);
 			if (!preparationResult.ok) throw preparationResult.error;
@@ -733,9 +729,10 @@ export class AgentHarness<
 				? { ok: true as const, value: provided }
 				: await compact(
 						preparation,
+						this.models,
 						model,
-						auth.apiKey,
-						auth.headers,
+						auth?.apiKey,
+						auth?.headers,
 						customInstructions,
 						undefined,
 						this.thinkingLevel,
@@ -792,12 +789,13 @@ export class AgentHarness<
 			if (!summaryText && options?.summarize && entries.length > 0) {
 				const model = this.model;
 				if (!model) throw new AgentHarnessError("invalid_state", "No model set for branch summary");
+				// Explicit auth wins; otherwise the request resolves through provider auth.
 				const auth = await this.getApiKeyAndHeaders?.(model);
-				if (!auth) throw new AgentHarnessError("auth", "No auth available for branch summary");
 				const branchSummary = await generateBranchSummary(entries, {
+					models: this.models,
 					model,
-					apiKey: auth.apiKey,
-					headers: auth.headers,
+					apiKey: auth?.apiKey,
+					headers: auth?.headers,
 					signal: new AbortController().signal,
 					customInstructions: hookResult?.customInstructions ?? options?.customInstructions,
 					replaceInstructions: hookResult?.replaceInstructions ?? options?.replaceInstructions,

packages/agent/src/harness/compaction/branch-summarization.ts

@@ -1,5 +1,5 @@
-import type { Model } from "@earendil-works/pi-ai/compat";
-import { completeSimple } from "@earendil-works/pi-ai/compat";
+import type { Model, Models } from "@earendil-works/pi-ai";
+
 import type { AgentMessage } from "../../types.ts";
 import {
 	convertToLlm,
@@ -49,10 +49,12 @@ export interface CollectEntriesResult {
 
 /** Options for generating a branch summary. */
 export interface GenerateBranchSummaryOptions {
+	/** Provider collection the summarization request goes through. */
+	models: Models;
 	/** Model used for summarization. */
 	model: Model<any>;
-	/** API key forwarded to the provider. */
-	apiKey: string;
+	/** Explicit API key; wins over provider-resolved auth. */
+	apiKey?: string;
 	/** Optional request headers forwarded to the provider. */
 	headers?: Record<string, string>;
 	/** Abort signal for the summarization request. */
@@ -202,7 +204,16 @@ export async function generateBranchSummary(
 	entries: SessionTreeEntry[],
 	options: GenerateBranchSummaryOptions,
 ): Promise<Result<BranchSummaryResult, BranchSummaryError>> {
-	const { model, apiKey, headers, signal, customInstructions, replaceInstructions, reserveTokens = 16384 } = options;
+	const {
+		models,
+		model,
+		apiKey,
+		headers,
+		signal,
+		customInstructions,
+		replaceInstructions,
+		reserveTokens = 16384,
+	} = options;
 	const contextWindow = model.contextWindow || 128000;
 	const tokenBudget = contextWindow - reserveTokens;
 
@@ -230,7 +241,7 @@ export async function generateBranchSummary(
 			timestamp: Date.now(),
 		},
 	];
-	const response = await completeSimple(
+	const response = await models.completeSimple(
 		model,
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		{ apiKey, headers, signal, maxTokens: 2048 },

packages/agent/src/harness/compaction/compaction.ts

@@ -1,5 +1,4 @@
-import type { AssistantMessage, ImageContent, Model, TextContent, Usage } from "@earendil-works/pi-ai/compat";
-import { completeSimple } from "@earendil-works/pi-ai/compat";
+import type { AssistantMessage, ImageContent, Model, Models, TextContent, Usage } from "@earendil-works/pi-ai";
 import type { AgentMessage, ThinkingLevel } from "../../types.ts";
 import {
 	convertToLlm,
@@ -455,9 +454,10 @@ Keep each section concise. Preserve exact file paths, function names, and error
 /** Generate or update a conversation summary for compaction. */
 export async function generateSummary(
 	currentMessages: AgentMessage[],
+	models: Models,
 	model: Model<any>,
 	reserveTokens: number,
-	apiKey: string,
+	apiKey?: string,
 	headers?: Record<string, string>,
 	signal?: AbortSignal,
 	customInstructions?: string,
@@ -493,7 +493,7 @@ export async function generateSummary(
 			? { maxTokens, signal, apiKey, headers, reasoning: thinkingLevel }
 			: { maxTokens, signal, apiKey, headers };
 
-	const response = await completeSimple(
+	const response = await models.completeSimple(
 		model,
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		completionOptions,
@@ -626,8 +626,9 @@ export { serializeConversation } from "./utils.ts";
 /** Generate compaction summary data from prepared session history. */
 export async function compact(
 	preparation: CompactionPreparation,
+	models: Models,
 	model: Model<any>,
-	apiKey: string,
+	apiKey?: string,
 	headers?: Record<string, string>,
 	customInstructions?: string,
 	signal?: AbortSignal,
@@ -655,6 +656,7 @@ export async function compact(
 			messagesToSummarize.length > 0
 				? generateSummary(
 						messagesToSummarize,
+						models,
 						model,
 						settings.reserveTokens,
 						apiKey,
@@ -667,6 +669,7 @@ export async function compact(
 				: Promise.resolve(ok<string, CompactionError>("No prior history.")),
 			generateTurnPrefixSummary(
 				turnPrefixMessages,
+				models,
 				model,
 				settings.reserveTokens,
 				apiKey,
@@ -681,6 +684,7 @@ export async function compact(
 	} else {
 		const summaryResult = await generateSummary(
 			messagesToSummarize,
+			models,
 			model,
 			settings.reserveTokens,
 			apiKey,
@@ -706,9 +710,10 @@ export async function compact(
 }
 async function generateTurnPrefixSummary(
 	messages: AgentMessage[],
+	models: Models,
 	model: Model<any>,
 	reserveTokens: number,
-	apiKey: string,
+	apiKey?: string,
 	headers?: Record<string, string>,
 	signal?: AbortSignal,
 	thinkingLevel?: ThinkingLevel,
@@ -728,7 +733,7 @@ async function generateTurnPrefixSummary(
 		},
 	];
 
-	const response = await completeSimple(
+	const response = await models.completeSimple(
 		model,
 		{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },
 		model.reasoning && thinkingLevel && thinkingLevel !== "off"

packages/agent/src/harness/types.ts

@@ -1,4 +1,4 @@
-import type { ImageContent, Model, SimpleStreamOptions, TextContent, Transport } from "@earendil-works/pi-ai";
+import type { ImageContent, Model, Models, SimpleStreamOptions, TextContent, Transport } from "@earendil-works/pi-ai";
 import type { AgentEvent, AgentMessage, AgentTool, QueueMode, ThinkingLevel } from "../index.ts";
 import type { Session } from "./session/session.ts";
 
@@ -802,6 +802,12 @@ export interface AgentHarnessOptions<
 > {
 	env: ExecutionEnv;
 	session: Session;
+	/**
+	 * Provider collection used for all model requests (turn streaming,
+	 * compaction, branch summarization). Auth resolves through the providers'
+	 * auth; explicit per-request values (`getApiKeyAndHeaders`) win per field.
+	 */
+	models: Models;
 	tools?: TTool[];
 	/**
 	 * Concrete resources available to explicit invocation methods and system-prompt callbacks.

packages/agent/src/types.ts

@@ -1,19 +1,22 @@
 import type {
+	Api,
 	AssistantMessage,
 	AssistantMessageEvent,
+	AssistantMessageEventStream,
+	Context,
 	ImageContent,
 	Message,
 	Model,
 	SimpleStreamOptions,
-	streamSimple,
 	TextContent,
 	Tool,
 	ToolResultMessage,
-} from "@earendil-works/pi-ai/compat";
+} from "@earendil-works/pi-ai";
 import type { Static, TSchema } from "typebox";
 
 /**
- * Stream function used by the agent loop.
+ * Stream function used by the agent loop. `Models.streamSimple` satisfies
+ * this shape.
  *
  * Contract:
  * - Must not throw or return a rejected promise for request/model/runtime failures.
@@ -22,8 +25,10 @@ import type { Static, TSchema } from "typebox";
  *   final AssistantMessage with stopReason "error" or "aborted" and errorMessage.
  */
 export type StreamFn = (
-	...args: Parameters<typeof streamSimple>
-) => ReturnType<typeof streamSimple> | Promise<ReturnType<typeof streamSimple>>;
+	model: Model<Api>,
+	context: Context,
+	options?: SimpleStreamOptions,
+) => AssistantMessageEventStream | Promise<AssistantMessageEventStream>;
 
 /**
  * Configuration for how tool calls from a single assistant message are executed.

packages/agent/test/harness/agent-harness-stream.test.ts

@@ -1,18 +1,27 @@
-import { fauxAssistantMessage, fauxToolCall, registerFauxProvider, type StreamOptions } from "@earendil-works/pi-ai";
-import { afterEach, describe, expect, it } from "vitest";
+import {
+	createModels,
+	type FauxProviderHandle,
+	fauxAssistantMessage,
+	fauxProvider,
+	fauxToolCall,
+	type StreamOptions,
+} from "@earendil-works/pi-ai";
+import { describe, expect, it } from "vitest";
 import { AgentHarness } from "../../src/harness/agent-harness.ts";
 import { NodeExecutionEnv } from "../../src/harness/env/nodejs.ts";
 import { InMemorySessionStorage } from "../../src/harness/session/memory-storage.ts";
 import { Session } from "../../src/harness/session/session.ts";
 import { calculateTool } from "../utils/calculate.ts";
 
-const registrations: Array<{ unregister(): void }> = [];
+/** Shared collection; each faux provider gets a unique id so coexisting fakes route correctly. */
+const models = createModels();
+let fauxCount = 0;
 
-afterEach(() => {
-	for (const registration of registrations.splice(0)) {
-		registration.unregister();
-	}
-});
+function newFaux(): FauxProviderHandle {
+	const faux = fauxProvider({ provider: `faux-${++fauxCount}` });
+	models.setProvider(faux.provider);
+	return faux;
+}
 
 function createHarness(options: ConstructorParameters<typeof AgentHarness>[0]): AgentHarness {
 	return new AgentHarness(options);
@@ -29,8 +38,7 @@ function captureOptions(options: StreamOptions | undefined): StreamOptions {
 describe("AgentHarness stream configuration", () => {
 	it("snapshots stream options and merges auth headers before provider request hooks", async () => {
 		let capturedOptions: StreamOptions | undefined;
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([
 			(_context, options) => {
 				capturedOptions = options;
@@ -40,6 +48,7 @@ describe("AgentHarness stream configuration", () => {
 
 		const session = new Session(new InMemorySessionStorage({ metadata: { id: "session-1", createdAt: "now" } }));
 		const harness = createHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session,
 			model: registration.getModel(),
@@ -81,8 +90,7 @@ describe("AgentHarness stream configuration", () => {
 
 	it("chains provider request patches and supports deletion semantics", async () => {
 		let capturedOptions: StreamOptions | undefined;
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([
 			(_context, options) => {
 				capturedOptions = options;
@@ -91,6 +99,7 @@ describe("AgentHarness stream configuration", () => {
 		]);
 
 		const harness = createHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -133,8 +142,7 @@ describe("AgentHarness stream configuration", () => {
 
 	it("uses updated stream options for save-point snapshots without mutating the active request", async () => {
 		const capturedOptions: StreamOptions[] = [];
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([
 			(_context, options) => {
 				capturedOptions.push(captureOptions(options));
@@ -149,6 +157,7 @@ describe("AgentHarness stream configuration", () => {
 		]);
 
 		const harness = createHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -174,8 +183,7 @@ describe("AgentHarness stream configuration", () => {
 	it("chains provider payload hooks", async () => {
 		const seenPayloads: unknown[] = [];
 		let finalPayload: unknown;
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([
 			async (_context, options, _state, model) => {
 				finalPayload = await options?.onPayload?.({ steps: ["provider"] }, model);
@@ -184,6 +192,7 @@ describe("AgentHarness stream configuration", () => {
 		]);
 
 		const harness = createHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),

packages/agent/test/harness/agent-harness.test.ts

@@ -1,5 +1,13 @@
-import { fauxAssistantMessage, fauxToolCall, getModel, registerFauxProvider } from "@earendil-works/pi-ai/compat";
-import { afterEach, describe, expect, it } from "vitest";
+import {
+	createModels,
+	type FauxProviderHandle,
+	fauxAssistantMessage,
+	fauxProvider,
+	fauxToolCall,
+	type RegisterFauxProviderOptions,
+} from "@earendil-works/pi-ai";
+import { getModel } from "@earendil-works/pi-ai/compat";
+import { describe, expect, it } from "vitest";
 import { AgentHarness } from "../../src/harness/agent-harness.ts";
 import { NodeExecutionEnv } from "../../src/harness/env/nodejs.ts";
 import { InMemorySessionStorage } from "../../src/harness/session/memory-storage.ts";
@@ -17,7 +25,15 @@ interface AppPromptTemplate extends PromptTemplate {
 	source: "project" | "user";
 }
 
-const registrations: Array<{ unregister(): void }> = [];
+/** Shared collection; each faux provider gets a unique id so coexisting fakes route correctly. */
+const models = createModels();
+let fauxCount = 0;
+
+function newFaux(options: RegisterFauxProviderOptions = {}): FauxProviderHandle {
+	const faux = fauxProvider({ provider: `faux-${++fauxCount}`, ...options });
+	models.setProvider(faux.provider);
+	return faux;
+}
 
 function textFromUserMessages(messages: Array<{ role: string; content: unknown }>): string[] {
 	return messages.flatMap((message) => {
@@ -44,18 +60,13 @@ function getReasoning(options: unknown): unknown {
 	return options.reasoning;
 }
 
-afterEach(() => {
-	for (const registration of registrations.splice(0)) {
-		registration.unregister();
-	}
-});
-
 describe("AgentHarness", () => {
 	it("constructs directly and exposes queue modes", () => {
 		const session = new Session(new InMemorySessionStorage());
 		const env = new NodeExecutionEnv({ cwd: process.cwd() });
 		const initialModel = getModel("anthropic", "claude-sonnet-4-5");
 		const harness = new AgentHarness({
+			models,
 			env,
 			session,
 			model: initialModel,
@@ -76,8 +87,7 @@ describe("AgentHarness", () => {
 	});
 
 	it("drains one queued steering message at a time and emits queue updates", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		const userCounts: number[] = [];
 		registration.setResponses([
 			(context) => {
@@ -94,6 +104,7 @@ describe("AgentHarness", () => {
 			},
 		]);
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -119,8 +130,7 @@ describe("AgentHarness", () => {
 	});
 
 	it("appends before_agent_start messages and persists them", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		let requestText: string[] = [];
 		registration.setResponses([
 			(context) => {
@@ -130,6 +140,7 @@ describe("AgentHarness", () => {
 		]);
 		const session = new Session(new InMemorySessionStorage());
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session,
 			model: registration.getModel(),
@@ -151,8 +162,7 @@ describe("AgentHarness", () => {
 	});
 
 	it("abort clears steer and follow-up queues but preserves next-turn messages", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		let releaseFirstResponse: (() => void) | undefined;
 		let abortedSignal: AbortSignal | undefined;
 		const firstResponseReleased = new Promise<void>((resolve) => {
@@ -171,6 +181,7 @@ describe("AgentHarness", () => {
 			},
 		]);
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -206,8 +217,7 @@ describe("AgentHarness", () => {
 	});
 
 	it("drains follow-up messages one at a time after the agent would otherwise stop", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		const userCounts: number[] = [];
 		registration.setResponses([
 			(context) => {
@@ -224,6 +234,7 @@ describe("AgentHarness", () => {
 			},
 		]);
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -249,11 +260,11 @@ describe("AgentHarness", () => {
 	});
 
 	it("settles thrown hook failures with persisted assistant error messages", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([() => fauxAssistantMessage("should not be used")]);
 		const session = new Session(new InMemorySessionStorage());
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session,
 			model: registration.getModel(),
@@ -280,13 +291,12 @@ describe("AgentHarness", () => {
 	});
 
 	it("refreshes model, thinking level, resources, system prompt, and active tools at save points", async () => {
-		const registration = registerFauxProvider({
+		const registration = newFaux({
 			models: [
 				{ id: "first", reasoning: true },
 				{ id: "second", reasoning: true },
 			],
 		});
-		registrations.push(registration);
 		const secondModel = registration.getModel("second");
 		if (!secondModel) throw new Error("missing second faux model");
 		const captured: Array<{ modelId: string; reasoning: unknown; systemPrompt: string; tools: string[] }> = [];
@@ -313,6 +323,7 @@ describe("AgentHarness", () => {
 			},
 		]);
 		const harness = new AgentHarness<Skill, PromptTemplate, AgentTool>({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -345,11 +356,11 @@ describe("AgentHarness", () => {
 	});
 
 	it("orders pending listener session writes after agent-emitted messages", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([() => fauxAssistantMessage("ok")]);
 		const session = new Session(new InMemorySessionStorage());
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session,
 			model: registration.getModel(),
@@ -376,11 +387,11 @@ describe("AgentHarness", () => {
 	});
 
 	it("waitForIdle waits for external run settlement and awaited listeners", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([() => fauxAssistantMessage("ok")]);
 		const barrier = deferred();
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session: new Session(new InMemorySessionStorage()),
 			model: registration.getModel(),
@@ -408,8 +419,7 @@ describe("AgentHarness", () => {
 	});
 
 	it("runs tool_call and tool_result hooks through the direct loop", async () => {
-		const registration = registerFauxProvider();
-		registrations.push(registration);
+		const registration = newFaux();
 		registration.setResponses([
 			() =>
 				fauxAssistantMessage(fauxToolCall("calculate", { expression: "2 + 2" }, { id: "call-1" }), {
@@ -418,6 +428,7 @@ describe("AgentHarness", () => {
 		]);
 		const session = new Session(new InMemorySessionStorage());
 		const harness = new AgentHarness({
+			models,
 			env: new NodeExecutionEnv({ cwd: process.cwd() }),
 			session,
 			model: registration.getModel(),
@@ -462,6 +473,7 @@ describe("AgentHarness", () => {
 		const inspectTool: AppTool = { ...calculateTool, name: "inspect", source: "builtin" };
 		const searchTool: AppTool = { ...calculateTool, name: "search", source: "extension" };
 		const harness = new AgentHarness<AppSkill, AppPromptTemplate, AppTool>({
+			models,
 			env,
 			session,
 			model,
@@ -530,11 +542,12 @@ describe("AgentHarness", () => {
 		const env = new NodeExecutionEnv({ cwd: process.cwd() });
 		const model = getModel("anthropic", "claude-sonnet-4-5");
 		expect(
-			() => new AgentHarness({ env, session, model, tools: [calculateTool], activeToolNames: ["missing"] }),
+			() => new AgentHarness({ env, session, models, model, tools: [calculateTool], activeToolNames: ["missing"] }),
 		).toThrow(/Unknown tool/);
 		expect(
 			() =>
 				new AgentHarness({
+					models,
 					env,
 					session,
 					model,
@@ -545,6 +558,7 @@ describe("AgentHarness", () => {
 		expect(
 			() =>
 				new AgentHarness({
+					models,
 					env,
 					session,
 					model,
@@ -558,7 +572,7 @@ describe("AgentHarness", () => {
 		const session = new Session(new InMemorySessionStorage());
 		const env = new NodeExecutionEnv({ cwd: process.cwd() });
 		const model = getModel("anthropic", "claude-sonnet-4-5");
-		const harness = new AgentHarness<AppSkill, AppPromptTemplate, AgentTool>({ env, session, model });
+		const harness = new AgentHarness<AppSkill, AppPromptTemplate, AgentTool>({ env, session, models, model });
 		const skill: AppSkill = {
 			name: "inspect",
 			description: "Inspect things",

packages/agent/test/harness/compaction.test.ts

@@ -1,13 +1,14 @@
 import {
 	type AssistantMessage,
-	type FauxProviderRegistration,
+	createModels,
+	type FauxProviderHandle,
 	fauxAssistantMessage,
+	fauxProvider,
 	type Message,
 	type Model,
-	registerFauxProvider,
 	type Usage,
 } from "@earendil-works/pi-ai";
-import { afterEach, beforeEach, describe, expect, it } from "vitest";
+import { beforeEach, describe, expect, it } from "vitest";
 import {
 	type CompactionPreparation,
 	calculateContextTokens,
@@ -121,11 +122,13 @@ function createModelChangeEntry(provider: string, modelId: string, parentId: str
 	};
 }
 
-function createFauxModel(
-	reasoning: boolean,
-	maxTokens = 8192,
-): { faux: FauxProviderRegistration; model: Model<string> } {
-	const faux = registerFauxProvider({
+/** Shared collection; each faux provider gets a unique id so coexisting fakes route correctly. */
+const models = createModels();
+let fauxCount = 0;
+
+function createFauxModel(reasoning: boolean, maxTokens = 8192): { faux: FauxProviderHandle; model: Model<string> } {
+	const faux = fauxProvider({
+		provider: `faux-${++fauxCount}`,
 		models: [
 			{
 				id: reasoning ? "reasoning-model" : "non-reasoning-model",
@@ -135,18 +138,10 @@ function createFauxModel(
 			},
 		],
 	});
-	fauxRegistrations.push(faux);
+	models.setProvider(faux.provider);
 	return { faux, model: faux.getModel() };
 }
 
-const fauxRegistrations: FauxProviderRegistration[] = [];
-
-afterEach(() => {
-	while (fauxRegistrations.length > 0) {
-		fauxRegistrations.pop()?.unregister();
-	}
-});
-
 describe("harness compaction", () => {
 	beforeEach(() => {
 		nextId = 0;
@@ -447,6 +442,7 @@ describe("harness compaction", () => {
 		getOrThrow(
 			await generateSummary(
 				messages,
+				models,
 				reasoningModel,
 				2000,
 				"test-key",
@@ -467,7 +463,18 @@ describe("harness compaction", () => {
 			},
 		]);
 		getOrThrow(
-			await generateSummary(messages, offModel, 2000, "test-key", undefined, undefined, undefined, undefined, "off"),
+			await generateSummary(
+				messages,
+				models,
+				offModel,
+				2000,
+				"test-key",
+				undefined,
+				undefined,
+				undefined,
+				undefined,
+				"off",
+			),
 		);
 		expect(seenOptions[1]).not.toHaveProperty("reasoning");
 
@@ -481,6 +488,7 @@ describe("harness compaction", () => {
 		getOrThrow(
 			await generateSummary(
 				messages,
+				models,
 				nonReasoningModel,
 				2000,
 				"test-key",
@@ -510,6 +518,7 @@ describe("harness compaction", () => {
 		const summary = getOrThrow(
 			await generateSummary(
 				messages,
+				models,
 				model,
 				2000,
 				"test-key",
@@ -529,7 +538,7 @@ describe("harness compaction", () => {
 		const messages: AgentMessage[] = [createUserMessage("Summarize this.")];
 		const { faux: errorFaux, model: errorModel } = createFauxModel(false);
 		errorFaux.setResponses([fauxAssistantMessage("", { stopReason: "error", errorMessage: "boom" })]);
-		const errorResult = await generateSummary(messages, errorModel, 2000, "test-key");
+		const errorResult = await generateSummary(messages, models, errorModel, 2000, "test-key");
 		expect(errorResult).toMatchObject({
 			ok: false,
 			error: { code: "summarization_failed", message: "Summarization failed: boom" },
@@ -537,7 +546,7 @@ describe("harness compaction", () => {
 
 		const { faux: abortedFaux, model: abortedModel } = createFauxModel(false);
 		abortedFaux.setResponses([fauxAssistantMessage("", { stopReason: "aborted", errorMessage: "stopped" })]);
-		const abortedResult = await generateSummary(messages, abortedModel, 2000, "test-key");
+		const abortedResult = await generateSummary(messages, models, abortedModel, 2000, "test-key");
 		expect(abortedResult).toMatchObject({ ok: false, error: { code: "aborted", message: "stopped" } });
 	});
 
@@ -565,7 +574,7 @@ describe("harness compaction", () => {
 			settings: { enabled: true, reserveTokens: 500000, keepRecentTokens: 20000 },
 		};
 
-		getOrThrow(await compact(preparation, model, "test-key"));
+		getOrThrow(await compact(preparation, models, model, "test-key"));
 
 		expect(seenOptions.map((options) => options?.maxTokens)).toEqual([128000, 128000]);
 	});
@@ -583,7 +592,7 @@ describe("harness compaction", () => {
 		};
 		const { faux: historyFaux, model: historyModel } = createFauxModel(false);
 		historyFaux.setResponses([fauxAssistantMessage("", { stopReason: "error", errorMessage: "history failed" })]);
-		expect(await compact(preparation, historyModel, "test-key")).toMatchObject({
+		expect(await compact(preparation, models, historyModel, "test-key")).toMatchObject({
 			ok: false,
 			error: { code: "summarization_failed", message: "Summarization failed: history failed" },
 		});
@@ -591,6 +600,7 @@ describe("harness compaction", () => {
 		const { model: invalidModel } = createFauxModel(false);
 		const invalidResult = await compact(
 			{ ...preparation, messagesToSummarize: [], firstKeptEntryId: "" },
+			models,
 			invalidModel,
 			"test-key",
 		);
@@ -617,7 +627,7 @@ describe("harness compaction", () => {
 			settings: { enabled: true, reserveTokens: 2000, keepRecentTokens: 20 },
 		};
 
-		getOrThrow(await compact(preparation, model, "test-key", undefined, undefined, undefined, "high"));
+		getOrThrow(await compact(preparation, models, model, "test-key", undefined, undefined, undefined, "high"));
 
 		expect(seenOptions[0]).toMatchObject({ reasoning: "high" });
 	});
@@ -636,14 +646,14 @@ describe("harness compaction", () => {
 		const { faux, model } = createFauxModel(false);
 		faux.setResponses([fauxAssistantMessage("", { stopReason: "error", errorMessage: "prefix failed" })]);
 
-		expect(await compact(preparation, model, "test-key")).toMatchObject({
+		expect(await compact(preparation, models, model, "test-key")).toMatchObject({
 			ok: false,
 			error: { code: "summarization_failed", message: "Turn prefix summarization failed: prefix failed" },
 		});
 
 		const { faux: abortedFaux, model: abortedModel } = createFauxModel(false);
 		abortedFaux.setResponses([fauxAssistantMessage("", { stopReason: "aborted", errorMessage: "prefix stopped" })]);
-		expect(await compact(preparation, abortedModel, "test-key")).toMatchObject({
+		expect(await compact(preparation, models, abortedModel, "test-key")).toMatchObject({
 			ok: false,
 			error: { code: "aborted", message: "prefix stopped" },
 		});
@@ -662,7 +672,7 @@ describe("harness compaction", () => {
 		expect(preparation).toBeDefined();
 		const { faux, model } = createFauxModel(false);
 		faux.setResponses([fauxAssistantMessage("## Goal\nTest summary")]);
-		const result = getOrThrow(await compact(preparation!, model, "test-key"));
+		const result = getOrThrow(await compact(preparation!, models, model, "test-key"));
 		expect(result.summary.length).toBeGreaterThan(0);
 		expect(result.firstKeptEntryId).toBeTruthy();
 		expect(result.details).toBeDefined();

packages/agent/test/scratch/simple.ts

@@ -1,6 +1,8 @@
 import { homedir } from "node:os";
 import { join } from "node:path";
-import { getModel } from "@earendil-works/pi-ai/compat";
+import { createModels } from "@earendil-works/pi-ai";
+import { getBuiltinModel } from "@earendil-works/pi-ai/providers/all";
+import { openaiProvider } from "@earendil-works/pi-ai/providers/openai";
 import { NodeExecutionEnv } from "../../src/harness/env/nodejs.ts";
 import { InMemorySessionStorage } from "../../src/harness/session/memory-storage.ts";
 import {
@@ -35,11 +37,15 @@ const { promptTemplates: sourcedPromptTemplates } = await loadSourcedPromptTempl
 	(promptTemplate, source) => ({ ...promptTemplate, source }),
 );
 
+const models = createModels();
+models.setProvider(openaiProvider());
+
 const session = new Session(new InMemorySessionStorage());
 const agent = new AgentHarness({
 	env,
 	session,
-	model: getModel("openai", "gpt-5.5"),
+	models,
+	model: getBuiltinModel("openai", "gpt-5.5"),
 	thinkingLevel: "low",
 	systemPrompt: ({ env, resources }) =>
 		[