@makerstufferupper/flypi 0.0.0-stage → 0.1.2

package/src/agent/permission-gate.ts

@@ -0,0 +1,70 @@
+/**
+ * The permission gate (docs/architecture.md Phase 5): a tool on flypi's ask-before-use list does not run
+ * until the user has put it there. `permissions.askBeforeUse` in `.flypi/agent/settings.json`, next to the
+ * `memory` block src/memory/config.ts reads — one place for what flypi asks pi to do.
+ *
+ * The plan said "using tool annotations". It cannot: pi hands a `tool_call` handler a tool *name* and its
+ * arguments, and the MCP annotations a server announces (`destructiveHint`, `openWorldHint`) stay inside
+ * pi's MCP extension, which exposes none of its connection state to the host (docs/pi-api.md). So the list
+ * is of tool names, and the status panel (src/agent/mcp-status.ts) is where the annotations surface — a
+ * user reads "this tool says it is destructive" there and writes its name here.
+ *
+ * The list is re-read on every call rather than captured at session start, so the loop closes without a
+ * restart: the model is told the call was gated, the user adds the name in the file, and the next call runs.
+ */
+import { readFileSync } from "node:fs";
+import type { ExtensionFactory } from "@earendil-works/pi-coding-agent";
+import type { FlypiPaths } from "./paths.ts";
+
+/** pi's own result shape for a blocked call (`ToolCallEventResult`). */
+export const blockedToolCall = (reason: string) => ({ block: true, reason });
+
+const BLOCKED = (toolName: string) =>
+	blockedToolCall(
+		`flypi did not run "${toolName}": it needs your approval first (it is on permissions.askBeforeUse in ` +
+		`.flypi/agent/settings.json). Ask the user whether this call may run; if they agree, they add "${toolName}" to that list and you can call it again.`,
+	);
+
+/**
+ * `permissions.askBeforeUse` out of a parsed settings.json. Anything that is not a non-empty string is
+ * dropped: this file is also pi's, and a stray key there must not become a gate that matches nothing.
+ */
+export function askBeforeUseOf(settings: unknown): string[] {
+	const permissions = (settings as { permissions?: { askBeforeUse?: unknown } } | null)?.permissions;
+	const list = Array.isArray(permissions?.askBeforeUse) ? permissions.askBeforeUse : [];
+	return list.filter((entry): entry is string => typeof entry === "string" && entry.trim() !== "").map((entry) => entry.trim());
+}
+
+/**
+ * Whether a tool is on the list. A name matches exactly; a name ending in `*` matches that prefix, which is
+ * how a whole server is gated (`mcp__filesystem__*`) without listing every tool it may grow.
+ */
+export function needsApproval(list: readonly string[], toolName: string): boolean {
+	return list.some((entry) => (entry.endsWith("*") ? toolName.startsWith(entry.slice(0, -1)) : entry === toolName));
+}
+
+/**
+ * The extension that enforces it. Registered by `createFlypiSession` like the other flypi extensions, so
+ * every session — the daemon's, `npm run serve`'s, a test's — is gated the same way.
+ */
+export function createPermissionGateExtension(paths: FlypiPaths): ExtensionFactory {
+	return (pi) => {
+		pi.on("tool_call", (event) => {
+			if (!needsApproval(readList(paths.settingsPath), event.toolName)) return;
+			return BLOCKED(event.toolName);
+		});
+	};
+}
+
+/**
+ * Read the list now, so an edit takes effect on the next call. A settings.json that will not parse asks for
+ * nothing: pi itself would have refused to start the conversation over that file, and a gate that guessed at
+ * a name would block tools nobody asked to gate.
+ */
+function readList(settingsPath: string): string[] {
+	try {
+		return askBeforeUseOf(JSON.parse(readFileSync(settingsPath, "utf8")));
+	} catch {
+		return [];
+	}
+}

package/src/agent/pi-mcp.ts

@@ -0,0 +1,97 @@
+/**
+ * pi's MCP OAuth machinery, reached at the file pi's own MCP extension loads it from.
+ *
+ * pi does not export any of this from its package entry: `createMcpExtension()` is the only public door to
+ * the code, and it runs sign-in solely through pi's TUI (`/mcp`, which draws a terminal dialog). flypi signs
+ * a user in from the settings view of a headless server, so it calls the same function the extension calls —
+ * `signInMcpServer` — with a prompt of its own: the URL goes to the browser, the redirect comes back as a
+ * paste (docs/architecture.md Phase 5, risk 5).
+ *
+ * The shapes below are copied from pi's `.d.ts` files, which the package's `exports` map keeps unreachable
+ * as well. They describe four pi functions and nothing more; re-check them against
+ * `dist/extensions/mcp/oauth.d.ts` after a pi upgrade (docs/pi-api.md).
+ */
+import { getPackageDir, type McpExtensionOptions, type McpServerConfig } from "@earendil-works/pi-coding-agent";
+import { join } from "node:path";
+import { pathToFileURL } from "node:url";
+import type { FlypiPaths } from "./paths.ts";
+
+/** pi's `McpOAuthCredentialStore`, which `createMcpExtension({ credentials })` takes. */
+export type PiMcpCredentials = NonNullable<McpExtensionOptions["credentials"]>;
+
+/** What an authorization needs: pi's `McpOAuthSettings`, i.e. the `oauth` block of an `mcp.json` entry. */
+export interface PiMcpOAuthSettings {
+	clientId?: string;
+	/** Already resolved — `${NAME}` and `!command` expanded. */
+	clientSecret?: string;
+	callbackPort?: number;
+	callbackUrl?: string;
+	scope?: string;
+	clientName?: string;
+	clientRegistration?: "dcr" | "cimd";
+	authServerMetadataUrl?: URL;
+}
+
+/** The two ways a sign-in finishes: the browser reaches the loopback callback, or the user pastes what it landed on. */
+export interface PiMcpSignInPrompt {
+	showAuthorizationUrl(url: URL): void;
+	/** `undefined` or "" cancels. Abandoned once the loopback callback arrives. */
+	promptForRedirectUrl(signal: AbortSignal): Promise<string | undefined>;
+}
+
+/** pi's `McpOAuthCredentialStore.forServer()` — the per-server state handle `signInMcpServer` takes. */
+export interface PiMcpServerStore {
+	load(): unknown;
+	save(state: { serverUrl: string; tokens?: { access_token?: string; token_type?: string } }): void | Promise<void>;
+}
+
+export interface PiMcpRuntime {
+	signInMcpServer(options: { serverUrl: string; store: PiMcpServerStore; settings: PiMcpOAuthSettings; prompt: PiMcpSignInPrompt }): Promise<void>;
+	McpOAuthCredentialStore: new (backend: unknown, lockDir: string) => PiMcpCredentials & {
+		/** What a sign-in stored for this server, or nothing. Synchronous, like pi's own. */
+		tokens(name: string, serverUrl: string): { access_token?: string } | undefined;
+		forServer(name: string, serverUrl: string): PiMcpServerStore;
+	};
+	FileAuthStorageBackend: new (filePath: string) => unknown;
+	McpSignInCancelledError: new (...args: never[]) => Error;
+	/** `${NAME}` / `!command` → the value, or a thrown error naming what is missing. */
+	resolveConfigValueOrThrow: (value: string, description: string) => string;
+}
+
+let cached: Promise<PiMcpRuntime> | undefined;
+
+/** Loaded once per process, and only when something actually needs it: this pulls in pi's MCP client. */
+export function loadPiMcpRuntime(): Promise<PiMcpRuntime> {
+	const dir = getPackageDir();
+	return (cached ??= Promise.all([
+		import(pathToFileURL(join(dir, "dist/extensions/mcp/runtime.js")).href),
+		import(pathToFileURL(join(dir, "dist/core/auth-storage.js")).href),
+		import(pathToFileURL(join(dir, "dist/core/resolve-config-value.js")).href),
+	]).then(([runtime, authStorage, configValue]) => ({
+		signInMcpServer: runtime.signInMcpServer,
+		McpOAuthCredentialStore: runtime.McpOAuthCredentialStore,
+		FileAuthStorageBackend: authStorage.FileAuthStorageBackend,
+		McpSignInCancelledError: runtime.McpSignInCancelledError,
+		resolveConfigValueOrThrow: configValue.resolveConfigValueOrThrow,
+	})));
+}
+
+/**
+ * The credential store, pointed at flypi's own agent dir. pi's default is `mcp-auth.json` in *its* agent dir
+ * (`$PI_CODING_AGENT_DIR`, else `~/.pi/agent`), which is the config bleed this repo refuses everywhere else
+ * (risk 1) — so every consumer passes this: a session's connections, the status probe, and a sign-in.
+ */
+export async function mcpCredentials(paths: FlypiPaths): Promise<PiMcpCredentials & { tokens(name: string, serverUrl: string): { access_token?: string } | undefined }> {
+	const pi = await loadPiMcpRuntime();
+	return new pi.McpOAuthCredentialStore(new pi.FileAuthStorageBackend(paths.mcpAuthPath), paths.agentDir);
+}
+
+/**
+ * Whether pi authenticates this entry with OAuth: an HTTP server that sends no `Authorization` header of its
+ * own and delegates to no provider. That is pi's own rule, copied because the probe and the sign-in have to
+ * agree with what a session will do — otherwise the panel offers a sign-in the file will ignore.
+ */
+export function usesMcpOAuth(config: McpServerConfig): boolean {
+	if (!("url" in config) || config.auth?.provider) return false;
+	return !Object.keys(config.headers ?? {}).some((header) => header.toLowerCase() === "authorization");
+}

package/src/agent/runtime.ts

@@ -0,0 +1,347 @@
+import { existsSync, mkdirSync, readFileSync } from "node:fs";
+import { join } from "node:path";
+import {
+	createAgentSession,
+	createMcpExtension,
+	createToolSearchExtension,
+	DefaultResourceLoader,
+	type LoadedMcpConfig,
+	type McpServerConfig,
+	type McpServerEntry,
+	ModelRuntime,
+	type CreateAgentSessionOptions,
+	type ExtensionUIContext,
+	SessionManager,
+	SettingsManager,
+} from "@earendil-works/pi-coding-agent";
+import { assertFlypiOwned, flypiPaths, type FlypiPaths } from "./paths.ts";
+import { mcpCredentials, usesMcpOAuth } from "./pi-mcp.ts";
+import { importEnvCredential, readAuth, resolveCredential } from "./onboarding.ts";
+import { memoryConfig } from "../memory/config.ts";
+import { createMemoryRecallExtension, memoryInstruction, memoryTool, recallServer } from "../memory/recall.ts";
+import { backgroundMount, reviewPrompt, type ReviewParent } from "../memory/review.ts";
+import { createPermissionGateExtension } from "./permission-gate.ts";
+import { createSecretsExtension } from "../secrets/extension.ts";
+import type { SecretStore } from "../secrets/vault.ts";
+
+/** Derived from pi's own option types, so no transitive pi-ai / pi-agent-core import is needed here. */
+export type FlypiModel = NonNullable<CreateAgentSessionOptions["model"]>;
+export type FlypiThinkingLevel = NonNullable<CreateAgentSessionOptions["thinkingLevel"]>;
+export type FlypiSession = NonNullable<Awaited<ReturnType<typeof createAgentSession>>["session"]>;
+
+/** A configuration problem the user can act on. probe.ts prints `message`, never a stack trace. */
+export class FlypiConfigError extends Error {}
+
+/**
+ * The chat renders markdown (ui/Chat.svelte, src/shared/markdown.ts) and nothing in pi's prompt asks
+ * for it — pi's only answer-shaping rules are "be concise" and "show file paths clearly" — so the
+ * format the renderer was built for is stated once, here. The loader's `appendSystemPrompt` becomes
+ * the prompt's addendum section, so it is byte-stable across turns like the rest of the system prompt.
+ */
+export const ANSWER_FORMAT_PROMPT =
+	"Answer in markdown — this interface renders it: headings for structure, fenced code blocks for code and command output, tables for comparisons, backticks for file paths and identifiers.";
+
+/**
+ * The one place flypi creates an `AgentSession`: probe, the WS server, and the tests all go
+ * through here, so the config-isolation invariant (docs/risks.md risk 1) is enforced once.
+ *
+ * `modelRuntime`/`model` are overridable so a test can pass the stub model
+ * (tests/helpers/stub-model.ts) instead of a real provider.
+ */
+export async function createFlypiSession(options: {
+	root?: string;
+	modelRuntime?: ModelRuntime;
+	model?: FlypiModel;
+	thinkingLevel?: FlypiThinkingLevel;
+	tools?: string[];
+	env?: Record<string, string | undefined>;
+	/** Supply an existing or opened session instead of starting a new one. */
+	sessionManager?: SessionManager;
+	/**
+	 * What MCP servers this session gets, instead of what `.flypi/agent/mcp.json` says. Only the probe's
+	 * `--no-memory` control needs it (docs/prompt-cache.md C1): renaming the real file aside to run the
+	 * control left the memory server unconfigured whenever the process was killed mid-run
+	 * (code-review.md finding 1), and the config is not something a measurement is allowed to edit.
+	 */
+	mcpConfig?: LoadedMcpConfig;
+	/**
+	 * What automatic recall put in front of the model this turn, handed to the caller as it happens. The block
+	 * itself is a session entry now (Phase 4.7 C2), but this is still how the debug panel learns *which* block
+	 * the turn in front of it was shown, and the one thing it can never read out of the file is the note a
+	 * failed search left on the message (docs/memory.md Phase 4).
+	 */
+	onMemoryBlock?: (block: string) => void;
+	/**
+	 * flypi's own id for the conversation, so the turns the memory store records are keyed by the same id the
+	 * UI and the registry use (docs/memory.md Phase 4). Left out, nothing is recorded about the turns.
+	 */
+	conversationId?: string;
+	/**
+	 * The vault (docs/secrets.md Phase 5.5). Given, the session can reach it and every tool result, memory body
+	 * and answer in it is redacted against it; omitted, the conversation has no secret tools at all — which is
+	 * what a process that never opened the vault (a bare `probe.ts`, a unit test) should get.
+	 */
+	vault?: SecretStore;
+	/**
+	 * This is a conversation a human is in, so it may spawn a review fork after every `memory.reviewInterval`
+	 * delivered answers (docs/memory.md Phase 4.8). Off unless asked for: a probe, a daemon job, a subagent and
+	 * a fork itself never spawn one, which is the whole suppression rule — there is no runner to fire.
+	 */
+	review?: boolean;
+	/**
+	 * The conversation's extension UI (src/server/extension-ui.ts): what `ctx.ui.notify/select/confirm/input`
+	 * do. Absent, pi binds its no-op context and an extension command's output goes nowhere — which is what a
+	 * process with no page attached (a probe, a fork) should get.
+	 */
+	extensionUi?: ExtensionUIContext;
+	/**
+	 * Automatic recall (src/memory/recall.ts). False for a review fork: it inherits the parent's memory block
+	 * rather than searching, logging a turn of its own, or auto-capturing over a conversation that is not its
+	 * own. The MCP server is still connected, so the fork can still write what it learned.
+	 */
+	memoryRecall?: boolean;
+} = {}): Promise<{ session: FlypiSession; paths: FlypiPaths; modelRuntime: ModelRuntime }> {
+	const paths = flypiPaths(options.root ?? process.cwd());
+	assertFlypiOwned(paths);
+
+	const modelRuntime =
+		options.modelRuntime ??
+		(await ModelRuntime.create({ authPath: paths.authPath, modelsPath: paths.modelsPath }));
+
+	const sessionManager = options.sessionManager ?? SessionManager.create(paths.root, paths.sessionsDir);
+	// A model the caller injects (the stub-model tests) wins. Otherwise the settings default is read — but
+	// it is *passed* only when the session has nothing to restore: createAgentSession restores a resumed
+	// session's own model from its `model_change` entries when no model is given, and answering over that
+	// is why a reopened chat went back to the default model while pi kept its thinking level. The settings
+	// default stays the fallback for a recorded model that can no longer be resolved, and pi's own
+	// findInitialModel applies it — so a session with no messages still starts where flypi starts one.
+	const resuming = sessionManager.buildSessionContext().messages.length > 0;
+	const model = options.model ?? (resuming ? undefined : resolveConfiguredModel(paths, modelRuntime, options.env));
+
+	// MCP is not part of an SDK session unless the host asks for it (docs/sdk.md), and pi's own MCP
+	// config loader reads `$PI_CODING_AGENT_DIR` (else `~/.pi/agent`) rather than the `agentDir` passed
+	// below — so both the extension and where it reads from are flypi's job (docs/risks.md risk 1).
+	const mcpConfig = options.mcpConfig ?? readFlypiMcpConfig(paths);
+	// pi keeps an OAuth server's tokens in `mcp-auth.json` in *its* agent dir, which for flypi would be the
+	// host `~/.pi/agent` — the same config bleed as the config file itself (risk 1). Only built when an entry
+	// could use it, because loading pi's OAuth machinery means loading its MCP client.
+	const mcpCredentialsStore = mcpConfig.servers.some((server) => server.config.enabled !== false && usesMcpOAuth(server.config))
+		? await mcpCredentials(paths)
+		: undefined;
+	// Resolved once, and used for both halves of the memory wiring: the instruction below and the recall
+	// extension have to agree about whether this session has a memory server at all.
+	const memoryServer = recallServer(mcpConfig);
+	// The review fork (docs/memory.md Phase 4.8): a conversation a human is in, a store to write to, and an
+	// interval that is not switched off. `session` is read from inside the closure, which runs turns later.
+	const reviewInterval = options.review && memoryServer ? memoryConfig(paths).review.interval : 0;
+	const review =
+		reviewInterval > 0 && memoryServer
+			? {
+					interval: reviewInterval,
+					fork: (parent: ReviewParent) =>
+						runReviewFork({ paths, modelRuntime, mcpConfig, memoryServer, vault: options.vault, tools: options.tools, parentSession: () => session }, parent),
+				}
+			: undefined;
+	const resourceLoader = new DefaultResourceLoader({
+		cwd: paths.root,
+		agentDir: paths.agentDir,
+		// The loader, not an extension: these say the same thing on every turn, and a per-turn
+		// `before_agent_start` handler could only repeat them. The memory sentence used to sit at the top of
+		// the recalled block; there it rode in the newest message, the one region a prompt cache re-pays at
+		// the turn boundary, while the prompt is byte-stable and at the front of every request
+		// (docs/prompt-cache.md Phase 4.6 B1). A session with no memory server is not told about one.
+		appendSystemPrompt: [
+			ANSWER_FORMAT_PROMPT,
+			...(memoryServer ? [memoryInstruction(memoryServer.name)] : []),
+		],
+		extensionFactories: [
+			// flypi's gate first: a tool the user has not approved is refused before any other extension
+			// gets to act on the call (src/agent/permission-gate.ts).
+			createPermissionGateExtension(paths),
+			createMcpExtension({
+				loadConfig: () => mcpConfig,
+				logPath: paths.mcpLogPath,
+				// Where a signed-in server's token comes from: `.flypi/agent/mcp-auth.json`, the file the settings
+				// view's sign-in writes (src/agent/mcp-signin.ts). Left out, pi would read the host's.
+				...(mcpCredentialsStore ? { credentials: mcpCredentialsStore } : {}),
+			}),
+			// The only route to an MCP server the model was not given its tools for: `exposure: "deferred"`,
+			// and pi's own default, `codemode` — flypi loads no codemode extension, so without this tool such
+			// a server is connected, answers, and can never be used (docs/mcp.md, "Control tool exposure").
+			// The settings view offers those exposures, so the sessions have to be able to load them.
+			createToolSearchExtension(),
+			...(options.memoryRecall === false
+				? []
+				: [createMemoryRecallExtension(memoryServer, options.onMemoryBlock, { conversationId: options.conversationId, ...(review ? { review } : {}) })]),
+			// Registered only where a vault was opened, and last: the redaction it installs rewrites what the
+			// other extensions produced on the way to the transcript (docs/secrets.md Phase 5.5).
+			...(options.vault ? [createSecretsExtension(options.vault, { paths })] : []),
+		],
+	});
+	await resourceLoader.reload();
+
+	mkdirSync(paths.sessionsDir, { recursive: true });
+	const { session } = await createAgentSession({
+		cwd: paths.root,
+		agentDir: paths.agentDir,
+		modelRuntime,
+		model,
+		thinkingLevel: options.thinkingLevel,
+		resourceLoader,
+		sessionManager,
+		settingsManager: SettingsManager.create(paths.root, paths.agentDir),
+		tools: options.tools,
+	});
+	// Emits `session_start`, which is what connects the MCP servers; without it the tools never arrive.
+	// Bound once: a second `bindExtensions` would emit `session_start` again and reconnect every server.
+	// `rpc`, not `tui`: there is a UI, but it is not a terminal, and extensions guard terminal-only calls
+	// (`ctx.ui.custom`, widgets, themes) on that value.
+	await session.bindExtensions(options.extensionUi ? { uiContext: options.extensionUi, mode: "rpc" } : {});
+	// Registering it is not enough: `tool_search` comes up inactive. A `deferred` server's tools happen to
+	// activate it when they register, and a default-exposure (`codemode`) server's do not — the first version
+	// of this line was deleted as redundant and only the second case in mcp-tool-search.test.ts noticed.
+	// An add-only loadout, so pi keeps the MCP tools still connecting pending and declares them when they
+	// arrive (`setActiveToolsByName` in pi's own comment).
+	session.setActiveToolsByName([...session.getActiveToolNames(), "tool_search"]);
+	return { session, paths, modelRuntime };
+}
+
+/**
+ * One review fork (docs/memory.md Phase 4.8): the parent's conversation replayed into a session that writes
+ * nothing to disk, asked one question — save what is worth saving, or say so — and disposed.
+ *
+ * It goes through `createFlypiSession` rather than `createAgentSession` so the config-isolation invariant
+ * (risk 1) is still enforced once, and so the request is built the same way the parent's was: same provider,
+ * model and thinking level, same system-prompt bytes, same advertised tools. That is what lets the fork extend
+ * a prefix the route has already read instead of paying ~30K tokens cold (docs/prompt-cache.md).
+ *
+ * What makes it safe is the session manager: `SessionManager.inMemory` seeded with the parent's entries is
+ * pi's own non-persisting session, so the fork can write memories and cannot write a line of the conversation
+ * a human is reading (risk 4). Its writes land `session_kind='background'` because its memory server is on the
+ * background mount, where the store refuses supersede, forget and `owner` (src/memory/mcp.ts).
+ */
+async function runReviewFork(
+	deps: {
+		paths: FlypiPaths;
+		modelRuntime: ModelRuntime;
+		mcpConfig: LoadedMcpConfig;
+		memoryServer: McpServerEntry;
+		vault?: SecretStore;
+		tools?: string[];
+		parentSession: () => FlypiSession;
+	},
+	parent: ReviewParent,
+): Promise<void> {
+	const source = deps.parentSession();
+	const { session } = await createFlypiSession({
+		root: deps.paths.root,
+		modelRuntime: deps.modelRuntime,
+		// The parent's own model and thinking level, not the settings default: a fork on another model is a
+		// different request, and extending the parent's prefix is the whole cost argument.
+		...(source.model ? { model: source.model } : {}),
+		...(source.thinkingLevel ? { thinkingLevel: source.thinkingLevel } : {}),
+		...(deps.tools ? { tools: deps.tools } : {}),
+		...(deps.vault ? { vault: deps.vault } : {}),
+		// Same servers, names and exposures — only the memory server moved to the background mount.
+		mcpConfig: backgroundMount(deps.mcpConfig, deps.memoryServer.name),
+		sessionManager: SessionManager.inMemory(deps.paths.root, undefined, parent.entries),
+		memoryRecall: false,
+	});
+	try {
+		// MCP tools arrive asynchronously, and asking before they land would advertise a smaller `tools[]` than
+		// the parent's: a different request, and a cold cache. Bounded, because a server that never answers must
+		// not hang a fork nobody is waiting on.
+		if (await toolsArrived(session, memoryTool(deps.memoryServer.name, "memory_write"))) {
+			await session.prompt(reviewPrompt(deps.memoryServer.name));
+		}
+	} finally {
+		session.dispose();
+	}
+}
+
+/** Whether the memory tools are declared yet. A fork with no memory tools has nothing to be for. */
+async function toolsArrived(session: FlypiSession, writeTool: string, deadlineMs = 10_000): Promise<boolean> {
+	const until = Date.now() + deadlineMs;
+	for (;;) {
+		if (session.getActiveToolNames().includes(writeTool)) return true;
+		if (Date.now() > until) return false;
+		await new Promise((resolve) => setTimeout(resolve, 50));
+	}
+}
+
+/**
+ * `.flypi/agent/mcp.json` as the MCP extension wants it. Values stay exactly as written: pi expands
+ * `${NAME}` in `env`/`headers` itself when it builds a transport, and the settings panel writes the
+ * same file (src/agent/config-files.ts), so there is one source of truth for the servers.
+ */
+export function readFlypiMcpConfig(paths: FlypiPaths): LoadedMcpConfig {
+	const servers: McpServerEntry[] = [];
+	const errors: string[] = [];
+	if (!existsSync(paths.mcpPath)) return { servers, errors };
+	let parsed: unknown;
+	try {
+		parsed = JSON.parse(readFileSync(paths.mcpPath, "utf8"));
+	} catch (error) {
+		return { servers, errors: [`${paths.mcpPath}: ${(error as Error).message}`] };
+	}
+	const entries = (parsed as { mcpServers?: Record<string, Record<string, unknown>> } | undefined)?.mcpServers ?? {};
+	for (const [name, config] of Object.entries(entries)) {
+		// The rule pi applies itself (docs/mcp.md): an invalid entry is reported, it does not take the
+		// rest of the file down with it.
+		if (!config || (config.url === undefined && config.command === undefined)) {
+			errors.push(`${paths.mcpPath}: "${name}" needs a url (HTTP) or a command (stdio)`);
+			continue;
+		}
+		servers.push({ name, config: config as unknown as McpServerConfig, source: paths.mcpPath, scope: "global" });
+	}
+	return { servers, errors };
+}
+
+/**
+ * Read `.flypi/agent/settings.json` for the default provider/model and check the credential,
+ * with the same actionable messages probe.ts prints. Imported env keys land in auth.json once.
+ */
+export function resolveConfiguredModel(
+	paths: FlypiPaths,
+	modelRuntime: ModelRuntime,
+	env: Record<string, string | undefined> = process.env,
+): FlypiModel {
+	let settings: Record<string, unknown>;
+	let models: Record<string, unknown>;
+	try {
+		settings = JSON.parse(readFileSync(paths.settingsPath, "utf8")) as Record<string, unknown>;
+		models = JSON.parse(readFileSync(paths.modelsPath, "utf8")) as Record<string, unknown>;
+	} catch (error) {
+		throw new FlypiConfigError(
+			`could not read ${paths.settingsPath} or ${paths.modelsPath}: ${(error as Error).message}`,
+		);
+	}
+	const provider = settings.defaultProvider as string | undefined;
+	const modelId = settings.defaultModel as string | undefined;
+	if (!provider || !modelId) {
+		throw new FlypiConfigError(
+			`${paths.settingsPath} must set defaultProvider and defaultModel (flypi has no built-in default).`,
+		);
+	}
+
+	importEnvCredential(paths.authPath, provider, env);
+	const credential = resolveCredential({
+		provider,
+		models: models as never,
+		auth: readAuth(paths.authPath),
+		env,
+	});
+	if (!credential.ok) throw new FlypiConfigError(credential.message);
+
+	const model = modelRuntime.getModel(provider, modelId);
+	if (!model) {
+		const known = modelRuntime
+			.getModels()
+			.map((m) => `${m.provider}/${m.id}`)
+			.join(", ");
+		throw new FlypiConfigError(
+			`model "${provider}/${modelId}" is not defined in ${paths.modelsPath}. Defined: ${known || "(none)"}`,
+		);
+	}
+	return model;
+}