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。

fix(pi): 模型切换后恢复会话思考档位,修复推理强度自动掉档 (#5733)

Yunjin Zhangceaf1c6f4cef
diff --git a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
index 66b2bbcdf5..114b05d512 100644
--- a/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
+++ b/packages/maker-core/src/agents/pi/__tests__/pi-provider-routing.test.ts
@@ -46,6 +46,7 @@ const captured = vi.hoisted(() => ({
remoteModels: new Map<string, string>(),
refreshAckMode: 'ok' as 'ok' | 'invalid',
hasRefreshCommand: true,
+ failThinkingLevel: null as null | 'reject' | 'throw',
requestHandler: undefined as
| undefined
| ((command: Record<string, unknown>) => Promise<{
@@ -131,6 +132,12 @@ vi.mock("../rpc-client.js", () => {
if (command.type === 'refresh_models' || command.type === 'set_compaction_reserve_tokens') {
return { success: false, error: `Unknown command: ${command.type}` };
}
+ if (command.type === 'set_thinking_level' && captured.failThinkingLevel === 'reject') {
+ return { success: false, error: 'mock reject' };
+ }
+ if (command.type === 'set_thinking_level' && captured.failThinkingLevel === 'throw') {
+ throw new Error('mock rpc timeout');
+ }
const response = captured.requestHandler
? await captured.requestHandler(command)
: command.type === "get_state"
@@ -279,6 +286,7 @@ describe("Pi provider-aware model routing", () => {
captured.remoteModels.clear();
captured.refreshAckMode = 'ok';
captured.hasRefreshCommand = true;
+ captured.failThinkingLevel = null;
captured.requestHandler = undefined;
agentHome = mkdtempSync(path.join(tmpdir(), "pi-provider-home-"));
cwd = mkdtempSync(path.join(tmpdir(), "pi-provider-cwd-"));
@@ -340,6 +348,143 @@ describe("Pi provider-aware model routing", () => {
expect(JSON.parse(query({ provider: 'relay', model: 'private-sol' })).fast).toBe(false);
});
+ it('re-asserts the session effort after a full-path model switch (pi resets it to the model default)', async () => {
+ const deps: AgentDeps = {
+ auth: {
+ getState: async () => ({ authenticated: true, identity: 'test', authSource: 'api-key' as const }),
+ triggerLogin: async () => ({ authenticated: true }),
+ logout: async () => {},
+ getAuthEnv: async () => ({}),
+ },
+ runtimeConfig: { endpoint: 'http://127.0.0.1:9' },
+ binaryPath: path.join(agentHome, 'pi'),
+ logger: noopLogger,
+ capabilityAdditions: {
+ availableModels: [
+ {
+ id: 'model-a',
+ displayName: 'Model A',
+ contextWindow: 200_000,
+ efforts: ['low', 'high', 'max'],
+ defaultEffort: 'high',
+ },
+ {
+ id: 'model-b',
+ displayName: 'Model B',
+ contextWindow: 200_000,
+ efforts: ['low', 'high', 'max'],
+ defaultEffort: 'high',
+ },
+ ],
+ },
+ resolvePiAgentHome: () => agentHome,
+ resolvePiNativeProviders: async () => ({
+ providers: [
+ { id: 'native-a', name: 'Native A', baseUrl: 'http://a.test', api: 'openai-completions', models: [{ id: 'model-a', reasoning: true, thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high', max: 'max' } }] },
+ { id: 'native-b', name: 'Native B', baseUrl: 'http://b.test', api: 'openai-completions', models: [{ id: 'model-b', reasoning: true, thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high', max: 'max' } }] },
+ ],
+ env: {},
+ }),
+ };
+ const agent = new PiAgent(deps);
+ const handle = await agent.startSession({
+ sessionId: 'model-switch-effort',
+ workingDir: cwd,
+ model: 'model-a',
+ providerId: 'native-a',
+ effort: 'max',
+ });
+ const thinkingCalls = () =>
+ captured.requests.filter((request) => request.type === 'set_thinking_level');
+ expect(thinkingCalls().at(-1)).toEqual({ type: 'set_thinking_level', level: 'max' });
+
+ // ① 带显式 effort 的跨模型切换:set_model 后下发该档。
+ const afterStartup = thinkingCalls().length;
+ await handle.setModel!('model-b', { providerId: 'native-b', effort: 'low' });
+ expect(thinkingCalls().slice(afterStartup)).toContainEqual({
+ type: 'set_thinking_level', level: 'low',
+ });
+
+ // ② 不带 effort 切回:恢复会话原档位(mutableEffort 已被①更新为 low)。
+ const beforeRestore = thinkingCalls().length;
+ await handle.setModel!('model-a', { providerId: 'native-a' });
+ expect(captured.requests.slice(beforeRestore)).toContainEqual({
+ type: 'set_model', provider: 'native-a', modelId: 'model-a',
+ });
+ expect(thinkingCalls().slice(beforeRestore)).toContainEqual({
+ type: 'set_thinking_level', level: 'low',
+ });
+
+ await handle.close();
+ });
+
+ it('effort 恢复失败不阻断已确认的模型切换,也不把被拒档位记成已生效', async () => {
+ const deps: AgentDeps = {
+ auth: {
+ getState: async () => ({ authenticated: true, identity: 'test', authSource: 'api-key' as const }),
+ triggerLogin: async () => ({ authenticated: true }),
+ logout: async () => {},
+ getAuthEnv: async () => ({}),
+ },
+ runtimeConfig: { endpoint: 'http://127.0.0.1:9' },
+ binaryPath: path.join(agentHome, 'pi'),
+ logger: noopLogger,
+ capabilityAdditions: {
+ availableModels: [
+ {
+ id: 'model-a',
+ displayName: 'Model A',
+ contextWindow: 200_000,
+ efforts: ['low', 'high', 'max'],
+ defaultEffort: 'high',
+ },
+ {
+ id: 'model-b',
+ displayName: 'Model B',
+ contextWindow: 200_000,
+ efforts: ['low', 'high', 'max'],
+ defaultEffort: 'high',
+ },
+ ],
+ },
+ resolvePiAgentHome: () => agentHome,
+ resolvePiNativeProviders: async () => ({
+ providers: [
+ { id: 'native-a', name: 'Native A', baseUrl: 'http://a.test', api: 'openai-completions', models: [{ id: 'model-a', reasoning: true, thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high', max: 'max' } }] },
+ { id: 'native-b', name: 'Native B', baseUrl: 'http://b.test', api: 'openai-completions', models: [{ id: 'model-b', reasoning: true, thinkingLevelMap: { low: 'low', medium: 'medium', high: 'high', max: 'max' } }] },
+ ],
+ env: {},
+ }),
+ };
+ const agent = new PiAgent(deps);
+ const handle = await agent.startSession({
+ sessionId: 'model-switch-effort-restore-fail',
+ workingDir: cwd,
+ model: 'model-a',
+ providerId: 'native-a',
+ effort: 'max',
+ });
+
+ // ① set_thinking_level 超时/链路抛错:模型路由已确认,setModel 必须照常完成,
+ // 不能把已成功的切换变成失败(否则子代理快照收尾被跳过,pending 卡死委派)。
+ captured.failThinkingLevel = 'throw';
+ await expect(handle.setModel!('model-b', { providerId: 'native-b', effort: 'low' }))
+ .resolves.toBeUndefined();
+ expect(handle.getEffort!()).toBe('max');
+
+ // ② set_thinking_level 被拒:mutableEffort 不跟随未生效的目标档。
+ captured.failThinkingLevel = 'reject';
+ await handle.setModel!('model-a', { providerId: 'native-a', effort: 'low' });
+ expect(handle.getEffort!()).toBe('max');
+
+ // ③ 成功路径不受影响:恢复生效后 getEffort 报新档。
+ captured.failThinkingLevel = null;
+ await handle.setModel!('model-b', { providerId: 'native-b', effort: 'low' });
+ expect(handle.getEffort!()).toBe('low');
+
+ await handle.close();
+ });
+
it("uses providerId as the primary key when duplicate model ids exist", async () => {
const authProviderIds: Array<string | null | undefined> = [];
const apiResolver = vi.fn((providerId: string | null | undefined) =>
diff --git a/packages/maker-core/src/agents/pi/index.ts b/packages/maker-core/src/agents/pi/index.ts
index db972c6dca..38c430a399 100644
--- a/packages/maker-core/src/agents/pi/index.ts
+++ b/packages/maker-core/src/agents/pi/index.ts
@@ -6972,6 +6972,41 @@ export class PiAgent extends BaseAgent {
mutablePiProviderId = provider;
activeEffortSnapshot = nextEffortSnapshot;
mutableProviderId = effectiveProviderId;
+ // pi CLI 在 set_model 后会把思考档位重置成新模型自己的默认(转录实证:声明了档位的
+ // 模型落到 defaultEffort 如 high,未声明的落到 low/medium),而本函数原本只在
+ // setOpts.effort 上做校验、从不下发。不在这里把目标档位重新下发的话,每次切模型
+ // 都会掉回默认档,表现为「推理强度自动变低」。目标档 = 显式 effort > 会话原档位
+ // (mutableEffort);新模型不支持(含未声明档位)时保持 pi 默认,与能力收敛一致。
+ const restoredEffort = setOpts?.effort ?? mutableEffort;
+ if (
+ restoredEffort &&
+ nextEffortSnapshot && nextEffortSnapshot.length > 0 &&
+ nextEffortSnapshot.includes(restoredEffort)
+ ) {
+ // 恢复请求失败(被拒或超时/链路抛错)都不能让 setModel 失败:模型路由此刻已确认,
+ // 抛错只会跳过下面的子代理快照收尾(pending 卡死委派),档位本身却救不回来。
+ try {
+ const restore = await proc.request({
+ type: 'set_thinking_level',
+ level: effortToPiThinkingLevel(restoredEffort),
+ });
+ if (restore.success) {
+ // mutableEffort 只跟随**已确认生效**的档位:它是对外汇报(getEffort)与后续
+ // 切换恢复的依据;被 Pi 拒绝的目标档不能当成已生效值。
+ mutableEffort = restoredEffort;
+ } else {
+ deps.logger.warn('pi: restore thinking level after model switch failed', {
+ effort: restoredEffort,
+ error: restore.error,
+ });
+ }
+ } catch (err) {
+ deps.logger.warn('pi: restore thinking level after model switch errored', {
+ effort: restoredEffort,
+ message: err instanceof Error ? err.message : String(err),
+ });
+ }
+ }
autoReviewDecisionCache.clear();
autoReviewUnavailableNotice.reset();
autoReviewConfirmUndeliveredNotice.reset();