diff --git a/src/cliproxy/catalog-cache.ts b/src/cliproxy/catalog-cache.ts index 10436887..998c86c3 100644 --- a/src/cliproxy/catalog-cache.ts +++ b/src/cliproxy/catalog-cache.ts @@ -264,9 +264,17 @@ export function mergeCatalog( mergedIds.add(remote.id.toLowerCase()); if (staticEntry) { + const mergedThinking = remoteEntry.thinking + ? { + ...remoteEntry.thinking, + maxLevel: remoteEntry.thinking.maxLevel ?? staticEntry.thinking?.maxLevel, + } + : staticEntry.thinking; + // Merge: remote overrides, static fills gaps mergedModels.push({ ...remoteEntry, + thinking: mergedThinking, // Preserve static-only fields tier: staticEntry.tier, broken: staticEntry.broken, diff --git a/src/cliproxy/config/thinking-config.ts b/src/cliproxy/config/thinking-config.ts index d9f2b39f..ef98234e 100644 --- a/src/cliproxy/config/thinking-config.ts +++ b/src/cliproxy/config/thinking-config.ts @@ -7,7 +7,7 @@ import type { CLIProxyProvider } from '../types'; import { DEFAULT_THINKING_TIER_DEFAULTS } from '../../config/unified-config-types'; import type { ThinkingConfig } from '../../config/unified-config-types'; import { getThinkingConfig } from '../../config/unified-config-loader'; -import { supportsThinking } from '../model-catalog'; +import { getModelThinkingSupport, supportsThinking } from '../model-catalog'; import { isThinkingOffValue, validateThinking } from '../thinking-validator'; import { normalizeModelIdForProvider } from '../model-id-normalizer'; import { warn } from '../../utils/ui'; @@ -86,7 +86,8 @@ function applyThinkingSuffixForProvider( const parenthesizedSuffixMatch = model.match(/\(([^)]+)\)$/); // Existing parenthesized suffix: - // - keep as-is for non-codex providers + // - keep as-is for non-codex providers unless the target model now expects + // named levels and we need to rewrite an old numeric suffix // - for codex effort levels, normalize to codex model suffix style if (parenthesizedSuffixMatch) { if (provider === 'codex') { @@ -95,6 +96,15 @@ function applyThinkingSuffixForProvider( return model.replace(/\([^)]+\)$/, `-${normalizedParensValue}`); } } + + if (provider) { + const normalizedBaseModel = normalizeModelForThinkingLookup(model, provider); + const thinking = getModelThinkingSupport(provider, normalizedBaseModel); + if (thinking?.type === 'levels') { + return model.replace(/\([^)]+\)$/, `(${thinkingValue})`); + } + } + return model; } diff --git a/src/cursor/cursor-anthropic-translator.ts b/src/cursor/cursor-anthropic-translator.ts index 6dbe7d92..416b4246 100644 --- a/src/cursor/cursor-anthropic-translator.ts +++ b/src/cursor/cursor-anthropic-translator.ts @@ -68,17 +68,30 @@ function toToolResultContent(content: unknown, label: string): string { return safeJsonStringify(content, TOOL_RESULT_SERIALIZATION_FALLBACK); } -function mapThinkingToReasoningEffort( - thinking: CursorAnthropicRequest['thinking'] -): string | undefined { +function mapAdaptiveEffortToCursorReasoningEffort(effort: string | undefined): string { + const normalized = effort?.trim().toLowerCase(); + if (!normalized || normalized === 'auto') { + return 'high'; + } + if (normalized === 'minimal' || normalized === 'low' || normalized === 'medium') { + return 'medium'; + } + return 'high'; +} + +function mapThinkingToReasoningEffort(request: CursorAnthropicRequest): string | undefined { + const thinking = request.thinking; if (!thinking) { return undefined; } if (thinking.type === 'disabled') { return undefined; } + if (thinking.type === 'adaptive') { + return mapAdaptiveEffortToCursorReasoningEffort(request.output_config?.effort); + } if (thinking.type !== 'enabled') { - throw new Error('thinking.type must be "enabled" or "disabled"'); + throw new Error('thinking.type must be "enabled", "adaptive", or "disabled"'); } return typeof thinking.budget_tokens === 'number' && thinking.budget_tokens >= 8192 ? 'high' @@ -209,7 +222,7 @@ export function translateAnthropicRequest(raw: unknown): TranslatedAnthropicRequ ? request.model : undefined, stream: request.stream === true, - reasoning_effort: mapThinkingToReasoningEffort(request.thinking), + reasoning_effort: mapThinkingToReasoningEffort(request), tools: Array.isArray(request.tools) ? request.tools : undefined, messages: translatedMessages, }; diff --git a/src/cursor/cursor-anthropic-types.ts b/src/cursor/cursor-anthropic-types.ts index 336fb75f..9374de5b 100644 --- a/src/cursor/cursor-anthropic-types.ts +++ b/src/cursor/cursor-anthropic-types.ts @@ -41,6 +41,9 @@ export interface CursorAnthropicRequest { system?: string | AnthropicTextBlock[]; stream?: boolean; tools?: CursorTool[]; + output_config?: { + effort?: string; + }; thinking?: { type?: string; budget_tokens?: number; diff --git a/src/proxy/transformers/request-transformer.ts b/src/proxy/transformers/request-transformer.ts index a8ccb98c..58e7969e 100644 --- a/src/proxy/transformers/request-transformer.ts +++ b/src/proxy/transformers/request-transformer.ts @@ -396,7 +396,7 @@ function mapThinkingToReasoning( }; } -const VALID_EFFORT_LEVELS = new Set(['low', 'medium', 'high', 'max']); +const VALID_EFFORT_LEVELS = new Set(['low', 'medium', 'high', 'xhigh', 'max']); function resolveOutputConfigEffort( outputConfig: AnthropicOutputConfig | undefined @@ -416,7 +416,7 @@ function resolveOutputConfigEffort( * for Codex; for generic OpenAI-compat providers we clamp to high. */ function toOpenAIEffort(effort: string): string { - return effort === 'max' ? 'high' : effort; + return effort === 'max' || effort === 'xhigh' ? 'high' : effort; } function transformMessages(messagesValue: unknown): OpenAIMessage[] { diff --git a/src/targets/droid-config-manager.ts b/src/targets/droid-config-manager.ts index 29a3943e..29f3c92d 100644 --- a/src/targets/droid-config-manager.ts +++ b/src/targets/droid-config-manager.ts @@ -9,6 +9,9 @@ import * as fs from 'fs'; import * as path from 'path'; import * as os from 'os'; import * as lockfile from 'proper-lockfile'; +import { getModelThinkingSupport } from '../cliproxy/model-catalog'; +import { validateThinking } from '../cliproxy/thinking-validator'; +import { stripModelConfigurationSuffixes } from '../shared/extended-context-utils'; const CCS_MODEL_PREFIX = 'ccs-'; const CCS_DISPLAY_PREFIX = 'CCS '; @@ -144,6 +147,24 @@ function toAnthropicBudget(value: string | number): number { return DROID_ANTHROPIC_BUDGET_BY_EFFORT[normalized] ?? DROID_ANTHROPIC_BUDGET_BY_EFFORT.high; } +function resolveAnthropicModelId(model: string): string { + return stripModelConfigurationSuffixes(model); +} + +function usesAnthropicAdaptiveThinking(model: string): boolean { + return getModelThinkingSupport('claude', resolveAnthropicModelId(model))?.type === 'levels'; +} + +function toAnthropicAdaptiveEffort(model: string, value: string | number): string | undefined { + const validation = validateThinking('claude', resolveAnthropicModelId(model), value); + if (isReasoningOffValue(validation.value)) { + return undefined; + } + + const normalized = String(validation.value).trim().toLowerCase(); + return normalized === 'auto' ? undefined : normalized; +} + function toReasoningEffort(value: string | number): string { if (typeof value === 'number') { if (value <= 4000) return 'low'; @@ -181,12 +202,35 @@ function applyReasoningOverride( if (isReasoningOffValue(reasoningOverride)) { delete extraArgs.thinking; + delete extraArgs.output_config; + } else if (usesAnthropicAdaptiveThinking(entry.model)) { + const thinking = isObject(extraArgs.thinking) ? { ...extraArgs.thinking } : {}; + const outputConfig = isObject(extraArgs.output_config) ? { ...extraArgs.output_config } : {}; + const effort = toAnthropicAdaptiveEffort(entry.model, reasoningOverride); + + thinking.type = 'adaptive'; + delete thinking.budget_tokens; + delete thinking.budgetTokens; + + if (effort) { + outputConfig.effort = effort; + } else { + delete outputConfig.effort; + } + + extraArgs.thinking = thinking; + if (Object.keys(outputConfig).length > 0) { + extraArgs.output_config = outputConfig; + } else { + delete extraArgs.output_config; + } } else { const thinking = isObject(extraArgs.thinking) ? { ...extraArgs.thinking } : {}; thinking.type = 'enabled'; thinking.budget_tokens = toAnthropicBudget(reasoningOverride); delete thinking.budgetTokens; extraArgs.thinking = thinking; + delete extraArgs.output_config; } } else if (provider === 'openai') { delete extraArgs.reasoning_effort; diff --git a/tests/unit/cliproxy/composite-thinking.test.ts b/tests/unit/cliproxy/composite-thinking.test.ts index 6fb3758b..6a9de3fd 100644 --- a/tests/unit/cliproxy/composite-thinking.test.ts +++ b/tests/unit/cliproxy/composite-thinking.test.ts @@ -538,6 +538,20 @@ describe('applyThinkingConfig - composite variant integration', () => { expect(result.ANTHROPIC_DEFAULT_OPUS_MODEL).toBe('claude-opus-4-6-thinking(32768)'); }); + it('rewrites legacy budget suffixes for claude level-based models', () => { + const envVars: NodeJS.ProcessEnv = { + ANTHROPIC_MODEL: 'claude-opus-4-7(32768)', + ANTHROPIC_DEFAULT_OPUS_MODEL: 'claude-opus-4-7(32768)', + ANTHROPIC_DEFAULT_SONNET_MODEL: 'claude-sonnet-4-6', + ANTHROPIC_DEFAULT_HAIKU_MODEL: 'claude-haiku-4-5-20251001', + }; + + const result = applyThinkingConfig(envVars, 'claude' as CLIProxyProvider, 'max'); + + expect(result.ANTHROPIC_MODEL).toBe('claude-opus-4-7(max)'); + expect(result.ANTHROPIC_DEFAULT_OPUS_MODEL).toBe('claude-opus-4-7(max)'); + }); + it('should use codex effort suffix style when provider is codex', () => { const envVars: NodeJS.ProcessEnv = { ANTHROPIC_MODEL: 'gpt-5.3-codex', diff --git a/tests/unit/cliproxy/model-catalog-compat.test.ts b/tests/unit/cliproxy/model-catalog-compat.test.ts index b6fc21cc..0590a44b 100644 --- a/tests/unit/cliproxy/model-catalog-compat.test.ts +++ b/tests/unit/cliproxy/model-catalog-compat.test.ts @@ -58,4 +58,24 @@ describe('model-catalog compatibility lookups', () => { expect(catalog?.models.map((model) => model.id)).toEqual(['gemini-2.5-pro']); }); + + it('preserves static maxLevel when live thinking metadata omits it', () => { + const catalog = mergeCatalog('claude', [ + { + id: 'claude-opus-4-7', + display_name: 'Claude Opus 4.7', + thinking: { + levels: ['low', 'medium', 'high', 'xhigh', 'max'], + dynamic_allowed: true, + }, + }, + ]); + + expect(catalog?.models[0]?.thinking).toMatchObject({ + type: 'levels', + levels: ['low', 'medium', 'high', 'xhigh', 'max'], + maxLevel: 'max', + dynamicAllowed: true, + }); + }); }); diff --git a/tests/unit/cursor/cursor-anthropic-translator.test.ts b/tests/unit/cursor/cursor-anthropic-translator.test.ts index 9da161fe..2fce792c 100644 --- a/tests/unit/cursor/cursor-anthropic-translator.test.ts +++ b/tests/unit/cursor/cursor-anthropic-translator.test.ts @@ -90,6 +90,16 @@ describe('translateAnthropicRequest', () => { ]); }); + it('maps adaptive anthropic thinking into Cursor reasoning effort', () => { + const translated = translateAnthropicRequest({ + thinking: { type: 'adaptive' }, + output_config: { effort: 'xhigh' }, + messages: [{ role: 'user', content: 'hello' }], + }); + + expect(translated.reasoning_effort).toBe('high'); + }); + it('preserves mixed user text around tool_result blocks in order', () => { const translated = translateAnthropicRequest({ messages: [ diff --git a/tests/unit/proxy/transformers/request-transformer-regressions.test.ts b/tests/unit/proxy/transformers/request-transformer-regressions.test.ts index 77b56294..4bb3192f 100644 --- a/tests/unit/proxy/transformers/request-transformer-regressions.test.ts +++ b/tests/unit/proxy/transformers/request-transformer-regressions.test.ts @@ -30,6 +30,17 @@ describe('ProxyRequestTransformer regressions', () => { expect(result.reasoning).toEqual({ enabled: true, effort: 'high' }); }); + it('explicitly normalizes anthropic xhigh adaptive effort for OpenAI-compatible upstreams', () => { + const result = new ProxyRequestTransformer().transform({ + messages: [{ role: 'user', content: 'hello' }], + thinking: { type: 'adaptive' }, + output_config: { effort: 'xhigh' }, + }); + + expect(result.reasoning_effort).toBe('high'); + expect(result.reasoning).toEqual({ enabled: true, effort: 'high' }); + }); + it('rejects unsupported thinking types instead of silently dropping them', () => { expect(() => new ProxyRequestTransformer().transform({ diff --git a/tests/unit/targets/droid-config-manager.test.ts b/tests/unit/targets/droid-config-manager.test.ts index 8f0efaa1..21128d55 100644 --- a/tests/unit/targets/droid-config-manager.test.ts +++ b/tests/unit/targets/droid-config-manager.test.ts @@ -143,6 +143,23 @@ describe('droid-config-manager', () => { expect(settings.customModels[0].extraArgs?.thinking?.budget_tokens).toBe(40960); }); + it('writes adaptive anthropic thinking for claude-opus-4-7 overrides', async () => { + await upsertCcsModel('claude', { + model: 'claude-opus-4-7', + displayName: 'CCS claude', + baseUrl: 'https://api.anthropic.com', + apiKey: 'anthropic-key', + provider: 'anthropic', + reasoningOverride: 'max', + }); + + const settingsPath = path.join(tmpDir, '.factory', 'settings.json'); + const settings = JSON.parse(fs.readFileSync(settingsPath, 'utf8')); + expect(settings.customModels[0].extraArgs?.thinking?.type).toBe('adaptive'); + expect(settings.customModels[0].extraArgs?.thinking?.budget_tokens).toBeUndefined(); + expect(settings.customModels[0].extraArgs?.output_config?.effort).toBe('max'); + }); + it('should clear prior reasoning config when override disables thinking', async () => { await upsertCcsModel('glm', { model: 'glm-4.7',