Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

Merge remote-tracking branch 'upstream/main' into fix/pi-thinking-level-on-model-switch

Zhang Yunjin58cf2d713a7e
diff --cc apps/desktop/src/main/maker-ipc/register.ts
index c64683204e,1b2d749226..b823efcd73
--- a/apps/desktop/src/main/maker-ipc/register.ts
+++ b/apps/desktop/src/main/maker-ipc/register.ts
@@@ -8661,7 -9152,8 +9171,9 @@@ export function registerMakerIpc(maker
const result = await applySessionRuntimeSelection(sessionId, intent.model, intent.providerId, {
effort: (intent.effort ?? null) as SessionRuntimeProfile['effort'],
fastMode: intent.fastMode ?? false,
+ ...(intent.thinking !== undefined ? { thinking: intent.thinking } : {}),
+ // 只有 Agent 在另一台电脑运行的任务会在意图里带确认过的目标窗口。
+ ...(intent.confirmedContextWindow ? { confirmedContextWindow: intent.confirmedContextWindow } : {}),
}, { source: 'user', sessionLockHeld: true, applyingUserSelectionOnSend: applyNow,
runtimeSource: intent.runtimeSource, configStaged: intent.configStaged === true,
assertSelectionCurrent });
diff --cc apps/desktop/src/main/maker-ipc/sessionAgentSwitchHandler.ts
index 83daf57005,2bfb330485..1a23757860
--- a/apps/desktop/src/main/maker-ipc/sessionAgentSwitchHandler.ts
+++ b/apps/desktop/src/main/maker-ipc/sessionAgentSwitchHandler.ts
@@@ -271,12 -299,13 +306,19 @@@ export interface PendingAgentSwitchInte
providerId: string | null | undefined;
effort?: string;
fastMode?: boolean;
+ /**
+ * Agent 在另一台电脑运行的任务:点选时按那台的目录确认过的目标上下文窗口。本机目录里没有
+ * 那台的模型,发送时据此做「换小窗口前先交接」。本机任务不设。
+ */
+ confirmedContextWindow?: number;
+ /** 选择同时换 Agent 所在电脑(null = 任务所在电脑);缺省 = 位置不变。 */
+ targetAgentDeviceId?: string | null;
+ /**
+ * Pi-only:目标模型的思考开关意图(renderer 记忆,缺省开)。发送时刻回放暂存
+ * 选择时必须一并应用,否则上一模型的 off 会残留——Pi 的 set_model 不重置
+ * thinking level。
+ */
+ thinking?: boolean;
/** resume 回落事务失败后的内部恢复载荷;下一次 send 先重试这笔原子事务。 */
resumeFallbackRecovery?: {
boundaryClientId: string | null;
@@@ -391,11 -412,11 +436,16 @@@ export async function performSessionAge
/** 目标引擎下应生效的 effort / fastMode(renderer 按目标目录与预设解析好带入)。 */
effort?: unknown;
fastMode?: unknown;
+ /**
+ * 选择的 Agent 位置(远程 Agent):undefined = 保持任务当前位置(内部调用与不认识该字段的
+ * 旧控制端都走这里,不做推断),null = 任务所在电脑,字符串 = 同账号另一台电脑。
+ */
+ agentDeviceId?: unknown;
+ /**
+ * Pi-only:目标模型的思考开关意图。重登记(意图期内改档位 / 改选 / 切思考
+ * 开关)时必须带上;缺省时由同引擎分支继承既有 intent,不能静默丢字段。
+ */
+ thinking?: unknown;
/**
* pending-apply 路径(send 事务派发前执行):跳过新引擎立即重建——send 随后
* 的 lazy-create 会按 DB 新值 spawn(reconcileCreateOptsWithDb 校正兜底),
@@@ -579,7 -594,7 +656,8 @@@
providerId: normalizedProviderId,
...(typeof params.effort === 'string' && params.effort ? { effort: params.effort } : {}),
...(typeof params.fastMode === 'boolean' ? { fastMode: params.fastMode } : {}),
+ ...(typeof thinking === 'boolean' ? { thinking } : {}),
+ ...(agentDeviceChanges ? { targetAgentDeviceId } : {}),
};
deps.pendingSwitches.set(sessionId, intent);
deps.onPendingSwitchChanged?.(sessionId, projectPendingAgentSwitchIntent(intent));
@@@ -980,7 -1002,7 +1075,8 @@@ export function applyPendingAgentSwitch
providerId: intent.providerId,
effort: intent.effort,
fastMode: intent.fastMode,
+ thinking: intent.thinking,
+ agentDeviceId: intent.targetAgentDeviceId,
// 普通 maker send 随后会按 DB 真源 lazy-create,无需在 apply 内多 spawn 一次;
// Goal / IM / scheduler 三条直发路径没有该 lazy-create 事务,必须先把新引擎
// bootstrap 好再拿 live session 发送,否则会继续持有已关闭的旧 session。
@@@ -1048,9 -1072,19 +1146,20 @@@ export function registerMakerSessionAge
providerId: unknown,
effort: unknown,
fastMode: unknown,
- thinking: unknown,
+ // 第 7 参(可选):{ agentDeviceId } —— 同时换 Agent 所在电脑。旧被控端忽略多余参数,
+ // 位置保持不变(与今天一致)。
+ options?: unknown,
++ thinking?: unknown,
) => {
const context = getDeviceLinkInvokeContext();
+ const agentDeviceId =
+ options && typeof options === 'object' && !Array.isArray(options) && 'agentDeviceId' in options
+ ? (options as { agentDeviceId?: unknown }).agentDeviceId
+ : undefined;
+ // 共享任务的访客只能在主人给的范围内改设置,不能把 Agent 挪到主人的其他电脑上。
+ if (agentDeviceId !== undefined && context?.sharedTask) {
+ throwIpcError('UNSUPPORTED_CAPABILITY', 'shared task guests cannot move the agent to another computer');
+ }
const guard = createSharedTaskSettingGuard(context?.sharedTask, String(sessionId), context?.sharedTaskSetting ?? { admitted: false });
const run = () => performSessionAgentSwitch(deps, {
sessionId,
@@@ -1059,7 -1093,7 +1168,8 @@@
providerId,
effort,
fastMode,
+ ...(agentDeviceId !== undefined ? { agentDeviceId } : {}),
+ thinking,
assertSelectionCurrent: guard.admit,
});
return typeof sessionId === 'string' && sessionId && deps.withSessionLock
diff --cc apps/desktop/src/preload/preload.ts
index efef445bb3,2397af7468..03c1521a1e
--- a/apps/desktop/src/preload/preload.ts
+++ b/apps/desktop/src/preload/preload.ts
@@@ -6729,7 -6920,8 +6922,9 @@@ contextBridge.exposeInMainWorld('electr
providerId?: string | null,
effort?: string,
fastMode?: boolean,
+ // 远程 Agent:同时换 Agent 所在电脑(null = 任务所在电脑)。不传 = 位置不变。
+ options?: { agentDeviceId?: string | null },
+ thinking?: boolean,
): Promise<{
switched: boolean;
agentKind: 'claude-code' | 'codex' | 'pi';
@@@ -6739,16 -6931,26 +6934,27 @@@
sameEngineRevision?: number;
sameEngineSuperseded?: boolean;
}> =>
- ipcRenderer.invoke(
- 'maker:switch-session-agent',
- sessionId,
- targetAgentKind,
- model,
- providerId,
- effort,
- fastMode,
- thinking,
- ),
- options
++ options || thinking !== undefined
+ ? ipcRenderer.invoke(
+ 'maker:switch-session-agent',
+ sessionId,
+ targetAgentKind,
+ model,
+ providerId,
+ effort,
+ fastMode,
+ options,
++ thinking,
+ )
+ : ipcRenderer.invoke(
+ 'maker:switch-session-agent',
+ sessionId,
+ targetAgentKind,
+ model,
+ providerId,
+ effort,
+ fastMode,
+ ),
// 读 main 权威的 pending 切换意图(内存态,不落库)。重开视图 / 远程会话重连后
// 用它恢复乐观显示——否则用户登记的意图在 UI 上凭空消失,下一条消息却按意图切换。
getSessionAgentSwitchIntent: (
diff --cc apps/desktop/src/renderer/components/new-chat/ChatInput.tsx
index 0885fb82c2,39d4e13799..65f802e13b
--- a/apps/desktop/src/renderer/components/new-chat/ChatInput.tsx
+++ b/apps/desktop/src/renderer/components/new-chat/ChatInput.tsx
@@@ -6798,7 -6651,11 +6673,12 @@@ export function ChatInput(
overrides?: {
effort?: Effort;
fastMode?: boolean;
+ /**
+ * 远程 Agent:这一行属于哪台电脑的目录(null = 本机)。与任务当前所在电脑不同 = 连 Agent
+ * 的运行位置一起换。不传 = 沿用已登记的换位置意图,没有就是位置不变。
+ */
+ agentDeviceId?: string | null;
+ thinking?: boolean;
},
): Promise<boolean> => {
// ★ 返回值 = **本端请求的完整配置真的落到会话上了没有**(2026-08-17 review 确立
@@@ -6885,15 -6756,24 +6787,27 @@@
const switchApi = deviceLinkDeviceId
? makerApiForDevice(deviceLinkDeviceId)
: makerApiFor(sourceSessionId);
- const result = await switchApi.switchSessionAgent(
- sourceSessionId,
- targetAgentKind,
- newModelId,
- providerId,
- newEffort,
- targetFast,
- targetThinking,
- );
+ const result = relocateTo !== undefined
+ ? await switchApi.switchSessionAgent(
+ sourceSessionId,
+ targetAgentKind,
+ newModelId,
+ providerId,
+ newEffort,
+ targetFast,
+ { agentDeviceId: relocateTo },
++ targetThinking,
+ )
+ : await switchApi.switchSessionAgent(
+ sourceSessionId,
+ targetAgentKind,
+ newModelId,
+ providerId,
+ newEffort,
+ targetFast,
++ undefined,
++ targetThinking,
+ );
// device-link 往返期间可以切到另一个任务:同一路由下 ChatInput 会带着新
// sessionId 继续渲染,sameEngineReselectRef 等闭包也已指向新会话。旧会话的
// 响应绝不能借最新的 ref 把模型/来源写进当前会话。
diff --cc apps/desktop/src/renderer/vite-env.d.ts
index 16708caa7c,d39976cb4e..d939ee3f7b
--- a/apps/desktop/src/renderer/vite-env.d.ts
+++ b/apps/desktop/src/renderer/vite-env.d.ts
@@@ -6143,7 -6255,8 +6255,9 @@@ interface ElectronAPI
providerId?: string | null,
effort?: string,
fastMode?: boolean,
+ /** 远程 Agent:同时换 Agent 所在电脑(null = 任务所在电脑)。不传 = 位置不变。 */
+ options?: { agentDeviceId?: string | null },
+ thinking?: boolean,
) => Promise<{
switched: boolean;
agentKind: 'claude-code' | 'codex' | 'pi';
diff --cc packages/maker-core/src/agents/base-agent.ts
index cc21839146,01f79fd0c0..1fc9325548
--- a/packages/maker-core/src/agents/base-agent.ts
+++ b/packages/maker-core/src/agents/base-agent.ts
@@@ -2319,9 -2380,15 +2380,15 @@@ export interface AgentSessionHandle
/** Host tool approval uses the same live intent/model/scope as native tool approval. */
reviewAutoPermissionAction?(action: ReviewableAction): Promise<AutoReviewDecision>;
- /** 运行时切换模型 —— 不支持时抛 NotSupportedError */
- setModel?(model: string, opts?: { providerId?: string | null; effort?: Effort }): Promise<void>;
+ /** 运行时切换模型 —— 不支持时抛 NotSupportedError。thinkingEnabled = 目标模型的思考开关意图(Pi 用它收敛切模后的 thinking level)。 */
+ setModel?(model: string, opts?: { providerId?: string | null; effort?: Effort; thinkingEnabled?: boolean }): Promise<void>;
+ /** Read-only Pi preflight before the host changes its persisted route or context. */
+ previewModelSwitch?(
+ model: string,
+ opts?: { providerId?: string | null },
+ ): Promise<PiModelSwitchPreview>;
+
/**
* 当前 provider handle 是否必须先关闭、再由同一业务任务 cold resume 才能应用目标模型。
* 缺省 false;用于 Codex 这类把部分模型配置冻结在 app-server spawn / thread resume
diff --cc packages/maker-core/src/agents/pi/index.ts
index a4a322a4f2,80d3530cd4..9ed7f9e285
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@@ -3819,11 -4016,8 +4016,12 @@@ export class PiAgent extends BaseAgent
let mutableProviderId: string | null | undefined = opts.providerId ?? authProviderId;
let activeEffortSnapshot = initialEffortSnapshot;
let mutableEffort: Effort | null = startupEffort ?? null;
+ // 用户明确关掉思考(setThinkingEnabled(false) / 切模显式 thinkingEnabled:false /
+ // 启动 opts.thinkingEnabled === false)的意图标记:无载体切模不得据此悄悄重新开启。
+ // 只跟随显式意图变化;目标不支持思考而下发 off 不算用户意图,不置位。
+ let thinkingExplicitlyOff = false;
let currentAutoReviewIntent: AutoReviewUserIntent = '';
+ let autoReviewIntentInitialized = false;
const autoReviewActionContext = createAutoReviewActionContext();
const autoReviewContext = () => activeTurnPermissionPolicy?.autoReviewContext
?? (activeTurnPermissionPolicy?.origin.kind === 'im'
@@@ -6072,265 -6319,455 +6328,455 @@@
'[PI_CATALOG_RELOAD_UNCONFIRMED] 模型目录重载未确认,已终止本任务。请重新打开任务后再切换模型。',
);
};
- const restoreNativeCatalog = async (
- previousProviders: PiNativeProviderSpec[],
- ): Promise<void> => {
- // 先写盘再改内存:写失败时磁盘仍是新目录,内存也保持新目录,避免分叉后再抛。
- const written = await this.writeModelsJson(
- configHome,
- previousProviders,
- retainedRuntimeModel,
- authProviderId,
- {
- remote,
- fileOps,
- contextWindow: ctx.contextWindow || startupContextWindow,
- workingContextWindow: ctx.workingContextWindow,
- piCompactionPct: sessionPiAutoCompactPct,
- packages: nativePackagePaths,
- disabledSkills: disabledSkillLaunch,
- },
- );
- nativeProviders = previousProviders;
- nativeProviderById.clear();
- nativeProviderBySourceId.clear();
- for (const spec of previousProviders) {
- if (spec.id === PI_PROVIDER_ID) continue;
- nativeProviderById.set(spec.id, spec);
- nativeProviderBySourceId.set(spec.sourceProviderId ?? spec.id, spec);
- }
- gatewayApiByModel.clear();
- for (const [key, value] of written.gatewayApiByModel) gatewayApiByModel.set(key, value);
- gatewayImageInputByModel.clear();
- for (const [key, value] of written.gatewayImageInputByModel) {
- gatewayImageInputByModel.set(key, value);
- }
- };
- const restoreNativeCatalogOrTerminate = async (
- previousProviders: PiNativeProviderSpec[],
- ): Promise<void> => {
+ let liveRuntimeSettings: PiNativeRuntimeSettings | undefined;
+ const queryNativeRuntime = async (
+ operation: 'inspect' | 'refresh',
+ env: Record<string, string> = {},
+ aliases: ReturnType<typeof nativeProviderAdapterAliases> = [],
+ ): Promise<{ ok: boolean; code?: string }> => {
+ // Unknown extension commands become model prompts. Check ownership first.
+ const commands = await proc.request({ type: 'get_commands' });
+ const entries = (commands.data as { commands?: Array<{ name?: unknown; source?: unknown }> } | undefined)?.commands;
+ if (!commands.success || !entries?.some((entry) =>
+ entry.name === 'cindy-native-provider-refresh' && entry.source === 'extension')) {
+ throw new Error('Cindy native provider refresh extension is unavailable; reopen this task with the current Cindy version.');
+ }
+ const nonce = randomBytes(24).toString('base64url');
+ let timer: ReturnType<typeof setTimeout> | undefined;
+ const ack = new Promise<{ ok: boolean; code?: string; settings?: unknown }>((resolve) => {
+ providerRefreshStages.set(nonce, { nonce, operation, env, aliases,
+ acknowledge: (ok, code, settings) => resolve({ ok, code, settings }) });
+ timer = setTimeout(() => resolve({ ok: false, code: 'ACK_TIMEOUT' }), 10_000);
+ });
try {
- await restoreNativeCatalog(previousProviders);
- } catch (err) {
- await terminateUnconfirmedCatalogReload(err);
+ const response = await proc.request({ type: 'prompt', message: `/cindy-native-provider-refresh ${nonce}` });
+ if (!response.success) return { ok: false, code: 'COMMAND_FAILED' };
+ const receipt = await ack;
+ if (!receipt.ok) return receipt;
+ const settings = parsePiNativeRuntimeSettings(receipt.settings);
+ if (!settings) return { ok: false, code: 'INVALID_SETTINGS' };
+ liveRuntimeSettings = settings;
+ return { ok: true };
+ } finally {
+ if (timer) clearTimeout(timer);
+ providerRefreshStages.delete(nonce);
}
};
- const refreshLiveXaiCatalog = async (
- modelId: string,
- providerId?: string | null,
- ): Promise<boolean> => {
- if (!this.deps.resolvePiNativeProviders) return false;
- try {
- const live = await this.deps.resolvePiNativeProviders({
- workingDir: opts.workingDir,
- remoteHostId: opts.remoteHostId,
- providerId,
- model: modelId,
- resumeSessionId: sdkSessionId || opts.resumeSessionId,
- });
- const sourceProviderId = providerId ?? 'xai';
- const liveXai = live?.providers.find(
- (provider) => (provider.sourceProviderId ?? provider.id) === sourceProviderId,
- );
- if (!live || !liveXai) return false;
- // 远端 hostProxyForward 依赖启动时注入的 session token / 代理登记。
- // 登录后才出现的 xAI 块带上隧道,请求会因没有 CINDY_PI_SESSION_TOKEN 失败。
- if (liveXai.hostProxyForward && !proxySessionToken) return false;
- const currentXai = nativeProviderForSource(sourceProviderId);
- if (currentXai) {
- if (currentXai.baseUrl !== liveXai.baseUrl) return false;
- if (!sameRecord(currentXai.headers, liveXai.headers)) return false;
- if (liveXai.hostProxyForward && !currentXai.hostProxyForward) return false;
+ const reserveForWindow = (window: number, source: string | null | undefined, model: string): number => {
+ const settings = JSON.parse(buildPiSettingsJsonContent(window, sessionPiAutoCompactPct, [],
+ this.deps.resolveModelContextLimit?.(source, model) ?? undefined));
+ return settings.compaction?.reserveTokens ?? 16_384;
+ };
+ const previewPiModelSwitch = async (
+ model: string,
+ target?: { providerId?: string | null },
+ ): Promise<PiModelSwitchPreview> => {
+ const requestedProviderId = target && Object.hasOwn(target, 'providerId')
+ ? target.providerId
+ : mutableProviderId;
+ const sameRoute = model === mutableModel && requestedProviderId === mutableProviderId;
+ let latestProviders = nativeProviders;
+ let latestEnv = nativeEnv;
+ const resolveLatestNativeProviders = sessionNativeProviderResolver();
+ if (resolveLatestNativeProviders) {
+ try {
+ const resolved = await resolveLatestNativeProviders({
+ workingDir: opts.workingDir,
+ remoteHostId: opts.remoteHostId,
+ providerId: requestedProviderId,
+ model,
+ resumeSessionId: sdkSessionId || opts.resumeSessionId,
+ purpose: 'preview',
+ });
+ if (resolved) {
+ latestProviders = resolved.providers;
+ latestEnv = resolved.env;
+ }
+ } catch (error) {
+ return {
+ action: 'unavailable', targetContextWindow: null, windowVerified: false,
+ reason: 'Pi provider catalog could not be read',
+ };
}
- const envKey = liveXai.apiKeyEnvVar;
- if (envKey) {
- if (!(envKey in nativeEnv)) return false;
- if (live.env[envKey] !== undefined && live.env[envKey] !== nativeEnv[envKey]) {
- return false;
+ }
+ const sourceId = requestedProviderId === undefined
+ ? (model.startsWith('xai/') ? 'xai' : model.startsWith('chatgpt/') ? 'openai' : null)
+ : requestedProviderId;
+ const gateway = !sourceId || sourceId === 'xd' || sourceId === PI_PROVIDER_ID;
+ let nativeModel: PiNativeModelSpec | undefined;
+ let inheritsNativeWindow = false;
+ if (!gateway) {
+ let native = latestProviders.find((entry) => (entry.sourceProviderId ?? entry.id) === sourceId);
+ if (sameRoute) {
+ const running = nativeProviderForSource(sourceId);
+ const nextNativeId = native?.modelIdAliases?.[model] ?? model;
+ // A removed/retired model stays available to its current task. Never
+ // rewrite that task's active provider to a different endpoint merely
+ // because a new public catalog no longer advertises it.
+ if (running && (!native || !native.models.some((candidate) => candidate.id === nextNativeId))) {
+ latestProviders = [...latestProviders.filter((entry) =>
+ (entry.sourceProviderId ?? entry.id) !== sourceId), running];
+ // Keep the credential paired with the retained endpoint, without
+ // retaining credentials for other providers removed from the catalog.
+ if (running.apiKeyEnvVar && nativeEnv[running.apiKeyEnvVar] !== undefined) {
+ latestEnv = { ...latestEnv, [running.apiKeyEnvVar]: nativeEnv[running.apiKeyEnvVar] };
+ }
+ native = running;
}
}
- nativeProviderById.set(liveXai.id, liveXai);
- nativeProviderBySourceId.set(liveXai.sourceProviderId ?? liveXai.id, liveXai);
- nativeProviders = [
- ...nativeProviders.filter(
- (provider) => (provider.sourceProviderId ?? provider.id) !== sourceProviderId,
- ),
- liveXai,
- ];
- const written = await this.writeModelsJson(
- configHome,
- nativeProviders,
- retainedRuntimeModel,
- authProviderId,
- {
- remote,
- fileOps,
- contextWindow: ctx.contextWindow || startupContextWindow,
- workingContextWindow: ctx.workingContextWindow,
- piCompactionPct: sessionPiAutoCompactPct,
- packages: nativePackagePaths,
- disabledSkills: disabledSkillLaunch,
- },
- );
- gatewayApiByModel.clear();
- for (const [key, value] of written.gatewayApiByModel) gatewayApiByModel.set(key, value);
- gatewayImageInputByModel.clear();
- for (const [key, value] of written.gatewayImageInputByModel) {
- gatewayImageInputByModel.set(key, value);
+ if (!native) {
+ return { action: 'unavailable', targetContextWindow: null, windowVerified: false,
+ reason: `Pi provider '${sourceId}' is not available in the current catalog` };
+ }
+ const nativeId = native.modelIdAliases?.[model] ?? model;
+ nativeModel = native.models.find((candidate) => candidate.id === nativeId);
+ if (!nativeModel && sameRoute) {
+ const retained = nativeProviderForSource(sourceId);
+ nativeModel = retained?.models.find((candidate) =>
+ candidate.id === (retained.modelIdAliases?.[model] ?? model));
+ }
+ if (!nativeModel) {
+ return { action: 'unavailable', targetContextWindow: null, windowVerified: false,
+ reason: `Pi provider '${sourceId}' does not offer model '${model}'` };
+ }
+ inheritsNativeWindow = native.inheritModels === true && nativeModel.api === undefined
+ && nativeModel.catalogAddition !== true && nativeModel.contextWindow === undefined;
+ }
+ const gatewaySpec = gateway
+ ? this.deps.resolvePiGatewayModelSpec?.(sourceId, model, { remote })
+ : undefined;
+ if (gateway && gatewaySpec === null) {
+ return { action: 'unavailable', targetContextWindow: null, windowVerified: false,
+ reason: `No supported Gateway Pi API is available for model '${model}'` };
+ }
+ const knownDescriptor = this.deps.resolvePiRuntimeModelDescriptor
+ ? this.deps.resolvePiRuntimeModelDescriptor(sourceId, model)
+ : this.capabilities.availableModels.find((candidate) => candidate.id === model);
+ const workingWindow = this.deps.resolveModelContextLimit?.(sourceId, model) ?? 0;
+ let configuredWindow = nativeModel ? nativeModel.contextWindow ?? 128_000
+ : knownDescriptor?.contextWindow ?? (sameRoute && gateway ? retainedLiveGatewayModel?.contextWindow : undefined);
+ if (inheritsNativeWindow) {
+ // 128K is only a default for models Cindy materializes. An inherited
+ // entry uses either our explicit model override or Pi's live catalog.
+ configuredWindow = workingWindow > 0 ? workingWindow : undefined;
+ if (!configuredWindow) {
+ try {
+ const available = await proc.request({ type: 'get_available_models' });
+ const entries = (available.data as { models?: Array<{ provider?: string; id?: string; contextWindow?: number }> } | undefined)?.models;
+ const provider = latestProviders.find((entry) => (entry.sourceProviderId ?? entry.id) === sourceId)?.id;
+ const wireId = nativeModel?.wireId ?? nativeModel?.id;
+ const actualWindow = available.success
+ ? entries?.find((entry) => entry.provider === provider && entry.id === wireId)?.contextWindow : undefined;
+ if (typeof actualWindow === 'number' && Number.isSafeInteger(actualWindow) && actualWindow > 0) {
+ configuredWindow = actualWindow;
+ }
+ } catch { /* No mutation: leave window resolution to the existing recovery path. */ }
+ if (!configuredWindow) {
+ return { action: 'rebuild', targetContextWindow: null, windowVerified: false,
+ reason: 'Pi must resolve the inherited model window before selecting this model' };
+ }
}
- return nativeOffersModel(liveXai.id, modelId);
- } catch (err) {
- this.deps.logger.warn('pi live xAI catalog refresh failed', {
- message: err instanceof Error ? err.message : String(err),
- });
- return false;
}
+ const targetContextWindow = configuredWindow ? Math.max(configuredWindow, workingWindow) : null;
+ // A refresh may change any provider block, so all credentials it references
+ // must be synchronised with the running Pi process before it is published.
+ // The private refresh bridge owns this update; this preview never handles a key.
+ const projected = await this.writeModelsJson(
+ configHome, latestProviders,
+ gateway ? (knownDescriptor ?? (sameRoute ? retainedLiveGatewayModel : undefined)
+ ?? retainedRuntimeModel) : retainedLiveGatewayModel,
+ gateway ? requestedProviderId : authProviderId,
+ { remote, fileOps, preview: true,
+ contextWindow: targetContextWindow ?? ctx.contextWindow ?? startupContextWindow,
+ workingContextWindow: this.deps.resolveModelContextLimit?.(sourceId, model) ?? undefined,
+ piCompactionPct: sessionPiAutoCompactPct,
+ packages: nativePackagePaths, disabledSkills: disabledSkillLaunch,
+ ...(guestOmitsGateway ? { omitGatewayModels: true } : {}) },
+ );
+ if (gateway && !projected.gatewayApiByModel.has(model)) {
+ return { action: 'unavailable', targetContextWindow, windowVerified: false,
+ reason: `Gateway Pi model '${model}' has no runnable protocol` };
+ }
+ const configChanged = projected.modelConfigHash !== appliedModelConfigHash;
+ const envChanged = Object.keys({ ...nativeEnv, ...latestEnv }).some(
+ (key) => nativeEnv[key] !== latestEnv[key],
+ );
+ const adaptersChanged = JSON.stringify(nativeProviderAdapterAliases(nativeProviders))
+ !== JSON.stringify(nativeProviderAdapterAliases(latestProviders));
+ let gatewayCredentialChanged = false;
+ if (remote && gateway) {
+ try {
+ const latestAuth = await this.deps.auth.getAuthEnv({
+ credentialMode: 'gateway-key', providerId: requestedProviderId,
+ });
+ gatewayCredentialChanged = latestAuth[PI_API_KEY_ENV] !== authEnv[PI_API_KEY_ENV];
+ } catch {
+ return { action: 'unavailable', targetContextWindow, windowVerified: false,
+ reason: 'The current Gateway credential could not be read' };
+ }
+ }
+ const needsRefresh = configChanged || envChanged || adaptersChanged || gatewayCredentialChanged;
+ const inspected = await queryNativeRuntime('inspect');
+ if (!inspected.ok || !liveRuntimeSettings) {
+ return { action: 'unavailable', targetContextWindow, windowVerified: false,
+ reason: 'Cindy could not read the live Pi compaction settings' };
+ }
+ const targetProvider = gateway ? PI_PROVIDER_ID
+ : latestProviders.find((entry) => (entry.sourceProviderId ?? entry.id) === sourceId)?.id;
+ const targetModel = nativeModel?.wireId ?? nativeModel?.id ?? model;
+ if (targetContextWindow && targetProvider &&
+ resolvePiNativeReserve(liveRuntimeSettings, targetProvider, targetModel) !==
+ reserveForWindow(targetContextWindow, sourceId, model)) {
+ return { action: 'rebuild', targetContextWindow, windowVerified: false,
+ reason: 'Pi must load the changed native compaction settings before selecting this model' };
+ }
+ return {
+ action: needsRefresh ? 'refresh' : 'hot',
+ targetContextWindow,
+ windowVerified: sameRoute && !needsRefresh && targetContextWindow === ctx.contextWindow,
+ };
};
- const switchModel = async (
+ const refreshLiveCatalog = async (
model: string,
- setOpts?: { providerId?: string | null; effort?: Effort; thinkingEnabled?: boolean },
+ requestedProviderId: string | null | undefined,
): Promise<void> => {
- const requestedProviderId = setOpts && Object.hasOwn(setOpts, 'providerId')
- ? setOpts.providerId
- : undefined;
- assertRemotePiContextProfileAvailable(opts.remoteHostId, model, requestedProviderId);
- // 心跳复用活进程会把当前 (provider, model) 再下发一次。
- // spawn 能靠 custom model id 跑在 Pi 自带目录没有的 SuperGrok
- // 路由上(grok-4.6);重复 set_model 反而 fail-closed。
- // 只对显式非 null 来源做 no-op:model-only 与 providerId=null
- // (钉回网关)仍走 RPC。
- const assertGatewayProtocolForModel = (
- nextModel: string,
- nextRequestedProviderId: string | null | undefined,
- ): void => {
- const nextProvider = resolveProviderForModel(nextModel, nextRequestedProviderId);
- if (nextProvider !== PI_PROVIDER_ID) return;
- const routeProviderId = nextRequestedProviderId !== undefined
- ? nextRequestedProviderId
- : mutableProviderId;
- const resolverContext = { remote };
- const resolvedSpec = this.deps.resolvePiGatewayModelSpec?.(
- routeProviderId,
- nextModel,
- resolverContext,
- );
- const resolvedApi = resolvedSpec?.api ?? this.deps.resolvePiGatewayModelApi?.(
- routeProviderId,
- nextModel,
- resolverContext,
- );
- if (
- resolvedSpec === null ||
- resolvedApi === null ||
- (resolvedApi !== undefined && !isPiGatewayApi(resolvedApi))
- ) {
- throw new Error(
- `No supported Gateway Pi API is available for model: ${nextModel}`,
- );
+ const live = await sessionNativeProviderResolver()?.({
+ workingDir: opts.workingDir,
+ remoteHostId: opts.remoteHostId,
+ providerId: requestedProviderId,
+ model,
+ resumeSessionId: sdkSessionId || opts.resumeSessionId,
+ purpose: 'live-refresh',
+ });
+ let nextProviders = live?.providers ?? nativeProviders;
+ let nextEnv = live?.env ?? nativeEnv;
+ const sameRoute = model === mutableModel && requestedProviderId === mutableProviderId;
+ const sourceId = requestedProviderId === undefined
+ ? (model.startsWith('xai/') ? 'xai' : model.startsWith('chatgpt/') ? 'openai' : null)
+ : requestedProviderId;
+ if (sameRoute && sourceId && sourceId !== 'xd' && sourceId !== PI_PROVIDER_ID) {
+ const next = nextProviders.find((entry) => (entry.sourceProviderId ?? entry.id) === sourceId);
+ const running = nativeProviderForSource(sourceId);
+ if (running && (!next || !next.models.some((candidate) =>
+ candidate.id === (next.modelIdAliases?.[model] ?? model)))) {
+ nextProviders = [...nextProviders.filter((entry) =>
+ (entry.sourceProviderId ?? entry.id) !== sourceId), running];
+ if (running.apiKeyEnvVar && nativeEnv[running.apiKeyEnvVar] !== undefined) {
+ nextEnv = { ...nextEnv, [running.apiKeyEnvVar]: nativeEnv[running.apiKeyEnvVar] };
+ }
}
- if (resolvedApi === undefined) {
- throw new Error(`Pi API is not configured for model: ${nextModel}`);
+ }
+ const nextDescriptor = this.deps.resolvePiRuntimeModelDescriptor
+ ? this.deps.resolvePiRuntimeModelDescriptor(sourceId, model)
+ : this.capabilities.availableModels.find((candidate) => candidate.id === model);
+ const nextGateway = !sourceId || sourceId === 'xd' || sourceId === PI_PROVIDER_ID;
+ const gatewayAuthEnv = remote && nextGateway
+ ? await this.deps.auth.getAuthEnv({ credentialMode: 'gateway-key', providerId: requestedProviderId })
+ : authEnv;
+ const nextAliases = nativeProviderAdapterAliases(nextProviders);
+ const mutableEnv = (value: Record<string, string>) => Object.fromEntries(
+ Object.entries(value).filter(([key]) => /^CINDY_PI_KEY_[A-Z0-9_]+$/.test(key)
+ || key === PI_API_KEY_ENV || key === PI_SESSION_TOKEN_ENV
+ || key === 'CINDY_PI_OPENAI_PROXY_KEY' || key === 'CINDY_PI_XAI_PROXY_API_KEY'));
+ const desiredEnv = mutableEnv({ ...nextEnv,
+ ...(gatewayAuthEnv[PI_API_KEY_ENV] ? { [PI_API_KEY_ENV]: gatewayAuthEnv[PI_API_KEY_ENV] } : {}),
+ ...(proxySessionToken ? { [PI_SESSION_TOKEN_ENV]: proxySessionToken } : {}),
+ });
+ // Preview has verified the Cindy-owned command and native settings.
+ // The extension uses Pi's public ModelRegistry.refresh; no custom Pi RPC.
+ const previousGatewayModel = retainedLiveGatewayModel;
+ const modelsPath = joinRemotePosixPath(configHome, 'models.json');
+ const settingsPath = joinRemotePosixPath(configHome, 'settings.json');
+ // Capture exact private bytes. Rebuilding from a live catalog would not be
+ // a rollback if the catalog changed after the session started.
+ const [oldModelsContent, oldSettingsContent] = await Promise.all([
+ fileOps ? fileOps.readFile(modelsPath, 8_000_000) : fs.readFile(modelsPath, 'utf8'),
+ fileOps ? fileOps.readFile(settingsPath, 1_000_000) : fs.readFile(settingsPath, 'utf8'),
+ ]);
+ if (fileOps) {
+ const [modelsHash, settingsHash] = await Promise.all([
+ fileOps.sha256File(modelsPath), fileOps.sha256File(settingsPath),
+ ]);
+ if (modelsHash !== createHash('sha256').update(oldModelsContent).digest('hex') ||
+ settingsHash !== createHash('sha256').update(oldSettingsContent).digest('hex')) {
+ throw new Error('Pi remote catalog backup was incomplete; live refresh was not started');
}
- const desiredApi = resolvedApi;
- const configuredApi = gatewayApiByModel.get(nextModel);
- if (configuredApi && configuredApi !== desiredApi) {
- throw new Error(
- `pi: provider switch for model '${nextModel}' requires API '${desiredApi}', but this session ` +
- `started with '${configuredApi}'; restart the Pi session to change provider API.`,
- );
+ }
+ const restoreFilesOrTerminate = async (): Promise<void> => {
+ try {
+ if (fileOps) {
+ await fileOps.writeFile(modelsPath, oldModelsContent);
+ await fileOps.writeFile(settingsPath, oldSettingsContent);
+ } else {
+ await fs.writeFile(modelsPath, oldModelsContent, { mode: 0o600 });
+ await fs.writeFile(settingsPath, oldSettingsContent, { mode: 0o600 });
+ }
+ } catch (error) {
+ await terminateUnconfirmedCatalogReload(error);
}
};
- if (
- model === mutableModel &&
- requestedProviderId !== undefined &&
- requestedProviderId !== null &&
- Object.is(requestedProviderId, mutableProviderId)
- ) {
- if (setOpts?.effort) {
- assertStartupEffortAllowed(activeEffortSnapshot, setOpts.effort);
+ let written: Awaited<ReturnType<PiAgent['writeModelsJson']>>;
+ try {
+ written = await this.writeModelsJson(
+ configHome, nextProviders,
+ nextGateway ? (nextDescriptor ?? retainedLiveGatewayModel ?? retainedRuntimeModel) : retainedLiveGatewayModel,
+ nextGateway ? requestedProviderId : authProviderId,
+ { remote, fileOps, contextWindow: nextDescriptor?.contextWindow ?? ctx.contextWindow,
+ workingContextWindow: this.deps.resolveModelContextLimit?.(sourceId, model) ?? undefined,
+ piCompactionPct: sessionPiAutoCompactPct, packages: nativePackagePaths,
+ disabledSkills: disabledSkillLaunch,
+ ...(guestOmitsGateway ? { omitGatewayModels: true } : {}) },
+ );
+ } catch (error) {
+ await restoreFilesOrTerminate();
+ throw error;
+ }
+ try {
+ const receipt = await queryNativeRuntime('refresh', desiredEnv, nextAliases);
+ if (!receipt.ok) {
+ if (receipt.code === 'INVALID_PAYLOAD') {
+ await restoreFilesOrTerminate();
+ throw new Error('Pi rejected the native provider refresh payload; the original route is unchanged.');
+ }
+ return await terminateUnconfirmedCatalogReload(receipt.code);
}
- assertGatewayProtocolForModel(model, requestedProviderId);
- // 同路由不打 set_model,但仍要重试子代理快照:初始写失败或确认写失败
- // 留下 pending 时,心跳复用活进程必须能把文件重建/清 pending,
- // 否则扩展会一直 fail-closed。不带 pending,路由已确认。
- const sameRouteProvider = resolveProviderForModel(model, requestedProviderId);
- const sameRouteWire = resolveWireModel(sameRouteProvider, model);
- if (!(await writeSubagentRuntimeFile({ model: sameRouteWire, provider: sameRouteProvider }))) {
- deps.logger.warn('pi: same-route setModel could not refresh subagent snapshot', {
- model,
- provider: sameRouteProvider,
- });
+ } catch (error) {
+ if (error instanceof Error && /rejected the native provider/.test(error.message)) throw error;
+ return await terminateUnconfirmedCatalogReload(error);
+ }
+ try {
+ const refreshed = await proc.request({ type: 'get_available_models' });
+ const result = refreshed.data as { models?: Array<{ provider?: string; id?: string }> } | undefined;
+ if (!refreshed.success) throw new Error('Pi did not expose the refreshed model catalog');
+ const desiredProvider = nextGateway ? PI_PROVIDER_ID : nextProviders.find(
+ (entry) => (entry.sourceProviderId ?? entry.id) === sourceId)?.id;
+ const desiredNative = nextProviders.find((entry) => entry.id === desiredProvider);
+ const desiredId = desiredNative?.models.find((candidate) =>
+ candidate.id === (desiredNative.modelIdAliases?.[model] ?? model))?.wireId
+ ?? desiredNative?.modelIdAliases?.[model] ?? model;
+ if (!desiredProvider || !result?.models?.some((candidate) =>
+ candidate.provider === desiredProvider && candidate.id === desiredId)) {
+ throw new Error('Pi refreshed the catalog but did not expose the selected provider/model');
+ }
+ } catch (error) {
+ // Pi may have partially replaced registry entries even on a negative
+ // response. A confirmed rollback would require another refresh+ACK;
+ // retiring is the only safe outcome for the live process.
+ await terminateUnconfirmedCatalogReload(error);
+ }
+ nativeProviders = nextProviders;
+ nativeEnv = nextEnv;
+ if (remote && nextGateway) authEnv = gatewayAuthEnv;
+ // Child processes are spawned by the Desktop host after this point. Their
+ // immutable startup env must match the parent Pi registry we just verified.
+ for (const key of Object.keys(durableSpawnEnv)) {
+ if (/^CINDY_PI_KEY_[A-Z0-9_]+$/.test(key)
+ || key === 'CINDY_PI_OPENAI_PROXY_KEY'
+ || key === 'CINDY_PI_XAI_PROXY_API_KEY'
+ || key === PI_API_KEY_ENV || key === PI_SESSION_TOKEN_ENV) {
+ delete durableSpawnEnv[key];
}
- deps.logger.debug('pi: setModel no-op; already on requested route', {
- model,
- providerId: requestedProviderId,
- });
- return;
}
- // null = 钉回网关,绝不按模型名推断 xAI。只有 undefined(旧会话未持久化来源)
- // 才允许 xai/ 前缀回退。`??` 会把 null 也吃掉,误进 live catalog 刷新。
- const sourceHint = requestedProviderId === undefined
- ? (model.startsWith('xai/') ? 'xai' : undefined)
- : (requestedProviderId ?? undefined);
- const liveProviderHint = sourceHint ? nativeProviderForSource(sourceHint) : undefined;
- const needsXaiCatalogReload = (sourceHint === 'xai' || Boolean(liveProviderHint?.hostProxyForward))
- && (!liveProviderHint || !nativeOffersModel(liveProviderHint.id, model));
- if (needsXaiCatalogReload) {
- const previousProviders = nativeProviders.slice();
- const refreshed = await refreshLiveXaiCatalog(model, requestedProviderId);
- if (refreshed && sdkSessionId) {
- // Claude 热切先 applyFlagSettings 扩白名单;Pi 的 set_model 不重读 models.json,
- // 但 switch_session 会 createRuntime → ModelConfig.load,等于无重启扩名单。
- // success:false = 确定没重载,回滚安全。reject/超时 = 不知道 Pi 侧有没有吃到新
- // models.json,回滚和放行都可能分叉,按未确认 set_model 一样终止会话。
- let reloaded;
- try {
- reloaded = await proc.request({
- type: 'switch_session',
- sessionPath: sdkSessionId,
- });
- } catch (err) {
- return await terminateUnconfirmedCatalogReload(err);
+ Object.assign(durableSpawnEnv, desiredEnv);
+ durableSpawnEnv.CINDY_PI_NATIVE_PROVIDER_ADAPTERS = JSON.stringify(nextAliases);
+ const currentSecretNames = JSON.parse(
+ durableSpawnEnv[PI_SECRET_ENV_NAMES_ENV] ?? '[]') as string[];
+ durableSpawnEnv[PI_SECRET_ENV_NAMES_ENV] = JSON.stringify(
+ [...new Set([...currentSecretNames, ...Object.keys(desiredEnv)])]);
+ if (subagentRoutingEnabled) {
+ for (const [id, candidates] of subagentModelRouteCandidates) {
+ for (const [routeKey, route] of candidates) {
+ if (route.provider !== PI_PROVIDER_ID || !written.gatewayApiByModel.has(id)) {
+ candidates.delete(routeKey);
+ }
}
- if (!reloaded.success) {
- await restoreNativeCatalogOrTerminate(previousProviders);
- throw new Error(
- `pi: failed to reload models after catalog update: ${reloaded.error ?? 'unknown'}`,
- );
+ if (candidates.size === 0) subagentModelRouteCandidates.delete(id);
+ }
+ for (const gatewayModelId of written.gatewayApiByModel.keys()) {
+ addSubagentModelRoute(gatewayModelId, PI_PROVIDER_ID, gatewayModelId,
+ gatewaySubagentProxyToken ? gatewaySubagentRouteKey : undefined,
+ Boolean(gatewaySubagentProxyToken));
+ }
+ for (const spec of nextProviders) {
+ if (spec.id === PI_PROVIDER_ID) continue;
+ const source = spec.sourceProviderId ?? spec.id;
+ const wasRegistered = subagentProxyTokens.has(source);
+ const token = proxyTokenForSubagentProvider(spec, source);
+ if (token && !wasRegistered && opts.sessionId && this.deps.registerPiProxySession) {
+ appendProxyDisposer(this.deps.registerPiProxySession(
+ opts.sessionId, token, () => source, { scope: 'subagent-route' }));
+ }
+ for (const candidate of spec.models) {
+ const wire = candidate.wireId ?? candidate.id;
+ addSubagentModelRoute(candidate.id, spec.id, wire, source, Boolean(token));
+ for (const [alias, mapped] of Object.entries(spec.modelIdAliases ?? {})) {
+ if (mapped === candidate.id) addSubagentModelRoute(alias, spec.id, wire, source, Boolean(token));
+ }
+ if (!candidate.id.includes('/')) {
+ const canonical = source === 'xai' ? `xai/${candidate.id}`
+ : source === 'openai' || candidate.api === 'openai-codex-responses'
+ ? `chatgpt/${candidate.id}` : undefined;
+ addSubagentModelRoute(canonical, spec.id, wire, source, Boolean(token));
+ }
}
- } else if (!refreshed) {
- await restoreNativeCatalogOrTerminate(previousProviders);
}
+ subagentModelRoutes = Object.fromEntries([...subagentModelRouteCandidates]
+ .map(([id, candidates]) => [id, [...candidates.values()]]));
}
- // 显式选一个启动快照 nativeProviderById 里“无法服务该 model”的 BYOM provider 时 fail
- // closed:要么该 provider 是会话启动后才新增的(不在快照),要么它虽在、但用户编辑
- // 配置后从中删/改了这个 model。两种都会让 resolveProviderForModel 静默回落 cindy 网关;
- // 若该 model id 也在网关目录里则 set_model “成功”、后续 prompt 发往网关而非用户选的
- // 本地/自定义端点(codex review P1)。提示重启会话以刷新启动快照,而不是静默换目的地。
- if (explicitByomUnresolvable(requestedProviderId, model)) {
- throw new Error(
- `pi: BYOM provider '${requestedProviderId}' cannot serve model '${model}' in this session's ` +
- 'startup provider set (provider not present, or it no longer offers this model); restart the ' +
- 'session to use it (refusing to fall back to the Cindy gateway).',
- );
+ nativeProviderById.clear();
+ nativeProviderBySourceId.clear();
+ for (const spec of nextProviders) {
+ if (spec.id === PI_PROVIDER_ID) continue;
+ nativeProviderById.set(spec.id, spec);
+ nativeProviderBySourceId.set(spec.sourceProviderId ?? spec.id, spec);
}
- const provider = resolveProviderForModel(model, requestedProviderId);
+ gatewayApiByModel.clear();
+ for (const [key, value] of written.gatewayApiByModel) gatewayApiByModel.set(key, value);
+ for (const gatewayModelId of written.gatewayApiByModel.keys()) {
+ const descriptor = this.deps.resolvePiGatewayModelDescriptor?.(requestedProviderId, gatewayModelId)
+ ?? this.deps.resolvePiRuntimeModels?.().find((candidate) => candidate.id === gatewayModelId)
+ ?? this.capabilities.availableModels.find((candidate) => candidate.id === gatewayModelId);
+ if (descriptor) appliedGatewayEfforts.set(gatewayModelId, descriptor.efforts);
+ }
+ gatewayImageInputByModel.clear();
+ for (const [key, value] of written.gatewayImageInputByModel) gatewayImageInputByModel.set(key, value);
+ appliedModelConfigHash = written.modelConfigHash;
+ // A full registry refresh can update the still-active A while a proposed
+ // switch to B later fails. Keep A pending until a confirmed set_model.
+ pendingCurrentDescriptorRefresh = true;
+ if (nextGateway) retainedLiveGatewayModel = nextDescriptor ?? previousGatewayModel;
+ };
+ const switchModel = async (
+ model: string,
- setOpts?: { providerId?: string | null; effort?: Effort },
++ setOpts?: { providerId?: string | null; effort?: Effort; thinkingEnabled?: boolean },
+ ): Promise<void> => {
+ const requestedProviderId = setOpts && Object.hasOwn(setOpts, 'providerId')
+ ? setOpts.providerId
+ : undefined;
+ assertRemotePiContextProfileAvailable(opts.remoteHostId, model, requestedProviderId);
+ const effectiveProviderId = requestedProviderId === undefined ? mutableProviderId : requestedProviderId;
+ const preview = await previewPiModelSwitch(model, { providerId: effectiveProviderId });
+ if (preview.action === 'unavailable') throw new Error(preview.reason ?? 'Pi model route is unavailable');
+ if (preview.action === 'rebuild') {
+ throw new Error(preview.reason ?? 'Pi model route needs a new process startup input');
+ }
+ if (preview.action === 'refresh') await refreshLiveCatalog(model, effectiveProviderId);
+ // Never let a requested native source silently fall through to Cindy's
+ // gateway when a provider disappeared or a model was removed.
+ if (explicitByomUnresolvable(effectiveProviderId, model)) {
+ throw new Error(`pi: provider '${effectiveProviderId}' cannot serve model '${model}'`);
+ }
+ const provider = resolveProviderForModel(model, effectiveProviderId);
const wireModel = resolveWireModel(provider, model);
- if (
- requestedProviderId
- && requestedProviderId !== PI_PROVIDER_ID
- && requestedProviderId !== 'xd'
- && provider === PI_PROVIDER_ID
- ) {
- throw new Error(
- `pi: provider '${requestedProviderId}' cannot serve model '${model}' even after reloading the live catalog; ` +
- 'restart the Pi session if the provider was added after this task started.',
- );
+ if (effectiveProviderId && effectiveProviderId !== PI_PROVIDER_ID &&
+ effectiveProviderId !== 'xd' && provider === PI_PROVIDER_ID) {
+ throw new Error(`pi: provider '${effectiveProviderId}' cannot serve model '${model}'`);
}
- assertGatewayProtocolForModel(model, requestedProviderId);
+ const routeUnchanged = model === mutableModel && provider === mutablePiProviderId &&
+ wireModel === mutableWireModel;
// effort 能力校验必须排在写路由快照**之前**:它会抛错中止本次切换,而快照一旦落盘就
// 指向了新 provider —— 那正是父子路由分叉的形状(upstream #1451 与本 PR 的合并点)。
- const nextEffortSnapshot = resolveStartupEffortSnapshot(provider, model);
+ const nextEffortSnapshot = resolveAppliedEffortSnapshot(provider, model);
if (setOpts?.effort) {
assertStartupEffortAllowed(nextEffortSnapshot, setOpts.effort);
}