feat: add codex fast mode and harden subagent launch

README.md

@@ -80,6 +80,12 @@ npm run dev
 
 所有设备相关模型配置都位于被 Git 忽略的 `data/` 或 `.env` 中。
 
+#### GPT Codex 快速模式
+
+选择 `openai-codex` 的 GPT-5.5 或 GPT-5.6 模型后，输入框底部会显示闪电按钮。开启后按钮显示 `Fast mode ON`，并将该模型的 `serviceTier` 设为 `priority`，请求会使用 Codex 的快速档位；关闭后恢复普通档位。该设置写入 `data/agent/models.json`，下次请求立即生效，无需重启；非 GPT Codex 模型不会显示此按钮。
+
+快速档位会增加费用：GPT-5.6 通常约为 2 倍，GPT-5.5 的 priority 费率约为 2.5 倍。Web UI 会在按钮提示中明确显示 `service_tier: priority` 和费用倍率。
+
 ### 可选环境变量
 
 需要可选服务时，复制模板：
@@ -147,6 +153,8 @@ data/workspaces/default/    默认工作目录
 
 Windows 的 Playwright 不下载独立 Chromium；首次 `setup` 只缓存 MCP 的 Node.js 包，浏览器执行使用系统 Edge。macOS 的 setup 不安装或启用 Playwright；如需浏览器自动化，可在 Web UI 的 MCP 面板中手动添加并配置。
 
+Windows 下的 `pi-subagents` 子进程由受管启动器自动处理：启动时将本地 `pi-web/node_modules/.bin` 加入子进程 PATH，并把 Pi Web 使用的 `pi-coding-agent` 包链接到受管 Profile，使子代理直接解析本地 `dist/cli.js`。这不会修改系统级 PATH，每次项目启动时会自动恢复。
+
 #### 视觉子代理（vision）
 
 DeepSeek 等纯文本模型不能接收图片。仓库内置 `vision` 子代理（`.agents/vision.md`）：它通过 `vision` 工具调用视觉模型读取图片，把完整 OCR、版式结构与语义描述返回给主模型，主模型基于文本继续推理。视觉后端可插拔（本地 Ollama 或任意 OpenAI 兼容视觉 API）——仓库**不预设默认后端**，首次使用前需自行选择并配置。
@@ -326,6 +334,12 @@ The repository contains no author model endpoint, API key, OAuth token, or defau
 
 Device-specific model configuration remains in ignored `data/` or `.env` files.
 
+#### GPT Codex fast mode
+
+When an `openai-codex` GPT-5.5 or GPT-5.6 model is selected, the lightning button appears beside the model selector. Enabling it shows `Fast mode ON` and sets that model's `serviceTier` to `priority`, using Codex's faster service tier; disabling it restores the default tier. The setting is written to `data/agent/models.json` and takes effect on the next request without a restart. Non-Codex models do not show the button.
+
+Fast mode costs more: GPT-5.6 is typically about 2x, while GPT-5.5 priority pricing is about 2.5x. The Web UI tooltip displays `service_tier: priority` and the applicable credit multiplier.
+
 ### Optional environment configuration
 
 macOS/Linux:
@@ -391,6 +405,8 @@ Versions are pinned in the platform defaults under `config/`: Windows uses `mcp.
 
 On Windows, Playwright never downloads a standalone Chromium: setup caches only its Node package and browser execution uses system Edge. On macOS, setup does not install or enable Playwright; add it manually through the MCP panel if browser automation is needed.
 
+On Windows, the managed launcher prepares `pi-subagents` child processes automatically: it prepends the local `pi-web/node_modules/.bin` directory to the child PATH and links the Pi Web `pi-coding-agent` package into the managed profile, allowing subagents to resolve the local `dist/cli.js` directly. This does not modify the system-wide PATH and is recreated on each project launch.
+
 #### Vision subagent
 
 Text-only models such as DeepSeek cannot receive image attachments. The repository ships a `vision` subagent (`.agents/vision.md`) that calls a vision model through the `vision` tool and returns a full OCR, layout, and semantic description the main model can reason over. The vision backend is pluggable (local Ollama or any OpenAI-compatible vision API) — the repository does **not** ship a default backend; pick and configure one before first use.

data/local-tools/pi-agent-web-supervisor.ps1

@@ -10,15 +10,16 @@ $supervisorLog = Join-Path $logsDirectory 'pi-web-supervisor.log'
 $restartRequestPath = Join-Path $projectRoot 'data\agent\restart-request.json'
 $webUrl = 'http://127.0.0.1:30141/'
 $restartRequestVersion = 1
+$maxResumeAttempts = 3
 
 New-Item -ItemType Directory -Path $logsDirectory -Force | Out-Null
 New-Item -ItemType Directory -Path (Split-Path -Parent $restartRequestPath) -Force | Out-Null
-Set-Content -LiteralPath $supervisorLog -Value "$(Get-Date -Format o) supervisor started"
+Set-Content -LiteralPath $supervisorLog -Value "$(Get-Date -Format o) supervisor started" -Encoding UTF8
 
 function Write-SupervisorLog {
     param([string]$Message)
 
-    Add-Content -LiteralPath $supervisorLog -Value "$(Get-Date -Format o) $Message"
+    Add-Content -LiteralPath $supervisorLog -Value "$(Get-Date -Format o) $Message" -Encoding UTF8
 }
 
 function Get-RestartRequest {
@@ -72,6 +73,20 @@ function Remove-RestartRequest {
     Remove-Item -LiteralPath $restartRequestPath -Force -ErrorAction SilentlyContinue
 }
 
+function Archive-RestartRequest {
+    param([object]$Request)
+
+    if (-not (Test-Path -LiteralPath $restartRequestPath -PathType Leaf)) {
+        return
+    }
+
+    $requestId = (Get-RequestField $Request 'requestId') -replace '[^A-Za-z0-9-]', '_'
+    $timestamp = Get-Date -Format 'yyyyMMdd-HHmmss'
+    $archivePath = Join-Path $logsDirectory "failed-restart-request-$timestamp-$requestId.json"
+    Move-Item -LiteralPath $restartRequestPath -Destination $archivePath -Force
+    Write-SupervisorLog "archived failed restart request at $archivePath"
+}
+
 function Stop-ProcessTree {
     param([System.Diagnostics.Process]$Process)
 
@@ -149,6 +164,7 @@ $testInstructions
 }
 
 $pendingRequest = Get-RestartRequest
+$resumeAttemptCount = 0
 
 while ($true) {
     $process = $null
@@ -176,10 +192,20 @@ while ($true) {
                     try {
                         Resume-AgentSession $pendingRequest
                         $pendingRequest = $null
+                        $resumeAttemptCount = 0
                     }
                     catch {
+                        $resumeAttemptCount += 1
                         Write-SupervisorLog "Agent resume failed: $($_.Exception.Message)"
-                        $nextResumeAttempt = (Get-Date).AddSeconds(3)
+                        if ($resumeAttemptCount -ge $maxResumeAttempts) {
+                            Write-SupervisorLog "Agent resume abandoned after $resumeAttemptCount attempts"
+                            Archive-RestartRequest $pendingRequest
+                            $pendingRequest = $null
+                            $resumeAttemptCount = 0
+                        }
+                        else {
+                            $nextResumeAttempt = (Get-Date).AddSeconds(3)
+                        }
                     }
                 }
             }
@@ -188,6 +214,7 @@ while ($true) {
                 $candidate = Get-RestartRequest
                 if ($candidate) {
                     $pendingRequest = $candidate
+                    $resumeAttemptCount = 0
                     Write-SupervisorLog "restart request $($candidate.requestId) detected"
                     Stop-ProcessTree $process
                     break

pi-web/app/api/models/route.ts

@@ -1,123 +1,180 @@
 import { stat } from "fs/promises";
 import { resolve } from "path";
-import { createAgentSessionServices, getAgentDir, type SettingsManager } from "@earendil-works/pi-coding-agent";
+import {
+	createAgentSessionServices,
+	getAgentDir,
+	type SettingsManager,
+} from "@earendil-works/pi-coding-agent";
 import { getSupportedThinkingLevels } from "@earendil-works/pi-ai";
-import { loadModelsWithCache, withModelRuntimeError, type ModelsData } from "@/lib/models-cache";
-import { getAllowedFileRoots, isExistingFilePathAllowed } from "@/lib/file-access";
+import {
+	loadModelsWithCache,
+	withModelRuntimeError,
+	type ModelsData,
+} from "@/lib/models-cache";
+import {
+	getAllowedFileRoots,
+	isExistingFilePathAllowed,
+} from "@/lib/file-access";
 import { projectTrustReloadOptions } from "@/lib/project-trust";
-import { createAppSettingsManager, getAppResourceLoaderOptions } from "@/lib/app-runtime";
+import {
+	createAppSettingsManager,
+	getAppResourceLoaderOptions,
+} from "@/lib/app-runtime";
 
 export const dynamic = "force-dynamic";
 
-const modelNameCollator = new Intl.Collator(undefined, { numeric: true, sensitivity: "base" });
+const modelNameCollator = new Intl.Collator(undefined, {
+	numeric: true,
+	sensitivity: "base",
+});
 
 function compareModelEntries(
-  a: { id: string; name: string; provider: string },
-  b: { id: string; name: string; provider: string }
+	a: { id: string; name: string; provider: string },
+	b: { id: string; name: string; provider: string },
 ): number {
-  return modelNameCollator.compare(a.name || a.id, b.name || b.id)
-    || modelNameCollator.compare(a.provider, b.provider)
-    || modelNameCollator.compare(a.id, b.id);
+	return (
+		modelNameCollator.compare(a.name || a.id, b.name || b.id) ||
+		modelNameCollator.compare(a.provider, b.provider) ||
+		modelNameCollator.compare(a.id, b.id)
+	);
 }
 
-const THINKING_SUFFIXES = new Set(["off", "minimal", "low", "medium", "high", "xhigh", "max"]);
+const THINKING_SUFFIXES = new Set([
+	"off",
+	"minimal",
+	"low",
+	"medium",
+	"high",
+	"xhigh",
+	"max",
+]);
 
 function stripThinkingSuffix(modelRef: string): string {
-  const trimmed = modelRef.trim();
-  const colonIndex = trimmed.lastIndexOf(":");
-  if (colonIndex === -1) return trimmed;
-  const suffix = trimmed.substring(colonIndex + 1);
-  return THINKING_SUFFIXES.has(suffix) ? trimmed.substring(0, colonIndex) : trimmed;
+	const trimmed = modelRef.trim();
+	const colonIndex = trimmed.lastIndexOf(":");
+	if (colonIndex === -1) return trimmed;
+	const suffix = trimmed.substring(colonIndex + 1);
+	return THINKING_SUFFIXES.has(suffix)
+		? trimmed.substring(0, colonIndex)
+		: trimmed;
 }
 
 function filterByExactEnabledModels<T extends { id: string; provider: string }>(
-  available: readonly T[],
-  enabledModels: string[] | undefined,
+	available: readonly T[],
+	enabledModels: string[] | undefined,
 ): readonly T[] {
-  if (!enabledModels || enabledModels.length === 0) return available;
+	if (!enabledModels || enabledModels.length === 0) return available;
 
-  const refs = new Set(enabledModels.map(stripThinkingSuffix).filter(Boolean));
-  const visible = available.filter((m) => refs.has(`${m.provider}/${m.id}`) || refs.has(m.id));
-  return visible.length > 0 ? visible : available;
+	const refs = new Set(enabledModels.map(stripThinkingSuffix).filter(Boolean));
+	const visible = available.filter(
+		(m) => refs.has(`${m.provider}/${m.id}`) || refs.has(m.id),
+	);
+	return visible.length > 0 ? visible : available;
 }
 
 async function loadModels(cwd: string): Promise<ModelsData> {
-  const nameMap = new Map<string, string>();
-  let modelList: { id: string; name: string; provider: string }[] = [];
-  let defaultModel: { provider: string; modelId: string } | null = null;
-  const thinkingLevels: Record<string, string[]> = {};
-  const thinkingLevelMaps: Record<string, Record<string, string | null>> = {};
+	const nameMap = new Map<string, string>();
+	let modelList: { id: string; name: string; provider: string }[] = [];
+	let defaultModel: { provider: string; modelId: string } | null = null;
+	const thinkingLevels: Record<string, string[]> = {};
+	const thinkingLevelMaps: Record<string, Record<string, string | null>> = {};
 
-  const agentDir = getAgentDir();
-  // Gate untrusted project extensions: enumerating models still imports and
-  // runs a repository's .pi/extensions factories, so honor project trust here
-  // too (see lib/project-trust.ts, #236).
-  const trustReloadOptions = projectTrustReloadOptions(cwd, agentDir);
-  const services = await createAgentSessionServices({
-    cwd,
-    agentDir,
-    settingsManager: createAppSettingsManager(cwd, agentDir),
-    resourceLoaderOptions: getAppResourceLoaderOptions(),
-    ...(trustReloadOptions ? { resourceLoaderReloadOptions: trustReloadOptions } : {}),
-  });
-  const available = await services.modelRuntime.getAvailable();
-  const modelError = services.modelRuntime.getError();
-  const settings: SettingsManager = services.settingsManager;
-  const enabledModels = settings.getEnabledModels();
-  const visible = filterByExactEnabledModels(available, enabledModels);
-  modelList = visible.map((m: { id: string; name: string; provider: string }) => ({
-    id: m.id,
-    name: m.name,
-    provider: m.provider,
-  })).sort(compareModelEntries);
-  for (const m of visible) {
-    const key = `${m.provider}:${m.id}`;
-    nameMap.set(key, m.name);
-    thinkingLevels[key] = getSupportedThinkingLevels(m);
-    if (m.thinkingLevelMap) thinkingLevelMaps[key] = m.thinkingLevelMap;
-  }
+	const agentDir = getAgentDir();
+	// Gate untrusted project extensions: enumerating models still imports and
+	// runs a repository's .pi/extensions factories, so honor project trust here
+	// too (see lib/project-trust.ts, #236).
+	const trustReloadOptions = projectTrustReloadOptions(cwd, agentDir);
+	const services = await createAgentSessionServices({
+		cwd,
+		agentDir,
+		settingsManager: createAppSettingsManager(cwd, agentDir),
+		resourceLoaderOptions: getAppResourceLoaderOptions(),
+		...(trustReloadOptions
+			? { resourceLoaderReloadOptions: trustReloadOptions }
+			: {}),
+	});
+	const available = await services.modelRuntime.getAvailable();
+	const modelError = services.modelRuntime.getError();
+	const settings: SettingsManager = services.settingsManager;
+	const enabledModels = settings.getEnabledModels();
+	const visible = filterByExactEnabledModels(available, enabledModels);
+	modelList = visible
+		.map(
+			(m: {
+				id: string;
+				name: string;
+				provider: string;
+				serviceTier?: string;
+			}) => ({
+				id: m.id,
+				name: m.name,
+				provider: m.provider,
+				serviceTier: m.serviceTier,
+			}),
+		)
+		.sort(compareModelEntries);
+	for (const m of visible) {
+		const key = `${m.provider}:${m.id}`;
+		nameMap.set(key, m.name);
+		thinkingLevels[key] = getSupportedThinkingLevels(m);
+		if (m.thinkingLevelMap) thinkingLevelMaps[key] = m.thinkingLevelMap;
+	}
 
-  const provider = settings.getDefaultProvider();
-  const modelId = settings.getDefaultModel();
-  if (provider && modelId && visible.some((m) => m.provider === provider && m.id === modelId)) {
-    defaultModel = { provider, modelId };
-  }
+	const provider = settings.getDefaultProvider();
+	const modelId = settings.getDefaultModel();
+	if (
+		provider &&
+		modelId &&
+		visible.some((m) => m.provider === provider && m.id === modelId)
+	) {
+		defaultModel = { provider, modelId };
+	}
 
-  return withModelRuntimeError(
-    { models: Object.fromEntries(nameMap), modelList, defaultModel, thinkingLevels, thinkingLevelMaps },
-    modelError,
-  );
+	return withModelRuntimeError(
+		{
+			models: Object.fromEntries(nameMap),
+			modelList,
+			defaultModel,
+			thinkingLevels,
+			thinkingLevelMaps,
+		},
+		modelError,
+	);
 }
 
 const EMPTY_MODELS: ModelsData = {
-  models: {},
-  modelList: [],
-  defaultModel: null,
-  thinkingLevels: {},
-  thinkingLevelMaps: {},
+	models: {},
+	modelList: [],
+	defaultModel: null,
+	thinkingLevels: {},
+	thinkingLevelMaps: {},
 };
 
 export async function GET(req: Request) {
-  const requestedCwd = new URL(req.url).searchParams.get("cwd") || process.cwd();
-  const cwd = resolve(requestedCwd);
+	const requestedCwd =
+		new URL(req.url).searchParams.get("cwd") || process.cwd();
+	const cwd = resolve(requestedCwd);
 
-  let cwdStat;
-  try {
-    cwdStat = await stat(cwd);
-  } catch {
-    return Response.json({ error: `Directory does not exist: ${cwd}` }, { status: 400 });
-  }
-  if (!cwdStat.isDirectory()) {
-    return Response.json({ error: `Not a directory: ${cwd}` }, { status: 400 });
-  }
-  const allowedRoots = await getAllowedFileRoots();
-  if (!isExistingFilePathAllowed(cwd, allowedRoots)) {
-    return Response.json({ error: "Access denied" }, { status: 403 });
-  }
+	let cwdStat;
+	try {
+		cwdStat = await stat(cwd);
+	} catch {
+		return Response.json(
+			{ error: `Directory does not exist: ${cwd}` },
+			{ status: 400 },
+		);
+	}
+	if (!cwdStat.isDirectory()) {
+		return Response.json({ error: `Not a directory: ${cwd}` }, { status: 400 });
+	}
+	const allowedRoots = await getAllowedFileRoots();
+	if (!isExistingFilePathAllowed(cwd, allowedRoots)) {
+		return Response.json({ error: "Access denied" }, { status: 403 });
+	}
 
-  try {
-    return Response.json(await loadModelsWithCache(cwd, () => loadModels(cwd)));
-  } catch {
-    return Response.json(EMPTY_MODELS);
-  }
+	try {
+		return Response.json(await loadModelsWithCache(cwd, () => loadModels(cwd)));
+	} catch {
+		return Response.json(EMPTY_MODELS);
+	}
 }

pi-web/lib/models-cache.ts

@@ -1,80 +1,94 @@
 export interface ModelsData {
-  models: Record<string, string>;
-  modelList: { id: string; name: string; provider: string }[];
-  defaultModel: { provider: string; modelId: string } | null;
-  thinkingLevels: Record<string, string[]>;
-  thinkingLevelMaps: Record<string, Record<string, string | null>>;
-  modelError?: string;
+	models: Record<string, string>;
+	modelList: {
+		id: string;
+		name: string;
+		provider: string;
+		serviceTier?: string;
+	}[];
+	defaultModel: { provider: string; modelId: string } | null;
+	thinkingLevels: Record<string, string[]>;
+	thinkingLevelMaps: Record<string, Record<string, string | null>>;
+	modelError?: string;
 }
 
 interface ModelsCacheState {
-  entries: Map<string, { data: ModelsData; expiresAt: number }>;
-  inFlight: Map<string, Promise<ModelsData>>;
-  generation: number;
+	entries: Map<string, { data: ModelsData; expiresAt: number }>;
+	inFlight: Map<string, Promise<ModelsData>>;
+	generation: number;
 }
 
 declare global {
-  var __piModelsCacheState: ModelsCacheState | undefined;
+	var __piModelsCacheState: ModelsCacheState | undefined;
 }
 
 const MODELS_CACHE_TTL_MS = 60_000;
 const MAX_MODELS_CACHE_ENTRIES = 32;
 
 function getModelsCacheState(): ModelsCacheState {
-  if (!globalThis.__piModelsCacheState) {
-    globalThis.__piModelsCacheState = {
-      entries: new Map(),
-      inFlight: new Map(),
-      generation: 0,
-    };
-  }
-  return globalThis.__piModelsCacheState;
+	if (!globalThis.__piModelsCacheState) {
+		globalThis.__piModelsCacheState = {
+			entries: new Map(),
+			inFlight: new Map(),
+			generation: 0,
+		};
+	}
+	return globalThis.__piModelsCacheState;
 }
 
 export function invalidateModelsCache(): void {
-  const state = getModelsCacheState();
-  state.generation += 1;
-  state.entries.clear();
-  state.inFlight.clear();
+	const state = getModelsCacheState();
+	state.generation += 1;
+	state.entries.clear();
+	state.inFlight.clear();
 }
 
-export function withModelRuntimeError(data: ModelsData, modelError: string | undefined): ModelsData {
-  return modelError ? { ...data, modelError } : data;
+export function withModelRuntimeError(
+	data: ModelsData,
+	modelError: string | undefined,
+): ModelsData {
+	return modelError ? { ...data, modelError } : data;
 }
 
-export function loadModelsWithCache(cwd: string, loader: () => Promise<ModelsData>): Promise<ModelsData> {
-  const state = getModelsCacheState();
-  const cached = state.entries.get(cwd);
-  if (cached) {
-    if (cached.expiresAt > Date.now()) return Promise.resolve(cached.data);
-    state.entries.delete(cwd);
-  }
+export function loadModelsWithCache(
+	cwd: string,
+	loader: () => Promise<ModelsData>,
+): Promise<ModelsData> {
+	const state = getModelsCacheState();
+	const cached = state.entries.get(cwd);
+	if (cached) {
+		if (cached.expiresAt > Date.now()) return Promise.resolve(cached.data);
+		state.entries.delete(cwd);
+	}
 
-  const existingLoad = state.inFlight.get(cwd);
-  if (existingLoad) return existingLoad;
+	const existingLoad = state.inFlight.get(cwd);
+	if (existingLoad) return existingLoad;
 
-  const generation = state.generation;
-  const loadPromise: Promise<ModelsData> = Promise.resolve()
-    .then(loader)
-    .then((data) => {
-      if (state.generation === generation && state.inFlight.get(cwd) === loadPromise) {
-        const now = Date.now();
-        for (const [key, entry] of state.entries) {
-          if (entry.expiresAt <= now) state.entries.delete(key);
-        }
-        while (state.entries.size >= MAX_MODELS_CACHE_ENTRIES) {
-          const oldestKey = state.entries.keys().next().value;
-          if (oldestKey === undefined) break;
-          state.entries.delete(oldestKey);
-        }
-        state.entries.set(cwd, { data, expiresAt: now + MODELS_CACHE_TTL_MS });
-      }
-      return data;
-    })
-    .finally(() => {
-      if (state.inFlight.get(cwd) === loadPromise) state.inFlight.delete(cwd);
-    });
+	const generation = state.generation;
+	const loadPromise: Promise<ModelsData> = Promise.resolve()
+		.then(loader)
+		.then((data) => {
+			if (
+				state.generation === generation &&
+				state.inFlight.get(cwd) === loadPromise
+			) {
+				const now = Date.now();
+				for (const [key, entry] of state.entries) {
+					if (entry.expiresAt <= now) state.entries.delete(key);
+				}
+				while (state.entries.size >= MAX_MODELS_CACHE_ENTRIES) {
+					const oldestKey = state.entries.keys().next().value;
+					if (oldestKey === undefined) break;
+					state.entries.delete(oldestKey);
+				}
+				state.entries.set(cwd, { data, expiresAt: now + MODELS_CACHE_TTL_MS });
+			}
+			return data;
+		})
+		.finally(() => {
+			if (state.inFlight.get(cwd) === loadPromise) state.inFlight.delete(cwd);
+		});
 
-  state.inFlight.set(cwd, loadPromise);
-  return loadPromise;
+	state.inFlight.set(cwd, loadPromise);
+	return loadPromise;
 }

pi/packages/ai/src/api/simple-options.ts

@@ -41,6 +41,7 @@ export function buildBaseOptions(
 		maxRetries: options?.maxRetries,
 		maxRetryDelayMs: options?.maxRetryDelayMs,
 		metadata: options?.metadata,
+		serviceTier: options?.serviceTier,
 		env: options?.env,
 	};
 }

pi/packages/ai/src/types.ts

@@ -113,6 +113,13 @@ export interface ProviderResponse {
 	headers: Record<string, string>;
 }
 
+/**
+ * Request-level service tier for providers that support it.
+ * OpenAI Codex / Responses: "priority" is fast mode (higher cost, lower latency), "flex" is the cheaper tier.
+ * See https://developers.openai.com/codex/speed for Codex fast mode semantics.
+ */
+export type ServiceTier = "auto" | "default" | "flex" | "priority" | "scale";
+
 export interface StreamOptions {
 	temperature?: number;
 	maxTokens?: number;
@@ -189,6 +196,12 @@ export interface StreamOptions {
 	 * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.
 	 */
 	metadata?: Record<string, unknown>;
+	/**
+	 * Request-level service tier. Providers that understand it (OpenAI Codex,
+	 * OpenAI Responses) send it as `service_tier`; other providers ignore it.
+	 * "priority" is Codex fast mode (1.5x speed, ~2x cost).
+	 */
+	serviceTier?: ServiceTier | null;
 	/**
 	 * Provider-scoped environment values. These take precedence over process.env for
 	 * provider configuration such as regional settings, endpoint placeholders, and
@@ -773,6 +786,12 @@ export interface Model<TApi extends Api> {
 	contextWindow: number;
 	maxTokens: number;
 	headers?: Record<string, string>;
+	/**
+	 * Optional default request-level service tier for this model (e.g. "priority"
+	 * for Codex fast mode). Configured via models.json; forwarded to providers
+	 * that support `service_tier`.
+	 */
+	serviceTier?: ServiceTier;
 	/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */
 	compat?: TApi extends "openai-completions"
 		? OpenAICompletionsCompat

pi/packages/coding-agent/src/core/model-config.ts

@@ -151,6 +151,13 @@ const ModelCostSchema = Type.Object({
 	tiers: Type.Optional(Type.Array(ModelCostTierSchema)),
 });
 
+const ServiceTierSchema = Type.Union([
+	Type.Literal("auto"),
+	Type.Literal("default"),
+	Type.Literal("priority"),
+	Type.Literal("flex"),
+]);
+
 const ModelDefinitionSchema = Type.Object({
 	id: Type.String({ minLength: 1 }),
 	name: Type.Optional(Type.String({ minLength: 1 })),
@@ -163,6 +170,7 @@ const ModelDefinitionSchema = Type.Object({
 	contextWindow: Type.Optional(Type.Number()),
 	maxTokens: Type.Optional(Type.Number()),
 	headers: Type.Optional(Type.Record(Type.String(), Type.String())),
+	serviceTier: Type.Optional(ServiceTierSchema),
 	compat: Type.Optional(ProviderCompatSchema),
 });
 
@@ -183,6 +191,7 @@ const ModelOverrideSchema = Type.Object({
 	contextWindow: Type.Optional(Type.Number()),
 	maxTokens: Type.Optional(Type.Number()),
 	headers: Type.Optional(Type.Record(Type.String(), Type.String())),
+	serviceTier: Type.Optional(ServiceTierSchema),
 	compat: Type.Optional(ProviderCompatSchema),
 });
 

pi/packages/coding-agent/src/core/provider-composer.ts

@@ -117,6 +117,7 @@ function applyModelOverride(model: Model<Api>, override: ModelsJsonModelOverride
 			: model.cost,
 		contextWindow: override.contextWindow ?? model.contextWindow,
 		maxTokens: override.maxTokens ?? model.maxTokens,
+		serviceTier: override.serviceTier ?? model.serviceTier,
 		compat: mergeCompat(model.compat, override.compat),
 	};
 }
@@ -154,6 +155,7 @@ function modelFromJson(
 		contextWindow: definition.contextWindow ?? 128000,
 		maxTokens: definition.maxTokens ?? 16384,
 		headers: undefined,
+		serviceTier: definition.serviceTier,
 		compat: mergeCompat(providerConfig.compat, definition.compat),
 	};
 }

pi/packages/coding-agent/src/core/sdk.ts

@@ -309,8 +309,13 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			const websocketConnectTimeoutMs =
 				options?.websocketConnectTimeoutMs ?? settingsManager.getWebSocketConnectTimeoutMs();
 			const headerRunner = extensionRunnerRef.current;
+			// Model-level service tier (models.json `serviceTier`, e.g. "priority" for
+			// Codex fast mode) is forwarded as request `service_tier`. Providers that
+			// don't understand it ignore the option.
+			const serviceTier = model.serviceTier;
 			return modelRuntime.streamSimple(model, context, {
 				...options,
+				...(serviceTier ? { serviceTier } : {}),
 				timeoutMs,
 				websocketConnectTimeoutMs,
 				maxRetries: options?.maxRetries ?? providerRetrySettings.maxRetries,

scripts/profile.mjs

@@ -1,7 +1,7 @@
 import { spawnSync } from "node:child_process";
-import { copyFileSync, existsSync, mkdirSync } from "node:fs";
+import { copyFileSync, existsSync, lstatSync, mkdirSync, symlinkSync } from "node:fs";
 import { platform } from "node:os";
-import { dirname, join, resolve } from "node:path";
+import { delimiter, dirname, join, resolve } from "node:path";
 import { loadEnvFile } from "node:process";
 import { fileURLToPath } from "node:url";
 
@@ -55,6 +55,56 @@ const seedFiles = [
 ];
 
 const persistedWindowsEnvironmentKeys = ["SEARXNG_TOKEN", "SEARXNG_URL"];
+const piWebBinDir = join(rootDir, "pi-web", "node_modules", ".bin");
+const piCodingAgentPackageDir = join(
+  rootDir,
+  "pi-web",
+  "node_modules",
+  "@earendil-works",
+  "pi-coding-agent",
+);
+const managedPeerPackageDir = join(
+  agentDir,
+  "npm",
+  "node_modules",
+  "@earendil-works",
+  "pi-coding-agent",
+);
+
+function ensurePiSubagentRuntime() {
+  // pi-subagents resolves the host CLI through its optional coding-agent peer.
+  // The managed profile is installed separately from pi-web, so expose the
+  // existing local package instead of installing a second coding-agent copy.
+  if (!existsSync(piCodingAgentPackageDir)) return;
+
+  mkdirSync(dirname(managedPeerPackageDir), { recursive: true });
+  if (existsSync(managedPeerPackageDir)) return;
+  try {
+    if (lstatSync(managedPeerPackageDir)) return;
+  } catch {
+    // The peer link is absent; create it below.
+  }
+
+  try {
+    symlinkSync(
+      piCodingAgentPackageDir,
+      managedPeerPackageDir,
+      platform() === "win32" ? "junction" : "dir",
+    );
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error);
+    console.error(`[subagents] could not link the local Pi CLI package: ${message}`);
+  }
+}
+
+function prependPiWebBinToPath(baseEnv) {
+  if (!existsSync(piWebBinDir)) return baseEnv.PATH ?? baseEnv.Path;
+  const pathKey = baseEnv.PATH !== undefined || baseEnv.Path === undefined ? "PATH" : "Path";
+  const currentPath = baseEnv[pathKey] ?? "";
+  const entries = currentPath.split(delimiter).filter(Boolean);
+  if (!entries.includes(piWebBinDir)) entries.unshift(piWebBinDir);
+  return entries.join(delimiter);
+}
 
 function readPersistedWindowsEnvironment(baseEnv) {
   if (platform() !== "win32") return {};
@@ -112,10 +162,12 @@ export function ensureProfile({ quiet = false } = {}) {
 
 export function managedEnvironment(baseEnv = process.env) {
   ensureProfile({ quiet: true });
+  ensurePiSubagentRuntime();
   const persistedEnvironment = readPersistedWindowsEnvironment(baseEnv);
   return {
     ...baseEnv,
     ...persistedEnvironment,
+    PATH: prependPiWebBinToPath({ ...baseEnv, ...persistedEnvironment }),
     PI_AGENT_MANAGED_RUNTIME: "1",
     PI_AGENT_APP_ROOT: rootDir,
     PI_AGENT_DATA_DIR: dataDir,