@oh-my-pi/pi-coding-agent 17.0.6 → 17.0.8

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-coding-agent",
-	"version": "17.0.6",
+	"version": "17.0.8",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -52,17 +52,17 @@
 		"@agentclientprotocol/sdk": "1.2.1",
 		"@babel/parser": "^7.29.7",
 		"@mozilla/readability": "^0.6.0",
-		"@oh-my-pi/hashline": "17.0.6",
-		"@oh-my-pi/omp-stats": "17.0.6",
-		"@oh-my-pi/pi-agent-core": "17.0.6",
-		"@oh-my-pi/pi-ai": "17.0.6",
-		"@oh-my-pi/pi-catalog": "17.0.6",
-		"@oh-my-pi/pi-mnemopi": "17.0.6",
-		"@oh-my-pi/pi-natives": "17.0.6",
-		"@oh-my-pi/pi-tui": "17.0.6",
-		"@oh-my-pi/pi-utils": "17.0.6",
-		"@oh-my-pi/pi-wire": "17.0.6",
-		"@oh-my-pi/snapcompact": "17.0.6",
+		"@oh-my-pi/hashline": "17.0.8",
+		"@oh-my-pi/omp-stats": "17.0.8",
+		"@oh-my-pi/pi-agent-core": "17.0.8",
+		"@oh-my-pi/pi-ai": "17.0.8",
+		"@oh-my-pi/pi-catalog": "17.0.8",
+		"@oh-my-pi/pi-mnemopi": "17.0.8",
+		"@oh-my-pi/pi-natives": "17.0.8",
+		"@oh-my-pi/pi-tui": "17.0.8",
+		"@oh-my-pi/pi-utils": "17.0.8",
+		"@oh-my-pi/pi-wire": "17.0.8",
+		"@oh-my-pi/snapcompact": "17.0.8",
 		"@opentelemetry/api": "^1.9.1",
 		"@opentelemetry/api-logs": "^0.220.0",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
@@ -79,7 +79,6 @@
 		"@xterm/headless": "^6.0.0",
 		"arktype": "2.2.3",
 		"chalk": "^5.6.2",
-		"diff": "^9.0.0",
 		"fast-xml-parser": "^5.9.3",
 		"handlebars": "^4.7.9",
 		"header-generator": "^2.1.82",

package/src/cli/models-cli.ts

@@ -17,8 +17,14 @@ import { formatNumber, getProjectDir } from "@oh-my-pi/pi-utils";
 import chalk from "chalk";
 import { ModelRegistry } from "../config/model-registry";
 import { Settings } from "../config/settings";
-import { discoverAndLoadExtensions, loadExtensions } from "../extensibility/extensions";
+import {
+	discoverAndLoadExtensions,
+	ExtensionRunner,
+	emitSessionShutdownEvent,
+	loadExtensions,
+} from "../extensibility/extensions";
 import { discoverAuthStorage } from "../sdk";
+import { SessionManager } from "../session/session-manager";
 import { EventBus } from "../utils/event-bus";
 
 export type ModelsAction = "ls" | "find" | "refresh";
@@ -303,25 +309,39 @@ export async function runModelsListing(options: RunModelsListingOptions): Promis
 				eventBus,
 				disabledExtensionIds,
 			);
+	const extensionRunner =
+		extensionsResult.extensions.length > 0
+			? new ExtensionRunner(
+					extensionsResult.extensions,
+					extensionsResult.runtime,
+					cwd,
+					SessionManager.inMemory(cwd),
+					modelRegistry,
+				)
+			: undefined;
 
-	for (const { path: extPath, error } of extensionsResult.errors) {
-		process.stderr.write(`Failed to load extension: ${extPath}: ${error}\n`);
-	}
+	try {
+		for (const { path: extPath, error } of extensionsResult.errors) {
+			process.stderr.write(`Failed to load extension: ${extPath}: ${error}\n`);
+		}
 
-	// Mirror sdk.ts: drain pending provider registrations into the registry.
-	const activeSources = extensionsResult.extensions.map(extension => extension.path);
-	modelRegistry.syncExtensionSources(activeSources);
-	for (const sourceId of new Set(activeSources)) {
-		modelRegistry.clearSourceRegistrations(sourceId);
-	}
-	for (const { name, config, sourceId } of extensionsResult.runtime.pendingProviderRegistrations) {
-		modelRegistry.registerProvider(name, config, sourceId);
-	}
-	extensionsResult.runtime.pendingProviderRegistrations = [];
-	// Discover runtime (extension) provider catalogs now that they are registered.
-	await modelRegistry.refreshRuntimeProviders(action === "refresh" ? "online" : "online-if-uncached");
+		// Mirror sdk.ts: drain pending provider registrations into the registry.
+		const activeSources = extensionsResult.extensions.map(extension => extension.path);
+		modelRegistry.syncExtensionSources(activeSources);
+		for (const sourceId of new Set(activeSources)) {
+			modelRegistry.clearSourceRegistrations(sourceId);
+		}
+		for (const { name, config, sourceId } of extensionsResult.runtime.pendingProviderRegistrations) {
+			modelRegistry.registerProvider(name, config, sourceId);
+		}
+		extensionsResult.runtime.pendingProviderRegistrations = [];
+		// Discover runtime (extension) provider catalogs now that they are registered.
+		await modelRegistry.refreshRuntimeProviders(action === "refresh" ? "online" : "online-if-uncached");
 
-	renderProviderModels(modelRegistry, action, pattern, json);
+		renderProviderModels(modelRegistry, action, pattern, json);
+	} finally {
+		await emitSessionShutdownEvent(extensionRunner);
+	}
 }
 
 /**

package/src/cli/update-cli.ts

@@ -556,6 +556,19 @@ async function pruneBunCacheAfterGlobalInstall(): Promise<BunInstallCachePruneRe
 	return await pruneBunInstallCache(cacheDir, packageNames.size === 0 ? undefined : packageNames);
 }
 
+/**
+ * Detect a musl-libc Linux host (Alpine, Void-musl) so self-update replaces a
+ * musl binary with the musl release asset instead of the glibc build, which
+ * would fail to start on the next run. Mirrors the detection in
+ * scripts/install.sh.
+ */
+function isMuslLinux(): boolean {
+	if (process.platform !== "linux") return false;
+	if (fs.existsSync("/etc/alpine-release")) return true;
+	const loaderArch = process.arch === "arm64" ? "aarch64" : "x86_64";
+	return fs.existsSync(`/lib/ld-musl-${loaderArch}.so.1`);
+}
+
 /**
  * Get the appropriate binary name for this platform.
  */
@@ -566,7 +579,7 @@ function getBinaryName(): string {
 	let os: string;
 	switch (platform) {
 		case "linux":
-			os = "linux";
+			os = isMuslLinux() ? "linux-musl" : "linux";
 			break;
 		case "darwin":
 			os = "darwin";

package/src/config/model-discovery.ts

@@ -10,6 +10,7 @@ import type { Api, Model, RemoteCompactionConfig } from "@oh-my-pi/pi-ai/types";
 import { buildModel } from "@oh-my-pi/pi-catalog/build";
 import {
 	getBundledModelReferenceIndex,
+	inheritReferenceThinking,
 	isQwenModelId,
 	resolveModelReference,
 	stripBracketedModelIdAffixes,
@@ -752,7 +753,7 @@ export async function discoverOpenAIModelsList(
 				provider: providerConfig.provider,
 				baseUrl,
 				reasoning: reference?.reasoning ?? false,
-				thinking: reference?.thinking,
+				thinking: inheritReferenceThinking(undefined, reference, providerConfig.provider),
 				input: nativeMetadataForModel?.input ?? reference?.input ?? ["text"],
 				...(providerConfig.discovery.type === "lm-studio" ? { imageInputDecoder: "stb" as const } : {}),
 				// Proxy/gateway pricing is provider-specific and rarely matches
@@ -907,7 +908,7 @@ export async function discoverProxyModels(
 				provider: providerConfig.provider,
 				baseUrl,
 				reasoning: reference?.reasoning ?? false,
-				thinking: reference?.thinking,
+				thinking: inheritReferenceThinking(undefined, reference, providerConfig.provider),
 				input: reference?.input ?? ["text"],
 				// Proxy pricing is provider-specific and usually does not match
 				// upstream bundled catalogs, so keep costs local-unknown even when

package/src/config/model-registry.ts

@@ -23,6 +23,7 @@ import { getBundledModels, getBundledProviders } from "@oh-my-pi/pi-catalog/mode
 import {
 	googleAntigravityModelManagerOptions,
 	googleGeminiCliModelManagerOptions,
+	type OpenAICodexAccount,
 	openaiCodexModelManagerOptions,
 	PROVIDER_DESCRIPTORS,
 } from "@oh-my-pi/pi-catalog/provider-models";
@@ -64,7 +65,11 @@ const BUILT_IN_DISCOVERY_NON_AUTHORITATIVE_RETRY_MS = 5 * 60 * 1000;
 import type { ApiKeyResolver, FetchImpl } from "@oh-my-pi/pi-ai";
 import { registerOAuthProvider, unregisterOAuthProviders } from "@oh-my-pi/pi-ai/oauth";
 import type { OAuthCredentials, OAuthLoginCallbacks } from "@oh-my-pi/pi-ai/oauth/types";
-import { getBundledModelReferenceIndex, resolveModelReference } from "@oh-my-pi/pi-catalog/identity";
+import {
+	getBundledModelReferenceIndex,
+	inheritReferenceThinking,
+	resolveModelReference,
+} from "@oh-my-pi/pi-catalog/identity";
 import { isBunTestRuntime, isRecord, logger, wrapFetchForExtraCa } from "@oh-my-pi/pi-utils";
 import { parseModelString, resolveProviderModelReference } from "../config/model-resolver";
 import type { AuthStorage, OAuthCredential } from "../session/auth-storage";
@@ -153,7 +158,7 @@ interface ProviderOverride {
 export function mergeDiscoveredModel<TApi extends Api>(
 	model: Model<TApi>,
 	existing: Model<Api> | undefined,
-	providerOverride?: Pick<ProviderOverride, "baseUrl" | "headers" | "remoteCompaction" | "transport">,
+	providerOverride?: Pick<ProviderOverride, "baseUrl" | "compat" | "headers" | "remoteCompaction" | "transport">,
 ): Model<TApi> {
 	if (existing) {
 		const supportsTools = model.supportsTools ?? existing.supportsTools;
@@ -167,7 +172,7 @@ export function mergeDiscoveredModel<TApi extends Api>(
 				providerOverride?.remoteCompaction,
 			),
 			...(supportsTools !== undefined ? { supportsTools } : {}),
-			compat: model.compatConfig,
+			compat: mergeCompat(model.compatConfig, providerOverride?.compat),
 		} as ModelSpec<TApi>);
 	}
 	if (providerOverride) {
@@ -180,7 +185,7 @@ export function mergeDiscoveredModel<TApi extends Api>(
 				model.remoteCompaction,
 				providerOverride.remoteCompaction,
 			),
-			compat: model.compatConfig,
+			compat: mergeCompat(model.compatConfig, providerOverride.compat),
 		} as ModelSpec<TApi>);
 	}
 	return model;
@@ -423,20 +428,35 @@ function getOAuthCredentialsForProvider(authStorage: AuthStorage, provider: stri
 	return entries.filter((entry): entry is OAuthCredential => entry.type === "oauth");
 }
 
-function resolveOAuthAccountIdForAccessToken(
+/**
+ * Resolve every configured Codex OAuth account for catalog discovery, refreshing
+ * each credential exactly once. Codex `/models` is account-scoped, so discovery
+ * must fetch per account and union the results; resolving a single access token
+ * (as before) hid models available only through a sibling account (#6265).
+ *
+ * Returns `null` when any stored account fails to resolve (e.g. a transient
+ * refresh failure): the Codex manager is authoritative, so unioning only the
+ * accounts that resolved would cache a partial catalog and hide the failed
+ * account's models for the cache TTL. Aborting keeps the previous/bundled
+ * catalog instead.
+ */
+async function resolveCodexDiscoveryAccounts(
 	authStorage: AuthStorage,
-	provider: string,
-	accessToken: string,
-): string | undefined {
-	const oauthCredentials = getOAuthCredentialsForProvider(authStorage, provider);
-	const matchingCredential = oauthCredentials.find(credential => credential.access === accessToken);
-	if (matchingCredential) {
-		return matchingCredential.accountId;
-	}
-	if (oauthCredentials.length === 1) {
-		return oauthCredentials[0].accountId;
+	resolvedAccessToken: string,
+): Promise<OpenAICodexAccount[] | null> {
+	const accesses = await authStorage.getOAuthAccesses("openai-codex");
+	const accounts: OpenAICodexAccount[] = [];
+	for (const access of accesses) {
+		if (!access.ok) return null;
+		accounts.push({ accessToken: access.accessToken, accountId: access.accountId });
+	}
+	if (!accounts.some(account => account.accessToken === resolvedAccessToken)) {
+		const matchingCredential = getOAuthCredentialsForProvider(authStorage, "openai-codex").find(
+			credential => credential.access === resolvedAccessToken,
+		);
+		accounts.push({ accessToken: resolvedAccessToken, accountId: matchingCredential?.accountId });
 	}
-	return undefined;
+	return accounts;
 }
 
 function mergeCompat<TBase extends object, TOverride extends object>(
@@ -667,7 +687,7 @@ function finalizeCustomModel(model: CustomModelOverlay, options: CustomModelBuil
 		provider: resolvedModel.provider,
 		baseUrl: resolvedModel.baseUrl,
 		reasoning: resolvedModel.reasoning ?? reference?.reasoning ?? (options.useDefaults ? false : undefined),
-		thinking: resolvedModel.thinking ?? reference?.thinking,
+		thinking: inheritReferenceThinking(resolvedModel.thinking, reference, resolvedModel.provider),
 		input: input as ("text" | "image")[],
 		...(supportsTools !== undefined ? { supportsTools } : {}),
 		cost,
@@ -1139,8 +1159,18 @@ export class ModelRegistry {
 					models.push(spec);
 					continue;
 				}
-				if (unrestorableHeaderIds.has(spec.id)) continue;
-				const bundledHeaders = bundledById?.get(spec.id)?.headers;
+				// Current unrestorable markers prove that neither same-id nor
+				// request-model bundled headers matched the live model. Only markers
+				// from the old id-only writer may recover through `requestModelId`.
+				const unrestorable = unrestorableHeaderIds.has(spec.id);
+				const bundledHeaders = (
+					unrestorable
+						? cache.legacyHeaderRestoreMarkers && spec.requestModelId
+							? bundledById?.get(spec.requestModelId)
+							: undefined
+						: (bundledById?.get(spec.id) ??
+							(spec.requestModelId ? bundledById?.get(spec.requestModelId) : undefined))
+				)?.headers;
 				if (!bundledHeaders) continue;
 				models.push({ ...spec, headers: bundledHeaders });
 			}
@@ -1719,14 +1749,11 @@ export class ModelRegistry {
 				providerId: "openai-codex",
 				authoritative: true,
 				resolveKey: value => value,
-				createOptions: accessToken => {
-					const accountId = resolveOAuthAccountIdForAccessToken(this.authStorage, "openai-codex", accessToken);
-					return openaiCodexModelManagerOptions({
-						accessToken,
-						accountId,
+				createOptions: accessToken =>
+					openaiCodexModelManagerOptions({
+						resolveAccounts: () => resolveCodexDiscoveryAccounts(this.authStorage, accessToken),
 						fetch: this.#fetch,
-					});
-				},
+					}),
 			},
 		];
 		const disabledProviders = getDisabledProviderIdsFromSettings();

package/src/config/model-resolver.ts

@@ -1598,11 +1598,33 @@ export function filterAvailableModelsByEnabledPatterns(
 
 	return includeSyntheticAllowedModels(available, allowedModels);
 }
+function findExactCliModel(
+	selector: string,
+	allModels: Model<Api>[],
+	availableModels: Model<Api>[],
+): Model<Api> | undefined {
+	// Explicit provider/id references stay authoritative against the full catalog.
+	const referenced = findExactModelReferenceMatch(selector, allModels);
+	if (referenced) return referenced;
+
+	// Flat-id (or full-selector-string) matches prefer authenticated providers,
+	// then fall back to catalog order. This covers aggregator-style flat ids
+	// that merely look provider-qualified (e.g. "openai/gpt-oss-120b" hosted on
+	// OpenRouter), where the provider/id decomposition above found nothing.
+	const lower = selector.toLowerCase();
+	const isFlatMatch = (model: Model<Api>) =>
+		model.id.toLowerCase() === lower || formatModelString(model).toLowerCase() === lower;
+	const preferred = availableModels.find(isFlatMatch);
+	if (preferred) return preferred;
+	return availableModels === allModels ? undefined : allModels.find(isFlatMatch);
+}
 
 export interface ResolveCliModelResult {
 	model: Model<Api> | undefined;
-	/** configuredPatterns is the full configured fallback chain when the selector resolves through a role. */
+	/** configuredPatterns contains the role's ordered primary candidates. */
 	configuredPatterns?: string[];
+	/** configuredRole identifies the role expanded into configuredPatterns. */
+	configuredRole?: string;
 	/** configuredPatternIndex identifies the configured role pattern that matched an available model. */
 	configuredPatternIndex?: number;
 	selector?: string;
@@ -1620,17 +1642,19 @@ export function resolveCliModel(options: {
 	cliProvider?: string;
 	cliModel?: string;
 	modelRegistry: CliModelRegistry;
+	/** Authenticated models to prefer for unqualified selectors; omit to preserve catalog-order behavior. */
+	availableModels?: Model<Api>[];
 	settings?: Settings;
 	preferences?: ModelMatchPreferences;
 }): ResolveCliModelResult {
-	const { cliProvider, cliModel, modelRegistry, settings, preferences } = options;
+	const { cliProvider, cliModel, modelRegistry, settings, preferences, availableModels: preferredModels } = options;
 
 	if (!cliModel) {
 		return { model: undefined, selector: undefined, warning: undefined, error: undefined };
 	}
 
-	const availableModels = modelRegistry.getAll();
-	if (availableModels.length === 0) {
+	const allModels = modelRegistry.getAll();
+	if (allModels.length === 0) {
 		return {
 			model: undefined,
 			selector: undefined,
@@ -1639,8 +1663,9 @@ export function resolveCliModel(options: {
 		};
 	}
 
+	const availableModels = preferredModels ?? allModels;
 	const providerMap = new Map<string, string>();
-	for (const model of availableModels) {
+	for (const model of allModels) {
 		providerMap.set(model.provider.toLowerCase(), model.provider);
 	}
 
@@ -1656,19 +1681,7 @@ export function resolveCliModel(options: {
 
 	const trimmedModel = cliModel.trim();
 	if (!provider) {
-		const lower = trimmedModel.toLowerCase();
-		// When input has provider/id format (e.g. "zai/glm-5"), prefer decomposed
-		// provider+id match over flat id match. Without this, a model with id
-		// "zai/glm-5" on provider "vercel-ai-gateway" wins over provider "zai"
-		// with id "glm-5", because Array.find returns the first catalog hit.
-		let exact = findExactModelReferenceMatch(trimmedModel, availableModels);
-		if (!exact) {
-			// Flat exact id (or full selector) by catalog order: CLI resolution
-			// stays deterministic across runs regardless of usage-based ranking.
-			exact = availableModels.find(
-				model => model.id.toLowerCase() === lower || `${model.provider}/${model.id}`.toLowerCase() === lower,
-			);
-		}
+		const exact = findExactCliModel(trimmedModel, allModels, availableModels);
 		if (exact) {
 			return {
 				model: exact,
@@ -1684,15 +1697,7 @@ export function resolveCliModel(options: {
 			MAX_THINKING_SUFFIX_OPTIONS,
 		);
 		if (exactThinkingLevel) {
-			let exactSuffixed = findExactModelReferenceMatch(exactBase, availableModels);
-			if (!exactSuffixed) {
-				const lowerExactBase = exactBase.toLowerCase();
-				exactSuffixed = availableModels.find(
-					model =>
-						model.id.toLowerCase() === lowerExactBase ||
-						`${model.provider}/${model.id}`.toLowerCase() === lowerExactBase,
-				);
-			}
+			const exactSuffixed = findExactCliModel(exactBase, allModels, availableModels);
 			if (exactSuffixed) {
 				return {
 					model: exactSuffixed,
@@ -1718,15 +1723,28 @@ export function resolveCliModel(options: {
 					? `${formatModelRoleAlias(bareRoleName)}${bareRoleThinkingLevel ? `:${bareRoleThinkingLevel}` : ""}`
 					: undefined;
 		if (roleSelector) {
+			const { base: roleAlias } = splitThinkingSuffix(
+				roleSelector,
+				modelRoleAliasPrefixLength(roleSelector) ?? -1,
+				MAX_THINKING_SUFFIX_OPTIONS,
+			);
+			const configuredRole = getModelRoleAlias(roleAlias, settings);
 			configuredPatterns = resolveConfiguredModelPatterns([roleSelector], settings);
-			const resolved = resolveModelRoleValue(roleSelector, availableModels, {
+			const availableResolved = resolveModelRoleValue(roleSelector, availableModels, {
 				settings,
 				matchPreferences: preferences,
 			});
+			const resolved = availableResolved.model
+				? availableResolved
+				: resolveModelRoleValue(roleSelector, allModels, {
+						settings,
+						matchPreferences: preferences,
+					});
 			if (resolved.model) {
 				return {
 					model: resolved.model,
 					selector: formatModelString(resolved.model),
+					configuredRole,
 					configuredPatterns,
 					configuredPatternIndex: resolved.matchedPatternIndex,
 					thinkingLevel: resolved.thinkingLevel,
@@ -1738,6 +1756,7 @@ export function resolveCliModel(options: {
 				return {
 					model: undefined,
 					configuredPatterns,
+					configuredRole,
 					selector: undefined,
 					thinkingLevel: undefined,
 					warning: resolved.warning,
@@ -1767,7 +1786,7 @@ export function resolveCliModel(options: {
 	}
 
 	if (provider) {
-		const exactProviderMatch = resolveProviderModelReference(provider, pattern, availableModels);
+		const exactProviderMatch = resolveProviderModelReference(provider, pattern, allModels);
 		if (exactProviderMatch) {
 			return {
 				model: exactProviderMatch,
@@ -1779,10 +1798,16 @@ export function resolveCliModel(options: {
 		}
 	}
 
-	const candidates = provider ? availableModels.filter(model => model.provider === provider) : availableModels;
-	const { model, thinkingLevel, warning, upstream } = parseModelPattern(pattern, candidates, preferences, {
+	const candidates = provider ? allModels.filter(model => model.provider === provider) : availableModels;
+	let parsed = parseModelPattern(pattern, candidates, preferences, {
 		allowInvalidThinkingSelectorFallback: false,
 	});
+	if (!parsed.model && !provider) {
+		parsed = parseModelPattern(pattern, allModels, preferences, {
+			allowInvalidThinkingSelectorFallback: false,
+		});
+	}
+	const { model, thinkingLevel, warning, upstream } = parsed;
 
 	if (!model) {
 		const display = provider ? `${provider}/${pattern}` : cliModel;

package/src/config/settings-schema.ts

@@ -2861,6 +2861,11 @@ export const SETTINGS_SCHEMA = {
 
 	"hindsight.debug": { type: "boolean", default: false },
 
+	"hindsight.requestTimeoutMs": { type: "number", default: 30_000 },
+	"hindsight.reflectTimeoutMs": { type: "number", default: 120_000 },
+	"hindsight.recallTimeoutMs": { type: "number", default: 30_000 },
+	"hindsight.retainTimeoutMs": { type: "number", default: 60_000 },
+
 	"hindsight.mentalModelsEnabled": {
 		type: "boolean",
 		default: true,

package/src/config/settings.ts

@@ -330,6 +330,8 @@ export class Settings {
 	#modified = new Set<string>();
 	/** Individual project model roles modified during this session */
 	#modifiedProjectModelRoles = new Set<string>();
+	/** Individual global model roles modified during this session (for partial save) */
+	#modifiedGlobalModelRoles = new Set<string>();
 	/**
 	 * Original process-wide model-role overrides captured before a project edit
 	 * temporarily replaced them via `#updateRuntimeModelRoleOverride`. Restored
@@ -551,7 +553,7 @@ export class Settings {
 		if (this.#projectSavePromise) {
 			await this.#projectSavePromise;
 		}
-		if (this.#modified.size > 0) {
+		if (this.#modified.size > 0 || this.#modifiedGlobalModelRoles.size > 0) {
 			await this.#saveNow();
 		}
 		if (this.#modifiedProjectModelRoles.size > 0) {
@@ -833,13 +835,22 @@ export class Settings {
 	 * stale skip in place.
 	 */
 	setModelRole(role: ModelRole | string, modelId: string | undefined): void {
+		const prev = this.get("modelRoles");
 		const current = this.#modelRolesFromLayer(this.#global);
 		if (modelId === undefined) {
 			delete current[role];
 		} else {
 			current[role] = modelId;
 		}
-		this.set("modelRoles", current);
+		// Persist per-role rather than marking the whole `modelRoles` path
+		// modified: #saveNow merges only the changed role into the re-read
+		// file, so a concurrent external edit to a sibling role is not
+		// clobbered by this process's stale in-memory snapshot.
+		setByPath(this.#global, ["modelRoles"], current);
+		this.#modifiedGlobalModelRoles.add(role);
+		this.#rebuildMerged();
+		this.#queueSave();
+		this.#fireEffectiveSettingChanged("modelRoles", this.get("modelRoles"), prev);
 		if (this.isProjectModelRoleRuntimeOverrideActive(role)) {
 			return;
 		}
@@ -1619,7 +1630,8 @@ export class Settings {
 		clearTimeout(this.#saveTimer);
 		this.#saveTimer = setTimeout(() => {
 			this.#saveTimer = undefined;
-			const savePromise = this.#saveNow();
+			const previousSave = this.#savePromise;
+			const savePromise = previousSave ? previousSave.then(() => this.#saveNow()) : this.#saveNow();
 			this.#savePromise = savePromise;
 			savePromise
 				.catch(err => {
@@ -1634,27 +1646,75 @@ export class Settings {
 	}
 
 	async #saveNow(): Promise<void> {
-		if (!this.#persist || !this.#configPath || this.#modified.size === 0) return;
+		if (!this.#persist || !this.#configPath) return;
+		if (this.#modified.size === 0 && this.#modifiedGlobalModelRoles.size === 0) return;
 
 		const configPath = this.#configPath;
 		const modifiedPaths = [...this.#modified];
+		const modifiedModelRoles = [...this.#modifiedGlobalModelRoles];
+		const globalRolesAtStart = this.#modelRolesFromLayer(this.#global);
 		this.#modified.clear();
+		this.#modifiedGlobalModelRoles.clear();
 
 		try {
 			await withFileLock(configPath, async () => {
 				// Re-read to preserve external changes
 				const current = await this.#loadYaml(configPath);
 
-				// Apply only our modified paths
+				// Apply only our modified whole-value paths
 				for (const modPath of modifiedPaths) {
 					const segments = modPath.split(".");
 					const value = getByPath(this.#global, segments);
 					setByPath(current, segments, value);
 				}
 
+				// Merge only the model roles captured by this save. Then retain
+				// any role changed while the async read/lock was pending before
+				// replacing #global, so the follow-up save still sees its value.
+				const latestGlobalRoles = this.#modelRolesFromLayer(this.#global);
+				const rolesToPreserve = new Set(this.#modifiedGlobalModelRoles);
+				for (const role in globalRolesAtStart) {
+					if (globalRolesAtStart[role] !== latestGlobalRoles[role]) {
+						rolesToPreserve.add(role);
+					}
+				}
+				for (const role in latestGlobalRoles) {
+					if (globalRolesAtStart[role] !== latestGlobalRoles[role]) {
+						rolesToPreserve.add(role);
+					}
+				}
+				if (modifiedModelRoles.length > 0 || rolesToPreserve.size > 0) {
+					const currentRoles = getByPath(current, ["modelRoles"]);
+					const mergedRoles: Record<string, unknown> = isRecord(currentRoles) ? { ...currentRoles } : {};
+					for (const role of modifiedModelRoles) {
+						if (Object.hasOwn(globalRolesAtStart, role)) {
+							mergedRoles[role] = globalRolesAtStart[role];
+						} else {
+							delete mergedRoles[role];
+						}
+					}
+					for (const role of rolesToPreserve) {
+						if (Object.hasOwn(latestGlobalRoles, role)) {
+							mergedRoles[role] = latestGlobalRoles[role];
+						} else {
+							delete mergedRoles[role];
+						}
+					}
+					setByPath(current, ["modelRoles"], mergedRoles);
+				}
+
 				// Update our global with any external changes we preserved
 				this.#global = current;
 				await Bun.write(configPath, YAML.stringify(this.#global, null, 2));
+				// These pending roles were included in this write. Remove each
+				// only if no newer local change arrived while Bun.write was in
+				// flight; a newer value still needs the queued follow-up save.
+				const globalRolesAfterWrite = this.#modelRolesFromLayer(this.#global);
+				for (const role of rolesToPreserve) {
+					if (latestGlobalRoles[role] === globalRolesAfterWrite[role]) {
+						this.#modifiedGlobalModelRoles.delete(role);
+					}
+				}
 			});
 		} catch (error) {
 			logger.warn("Settings: save failed", { error: String(error) });
@@ -1662,6 +1722,9 @@ export class Settings {
 			for (const p of modifiedPaths) {
 				this.#modified.add(p);
 			}
+			for (const role of modifiedModelRoles) {
+				this.#modifiedGlobalModelRoles.add(role);
+			}
 		}
 
 		this.#rebuildMerged();