dsh-plugin-subscriptions 0.4.2 → 0.5.0

package/lib/index.js

@@ -1,5 +1,5 @@
 import z from "@deepseek-ai/schemastery";
-import { CONTEXT_WINDOW_EXCEEDED_CODE, CallId, EMPTY_RESPONSE_CODE, LlmAdapter, LlmError, QUOTA_EXCEEDED_CODE, ReasoningEffortId, attributionHeaders, createUserMessage, errorChain, isContextWindowExceededError, isQuotaExceededError, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
+import { CONTEXT_WINDOW_EXCEEDED_CODE, CallId, EMPTY_RESPONSE_CODE, LlmAdapter, LlmError, QUOTA_EXCEEDED_CODE, ReasoningEffortId, attributionHeaders, errorChain, isContextWindowExceededError, isQuotaExceededError, resolveRetryPolicy } from "@deepseek-ai/dsh-llm";
 import { createServer } from "node:http";
 import { createHash, randomBytes, randomUUID } from "node:crypto";
 import { chmod, mkdir, readFile, rename, rm, writeFile } from "node:fs/promises";
@@ -595,6 +595,12 @@ function readString(payload, field) {
 	if (typeof value !== "string" || value.length === 0) throw new BadRequest(`payload.${field} must be a non-empty string`);
 	return value;
 }
+/** Validate the `setSpeed` endpoint's tier. */
+function readSpeedTier(payload) {
+	const tier = payload.tier;
+	if (tier !== "standard" && tier !== "fast") throw new BadRequest("payload.tier must be \"standard\" or \"fast\"");
+	return tier;
+}
 /** Validate the `image` endpoint's payload into a full attachment reference. */
 function readImageRef(payload) {
 	if (typeof payload !== "object" || payload === null) throw new BadRequest("payload must be an object");
@@ -633,7 +639,12 @@ function readVideoName(payload) {
 	if (typeof name$1 !== "string" || !VIDEO_NAME_PATTERN.test(name$1)) throw new BadRequest("payload.name must be a bare .mp4 file name");
 	return name$1;
 }
-async function dispatch(controller, endpoint, payload, signal) {
+/** Validate the session id both speed endpoints carry. */
+function readSessionId(payload) {
+	if (typeof payload !== "object" || payload === null) throw new BadRequest("payload must be an object");
+	return readString(payload, "sessionId");
+}
+async function dispatch(controller, speed, endpoint, payload, signal) {
 	switch (endpoint) {
 		case "status": {
 			const entries = await Promise.all(PROVIDER_IDS.map(async (provider) => [provider, await controller.status(provider)]));
@@ -654,6 +665,10 @@ async function dispatch(controller, endpoint, payload, signal) {
 		case "usage": return ok(await controller.usage(readProvider(payload), signal));
 		case "image": return ok(await controller.readImage(readImageRef(payload), signal));
 		case "video": return ok(await controller.readVideo(readVideoName(payload), signal));
+		case "speed": return ok(await speed.speed(readSessionId(payload)));
+		case "setSpeed":
+			await speed.setSpeed(readSessionId(payload), readSpeedTier(payload));
+			return ok({ ok: true });
 		default: throw new BadRequest(`unknown /subscriptions-auth endpoint "${endpoint}"`);
 	}
 }
@@ -661,13 +676,14 @@ async function dispatch(controller, endpoint, payload, signal) {
 * Register the `/subscriptions-auth` RPC channel when a host connection exists.
 * @param ctx - the plugin context (headless profiles have no `connection`).
 * @param controller - the auth operations backing the endpoints.
+* @param speed - the per-session speed-tier state backing the Speed toggle.
 */
-function registerAuthRpc(ctx, controller) {
+function registerAuthRpc(ctx, controller, speed) {
 	ctx.inject(["connection"], (ctx$1) => {
 		const connection = ctx$1.get("connection");
 		ctx$1.effect(() => connection.rpc.handle(SUBSCRIPTIONS_AUTH_CHANNEL, async (endpoint, payload, signal) => {
 			try {
-				return await dispatch(controller, endpoint, payload, signal);
+				return await dispatch(controller, speed, endpoint, payload, signal);
 			} catch (error) {
 				return failure(error);
 			}
@@ -1013,6 +1029,8 @@ function sanitizeModel(value) {
 	if (raw.reasoning !== void 0 && reasoning === void 0) return void 0;
 	const thinkingType = raw.thinkingType;
 	if (thinkingType !== void 0 && thinkingType !== "enabled" && thinkingType !== "adaptive") return void 0;
+	const fastTier = raw.fastTier;
+	if (fastTier !== void 0 && typeof fastTier !== "boolean") return void 0;
 	return {
 		id: raw.id,
 		name: raw.name,
@@ -1020,7 +1038,8 @@ function sanitizeModel(value) {
 		...raw.contextWindow === void 0 ? {} : { contextWindow: raw.contextWindow },
 		...raw.priority === void 0 ? {} : { priority: raw.priority },
 		...reasoning === void 0 ? {} : { reasoning },
-		...thinkingType === void 0 ? {} : { thinkingType }
+		...thinkingType === void 0 ? {} : { thinkingType },
+		...fastTier === void 0 ? {} : { fastTier }
 	};
 }
 /**
@@ -1584,6 +1603,14 @@ const CODEX_EFFORTS = [
 const CODEX_DEFAULT_EFFORT = ReasoningEffortId("high");
 /** Every gpt-5.x codex model accepts image input. */
 const CODEX_MODALITIES = ["text", "image"];
+/**
+* Fast tier (the codex CLI's "fast mode"): the Responses `service_tier` wire
+* value for priority processing, mirroring codex-rs
+* `ServiceTier::Fast.request_value()`. The legacy catalog spelling is the
+* `additional_speed_tiers` entry "fast".
+*/
+const CODEX_FAST_SERVICE_TIER = "priority";
+const CODEX_FAST_SPEED_TIER = "fast";
 /** Static codex flow facts for the OAuth flow engine. */
 const codexFlow = {
 	callbackPath: CODEX_CALLBACK_PATH,
@@ -1795,6 +1822,14 @@ function effortName(effort) {
 	return effort === "xhigh" ? "Extra High" : effort.charAt(0).toUpperCase() + effort.slice(1);
 }
 /**
+* Whether a catalog entry advertises the fast tier. Mirrors codex-rs
+* `ModelPreset::supports_fast_mode`: a `service_tiers` id matching the fast
+* wire value, or the legacy `additional_speed_tiers` "fast" entry.
+*/
+function supportsFastTier(entry) {
+	return (entry.service_tiers ?? []).some((tier) => tier.id === CODEX_FAST_SERVICE_TIER) || (entry.additional_speed_tiers ?? []).includes(CODEX_FAST_SPEED_TIER);
+}
+/**
 * Fetch the live codex model catalog with the session's auth headers.
 * @param session - the stored session (used as-is; never refreshed here).
 * @param fetchFn - fetch implementation (injectable for tests).
@@ -1821,7 +1856,7 @@ async function fetchCodexModels(session, fetchFn = fetch) {
 			...level.description === void 0 ? {} : { description: level.description }
 		}));
 		const defaultEffort = typeof entry.default_reasoning_level === "string" && entry.default_reasoning_level.length > 0 && efforts.some((effort) => effort.id === ReasoningEffortId(entry.default_reasoning_level)) ? ReasoningEffortId(entry.default_reasoning_level) : void 0;
-		discovered.push({
+		const model = {
 			id: entry.slug,
 			name: typeof entry.display_name === "string" && entry.display_name.length > 0 ? entry.display_name : entry.slug,
 			...typeof entry.description === "string" && entry.description.length > 0 ? { description: entry.description } : {},
@@ -1830,13 +1865,40 @@ async function fetchCodexModels(session, fetchFn = fetch) {
 			...efforts.length > 0 ? { reasoning: {
 				efforts,
 				...defaultEffort === void 0 ? {} : { defaultEffort }
-			} } : {}
-		});
+			} } : {},
+			...supportsFastTier(entry) ? { fastTier: true } : {}
+		};
+		discovered.push(model);
 	}
 	discovered.sort((a, b) => (a.priority ?? Number.MAX_SAFE_INTEGER) - (b.priority ?? Number.MAX_SAFE_INTEGER));
 	if (discovered.length === 0) throw new Error(`codex models endpoint returned an empty catalog (client_version ${CODEX_CLIENT_VERSION})`);
 	return discovered;
 }
+/**
+* The Responses request body for one generation. A fast-tier request (the
+* composer Speed toggle, the codex CLI's fast mode) carries
+* `service_tier: priority`; the tier field is omitted entirely otherwise,
+* matching the CLI (it never sends an explicit standard tier).
+*/
+function codexRequestBody(options, resolved, fast) {
+	return {
+		model: options.model,
+		instructions: resolved.instructions ?? DEFAULT_CODEX_INSTRUCTIONS,
+		input: resolved.input,
+		...options.tools !== void 0 && options.tools.length > 0 ? { tools: toResponsesTools(options.tools) } : {},
+		tool_choice: "auto",
+		parallel_tool_calls: true,
+		...options.reasoningEffort !== void 0 ? { reasoning: {
+			effort: String(options.reasoningEffort),
+			summary: "auto"
+		} } : {},
+		store: false,
+		stream: true,
+		include: ["reasoning.encrypted_content"],
+		...options.sessionId !== void 0 ? { prompt_cache_key: String(options.sessionId) } : {},
+		...fast ? { service_tier: CODEX_FAST_SERVICE_TIER } : {}
+	};
+}
 /** Codex wire adapter: one instance serves the `codex` provider route. */
 var CodexAdapter = class extends LlmAdapter {
 	catalog;
@@ -1892,6 +1954,16 @@ var CodexAdapter = class extends LlmAdapter {
 		if (!this.options.discovery) return void 0;
 		return (await this.catalog.resolve(() => this.fetchCatalog()))?.find((entry) => entry.id === model);
 	}
+	/** Whether the discovered catalog advertises a fast tier for this model. */
+	async supportsFastTier(model) {
+		return (await this.discovered(model))?.fastTier === true;
+	}
+	/** Ids of every discovered model with a fast tier (the Speed toggle's visibility list). */
+	async fastCapableModels() {
+		if (!this.options.discovery) return [];
+		if (await this.options.tokens.peek() === void 0) return [];
+		return (await this.catalog.resolve(() => this.fetchCatalog()) ?? []).filter((model) => model.fastTier === true).map((model) => model.id);
+	}
 	async resolveModel(provider, model) {
 		const discovered = await this.discovered(model);
 		const configured = this.options.models.find((entry) => entry.id === model);
@@ -1930,23 +2002,9 @@ var CodexAdapter = class extends LlmAdapter {
 		}
 	}
 	async request(options, session, signal) {
-		const { instructions, input } = toResponsesInput(await resolveImages(options.messages, this.options.resolveAttachments?.(), signal), options.system);
-		const body = {
-			model: options.model,
-			instructions: instructions ?? DEFAULT_CODEX_INSTRUCTIONS,
-			input,
-			...options.tools !== void 0 && options.tools.length > 0 ? { tools: toResponsesTools(options.tools) } : {},
-			tool_choice: "auto",
-			parallel_tool_calls: true,
-			...options.reasoningEffort !== void 0 ? { reasoning: {
-				effort: String(options.reasoningEffort),
-				summary: "auto"
-			} } : {},
-			store: false,
-			stream: true,
-			include: ["reasoning.encrypted_content"],
-			...options.sessionId !== void 0 ? { prompt_cache_key: String(options.sessionId) } : {}
-		};
+		const messages = await resolveImages(options.messages, this.options.resolveAttachments?.(), signal);
+		const fast = this.options.speedFor !== void 0 && await this.options.speedFor(options.sessionId, options.model);
+		const body = codexRequestBody(options, toResponsesInput(messages, options.system), fast);
 		return fetch(CODEX_API_URL, {
 			method: "POST",
 			headers: {
@@ -2355,7 +2413,11 @@ function detectClaudeVersion() {
 	} catch {}
 	return CLAUDE_CLI_FALLBACK_VERSION;
 }
-const CLAUDE_CLI_USER_AGENT = `claude-cli/${detectClaudeVersion()} (external, cli)`;
+let claudeCliUserAgent;
+function getClaudeCliUserAgent() {
+	if (claudeCliUserAgent === void 0) claudeCliUserAgent = `claude-cli/${detectClaudeVersion()} (external, cli)`;
+	return claudeCliUserAgent;
+}
 const CLAUDE_BETA_FALLBACK = [
 	"claude-code-20250219",
 	"oauth-2025-04-20",
@@ -2508,7 +2570,7 @@ async function fetchClaudeUsage(session, fetchFn = fetch, signal) {
 		headers: {
 			"authorization": `Bearer ${session.accessToken}`,
 			"anthropic-beta": "oauth-2025-04-20",
-			"user-agent": CLAUDE_CLI_USER_AGENT,
+			"user-agent": getClaudeCliUserAgent(),
 			"accept": "application/json"
 		},
 		...signal === void 0 ? {} : { signal }
@@ -2560,7 +2622,7 @@ async function fetchClaudeModels(session, fetchFn = fetch) {
 	const response = await fetchFn(CLAUDE_MODELS_URL, { headers: {
 		"authorization": `Bearer ${session.accessToken}`,
 		"anthropic-version": "2023-06-01",
-		"user-agent": CLAUDE_CLI_USER_AGENT,
+		"user-agent": getClaudeCliUserAgent(),
 		"anthropic-dangerous-direct-browser-access": "true",
 		"accept": "application/json"
 	} });
@@ -2727,7 +2789,7 @@ var ClaudeAdapter = class extends LlmAdapter {
 				"authorization": `Bearer ${session.accessToken}`,
 				"anthropic-version": "2023-06-01",
 				"anthropic-beta": CLAUDE_BETA_FLAGS,
-				"user-agent": CLAUDE_CLI_USER_AGENT,
+				"user-agent": getClaudeCliUserAgent(),
 				"x-app": "cli",
 				"anthropic-dangerous-direct-browser-access": "true",
 				"accept": "text/event-stream",
@@ -3697,19 +3759,11 @@ function createImageGenerateTool(options) {
 				});
 			}
 			const revisedPrompt = images.find((image) => image.revisedPrompt !== void 0)?.revisedPrompt;
-			const value = {
+			return {
 				paths,
 				...refs.length > 0 ? { images: refs } : {},
 				...revisedPrompt === void 0 ? {} : { revisedPrompt }
 			};
-			if (exec.parent !== void 0 && refs.length > 0) exec.deferContext(createUserMessage({
-				content: imageGenerateContent(value),
-				source: {
-					kind: "plugin",
-					plugin: "dsh-plugin-subscriptions"
-				}
-			}));
-			return value;
 		}
 	});
 }
@@ -4192,6 +4246,8 @@ function apply(ctx, config) {
 	let claudeTokens;
 	let grokTokens;
 	const usageFetchers = {};
+	const speedBySession = /* @__PURE__ */ new Map();
+	let codexAdapter;
 	for (const provider of providers) switch (provider) {
 		case "codex": {
 			const tokens = new TokenManager({
@@ -4208,15 +4264,19 @@ function apply(ctx, config) {
 			});
 			codexTokens = tokens;
 			usageFetchers.codex = async (signal) => fetchCodexUsage(await tokens.session(), fetch, signal);
-			handles.set("codex", ctx.llm.registerAdapter(["codex"], new CodexAdapter({
+			let adapter;
+			adapter = new CodexAdapter({
 				models: catalog.codex,
 				streamIdleTimeoutMs,
 				tokens,
 				discovery: !overridden.has("codex"),
 				onWarn,
 				resolveAttachments,
-				catalogStore: catalogStore("codex")
-			})));
+				catalogStore: catalogStore("codex"),
+				speedFor: (sessionId, model) => sessionId !== void 0 && speedBySession.get(sessionId) === "fast" && adapter.supportsFastTier(model)
+			});
+			codexAdapter = adapter;
+			handles.set("codex", ctx.llm.registerAdapter(["codex"], adapter));
 			break;
 		}
 		case "claude": {
@@ -4273,7 +4333,18 @@ function apply(ctx, config) {
 			break;
 		}
 	}
-	registerAuthRpc(ctx, new SubscriptionsAuthController(flows, authChanged, resolveAttachments, usageFetchers));
+	registerAuthRpc(ctx, new SubscriptionsAuthController(flows, authChanged, resolveAttachments, usageFetchers), {
+		async speed(sessionId) {
+			return {
+				tier: speedBySession.get(sessionId) ?? "standard",
+				fastModels: await codexAdapter?.fastCapableModels() ?? []
+			};
+		},
+		async setSpeed(sessionId, tier) {
+			if (tier === "standard") speedBySession.delete(sessionId);
+			else speedBySession.set(sessionId, tier);
+		}
+	});
 	if (claudeTokens !== void 0) {
 		const syncTimer = setInterval(() => {
 			claudeTokens?.session().catch(() => {});

package/lib/providers/catalog-store.js

@@ -74,6 +74,9 @@ function sanitizeModel(value) {
     const thinkingType = raw.thinkingType;
     if (thinkingType !== undefined && thinkingType !== 'enabled' && thinkingType !== 'adaptive')
         return undefined;
+    const fastTier = raw.fastTier;
+    if (fastTier !== undefined && typeof fastTier !== 'boolean')
+        return undefined;
     return {
         id: raw.id,
         name: raw.name,
@@ -82,6 +85,7 @@ function sanitizeModel(value) {
         ...raw.priority === undefined ? {} : { priority: raw.priority },
         ...reasoning === undefined ? {} : { reasoning },
         ...thinkingType === undefined ? {} : { thinkingType: thinkingType },
+        ...fastTier === undefined ? {} : { fastTier },
     };
 }
 /**

package/lib/providers/claude.js

@@ -36,7 +36,17 @@ export function detectClaudeVersion() {
     catch { }
     return CLAUDE_CLI_FALLBACK_VERSION;
 }
-const CLAUDE_CLI_USER_AGENT = `claude-cli/${detectClaudeVersion()} (external, cli)`;
+// Lazy + memoized: detectClaudeVersion() shells out to `claude --version`,
+// so this must not run at module-evaluation time (it would fire for every
+// consumer of this module regardless of whether Claude is a configured
+// provider). Computed on first use of getClaudeCliUserAgent() instead.
+let claudeCliUserAgent;
+function getClaudeCliUserAgent() {
+    if (claudeCliUserAgent === undefined) {
+        claudeCliUserAgent = `claude-cli/${detectClaudeVersion()} (external, cli)`;
+    }
+    return claudeCliUserAgent;
+}
 export const CLAUDE_BETA_FALLBACK = [
     'claude-code-20250219',
     'oauth-2025-04-20',
@@ -234,7 +244,7 @@ export async function fetchClaudeUsage(session, fetchFn = fetch, signal) {
             'anthropic-beta': 'oauth-2025-04-20',
             // Unrecognized clients are aggressively rate-limited on this endpoint,
             // so it presents as the CLI like every other subscription request.
-            'user-agent': CLAUDE_CLI_USER_AGENT,
+            'user-agent': getClaudeCliUserAgent(),
             'accept': 'application/json',
         },
         ...signal === undefined ? {} : { signal },
@@ -283,7 +293,7 @@ export async function fetchClaudeModels(session, fetchFn = fetch) {
         headers: {
             'authorization': `Bearer ${session.accessToken}`,
             'anthropic-version': '2023-06-01',
-            'user-agent': CLAUDE_CLI_USER_AGENT,
+            'user-agent': getClaudeCliUserAgent(),
             'anthropic-dangerous-direct-browser-access': 'true',
             'accept': 'application/json',
         },
@@ -473,7 +483,7 @@ export class ClaudeAdapter extends LlmAdapter {
                 'authorization': `Bearer ${session.accessToken}`,
                 'anthropic-version': '2023-06-01',
                 'anthropic-beta': CLAUDE_BETA_FLAGS,
-                'user-agent': CLAUDE_CLI_USER_AGENT,
+                'user-agent': getClaudeCliUserAgent(),
                 'x-app': 'cli',
                 'anthropic-dangerous-direct-browser-access': 'true',
                 'accept': 'text/event-stream',

package/lib/providers/codex.d.ts

@@ -8,6 +8,7 @@ import type { GenerateOptions, LlmModelInfo, LlmProviderInfo, LlmResolvedModelIn
 import type { FlowSpec } from '../auth/oauth-flow.js';
 import type { CodexSession } from '../auth/store.js';
 import type { AttachmentStore } from '@deepseek-ai/dsh-attachment';
+import type { ResponsesRequestInput } from '../translate/responses.js';
 import { TokenManager } from './common.js';
 import type { CatalogPersistence, DiscoveredModel, FetchFn, ModelEntry, ProviderUsage } from './common.js';
 export declare const CODEX_CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann";
@@ -16,6 +17,15 @@ export declare const CODEX_TOKEN_URL = "https://auth.openai.com/oauth/token";
 export declare const CODEX_API_URL = "https://chatgpt.com/backend-api/codex/responses";
 /** Refresh when the access token has less than this much life left. */
 export declare const CODEX_PREEMPT_MS: number;
+/**
+ * Fast tier (the codex CLI's "fast mode"): the Responses `service_tier` wire
+ * value for priority processing, mirroring codex-rs
+ * `ServiceTier::Fast.request_value()`. The legacy catalog spelling is the
+ * `additional_speed_tiers` entry "fast".
+ */
+export declare const CODEX_FAST_SERVICE_TIER = "priority";
+/** One session's speed choice: standard routing or the fast (priority) tier. */
+export type CodexSpeedTier = 'standard' | 'fast';
 /** Static codex flow facts for the OAuth flow engine. */
 export declare const codexFlow: FlowSpec;
 /** User identity claims decoded from a codex id token. */
@@ -98,7 +108,20 @@ export interface CodexAdapterOptions {
     resolveAttachments?: () => AttachmentStore | undefined;
     /** Durable catalog store seeding capability metadata across restarts. */
     catalogStore?: CatalogPersistence;
+    /**
+     * Per-request speed lookup (the composer Speed toggle's host half). Returns
+     * whether this session's current choice sends the model on the fast tier;
+     * absent means every request stays on standard routing.
+     */
+    speedFor?: (sessionId: string | undefined, model: string) => Promise<boolean> | boolean;
 }
+/**
+ * The Responses request body for one generation. A fast-tier request (the
+ * composer Speed toggle, the codex CLI's fast mode) carries
+ * `service_tier: priority`; the tier field is omitted entirely otherwise,
+ * matching the CLI (it never sends an explicit standard tier).
+ */
+export declare function codexRequestBody(options: GenerateOptions, resolved: ResponsesRequestInput, fast: boolean): Record<string, unknown>;
 /** Codex wire adapter: one instance serves the `codex` provider route. */
 export declare class CodexAdapter extends LlmAdapter {
     private readonly options;
@@ -117,6 +140,10 @@ export declare class CodexAdapter extends LlmAdapter {
      * call — just because the TTL lapsed mid-turn.
      */
     private discovered;
+    /** Whether the discovered catalog advertises a fast tier for this model. */
+    supportsFastTier(model: string): Promise<boolean>;
+    /** Ids of every discovered model with a fast tier (the Speed toggle's visibility list). */
+    fastCapableModels(): Promise<string[]>;
     resolveModel(provider: string, model: string): Promise<LlmResolvedModelInfo>;
     stream(options: GenerateOptions): AsyncIterable<StreamChunk>;
     private request;

package/lib/providers/codex.js

@@ -39,6 +39,14 @@ const CODEX_EFFORTS = [
 const CODEX_DEFAULT_EFFORT = ReasoningEffortId('high');
 /** Every gpt-5.x codex model accepts image input. */
 const CODEX_MODALITIES = ['text', 'image'];
+/**
+ * Fast tier (the codex CLI's "fast mode"): the Responses `service_tier` wire
+ * value for priority processing, mirroring codex-rs
+ * `ServiceTier::Fast.request_value()`. The legacy catalog spelling is the
+ * `additional_speed_tiers` entry "fast".
+ */
+export const CODEX_FAST_SERVICE_TIER = 'priority';
+const CODEX_FAST_SPEED_TIER = 'fast';
 /** Static codex flow facts for the OAuth flow engine. */
 export const codexFlow = {
     callbackPath: CODEX_CALLBACK_PATH,
@@ -286,6 +294,15 @@ export const CODEX_CLIENT_VERSION = '0.147.0';
 function effortName(effort) {
     return effort === 'xhigh' ? 'Extra High' : effort.charAt(0).toUpperCase() + effort.slice(1);
 }
+/**
+ * Whether a catalog entry advertises the fast tier. Mirrors codex-rs
+ * `ModelPreset::supports_fast_mode`: a `service_tiers` id matching the fast
+ * wire value, or the legacy `additional_speed_tiers` "fast" entry.
+ */
+function supportsFastTier(entry) {
+    return (entry.service_tiers ?? []).some(tier => tier.id === CODEX_FAST_SERVICE_TIER)
+        || (entry.additional_speed_tiers ?? []).includes(CODEX_FAST_SPEED_TIER);
+}
 /**
  * Fetch the live codex model catalog with the session's auth headers.
  * @param session - the stored session (used as-is; never refreshed here).
@@ -328,7 +345,7 @@ export async function fetchCodexModels(session, fetchFn = fetch) {
             && efforts.some(effort => effort.id === ReasoningEffortId(entry.default_reasoning_level))
             ? ReasoningEffortId(entry.default_reasoning_level)
             : undefined;
-        discovered.push({
+        const model = {
             id: entry.slug,
             name: typeof entry.display_name === 'string' && entry.display_name.length > 0
                 ? entry.display_name
@@ -343,7 +360,9 @@ export async function fetchCodexModels(session, fetchFn = fetch) {
             ...efforts.length > 0
                 ? { reasoning: { efforts, ...defaultEffort === undefined ? {} : { defaultEffort } } }
                 : {},
-        });
+            ...supportsFastTier(entry) ? { fastTier: true } : {},
+        };
+        discovered.push(model);
     }
     discovered.sort((a, b) => (a.priority ?? Number.MAX_SAFE_INTEGER) - (b.priority ?? Number.MAX_SAFE_INTEGER));
     // An empty catalog from a 200 response means the backend gated us out (e.g.
@@ -354,6 +373,32 @@ export async function fetchCodexModels(session, fetchFn = fetch) {
     }
     return discovered;
 }
+/**
+ * The Responses request body for one generation. A fast-tier request (the
+ * composer Speed toggle, the codex CLI's fast mode) carries
+ * `service_tier: priority`; the tier field is omitted entirely otherwise,
+ * matching the CLI (it never sends an explicit standard tier).
+ */
+export function codexRequestBody(options, resolved, fast) {
+    return {
+        model: options.model,
+        instructions: resolved.instructions ?? DEFAULT_CODEX_INSTRUCTIONS,
+        input: resolved.input,
+        ...options.tools !== undefined && options.tools.length > 0
+            ? { tools: toResponsesTools(options.tools) }
+            : {},
+        tool_choice: 'auto',
+        parallel_tool_calls: true,
+        ...options.reasoningEffort !== undefined
+            ? { reasoning: { effort: String(options.reasoningEffort), summary: 'auto' } }
+            : {},
+        store: false,
+        stream: true,
+        include: ['reasoning.encrypted_content'],
+        ...options.sessionId !== undefined ? { prompt_cache_key: String(options.sessionId) } : {},
+        ...fast ? { service_tier: CODEX_FAST_SERVICE_TIER } : {},
+    };
+}
 /** Codex wire adapter: one instance serves the `codex` provider route. */
 export class CodexAdapter extends LlmAdapter {
     options;
@@ -423,6 +468,22 @@ export class CodexAdapter extends LlmAdapter {
         const models = await this.catalog.resolve(() => this.fetchCatalog());
         return models?.find(entry => entry.id === model);
     }
+    /** Whether the discovered catalog advertises a fast tier for this model. */
+    async supportsFastTier(model) {
+        return (await this.discovered(model))?.fastTier === true;
+    }
+    /** Ids of every discovered model with a fast tier (the Speed toggle's visibility list). */
+    async fastCapableModels() {
+        if (!this.options.discovery)
+            return [];
+        // Not logged in → no fast models, so the Speed toggle hides after logout
+        // (mirrors the listModels guard above).
+        const session = await this.options.tokens.peek();
+        if (session === undefined)
+            return [];
+        const models = await this.catalog.resolve(() => this.fetchCatalog());
+        return (models ?? []).filter(model => model.fastTier === true).map(model => model.id);
+    }
     async resolveModel(provider, model) {
         // Discovered metadata (when discovery is on) wins over the static entry;
         // the static entry wins over the built-in defaults.
@@ -465,24 +526,9 @@ export class CodexAdapter extends LlmAdapter {
     }
     async request(options, session, signal) {
         const messages = await resolveImages(options.messages, this.options.resolveAttachments?.(), signal);
-        const { instructions, input } = toResponsesInput(messages, options.system);
-        const body = {
-            model: options.model,
-            instructions: instructions ?? DEFAULT_CODEX_INSTRUCTIONS,
-            input,
-            ...options.tools !== undefined && options.tools.length > 0
-                ? { tools: toResponsesTools(options.tools) }
-                : {},
-            tool_choice: 'auto',
-            parallel_tool_calls: true,
-            ...options.reasoningEffort !== undefined
-                ? { reasoning: { effort: String(options.reasoningEffort), summary: 'auto' } }
-                : {},
-            store: false,
-            stream: true,
-            include: ['reasoning.encrypted_content'],
-            ...options.sessionId !== undefined ? { prompt_cache_key: String(options.sessionId) } : {},
-        };
+        const fast = this.options.speedFor !== undefined
+            && await this.options.speedFor(options.sessionId, options.model);
+        const body = codexRequestBody(options, toResponsesInput(messages, options.system), fast);
         return fetch(CODEX_API_URL, {
             method: 'POST',
             headers: {

package/lib/providers/common.d.ts

@@ -177,6 +177,8 @@ export interface DiscoveredModel {
     };
     /** Claude-specific: which extended-thinking wire shape this model accepts. */
     thinkingType?: 'enabled' | 'adaptive';
+    /** Codex-specific: the catalog advertises a fast (priority) service tier. */
+    fastTier?: boolean;
 }
 /** How long a discovered catalog is trusted before re-fetching. */
 export declare const DISCOVERY_TTL_MS: number;

package/lib/tools/image-generate.js

@@ -16,7 +16,6 @@ import { mkdir, writeFile } from 'node:fs/promises';
 import { basename, join } from 'node:path';
 import { dshHomePath } from '@deepseek-ai/dsh-home-paths';
 import { AttachmentId } from '@deepseek-ai/dsh-attachment';
-import { createUserMessage } from '@deepseek-ai/dsh-llm';
 import { defineTool } from '@deepseek-ai/dsh-tools';
 import { httpLlmError, TokenManager } from '../providers/common.js';
 /** Endpoint the codex generation request is posted to. */
@@ -334,15 +333,9 @@ export function createImageGenerateTool(options) {
                 ...refs.length > 0 ? { images: refs } : {},
                 ...revisedPrompt === undefined ? {} : { revisedPrompt },
             };
-            // Nested (Code Mode) dispatches have no card: defer the image content as
-            // a user message so the next model request still sees it (read_image's
-            // pattern).
-            if (exec.parent !== undefined && refs.length > 0) {
-                exec.deferContext(createUserMessage({
-                    content: imageGenerateContent(value),
-                    source: { kind: 'plugin', plugin: 'dsh-plugin-subscriptions' },
-                }));
-            }
+            // Nested (Code Mode) dispatches need no defer here: the harness's code
+            // mode already defers any image-bearing sub-result as a user message, so
+            // deferring again would inject the same attachment twice.
             return value;
         },
     });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-plugin-subscriptions",
-  "version": "0.4.2",
+  "version": "0.5.0",
   "description": "Use ChatGPT (Codex), Claude, and Grok (X Premium) subscriptions as DeepSeek Harness LLM providers, with OAuth login from the web Settings page",
   "license": "MIT",
   "repository": {
@@ -48,6 +48,12 @@
       ]
     }
   },
+  "scripts": {
+    "build": "tsc && tsdown",
+    "test": "tsc -p tsconfig.test.json && node --test lib-test/test/",
+    "prepare": "tsdown -c tsdown.prepare.config.ts",
+    "prepublishOnly": "pnpm build && pnpm test"
+  },
   "peerDependencies": {
     "@deepseek-ai/cordis": "^4.0.1",
     "@deepseek-ai/dsh-attachment": "^0.1.0-rc.5",
@@ -64,6 +70,8 @@
     "@deepseek-ai/dsh-client-locale": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/locale",
     "@deepseek-ai/dsh-client-runtime": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/runtime",
     "@deepseek-ai/dsh-client-ui-settings": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/ui-settings",
+    "@deepseek-ai/dsh-client-ui-conversation": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/ui-conversation",
+    "@deepseek-ai/dsh-client-ui-commands": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/ui-commands",
     "@deepseek-ai/dsh-client-ui-slots": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/client/ui-slots",
     "@deepseek-ai/dsh-home-paths": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/util/home-paths",
     "@deepseek-ai/dsh-host-apiproxy": "link:/Users/v1ki/Documents/projs/source/deepseek-harness/packages/host/apiproxy",
@@ -75,9 +83,5 @@
     "react": "^18.2.0",
     "tsdown": "^0.15.0",
     "typescript": "^5.8.0"
-  },
-  "scripts": {
-    "build": "tsc && tsdown",
-    "test": "tsc -p tsconfig.test.json && node --test lib-test/test/"
   }
-}
+}