dsh-subscriptions 0.2.0 → 0.2.3

package/lib/http.js

@@ -1,5 +1,5 @@
 /** Host-managed network routing for subscription requests. */
-import { getDefaultAutoSelectFamilyAttemptTimeout, setDefaultAutoSelectFamilyAttemptTimeout } from 'node:net';
+import { getDefaultAutoSelectFamilyAttemptTimeout, setDefaultAutoSelectFamilyAttemptTimeout, } from 'node:net';
 /** Keep a sufficient per-address connection budget on high-latency networks. */
 export const MIN_CONNECT_ATTEMPT_TIMEOUT_MS = 1500;
 export function ensureConnectAttemptTimeout(minMs = MIN_CONNECT_ATTEMPT_TIMEOUT_MS) {

package/lib/index.d.ts

@@ -13,13 +13,13 @@ import type { ModelEntry } from './providers/common.js';
 import type { RateLimitConfig } from './providers/rate-limit.js';
 import type { PromptCacheTtl } from './translate/anthropic.js';
 import type { CliVersion, NpmCliVersionCache } from './providers/npm-cli-version.js';
-export type { AccountPreferences, ProviderPreferences } from './provider-settings.js';
+export type { AccountPreferences, ProviderPreferences, } from './provider-settings.js';
 import type { PoolMemberRef } from './providers/pool-family.js';
 import type { AntigravityRuntimeConfig } from './providers/antigravity.js';
-export type { ModelEntry, ProviderUsage, UsageWindow } from './providers/common.js';
+export type { ModelEntry, ProviderUsage, UsageWindow, } from './providers/common.js';
 export type { RateLimitConfig, RateLimitWait } from './providers/rate-limit.js';
 export type { ProviderStatus } from './auth/rpc.js';
-export type { AntigravitySession, ClaudeSession, CodexSession, CopilotSession, GrokSession, ProviderId } from './auth/store.js';
+export type { AntigravitySession, ClaudeSession, CodexSession, CopilotSession, GrokSession, ProviderId, } from './auth/store.js';
 export declare const name = "dsh-subscriptions";
 export declare const inject: string[];
 /** Default maximum provider idle time while one stream read is outstanding. */

package/lib/index.js

@@ -1205,7 +1205,7 @@ var DeviceFlowManager = class {
 		const response = await fetchFn(spec.deviceCodeUrl, {
 			method: "POST",
 			headers: {
-				"accept": "application/json",
+				accept: "application/json",
 				"content-type": "application/x-www-form-urlencoded"
 			},
 			body: new URLSearchParams({
@@ -1244,7 +1244,7 @@ var DeviceFlowManager = class {
 				const pollResponse = await fetchFn(spec.tokenUrl, {
 					method: "POST",
 					headers: {
-						"accept": "application/json",
+						accept: "application/json",
 						"content-type": "application/x-www-form-urlencoded"
 					},
 					body: new URLSearchParams({
@@ -3064,8 +3064,8 @@ function withoutImages(messages) {
 }
 function discoveryHeaders(apiKey) {
 	return {
-		"accept": "application/json",
-		...apiKey === void 0 ? {} : { "authorization": `Bearer ${apiKey}` },
+		accept: "application/json",
+		...apiKey === void 0 ? {} : { authorization: `Bearer ${apiKey}` },
 		...attributionHeaders()
 	};
 }
@@ -3340,8 +3340,8 @@ var OllamaAdapter = class extends LlmAdapter {
 			const response = await (this.options.fetchFn ?? hostFetch)(`${chatBase}/chat/completions`, {
 				method: "POST",
 				headers: {
-					"authorization": `Bearer ${apiKey}`,
-					"accept": "text/event-stream",
+					authorization: `Bearer ${apiKey}`,
+					accept: "text/event-stream",
 					"content-type": "application/json",
 					...attributionHeaders()
 				},
@@ -3377,7 +3377,7 @@ function valueOf(value) {
 const DEFAULT_REFS = {
 	"opencode-go": "OPENCODE_GO_API_KEY",
 	"kimi-code": "KIMI_CODING_API_KEY",
-	"minimax": "MINIMAX_API_KEY",
+	minimax: "MINIMAX_API_KEY",
 	"minimax-cn": "MINIMAX_CN_API_KEY",
 	"ollama-cloud": DEFAULT_OLLAMA_API_KEY_REF
 };
@@ -3834,7 +3834,13 @@ async function dispatch(controller, speed, modelDefaults, endpoint, payload, sig
 				accounts: [],
 				detail: error instanceof Error ? error.message : String(error)
 			}))]));
-			return ok({ providers: Object.fromEntries(entries) });
+			return ok({ providers: Object.fromEntries(entries.map(([provider, status]) => {
+				const visibleModels = providerSettings?.visibleModels?.(provider);
+				return [provider, {
+					...status,
+					...visibleModels === void 0 ? {} : { visibleModels }
+				}];
+			})) });
 		}
 		case "login": {
 			const provider = readProvider(payload);
@@ -3931,7 +3937,7 @@ async function dispatch(controller, speed, modelDefaults, endpoint, payload, sig
 			return ok({ ok: true });
 		case "image": return ok(await controller.readImage(readImageRef(payload), signal));
 		case "video": return ok(await controller.readVideo(readVideoName(payload), signal));
-		case "speed": return ok(await speed.speed(readSessionId(payload)));
+		case "speed": return ok(await speed.speed(readSessionId(payload), payload.model === void 0 ? void 0 : readString(payload, "model")));
 		case "setSpeed":
 			await speed.setSpeed(readSessionId(payload), readSpeedTier(payload));
 			return ok({ ok: true });
@@ -6950,8 +6956,8 @@ async function postJson(url, apiKey, payload, options, signal) {
 				method: "POST",
 				headers: {
 					"content-type": "application/json",
-					"accept": "application/json",
-					"authorization": `Bearer ${apiKey}`,
+					accept: "application/json",
+					authorization: `Bearer ${apiKey}`,
 					...attributionHeaders()
 				},
 				body: JSON.stringify(payload),
@@ -8008,13 +8014,39 @@ async function readCatalogFile(path) {
 		return {};
 	}
 }
+/**
+* Windows refuses to replace a file another handle holds open without
+* `FILE_SHARE_DELETE` (an indexer, antivirus scan, or a second dsh profile
+* reading the cache), surfacing as a transient `EPERM`/`EBUSY`/`EACCES` on
+* `rename` (V1ki/dsh-plugin-subscriptions#129). On Windows only, retry
+* briefly before giving the write up: on POSIX those codes are real
+* failures, so retrying would only delay surfacing them.
+*/
+const RENAME_RETRY_CODES = new Set([
+	"EPERM",
+	"EBUSY",
+	"EACCES"
+]);
+const RENAME_RETRIES = 5;
+const RENAME_RETRY_DELAY_MS = 20;
+/** `rename` with a short bounded retry on Windows share-mode conflicts. */
+async function renameWithRetry(from, to) {
+	for (let attempt = 0;; attempt++) try {
+		await rename(from, to);
+		return;
+	} catch (error) {
+		const code = error.code;
+		if (process.platform !== "win32" || attempt >= RENAME_RETRIES || code === void 0 || !RENAME_RETRY_CODES.has(code)) throw error;
+		await new Promise((resolve$2) => setTimeout(resolve$2, RENAME_RETRY_DELAY_MS * (attempt + 1)));
+	}
+}
 /** Persist the whole file atomically (tmp file + rename). */
 async function writeCatalogFile(store, path) {
 	await mkdir(dirname(path), { recursive: true });
 	const tmp = `${path}.tmp-${process.pid}-${Math.random().toString(36).slice(2)}`;
 	try {
 		await writeFile(tmp, JSON.stringify(store, null, 2));
-		await rename(tmp, path);
+		await renameWithRetry(tmp, path);
 	} catch (error) {
 		await rm(tmp, { force: true });
 		throw error;
@@ -8550,10 +8582,10 @@ async function fetchClaudeUsage(session, fetchFn = hostFetch, signal, cliVersion
 	const userAgent = claudeCliUserAgent(await cliVersion());
 	const init = {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"anthropic-beta": "oauth-2025-04-20",
 			"user-agent": userAgent,
-			"accept": "application/json"
+			accept: "application/json"
 		},
 		...signal === void 0 ? {} : { signal }
 	};
@@ -8625,11 +8657,11 @@ async function fetchClaudeModels(session, fetchFn = hostFetch, signal, cliVersio
 	const userAgent = claudeCliUserAgent(await cliVersion());
 	const response = await fetchFn(CLAUDE_MODELS_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"anthropic-version": "2023-06-01",
 			"user-agent": userAgent,
 			"anthropic-dangerous-direct-browser-access": "true",
-			"accept": "application/json"
+			accept: "application/json"
 		},
 		...signal === void 0 ? {} : { signal }
 	});
@@ -8697,8 +8729,9 @@ const CLAUDE_MODALITIES = ["text", "image"];
 * @param effort - the reasoning effort, when the model advertises efforts.
 * @param cacheTtl - lifetime of every cache breakpoint; defaults to five minutes.
 * Requests that carry a `purpose` (compaction, session titles) always use five
-* minutes to preserve their existing auxiliary-request behavior. A five-minute
-* mark can still read an entry a one-hour mark wrote.
+* minutes to preserve their existing auxiliary-request behavior — the override
+* changes the TTL for ordinary turns only, never the auxiliary floor. A
+* five-minute mark can still read an entry a one-hour mark wrote.
 * @returns the JSON body to POST.
 */
 function claudeRequestBody(options, messages, maxTokens, thinking, effort, cacheTtl) {
@@ -8913,18 +8946,18 @@ var ClaudeAdapter = class extends LlmAdapter {
 		}
 		const disc = await this.discovered(options.model);
 		const maxTokens = options.maxTokens ?? claudeMaxTokens(this.options.models.find((entry) => entry.id === options.model), disc);
-		const body = claudeRequestBody(options, messages, maxTokens, this.thinkingParam(disc?.thinkingType, maxTokens), options.reasoningEffort !== void 0 && disc?.reasoning !== void 0 ? String(options.reasoningEffort) : void 0, this.options.promptCacheTtl);
+		const body = claudeRequestBody(options, messages, maxTokens, this.thinkingParam(disc?.thinkingType, maxTokens), options.reasoningEffort !== void 0 && disc?.reasoning !== void 0 ? String(options.reasoningEffort) : void 0, this.options.promptCacheTtlOverride?.() ?? this.options.promptCacheTtl);
 		const cliVersion = await (this.options.resolveCliVersion ?? localClaudeCliVersion)();
 		return hostFetch(CLAUDE_API_URL, {
 			method: "POST",
 			headers: {
-				"authorization": `Bearer ${session.accessToken}`,
+				authorization: `Bearer ${session.accessToken}`,
 				"anthropic-version": "2023-06-01",
 				"anthropic-beta": CLAUDE_BETA_FLAGS,
 				"user-agent": claudeCliUserAgent(cliVersion),
 				"x-app": "cli",
 				"anthropic-dangerous-direct-browser-access": "true",
-				"accept": "text/event-stream",
+				accept: "text/event-stream",
 				"content-type": "application/json"
 			},
 			body: JSON.stringify(body),
@@ -9653,10 +9686,10 @@ function codexUsageWindow(value, fallbackKind) {
 async function fetchCodexUsage(session, fetchFn = hostFetch, signal) {
 	const response = await fetchFn(CODEX_USAGE_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"chatgpt-account-id": session.accountId,
-			"originator": "codex_cli_rs",
-			"accept": "application/json",
+			originator: "codex_cli_rs",
+			accept: "application/json",
 			...attributionHeaders()
 		},
 		...signal === void 0 ? {} : { signal }
@@ -9707,10 +9740,10 @@ async function fetchCodexModels(session, fetchFn = hostFetch, signal, clientVers
 	const version = codexClientVersion(clientVersion);
 	const response = await fetchFn(`${CODEX_MODELS_URL}?client_version=${encodeURIComponent(version)}`, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"chatgpt-account-id": session.accountId,
-			"originator": "codex_cli_rs",
-			"accept": "application/json",
+			originator: "codex_cli_rs",
+			accept: "application/json",
 			...attributionHeaders()
 		},
 		...signal === void 0 ? {} : { signal }
@@ -10101,11 +10134,11 @@ var CodexAdapter = class extends LlmAdapter {
 		return hostFetch(CODEX_API_URL, {
 			method: "POST",
 			headers: {
-				"authorization": `Bearer ${session.accessToken}`,
+				authorization: `Bearer ${session.accessToken}`,
 				"chatgpt-account-id": session.accountId,
-				"originator": "codex_cli_rs",
+				originator: "codex_cli_rs",
 				"session-id": deterministicSessionId(options.sessionId),
-				"accept": "text/event-stream",
+				accept: "text/event-stream",
 				"content-type": "application/json",
 				...attributionHeaders()
 			},
@@ -10185,10 +10218,10 @@ var CodexWebSearchProvider = class {
 				response = await fetchFn(CODEX_SEARCH_URL, {
 					method: "POST",
 					headers: {
-						"authorization": `Bearer ${session.accessToken}`,
+						authorization: `Bearer ${session.accessToken}`,
 						"chatgpt-account-id": session.accountId,
 						"content-type": "application/json",
-						"originator": "codex_cli_rs",
+						originator: "codex_cli_rs",
 						...attributionHeaders()
 					},
 					body: JSON.stringify(body),
@@ -10769,7 +10802,7 @@ var PoolAdapter = class extends LlmAdapter {
 		if (definition === void 0) throw new LlmError(`unknown pool model "${options.model}"`, "NO_ADAPTER");
 		const members = await this.concrete(definition.members);
 		const candidates = await this.select(options.model, members, options.sessionId);
-		if (candidates.length === 0) throw this.exhausted(options.model, members);
+		if (candidates.length === 0) throw this.exhaustedWithWarn(options.model, members);
 		const failures = [];
 		for (const member of candidates) {
 			const adapter = this.options.adapters[member.provider];
@@ -10804,7 +10837,7 @@ var PoolAdapter = class extends LlmAdapter {
 			}
 			return;
 		}
-		throw this.exhausted(options.model, members, failures);
+		throw this.exhaustedWithWarn(options.model, members, failures);
 	}
 	/**
 	* Order the candidates for one request. Health filters both strategies;
@@ -10845,6 +10878,18 @@ var PoolAdapter = class extends LlmAdapter {
 		}
 		this.sticky.set(key, memberKey(member.provider, member.account, member.model));
 	}
+	/**
+	* Build the terminal exhaustion error and log it. Per-member failovers are
+	* already warned above; without this line dsh.log shows the failover chain
+	* but not the exhaustion that ends it. The message carries only the
+	* LlmError's own sanitized text and code — raw provider payloads never
+	* reach here (see the note in {@link exhausted}).
+	*/
+	exhaustedWithWarn(model, pool, failures = []) {
+		const error = this.exhausted(model, pool, failures);
+		this.options.onWarn(`pool "${model}": exhausted (${error.code}: ${error.message})`);
+		return error;
+	}
 	/** Report attempt failures or this pool's active cooldowns, never unrelated health records. */
 	exhausted(model, pool, failures = []) {
 		if (pool.length === 0) return new LlmError(`pool "${model}" has no usable member: no account could be resolved`, "NO_ADAPTER");
@@ -11647,9 +11692,9 @@ function grokResetsAt(value) {
 async function fetchGrokUsage(session, fetchFn = hostFetch, signal) {
 	const response = await fetchFn(GROK_BILLING_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"x-xai-token-auth": "xai-grok-cli",
-			"accept": "application/json",
+			accept: "application/json",
 			...attributionHeaders()
 		},
 		...signal === void 0 ? {} : { signal }
@@ -11726,9 +11771,9 @@ function grokCliReasoning(entry) {
 async function fetchGrokCliCatalog(session, fetchFn = hostFetch, signal) {
 	const response = await fetchFn(GROK_CLI_MODELS_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
+			authorization: `Bearer ${session.accessToken}`,
 			"x-xai-token-auth": "xai-grok-cli",
-			"accept": "application/json",
+			accept: "application/json",
 			...attributionHeaders()
 		},
 		...signal === void 0 ? {} : { signal }
@@ -11792,8 +11837,8 @@ async function fetchGrokModels(session, fetchFn = hostFetch, onWarn, previous, s
 	const previousById = previous === void 0 || previous.length === 0 ? void 0 : new Map(previous.map((model) => [model.id, model]));
 	const [response, cliCatalog] = await Promise.all([fetchFn(GROK_MODELS_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
-			"accept": "application/json",
+			authorization: `Bearer ${session.accessToken}`,
+			accept: "application/json",
 			...attributionHeaders()
 		},
 		...signal === void 0 ? {} : { signal }
@@ -12017,8 +12062,8 @@ var GrokAdapter = class extends LlmAdapter {
 		return hostFetch(GROK_API_URL, {
 			method: "POST",
 			headers: {
-				"authorization": `Bearer ${session.accessToken}`,
-				"accept": "text/event-stream",
+				authorization: `Bearer ${session.accessToken}`,
+				accept: "text/event-stream",
 				"content-type": "application/json",
 				...attributionHeaders()
 			},
@@ -12131,8 +12176,8 @@ function copilotHeaders(hasVision = false, vscodeVersion = FALLBACK_VSCODE_VERSI
 */
 async function exchangeCopilotToken(githubToken, fetchFn = hostFetch) {
 	const response = await fetchFn(COPILOT_TOKEN_URL, { headers: {
-		"authorization": `Bearer ${githubToken}`,
-		"accept": "application/json",
+		authorization: `Bearer ${githubToken}`,
+		accept: "application/json",
 		...copilotHeaders(false, await latestVsCodeVersion(fetchFn))
 	} });
 	if (!response.ok) throw await oauthEndpointError(response, "copilot");
@@ -12155,8 +12200,8 @@ async function completeCopilotLogin(githubToken, fetchFn = hostFetch) {
 	let account;
 	try {
 		const response = await fetchFn(GITHUB_USER_URL, { headers: {
-			"authorization": `Bearer ${githubToken}`,
-			"accept": "application/json",
+			authorization: `Bearer ${githubToken}`,
+			accept: "application/json",
 			"user-agent": "GitHubCopilotChat/0.35.0"
 		} });
 		if (response.ok) {
@@ -12240,8 +12285,8 @@ function copilotReasoning(entry) {
 async function fetchCopilotModels(session, fetchFn = hostFetch, signal) {
 	const response = await fetchFn(COPILOT_MODELS_URL, {
 		headers: {
-			"authorization": `Bearer ${session.accessToken}`,
-			"accept": "application/json",
+			authorization: `Bearer ${session.accessToken}`,
+			accept: "application/json",
 			...copilotHeaders(false, await latestVsCodeVersion(fetchFn))
 		},
 		...signal === void 0 ? {} : { signal }
@@ -12757,8 +12802,8 @@ var CopilotAdapter = class CopilotAdapter extends LlmAdapter {
 		return hostFetch(wire === "responses" ? COPILOT_RESPONSES_URL : COPILOT_API_URL, {
 			method: "POST",
 			headers: {
-				"authorization": `Bearer ${session.accessToken}`,
-				"accept": "text/event-stream",
+				authorization: `Bearer ${session.accessToken}`,
+				accept: "text/event-stream",
 				"content-type": "application/json",
 				...copilotHeaders(hasVision, await latestVsCodeVersion(this.options.fetchFn ?? hostFetch))
 			},
@@ -12999,9 +13044,17 @@ function toAntigravityTools(tools, model = "gemini") {
 	})) }];
 }
 const SKIP_THOUGHT_SIGNATURE_VALIDATOR = "skip_thought_signature_validator";
-/** Gemini 3 validates the first function call in every model tool-call step. */
+/**
+* Whether this model validates thought signatures on replayed function calls.
+*
+* Gemini 3 and 4 require a signature on the first function call of each step;
+* foreign history uses the documented placeholder when no signature can be reused.
+* https://ai.google.dev/gemini-api/docs/generate-content/thought-signatures
+* Keep this scoped to these model families; future versions and unversioned aliases
+* need verification before enabling the same fallback.
+*/
 function requiresThoughtSignatures(model) {
-	return model !== void 0 && /^gemini-3(?:[.-]|$)/i.test(model);
+	return model !== void 0 && /^gemini-[34](?:[.-]|$)/i.test(model.trim());
 }
 /**
 * Convert resolved harness messages into Gemini contents. Function response
@@ -13432,7 +13485,7 @@ function antigravityFlow(oauth) {
 /** Shared Antigravity API headers. */
 function antigravityHeaders(accessToken, userAgent = ANTIGRAVITY_DEFAULT_USER_AGENT) {
 	return {
-		"authorization": `Bearer ${accessToken}`,
+		authorization: `Bearer ${accessToken}`,
 		"content-type": "application/json",
 		"user-agent": userAgent
 	};
@@ -13951,9 +14004,9 @@ function createXSearchTool(options) {
 			const response = await (options.fetchFn ?? hostFetch)(X_SEARCH_URL, {
 				method: "POST",
 				headers: {
-					"authorization": `Bearer ${session.accessToken}`,
+					authorization: `Bearer ${session.accessToken}`,
 					"content-type": "application/json",
-					"accept": "application/json"
+					accept: "application/json"
 				},
 				body: JSON.stringify({
 					model: X_SEARCH_MODEL,
@@ -14033,7 +14086,7 @@ const GROK_ASPECT_RATIOS = {
 	"1024x1024": "1:1",
 	"1024x1536": "2:3",
 	"1536x1024": "3:2",
-	"auto": "auto"
+	auto: "auto"
 };
 /**
 * Assemble the grok request body from the same tool arguments: `size` maps
@@ -14319,11 +14372,11 @@ function createImageGenerateTool(options) {
 				send: (session) => fetchFn(references === void 0 ? IMAGE_GENERATE_URL : IMAGE_EDIT_URL, {
 					method: "POST",
 					headers: {
-						"authorization": `Bearer ${session.accessToken}`,
+						authorization: `Bearer ${session.accessToken}`,
 						"chatgpt-account-id": session.accountId,
-						"originator": "codex_cli_rs",
+						originator: "codex_cli_rs",
 						"content-type": "application/json",
-						"accept": "application/json"
+						accept: "application/json"
 					},
 					body: JSON.stringify({
 						...buildImageGenerateBody(args),
@@ -14341,9 +14394,9 @@ function createImageGenerateTool(options) {
 				send: (session) => fetchFn(references === void 0 ? GROK_IMAGE_GENERATE_URL : GROK_IMAGE_EDIT_URL, {
 					method: "POST",
 					headers: {
-						"authorization": `Bearer ${session.accessToken}`,
+						authorization: `Bearer ${session.accessToken}`,
 						"content-type": "application/json",
-						"accept": "application/json"
+						accept: "application/json"
 					},
 					body: JSON.stringify({
 						...buildGrokImageGenerateBody(args),
@@ -14602,8 +14655,8 @@ function createVideoGenerateTool(options) {
 			const session = await options.tokens.session();
 			const fetchFn = options.fetchFn ?? hostFetch;
 			const headers = {
-				"authorization": `Bearer ${session.accessToken}`,
-				"accept": "application/json"
+				authorization: `Bearer ${session.accessToken}`,
+				accept: "application/json"
 			};
 			const submit = await fetchFn(VIDEO_GENERATE_URL, {
 				method: "POST",
@@ -14705,6 +14758,20 @@ function validatePreferences(provider, input) {
 			result.contextWindows[model] = value;
 		}
 	}
+	if (raw.promptCacheTtl !== void 0) {
+		if (provider !== "claude") throw new Error("the prompt-cache TTL override is supported only for Claude");
+		if (raw.promptCacheTtl !== "5m" && raw.promptCacheTtl !== "1h") throw new Error("promptCacheTtl must be \"5m\" or \"1h\"");
+		result.promptCacheTtl = raw.promptCacheTtl;
+	}
+	if (raw.subagentSpeeds !== void 0) {
+		if (provider !== "codex") throw new Error("subagent speeds are currently supported only for Codex");
+		if (!raw.subagentSpeeds || typeof raw.subagentSpeeds !== "object" || Array.isArray(raw.subagentSpeeds)) throw new Error("subagent speeds must be a model-to-speed map");
+		result.subagentSpeeds = Object.create(null);
+		for (const [model, tier] of Object.entries(raw.subagentSpeeds)) {
+			if (!model.trim() || tier !== "standard" && tier !== "fast") throw new Error("subagent speeds require a model id and either \"standard\" or \"fast\"");
+			result.subagentSpeeds[model] = tier;
+		}
+	}
 	if (raw.tools !== void 0) {
 		if (!raw.tools || typeof raw.tools !== "object" || Array.isArray(raw.tools)) throw new Error("tools must be an object");
 		result.tools = {};
@@ -14753,6 +14820,19 @@ var ProviderSettingsStore = class {
 		const windows = this.current.providers[provider]?.contextWindows;
 		return windows && Object.hasOwn(windows, model) ? windows[model] : void 0;
 	}
+	/** Read on every request so saved changes apply to the next subagent request. */
+	subagentSpeed(model) {
+		const speeds = this.current.providers.codex?.subagentSpeeds;
+		return speeds && Object.hasOwn(speeds, model) ? speeds[model] : "standard";
+	}
+	/**
+	* The Claude route's prompt-cache TTL override, read live so a saved change
+	* takes effect from the next request without a restart. Absent means follow
+	* the `claudePromptCacheTtl` config.
+	*/
+	promptCacheTtl() {
+		return this.current.providers.claude?.promptCacheTtl;
+	}
 	/** Creation-time policy survives restarts and never changes an existing session. */
 	toolEnabled(provider, tool, createdAt = Date.now()) {
 		let enabled = true;
@@ -15064,7 +15144,7 @@ var SubscriptionsAuthController = class {
 	* cannot be read off the flow manager.
 	*/
 	claims = /* @__PURE__ */ new Map();
-	constructor(flows, deviceFlows, onAuthChanged, resolveAttachments, usageFetchers = {}, readClaudeCreds = readClaudeCodeCredentials, poolUsage = void 0, antigravityConfig = {}, clientVersions = {}, resetRedemption) {
+	constructor(flows, deviceFlows, onAuthChanged, resolveAttachments, usageFetchers = {}, readClaudeCreds = readClaudeCodeCredentials, poolUsage = void 0, antigravityConfig = {}, clientVersions = {}, resetRedemption, onUsageForced) {
 		this.flows = flows;
 		this.deviceFlows = deviceFlows;
 		this.onAuthChanged = onAuthChanged;
@@ -15075,6 +15155,7 @@ var SubscriptionsAuthController = class {
 		this.antigravityConfig = antigravityConfig;
 		this.clientVersions = clientVersions;
 		this.resetRedemption = resetRedemption;
+		this.onUsageForced = onUsageForced;
 	}
 	prepareReset(account, signal) {
 		if (!this.resetRedemption) throw new BadRequest("Reset redemption unavailable");
@@ -15085,6 +15166,7 @@ var SubscriptionsAuthController = class {
 		await this.resetRedemption.redeem(account, ticket, AbortSignal.any([signal, AbortSignal.timeout(2e4)]));
 	}
 	usage(provider, account, signal, force = false) {
+		if (force) this.onUsageForced?.(provider, account);
 		const fetcher = this.usageFetchers[provider];
 		if (fetcher === void 0) return Promise.resolve({ supported: false });
 		if (this.poolUsage === void 0) return fetcher(account, signal);
@@ -15347,6 +15429,13 @@ function apply(ctx, config) {
 	let grokTokens;
 	const usageFetchers = {};
 	const speedBySession = /* @__PURE__ */ new Map();
+	const selectedSpeed = (sessionId, model) => {
+		const explicit = speedBySession.get(sessionId);
+		if (explicit !== void 0) return explicit;
+		const agent = ctx.get("agents")?.get(sessionId);
+		if (model !== void 0 && agent?.session.header.origin === "subagent") return preferences.subagentSpeed(parseAccountModelId(model)?.model ?? model);
+		return "standard";
+	};
 	let codexAdapter;
 	let claudeAdapter;
 	let copilotAdapter;
@@ -15397,8 +15486,7 @@ function apply(ctx, config) {
 					};
 				}
 			};
-			let adapter;
-			adapter = new CodexAdapter({
+			const adapter = new CodexAdapter({
 				...config.codexClientVersion === void 0 ? {} : { clientVersion: config.codexClientVersion },
 				resolveClientVersion: () => codexVersion.resolve(),
 				models: catalog.codex,
@@ -15413,7 +15501,7 @@ function apply(ctx, config) {
 				defaultEffortOf: (model) => defaultEffortOf("codex", model),
 				contextWindowOf: (model) => preferences.contextWindow(model),
 				pool: () => poolAdapter,
-				speedFor: (sessionId, model) => sessionId !== void 0 && speedBySession.get(sessionId) === "fast" && adapter.supportsFastTier(model)
+				speedFor: (sessionId, model) => sessionId !== void 0 && selectedSpeed(sessionId, model) === "fast" && adapter.supportsFastTier(model)
 			});
 			codexAdapter = adapter;
 			adapters.set("codex", adapter);
@@ -15449,6 +15537,7 @@ function apply(ctx, config) {
 				contextWindowOf: (model) => preferences.contextWindow(model, "claude"),
 				resolveCliVersion: () => claudeVersion.resolve(),
 				promptCacheTtl: claudePromptCacheTtl,
+				promptCacheTtlOverride: () => preferences.promptCacheTtl(),
 				pool: () => poolAdapter
 			});
 			claudeAdapter = adapter;
@@ -15628,24 +15717,23 @@ function apply(ctx, config) {
 		adapter.listModels = async (route) => (await list(route)).filter((model) => preferences.visible(provider, model.id));
 	}
 	const speed = {
-		async speed(sessionId) {
+		async speed(sessionId, model) {
 			const fastModels = await codexAdapter?.fastCapableModels() ?? [];
 			const accountPreferences = preferences.get("codex").accounts;
 			for (const { key } of await codexTokens?.list() ?? []) {
 				if (!accountPreferences || !Object.hasOwn(accountPreferences, key) || accountPreferences[key].independentEntry !== true) continue;
-				for (const model of [...fastModels]) {
-					if (parseAccountModelId(model)) continue;
-					if (await codexAdapter?.supportsFastTier(model, key)) fastModels.push(accountModelId(key, model));
+				for (const model$1 of [...fastModels]) {
+					if (parseAccountModelId(model$1)) continue;
+					if (await codexAdapter?.supportsFastTier(model$1, key)) fastModels.push(accountModelId(key, model$1));
 				}
 			}
 			return {
-				tier: speedBySession.get(sessionId) ?? "standard",
+				tier: selectedSpeed(sessionId, model),
 				fastModels
 			};
 		},
 		async setSpeed(sessionId, tier) {
-			if (tier === "standard") speedBySession.delete(sessionId);
-			else speedBySession.set(sessionId, tier);
+			speedBySession.set(sessionId, tier);
 		}
 	};
 	const modelDefaults = {
@@ -15661,7 +15749,7 @@ function apply(ctx, config) {
 			const catalog$1 = [];
 			for (const provider of PROVIDER_IDS) {
 				if (!visible.has(provider)) continue;
-				let models = [];
+				let models;
 				try {
 					models = await ctx.llm.listModels(provider);
 				} catch {
@@ -15776,7 +15864,8 @@ function apply(ctx, config) {
 			source: "config"
 		}) } : {},
 		...providers.includes("claude") ? { claude: presentedVersion(claudeVersion) } : {}
-	}, resetRedemption), speed, modelDefaults, {
+	}, resetRedemption, (provider, account) => poolHealth?.clear(provider, account)), speed, modelDefaults, {
+		visibleModels: (provider) => preferences.get(provider).visibleModels,
 		async get(provider, force) {
 			await loadModelDefaults();
 			const adapter = adapters.get(provider);
@@ -15795,6 +15884,7 @@ function apply(ctx, config) {
 				models: await withTimeout((signal) => adapter.listOwnModels(provider, account.key, signal), DISCOVERY_TIMEOUT_MS).catch(() => void 0)
 			})));
 			const tierIds = new Set((await poolAdapter?.modelsForProvider(provider).catch(() => []) ?? []).map((model) => model.id));
+			const fastModels = new Set(provider === "codex" ? await codexAdapter?.fastCapableModels() ?? [] : []);
 			const rows = await Promise.all(models.map(async (model) => {
 				const contexts = [];
 				if (provider === "codex" && codexAdapter) {
@@ -15819,6 +15909,7 @@ function apply(ctx, config) {
 						name: name$1
 					})) ?? [],
 					configured: defaultEffortOf(provider, model.id),
+					...provider === "codex" ? { fastTier: !tierIds.has(model.id) && fastModels.has(model.id) } : {},
 					...contexts.length ? {
 						defaultContextWindow: Math.min(...contexts.map((entry) => entry.default)),
 						maxContextWindow: Math.min(...contexts.map((entry) => entry.max))

package/lib/model-defaults.js

@@ -163,7 +163,9 @@ export function defaultEffortOf(provider, model) {
     // plain index would inherit from Object.prototype for names like
     // `toString`, handing a *function* to mergeReasoning (which then throws and
     // breaks that model's resolution).
-    return Object.prototype.hasOwnProperty.call(section, model) ? section[model] : undefined;
+    return Object.prototype.hasOwnProperty.call(section, model)
+        ? section[model]
+        : undefined;
 }
 /**
  * Detached snapshot for store assertions in tests; no production RPC uses it.
@@ -193,7 +195,7 @@ export function setDefaultEffort(provider, model, effort) {
     // shared chain.
     const run = writeChain.then(async () => {
         await ensureReady();
-        const section = { ...sectionOf(current, provider) ?? {} };
+        const section = { ...(sectionOf(current, provider) ?? {}) };
         if (effort === undefined) {
             delete section[model];
         }