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(claude-code): GLM-5.3 思考档位按模型能力下发,智谱预设补齐档位 (#5408)

Dash97374e1888a0
diff --git a/apps/desktop/src/main/maker-host/__tests__/modelContextSettings.test.ts b/apps/desktop/src/main/maker-host/__tests__/modelContextSettings.test.ts
index 60ab947fd4..0c56153667 100644
--- a/apps/desktop/src/main/maker-host/__tests__/modelContextSettings.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/modelContextSettings.test.ts
@@ -18,7 +18,7 @@ vi.mock('../model-context-limit-store.js', () => ({
state.limits.get(`${agent}:${provider}:${model}`) ?? null,
}));
-import { resolveConfiguredContextWindow, resolveDesktopModelContextProviderId } from '../model-context-settings.js';
+import { resolveConfiguredContextWindow, resolveDesktopModelContextProviderId, resolveDesktopModelEfforts } from '../model-context-settings.js';
import { shouldRebuildForModelWindowSwitch } from '../../maker-ipc/contextOverflowRollover.js';
function dualCatalog(agent: AgentKind): Catalog {
@@ -118,3 +118,22 @@ describe('implicit context settings use the startup source in history protection
expect(resolveConfiguredContextWindow(catalog, 'codex', null, 'codex/shared-model')).toBe(100_000);
});
});
+
+describe('route efforts follow the session source', () => {
+ it('reads the selected provider instead of another same-ID route (#5402)', () => {
+ const catalog = dualCatalog('claude-code');
+ for (const provider of catalog.providers) {
+ for (const model of provider.models['claude-code'] ?? []) {
+ model.efforts = provider.id === 'xd' ? ['low', 'medium', 'high', 'max'] : ['low', 'high', 'xhigh', 'max'];
+ }
+ }
+ expect(resolveDesktopModelEfforts(catalog, 'claude-code', 'anthropic', 'shared-model'))
+ .toEqual(['low', 'high', 'xhigh', 'max']);
+ expect(resolveDesktopModelEfforts(catalog, 'claude-code', 'xd', 'shared-model'))
+ .toEqual(['low', 'medium', 'high', 'max']);
+ // Implicit sessions use the same startup source as the context settings (gateway here).
+ expect(resolveDesktopModelEfforts(catalog, 'claude-code', null, 'shared-model'))
+ .toEqual(['low', 'medium', 'high', 'max']);
+ expect(resolveDesktopModelEfforts(catalog, 'claude-code', 'missing', 'shared-model')).toBeNull();
+ });
+});
diff --git a/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts b/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
index 6c44da83a3..73b0fb02e9 100644
--- a/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
+++ b/apps/desktop/src/main/maker-host/__tests__/piProviderTransport.test.ts
@@ -126,6 +126,25 @@ describe('Pi-owned transport for Cindy harnesses', () => {
if (providerId === 'together') expect(sent).toMatchObject({ reasoning: { enabled: true } });
expect(sent!.max_tokens ?? sent!.max_completion_tokens).toBe(1024);
});
+
+ it('does not ask an always-thinking GLM-5.3 to disable thinking when the request has no effort (#5402)', async () => {
+ const glm = PROVIDER_MODEL_CATALOG.providers['zai-coding-cn'].find(row => row.id === 'glm-5.3')!;
+ const { off: _off, ...optionalOffLevels } = glm.execution.pi.thinkingLevelMap ?? {};
+ const optionalOff = { ...glm, execution: { pi: { ...glm.execution.pi, thinkingLevelMap: optionalOffLevels } } };
+ const sentFor = async (row: typeof glm) => {
+ let sent: Record<string, unknown> | undefined;
+ const send = createPiProviderFetch({ row, providerId: 'zai-coding-cn', apiKey: 'fixture-provider-key', fetchImpl: async (_url, init) => {
+ sent = JSON.parse(String(init?.body));
+ return new Response(reply, { headers: { 'content-type': 'text/event-stream' } });
+ } });
+ await (await send('https://unused.invalid', { body: JSON.stringify({ model: row.id, input: 'hello', stream: true }) })).text();
+ return sent;
+ };
+
+ expect(glm.execution.pi.thinkingLevelMap?.off).toBeNull();
+ expect(await sentFor(glm)).toMatchObject({ thinking: { type: 'enabled' }, reasoning_effort: 'high' });
+ expect(await sentFor(optionalOff)).toMatchObject({ thinking: { type: 'disabled' } });
+ });
});
it('keeps native Gemini tool signatures across two turns through Responses history', async () => {
diff --git a/apps/desktop/src/main/maker-host/__tests__/zhipuPresetGlmEfforts.test.ts b/apps/desktop/src/main/maker-host/__tests__/zhipuPresetGlmEfforts.test.ts
new file mode 100644
index 0000000000..d4182b2a60
--- /dev/null
+++ b/apps/desktop/src/main/maker-host/__tests__/zhipuPresetGlmEfforts.test.ts
@@ -0,0 +1,40 @@
+import { afterEach, describe, expect, it } from 'vitest';
+import { BUNDLED_CATALOG, type CustomProviderConfig } from '@cindy/model-providers';
+import { getActiveCatalog, setActiveCatalog, setCustomProviderConfigs, setCustomProviders } from '../active-catalog.js';
+
+afterEach(() => {
+ setCustomProviders([]);
+ setActiveCatalog(BUNDLED_CATALOG);
+});
+
+describe('Zhipu preset GLM-5.3 thinking levels', () => {
+ // GLM-5.3 thinks on every request; an empty list left preset connections without a
+ // picker or default while GLM-5.3-Flash had both (#5402).
+ it.each(['zhipu-glm-cn', 'zhipu-glm-global'])('%s inherits the public GLM-5.3 efforts like Flash', (presetId) => {
+ setActiveCatalog(BUNDLED_CATALOG);
+ const preset = BUNDLED_CATALOG.presets!.find((p) => p.id === presetId)!;
+ const config: CustomProviderConfig = {
+ id: `custom:${presetId}`,
+ name: presetId,
+ runtimes: Object.fromEntries((['claude-code', 'codex'] as const).map((agent) => {
+ const runtime = preset.runtimes[agent]!;
+ return [agent, {
+ baseUrl: runtime.baseUrl,
+ ...(runtime.wireProtocol ? { wireProtocol: runtime.wireProtocol } : {}),
+ catalogPresetId: presetId,
+ models: runtime.models.filter((model) => model.id.startsWith('glm-5.3')),
+ }];
+ })),
+ };
+ setCustomProviderConfigs([config]);
+ const provider = getActiveCatalog().providers.find((p) => p.id === config.id)!;
+ for (const agent of ['claude-code', 'codex'] as const) {
+ for (const id of ['glm-5.3', 'glm-5.3-flash']) {
+ expect(provider.models[agent]?.find((model) => model.id === id), `${agent}/${id}`).toMatchObject({
+ efforts: ['low', 'high', 'max'],
+ defaultEffort: 'high',
+ });
+ }
+ }
+ });
+});
diff --git a/apps/desktop/src/main/maker-host/index.ts b/apps/desktop/src/main/maker-host/index.ts
index 9a43e79fc9..02e19fb4ea 100644
--- a/apps/desktop/src/main/maker-host/index.ts
+++ b/apps/desktop/src/main/maker-host/index.ts
@@ -184,7 +184,7 @@ import {
resolveModelDefaultContextWindow,
} from './catalog-to-descriptors.js';
import { readModelContextLimit } from './model-context-limit-store.js';
-import { resolveDesktopModelContextProviderId } from './model-context-settings.js';
+import { resolveDesktopModelContextProviderId, resolveDesktopModelEfforts } from './model-context-settings.js';
import {
prepareCodexCustomContextCatalog,
} from './codex-custom-context-catalog.js';
@@ -1274,6 +1274,8 @@ export function getMaker(): Maker {
resolveVerifiedContextWindow: (providerId, modelId) =>
resolveVerifiedContextWindow(getDesktopSelectableCatalog(), 'claude-code',
resolveDesktopModelContextProviderId(getDesktopSelectableCatalog(), 'claude-code', providerId, modelId), modelId),
+ resolveModelEfforts: (providerId, modelId) =>
+ resolveDesktopModelEfforts(getDesktopSelectableCatalog(), 'claude-code', providerId, modelId),
// SDK PreToolUse / PostToolUse 等 in-process hook 注入点。host 自己定义 hook
// 实现 (./claude-hooks/*.ts), maker-core 不感知具体逻辑。
//
diff --git a/apps/desktop/src/main/maker-host/model-context-settings.ts b/apps/desktop/src/main/maker-host/model-context-settings.ts
index 48768b31d6..f2cd37ba35 100644
--- a/apps/desktop/src/main/maker-host/model-context-settings.ts
+++ b/apps/desktop/src/main/maker-host/model-context-settings.ts
@@ -1,4 +1,4 @@
-import type { AgentKind, Catalog } from '@cindy/model-providers';
+import type { AgentKind, Catalog, Effort } from '@cindy/model-providers';
import { desktopCodexAuthAdapter, readClaudeApiKey } from './auth-adapters.js';
import { hasClaudeNativeLogin } from './claude-native-auth.js';
@@ -28,6 +28,22 @@ export function resolveDesktopModelContextProviderId(
return resolveModelContextProviderId(catalog, agent, providerId, modelId, defaultSource);
}
+/**
+ * Efforts declared by the route this session actually uses. Same-ID models from different
+ * providers can declare different efforts; null means the route is unknown or ambiguous.
+ */
+export function resolveDesktopModelEfforts(
+ catalog: Pick<Catalog, 'providers'>,
+ agent: AgentKind,
+ providerId: string | null | undefined,
+ modelId: string,
+): readonly Effort[] | null {
+ const source = resolveDesktopModelContextProviderId(catalog, agent, providerId, modelId);
+ if (!source) return null;
+ return catalog.providers.find((provider) => provider.id === source)
+ ?.models[agent]?.find((model) => model.id === modelId)?.efforts ?? null;
+}
+
/** Working budgets can tighten history protection, but never raise its verified ceiling. */
export function resolveConfiguredContextWindow(
catalog: Pick<Catalog, 'providers'>,
diff --git a/apps/desktop/src/main/maker-host/pi-provider-transport.ts b/apps/desktop/src/main/maker-host/pi-provider-transport.ts
index 543b72cc91..59cd5f8459 100644
--- a/apps/desktop/src/main/maker-host/pi-provider-transport.ts
+++ b/apps/desktop/src/main/maker-host/pi-provider-transport.ts
@@ -185,6 +185,18 @@ function cloudflareGatewayHeaders(
};
}
+/**
+ * Pi adapters read a missing level as "thinking off" (Z.AI sends `thinking.type=disabled`).
+ * Models whose catalog marks `off` unsupported reject that: GLM-5.3 answers 400/1210 (#5402).
+ * For those, a request without an effort uses the declared default instead.
+ */
+function nativeThinkingLevel(requested: unknown, row: ProviderModelRecord, model: Model<Api>): ThinkingLevel | undefined {
+ const efforts = row.efforts ?? [];
+ const effort = reconcileOutboundReasoningEffort(requested, efforts);
+ if (effort || !model.reasoning || model.thinkingLevelMap?.off !== null) return effort as ThinkingLevel | undefined;
+ return reconcileOutboundReasoningEffort(row.defaultEffort, efforts) as ThinkingLevel | undefined;
+}
+
function nativeInvocationModel(options: PiProviderTransportOptions, modelId: string): Model<Api> {
const row = options.row;
const destination = options.upstream ?? row.upstream;
@@ -307,7 +319,7 @@ export function createPiProviderFetch(options: PiProviderTransportOptions): type
['openai-responses', 'azure-openai-responses', 'openai-completions'].includes(model.api)
? { onPayload: (payload: unknown) => payload && typeof payload === 'object' && !Array.isArray(payload)
? { ...payload, service_tier: 'priority' } : payload } : {}),
- reasoning: reconcileOutboundReasoningEffort(request.reasoning?.effort, options.row.efforts ?? []) as ThinkingLevel | undefined,
+ reasoning: nativeThinkingLevel(request.reasoning?.effort, options.row, model),
maxTokens: typeof request.max_output_tokens === 'number' ? Math.min(request.max_output_tokens, model.maxTokens) : model.maxTokens,
});
const iterator = events[Symbol.asyncIterator]();
diff --git a/packages/maker-core/src/agents/base-agent.ts b/packages/maker-core/src/agents/base-agent.ts
index bd6f2961d4..6394b07157 100644
--- a/packages/maker-core/src/agents/base-agent.ts
+++ b/packages/maker-core/src/agents/base-agent.ts
@@ -997,6 +997,17 @@ export interface AgentDeps {
modelId: string,
) => number | null;
+ /**
+ * Resolve the declared efforts of a concrete (provider, model) route, used to
+ * narrow an outgoing effort to what that route accepts. Return null for unknown
+ * or ambiguous routes. Same-ID models from different providers can declare
+ * different efforts, so do not use capabilities.availableModels for this.
+ */
+ resolveModelEfforts?: (
+ providerId: string | null | undefined,
+ modelId: string,
+ ) => readonly Effort[] | null;
+
/** Local disk-auth policy, independent of the actual Provider credential mode. */
resolveCodexLocalAuthPolicy?: (
providerId: string | null | undefined,
diff --git a/packages/maker-core/src/agents/claude-code/__tests__/rewind-runtime-settings.test.ts b/packages/maker-core/src/agents/claude-code/__tests__/rewind-runtime-settings.test.ts
index 2e88cd699b..dd4eb60288 100644
--- a/packages/maker-core/src/agents/claude-code/__tests__/rewind-runtime-settings.test.ts
+++ b/packages/maker-core/src/agents/claude-code/__tests__/rewind-runtime-settings.test.ts
@@ -25,6 +25,7 @@ import type { AuthAdapter } from '../../../interfaces/auth-adapter.js';
import type { AgentEvent } from '../../../types/events.js';
import type { Logger } from '../../../interfaces/logger.js';
import type { ModelDescriptor } from '../../../types/capabilities.js';
+import type { Effort } from '../../../types/common.js';
const sdkMock = vi.hoisted(() => ({
forkSession: vi.fn(),
@@ -206,8 +207,10 @@ async function startRewindableSession(
idleTimeoutMs?: number;
remoteHostId?: string;
model?: string;
+ effort?: Effort;
availableModels?: ModelDescriptor[];
resolveModelContextLimit?: AgentDeps['resolveModelContextLimit'];
+ resolveModelEfforts?: AgentDeps['resolveModelEfforts'];
shouldHandoffAfterContextAssessment?: (tokens: number, window: number) => boolean;
} = {},
) {
@@ -240,11 +243,13 @@ async function startRewindableSession(
),
capabilityAdditions: { availableModels: options.availableModels ?? TEST_MODELS },
resolveModelContextLimit: options.resolveModelContextLimit,
+ ...(options.resolveModelEfforts ? { resolveModelEfforts: options.resolveModelEfforts } : {}),
...(remoteCcQueryFactory ? { remoteCcQueryFactory } : {}),
});
const handle = await agent.startSession({
sessionId: 'session-rewind',
model: options.model ?? 'claude-opus-4-6',
+ ...(options.effort ? { effort: options.effort } : {}),
workingDir,
permissionMode: 'acceptEdits',
...(options.remoteHostId ? { remoteHostId: options.remoteHostId } : {}),
@@ -677,6 +682,194 @@ describe('ClaudeCodeAgent runtime settings during rewind window', () => {
await handle.close();
});
+ it('clamps a carried-over xhigh to the target model efforts on start and live change (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'z-ai/glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const { handle, firstQuery } = await startRewindableSession({
+ model: glm.id,
+ effort: 'xhigh',
+ availableModels: [...TEST_MODELS, glm],
+ });
+
+ expect(sdkMock.query.mock.calls[0]?.[0]?.options?.effort).toBe('high');
+
+ await handle.setEffort?.('xhigh');
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'high' });
+
+ await handle.setModel?.('claude-sonnet-5');
+ await handle.setEffort?.('xhigh');
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'xhigh' });
+
+ await handle.close();
+ });
+
+ it('re-applies the clamped effort on a hot model switch without a separate setEffort (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'z-ai/glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const { handle, firstQuery } = await startRewindableSession({
+ model: 'claude-sonnet-5',
+ effort: 'xhigh',
+ availableModels: [...TEST_MODELS, glm],
+ });
+
+ // The live Query keeps the sticky xhigh; the switch itself must narrow it.
+ await handle.setModel?.(glm.id);
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'high' });
+
+ // Switching back restores the selected xhigh for the model that supports it.
+ await handle.setModel?.('claude-sonnet-5');
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'xhigh' });
+
+ // An effort passed with the switch wins over the previous one.
+ await handle.setModel?.(glm.id, { effort: 'max' });
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'max' });
+ expect(handle.getEffort?.()).toBe('max');
+
+ await handle.close();
+ });
+
+ it('re-applies on a hot switch even when the selected effort is unchanged, since the runtime may have fallen back (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'z-ai/glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const { handle, firstQuery } = await startRewindableSession({
+ model: 'claude-sonnet-5',
+ availableModels: [...TEST_MODELS, glm],
+ });
+ // An older runtime rejects max; the live Query actually keeps xhigh while the session keeps max.
+ firstQuery.applyFlagSettings
+ .mockRejectedValueOnce(new Error('invalid effortLevel: max'))
+ .mockResolvedValueOnce(undefined);
+ await handle.setEffort?.('max');
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'xhigh' });
+
+ // Same selected max on GLM-5.3: the sticky xhigh must still be replaced.
+ await handle.setModel?.(glm.id);
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'max' });
+
+ await handle.close();
+ });
+
+ it('re-applies the narrowed effort when the model drifts during query rebuild (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'z-ai/glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const { handle } = await startRewindableSession({
+ model: 'claude-sonnet-5',
+ effort: 'xhigh',
+ availableModels: [...TEST_MODELS, glm],
+ });
+ await handle.commitRewindFiles?.('user-uuid-1', 'assistant-uuid-1');
+
+ const secondQuery = createFakeQuery();
+ sdkMock.query.mockImplementationOnce(() => {
+ // Only the model drifts; the selected effort stays xhigh.
+ void handle.setModel?.(glm.id);
+ return secondQuery;
+ });
+ await handle.send({ type: 'user', content: 'switch to glm during rebuild' });
+
+ expect(secondQuery.setModel).toHaveBeenCalled();
+ expect(secondQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'high' });
+
+ await handle.close();
+ });
+
+ it('narrows with the efforts of the session route, not the first same-ID descriptor (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const resolveModelEfforts = vi.fn((providerId: string | null | undefined) =>
+ providerId === 'custom-glm' ? (['low', 'max'] as const) : null);
+ const { handle, firstQuery } = await startRewindableSession({
+ model: 'claude-sonnet-5',
+ effort: 'xhigh',
+ availableModels: [...TEST_MODELS, glm],
+ resolveModelEfforts,
+ });
+
+ await handle.setModel?.(glm.id, { providerId: 'custom-glm' });
+ expect(resolveModelEfforts).toHaveBeenCalledWith('custom-glm', glm.id);
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'low' });
+
+ // An unknown or ambiguous route keeps the selected effort instead of borrowing a descriptor.
+ await handle.setModel?.(glm.id, { providerId: 'other' });
+ expect(firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'xhigh' });
+
+ await handle.close();
+ });
+
+ it('uses the target route for the max fallback and replays a provider-only drift (#5402)', async () => {
+ const glm: ModelDescriptor = {
+ id: 'glm-5.3',
+ displayName: 'GLM-5.3',
+ contextWindow: 1_000_000,
+ efforts: ['low', 'medium', 'high', 'max'],
+ defaultEffort: 'high',
+ };
+ const routes: Record<string, readonly Effort[]> = {
+ preset: ['low', 'medium', 'high', 'max'],
+ custom: ['low', 'max'],
+ };
+ const resolveModelEfforts = (providerId: string | null | undefined) =>
+ (providerId ? routes[providerId] : undefined) ?? null;
+
+ // An older runtime rejects max: the retry must come from the custom route, not the first descriptor.
+ const live = await startRewindableSession({
+ model: glm.id,
+ effort: 'high',
+ availableModels: [...TEST_MODELS, glm],
+ resolveModelEfforts,
+ });
+ live.firstQuery.applyFlagSettings.mockImplementation(async (settings: Record<string, unknown>) => {
+ if (settings.effortLevel === 'max') throw new Error('invalid effortLevel: max');
+ });
+ await live.handle.setModel?.(glm.id, { providerId: 'custom', effort: 'max' });
+ expect(live.firstQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'low' });
+ await live.handle.close();
+
+ // Same model ID switched to another provider while the Query is rebuilding.
+ sdkMock.query.mockReset();
+ const { handle } = await startRewindableSession({
+ model: glm.id,
+ effort: 'high',
+ availableModels: [...TEST_MODELS, glm],
+ resolveModelEfforts,
+ });
+ await handle.commitRewindFiles?.('user-uuid-1', 'assistant-uuid-1');
+ const secondQuery = createFakeQuery();
+ sdkMock.query.mockImplementationOnce(() => {
+ void handle.setModel?.(glm.id, { providerId: 'custom' });
+ return secondQuery;
+ });
+ await handle.send({ type: 'user', content: 'switch provider during rebuild' });
+ expect(secondQuery.applyFlagSettings).toHaveBeenLastCalledWith({ effortLevel: 'low' });
+
+ await handle.close();
+ });
+
it('falls back to the model-supported xhigh when an older runtime rejects max', async () => {
const { handle, firstQuery } = await startRewindableSession();
@@ -2608,12 +2801,12 @@ describe('ClaudeCodeAgent runtime settings during rewind window', () => {
await handle.send({ type: 'user', content: 'long bridge compact' });
await expect(handle.setModel?.('claude-opus-4-6')).resolves.toBeUndefined();
- await expect(handle.setEffort?.('xhigh')).resolves.toBeUndefined();
+ await expect(handle.setEffort?.('max')).resolves.toBeUndefined();
await expect(handle.setFastMode?.(true)).resolves.toBeUndefined();
await expect(handle.setPermissionMode?.('auto')).resolves.toBeUndefined();
expect(secondQuery.setModel).toHaveBeenCalledWith('claude-opus-4-6[1m]');
- expect(secondQuery.applyFlagSettings).toHaveBeenCalledWith({ effortLevel: 'xhigh' });
+ expect(secondQuery.applyFlagSettings).toHaveBeenCalledWith({ effortLevel: 'max' });
expect(secondQuery.applyFlagSettings).toHaveBeenCalledWith({ fastMode: true });
// Cindy 档 'auto' 映射到 SDK 'default'(见 toSdkPermissionMode)。
expect(secondQuery.setPermissionMode).toHaveBeenCalledWith('default');
diff --git a/packages/maker-core/src/agents/claude-code/index.ts b/packages/maker-core/src/agents/claude-code/index.ts
index 608acc6846..41a4b190ac 100644
--- a/packages/maker-core/src/agents/claude-code/index.ts
+++ b/packages/maker-core/src/agents/claude-code/index.ts
@@ -174,6 +174,7 @@ import { isClaudeResumeSessionNotFound } from './invalid-resume.js';
import { translateSdkMessage, newRuntimeState, type TurnState, type RuntimeState } from './translator.js';
import { resetClaudeGenerationTiming } from './generation-timing.js';
import type { Effort, PermissionMode } from '../../types/common.js';
+import { clampEffortToSupported } from '@cindy/model-providers/effort-resolution';
import type {
ScanAtResourcesOptions,
ScanAtResourcesResult,
@@ -949,16 +950,37 @@ export class ClaudeCodeAgent extends BaseAgent {
this.capabilities = this.buildCapabilities(CAPABILITIES);
}
- private sdkEffortForModel(model: string, effort: Effort): ClaudeSdkEffort | undefined {
- const descriptor = this.capabilities.availableModels.find((m) => m.id === model);
- if (descriptor && descriptor.efforts.length === 0) return undefined;
- return clampEffortForClaude(effort);
+ /**
+ * 目标路由声明的档位。availableModels 按 ID 跨来源去重,同名模型的档位可能来自别的来源;
+ * 有 host 解析器时按实际来源取,来源未知或不唯一返回 null(调用方不得借用首见 descriptor)。
+ * 主档位收窄与 max 兼容回退共用这一判据。
+ */
+ private routeEffortsForModel(model: string, providerId?: string | null): readonly Effort[] | null {
+ if (this.deps.resolveModelEfforts) return this.deps.resolveModelEfforts(providerId, model);
+ return this.capabilities.availableModels.find((m) => m.id === model)?.efforts ?? null;
}
- private sdkMaxEffortFallbackForModel(model: string): Exclude<ClaudeSdkEffort, 'max'> {
+ private sdkEffortForModel(
+ model: string,
+ effort: Effort,
+ providerId?: string | null,
+ ): ClaudeSdkEffort | undefined {
const descriptor = this.capabilities.availableModels.find((m) => m.id === model);
- if (!descriptor) return 'xhigh';
- const supported = new Set(descriptor.efforts.map(clampEffortForClaude));
+ // 来源不明(null)时不收窄,只保留原有的「无档位模型不下发」判断。
+ const routeEfforts = this.routeEffortsForModel(model, providerId);
+ if ((routeEfforts ?? descriptor?.efforts)?.length === 0) return undefined;
+ // 会话档位可能来自上一个模型(如 Claude 的 xhigh),原样下发会被只认目录档位的
+ // 上游拒绝(GLM-5.3 收到 xhigh 回 400/1210,#5402)。按目标来源声明的档位收窄。
+ return clampEffortForClaude((clampEffortToSupported(effort, routeEfforts ?? undefined) as Effort | undefined) ?? effort);
+ }
+
+ private sdkMaxEffortFallbackForModel(
+ model: string,
+ providerId?: string | null,
+ ): Exclude<ClaudeSdkEffort, 'max'> {
+ const efforts = this.routeEffortsForModel(model, providerId);
+ if (!efforts) return 'xhigh';
+ const supported = new Set(efforts.map(clampEffortForClaude));
for (const candidate of ['xhigh', 'high', 'medium', 'low'] as const) {
if (supported.has(candidate)) return candidate;
}
@@ -1394,7 +1416,7 @@ export class ClaudeCodeAgent extends BaseAgent {
const resolveRemoteClaudeRoute = this.deps.resolveRemoteClaudeRoute?.bind(this.deps);
const getAuthEnv = this.deps.auth.getAuthEnv.bind(this.deps.auth);
const sdkModel = sdkModelFor(opts.model);
- const initialSdkEffort = this.sdkEffortForModel(opts.model, opts.effort ?? 'high');
+ const initialSdkEffort = this.sdkEffortForModel(opts.model, opts.effort ?? 'high', opts.providerId);
const binaryPath = this.deps.binaryPath;
const providerRoutedModels = this.capabilities.availableModels.filter((model) =>
isProviderRoutedModel(model.id),
@@ -2663,10 +2685,10 @@ export class ClaudeCodeAgent extends BaseAgent {
// file checkpointing 与 capability 强绑定 —— 声明 rewind 能力时必须开此开关,
// 否则 SDK rewindFiles() 报 "no checkpoint"。
const enableFileCheckpointing = this.capabilities.rewind.supported;
- const getSdkEffortForModel = (model: string, effort: Effort) =>
- this.sdkEffortForModel(model, effort);
- const getSdkMaxEffortFallbackForModel = (model: string) =>
- this.sdkMaxEffortFallbackForModel(model);
+ const getSdkEffortForModel = (model: string, effort: Effort, providerId: string | null) =>
+ this.sdkEffortForModel(model, effort, providerId);
+ const getSdkMaxEffortFallbackForModel = (model: string, providerId: string | null) =>
+ this.sdkMaxEffortFallbackForModel(model, providerId);
// memoryOverride 闭包以前抽过 getter, buildSettings 接管后直接读 this.memoryOverride。
@@ -3341,7 +3363,7 @@ export class ClaudeCodeAgent extends BaseAgent {
appliedContextWindow = workingWindow;
applyClaudeContextWindow(env, workingWindow, this.deps.runtimeConfig.autoCompactThresholdPct);
if (remoteEnv) applyClaudeContextWindow(remoteEnv, workingWindow, this.deps.runtimeConfig.autoCompactThresholdPct);
- const currentSdkEffort = getSdkEffortForModel(mutableModel, mutableEffort);
+ const currentSdkEffort = getSdkEffortForModel(mutableModel, mutableEffort, mutableProviderId);
const baseResumeAt = vo.resumeSessionAt as string | undefined;
const baseFork = vo.forkSession as boolean | undefined;
const finalResumeAt = extra?.fresh ? undefined : (extra?.resumeSessionAt ?? baseResumeAt);
@@ -5707,6 +5729,7 @@ export class ClaudeCodeAgent extends BaseAgent {
// 避免新 query 带旧 model/flags 起跑而 handle getter 报新值。
const runtimeSnapshot: QueryRuntimeSnapshot = {
model: mutableModel,
+ providerId: mutableProviderId,
effort: mutableEffort,
fastMode: mutableFastMode,
sdkPermissionMode: currentTurnSdkPermissionMode(),
@@ -5815,10 +5838,24 @@ export class ClaudeCodeAgent extends BaseAgent {
canceledBridgeQueries.has(q);
type QueryRuntimeSnapshot = {
model: string;
+ providerId: string | null;
effort: Effort;
fastMode: boolean;
sdkPermissionMode: SdkPermissionMode;
};
+ /**
+ * 按目标路由收窄档位并写入当前 Query。effortLevel 是 sticky flag,且旧 runtime 拒绝
+ * max 时实际生效的是回退档,与 mutableEffort 不一定相同;因此切模(含重建回放)后一律
+ * 重下发,不按「新旧档位相同」跳过,否则上一个模型的 xhigh 会打给 GLM-5.3(1210,#5402)。
+ */
+ async function applyRouteEffort(model: string, effort: Effort, providerId: string | null): Promise<void> {
+ const sdkEffort = getSdkEffortForModel(model, effort, providerId);
+ if (!sdkEffort) return;
+ const appliedEffort = await applyClaudeEffortFlagSettings(
+ q, sdkEffort, getSdkMaxEffortFallbackForModel(model, providerId),
+ );
+ log.debug('applied route effort', { model, effort, sdk: appliedEffort, downgraded: appliedEffort !== sdkEffort });
+ }
async function replayRuntimeDrift(snapshot: QueryRuntimeSnapshot, label: string): Promise<void> {
for (let pass = 0; pass < 5; pass += 1) {
let replayed = false;
@@ -5834,31 +5871,29 @@ export class ClaudeCodeAgent extends BaseAgent {
await q.setModel(sdkModelFor(targetModel));
snapshot.model = targetModel;
log.debug(`${label}: replayed setModel`, { model: targetModel });
+ // 新 Query 的档位按旧模型构建;切模漂移时同 live setModel 一样按目标路由重下发。
+ const targetEffort = mutableEffort;
+ const targetProviderId = mutableProviderId;
+ await applyRouteEffort(targetModel, targetEffort, targetProviderId);
+ snapshot.effort = targetEffort;
+ snapshot.providerId = targetProviderId;
} catch (e) {
log.warn(`${label}: replay setModel failed`, { error: String(e) });
}
}
- if (mutableEffort !== snapshot.effort) {
+ // 同一 model ID 换来源(受阻 setModel 只改 mutableProviderId)时档位能力也可能不同。
+ if (mutableEffort !== snapshot.effort || mutableProviderId !== snapshot.providerId) {
replayed = true;
const targetEffort = mutableEffort;
- const sdkEffort = getSdkEffortForModel(mutableModel, targetEffort);
- if (sdkEffort) {
- try {
- const appliedEffort = await applyClaudeEffortFlagSettings(
- q,
- sdkEffort,
- getSdkMaxEffortFallbackForModel(mutableModel),
- );
- log.debug(`${label}: replayed setEffort`, {
- effort: targetEffort,
- sdk: appliedEffort,
- downgraded: appliedEffort !== sdkEffort,
- });
- } catch (e) {
- log.warn(`${label}: replay setEffort failed`, { error: String(e) });
- }
+ const targetProviderId = mutableProviderId;
+ try {
+ await applyRouteEffort(mutableModel, targetEffort, targetProviderId);
+ log.debug(`${label}: replayed setEffort`, { effort: targetEffort });
+ } catch (e) {
+ log.warn(`${label}: replay setEffort failed`, { error: String(e) });
}
snapshot.effort = targetEffort;
+ snapshot.providerId = targetProviderId;
}
if (mutableFastMode !== snapshot.fastMode) {
replayed = true;
@@ -5905,6 +5940,7 @@ export class ClaudeCodeAgent extends BaseAgent {
const staleQuery = q;
const runtimeSnapshot: QueryRuntimeSnapshot = {
model: mutableModel,
+ providerId: mutableProviderId,
effort: mutableEffort,
fastMode: mutableFastMode,
sdkPermissionMode: currentTurnSdkPermissionMode(),
@@ -6227,6 +6263,7 @@ export class ClaudeCodeAgent extends BaseAgent {
// 闭包值; await 期间若有切换到达(被 controlRequestsBlocked() 短路成"只更新闭包"),
// 下方 diff 重放据此识别漂移项。
const snapModel = mutableModel;
+ const snapProviderId = mutableProviderId;
const snapEffort = mutableEffort;
const snapFastMode = mutableFastMode;
// 用 turn-scoped 档快照 (planTurnActive + mutablePermissionMode), 不含 mutablePlanMode
@@ -6288,6 +6325,7 @@ export class ClaudeCodeAgent extends BaseAgent {
clearBridgeState();
runtimeReplaySnapshot = {
model: snapModel,
+ providerId: snapProviderId,
effort: snapEffort,
fastMode: snapFastMode,
sdkPermissionMode: snapSdkPermissionMode,
@@ -6967,7 +7005,7 @@ export class ClaudeCodeAgent extends BaseAgent {
return expected !== liveEnv?.CLAUDE_CODE_AUTO_COMPACT_WINDOW;
},
- async setModel(newModel: string, setModelOpts?: { providerId?: string | null }) {
+ async setModel(newModel: string, setModelOpts?: { providerId?: string | null; effort?: Effort }) {
if (reviewMode) return;
const targetProviderId = setModelOpts?.providerId !== undefined
? setModelOpts.providerId
@@ -7085,7 +7123,16 @@ export class ClaudeCodeAgent extends BaseAgent {
availableModels: currentAvailableSdkModels(newModel),
});
await q.setModel(sdkModel);
+ // 切模已生效:档位重下发失败只 warn,不把整个 setModel 报成失败(同下方 auto 审查重配)。
+ await applyRouteEffort(newModel, setModelOpts?.effort ?? mutableEffort, targetProviderId ?? null)
+ .catch((e) => {
+ log.warn('setModel: effort reapply for target model failed; model switch kept', {
+ model: newModel,
+ error: String(e),
+ });
+ });
}
+ if (setModelOpts?.effort) mutableEffort = setModelOpts.effort;
const usedNativeAutoReview = usesNativeClaudeAutoReview();
mutableProviderId = targetProviderId ?? null;
mutableAutoReviewCredentialMode = resolveEffectiveCredentialModeFromAuthSource(
@@ -7158,7 +7205,7 @@ export class ClaudeCodeAgent extends BaseAgent {
if (reviewMode) return;
// maker 的 minimal / ultra 先归一成 Claude 的 low / max;2.1.219 起
// applyFlagSettings 可原样接收 max,不能再静默降成 xhigh。
- const sdkEffort = getSdkEffortForModel(mutableModel, newEffort);
+ const sdkEffort = getSdkEffortForModel(mutableModel, newEffort, mutableProviderId);
const isControlBlocked = controlRequestsBlocked();
log.debug('setEffort', { from: mutableEffort, to: newEffort, sdk: sdkEffort, controlRequestsBlocked: isControlBlocked });
if (!sdkEffort) {
@@ -7169,7 +7216,7 @@ export class ClaudeCodeAgent extends BaseAgent {
const appliedEffort = await applyClaudeEffortFlagSettings(
q,
sdkEffort,
- getSdkMaxEffortFallbackForModel(mutableModel),
+ getSdkMaxEffortFallbackForModel(mutableModel, mutableProviderId),
);
if (appliedEffort !== sdkEffort) {
log.warn('setEffort: runtime rejected max; applied model-compatible fallback', {
diff --git a/packages/model-providers/catalog/model-registry.json b/packages/model-providers/catalog/model-registry.json
index 290e24ca00..0212b9c057 100644
--- a/packages/model-providers/catalog/model-registry.json
+++ b/packages/model-providers/catalog/model-registry.json
@@ -1,6 +1,6 @@
{
"schemaVersion": 5,
- "updatedAt": "2026-09-29T19:00:00.002Z",
+ "updatedAt": "2026-10-03T12:00:00.000Z",
"models": [
{
"id": "openai/gpt-6.1-sol",
@@ -5465,7 +5465,7 @@
},
{
"id": "z-ai/glm-5.3",
- "aliases": [],
+ "aliases": ["glm-5.3"],
"defaults": {
"name": "GLM-5.3",
"description": "Zhipu GLM-5.3 agentic coding model; 1M context",