Merge pull request #135 from Askhz/feat/costrict-cheapest-model-for-haiku-tasks

feat(costrict): use cheapest model by creditConsumption for haiku-tier auxiliary tasks

src/costrict/provider/modelMapping.ts

@@ -3,7 +3,7 @@
  * 将 Anthropic 模型名映射到 CoStrict 模型名
  */
 
-import { getCachedCoStrictModels } from './models.js'
+import { getCachedCoStrictModels, getCheapestCoStrictModel } from './models.js'
 
 function getModelFamily(model: string): 'haiku' | 'sonnet' | 'opus' | null {
   if (/haiku/i.test(model)) return 'haiku'
@@ -34,6 +34,11 @@ export function resolveCoStrictModel(anthropicModel: string): string {
     const envVar = `COSTRICT_DEFAULT_${family.toUpperCase()}_MODEL`
     const override = process.env[envVar]
     if (override) return override
+    // haiku 族无 env var 时，自动选取 creditConsumption 最低的模型
+    if (family === 'haiku') {
+      const cheapest = getCheapestCoStrictModel()
+      if (cheapest) return cheapest
+    }
   }
 
   // 优先级 3: 直接透传原始模型名（用户手动配置的模型名优先于环境变量）

src/costrict/provider/models.ts

@@ -90,3 +90,17 @@ export function clearModelCache(): void {
 export function getCachedCoStrictModels(): CoStrictModel[] {
   return modelCache?.models ?? []
 }
+
+/**
+ * 从缓存模型列表中选取 creditConsumption 最低的模型，
+ * 作为 Anthropic Haiku 的等价物用于轻量级辅助请求。
+ * 缓存为空时返回 undefined（调用方应 fallback 到主模型）。
+ */
+export function getCheapestCoStrictModel(): string | undefined {
+  const cached = getCachedCoStrictModels()
+  if (cached.length === 0) return undefined
+  const sorted = [...cached].sort(
+    (a, b) => (a.creditConsumption ?? Infinity) - (b.creditConsumption ?? Infinity),
+  )
+  return sorted[0]?.id
+}

src/memdir/findRelevantMemories.ts

@@ -1,7 +1,8 @@
 import { feature } from 'bun:bundle'
 import { logForDebugging } from '../utils/debug.js'
 import { errorMessage } from '../utils/errors.js'
-import { getDefaultSonnetModel } from '../utils/model/model.js'
+import { getDefaultSonnetModel, getSmallFastModel } from '../utils/model/model.js'
+import { getAPIProvider } from '../utils/model/providers.js'
 import { sideQuery } from '../utils/sideQuery.js'
 import type { LangfuseSpan } from '../services/langfuse/index.js'
 import { jsonParse } from '../utils/slowOperations.js'
@@ -99,8 +100,10 @@ async function selectRelevantMemories(
       : ''
 
   try {
+    const selectionModel =
+      getAPIProvider() === 'costrict' ? getSmallFastModel() : getDefaultSonnetModel()
     const result = await sideQuery({
-      model: getDefaultSonnetModel(),
+      model: selectionModel,
       system: SELECT_MEMORIES_SYSTEM_PROMPT,
       skipSystemPromptPrefix: true,
       messages: [

src/services/SessionMemory/sessionMemory.ts

@@ -25,6 +25,8 @@ import {
   runForkedAgent,
 } from '../../utils/forkedAgent.js'
 import { getFsImplementation } from '../../utils/fsOperations.js'
+import { getSmallFastModel } from '../../utils/model/model.js'
+import { getAPIProvider } from '../../utils/model/providers.js'
 import {
   type REPLHookContext,
   registerPostSamplingHook,
@@ -322,13 +324,19 @@ const extractSessionMemory = sequential(async function (
   // Run session memory extraction using runForkedAgent for prompt caching
   // runForkedAgent creates an isolated context to prevent mutation of parent state
   // Pass setupContext.readFileState so the forked agent can edit the memory file
+  const cacheSafeParams = createCacheSafeParams(context)
+  const smallModel = getSmallFastModel()
+  const modelOverride =
+    getAPIProvider() === 'costrict'
+      ? { options: { ...cacheSafeParams.toolUseContext.options, mainLoopModel: smallModel, model: smallModel } }
+      : {}
   await runForkedAgent({
     promptMessages: [createUserMessage({ content: userPrompt })],
-    cacheSafeParams: createCacheSafeParams(context),
+    cacheSafeParams,
     canUseTool: createMemoryFileCanUseTool(memoryPath),
     querySource: 'session_memory',
     forkLabel: 'session_memory',
-    overrides: { readFileState: setupContext.readFileState },
+    overrides: { readFileState: setupContext.readFileState, ...modelOverride },
   })
 
   // Log extraction event for tracking frequency
@@ -424,6 +432,11 @@ export async function manuallyExtractSessionMemory(
     const systemPrompt = asSystemPrompt(rawSystemPrompt)
 
     // Run session memory extraction using runForkedAgent
+    const smallModelManual = getSmallFastModel()
+    const modelOverrideManual =
+      getAPIProvider() === 'costrict'
+        ? { options: { ...setupContext.options, mainLoopModel: smallModelManual, model: smallModelManual } }
+        : {}
     await runForkedAgent({
       promptMessages: [createUserMessage({ content: userPrompt })],
       cacheSafeParams: {
@@ -436,7 +449,7 @@ export async function manuallyExtractSessionMemory(
       canUseTool: createMemoryFileCanUseTool(memoryPath),
       querySource: 'session_memory',
       forkLabel: 'session_memory_manual',
-      overrides: { readFileState: setupContext.readFileState },
+      overrides: { readFileState: setupContext.readFileState, ...modelOverrideManual },
     })
 
     // Log manual extraction event

src/services/extractMemories/extractMemories.ts

@@ -55,6 +55,8 @@ import {
 import { getFeatureValue_CACHED_MAY_BE_STALE } from '../analytics/growthbook.js'
 import { logEvent } from '../analytics/index.js'
 import { sanitizeToolNameForAnalytics } from '../analytics/metadata.js'
+import { getSmallFastModel } from '../../utils/model/model.js'
+import { getAPIProvider } from '../../utils/model/providers.js'
 import {
   buildExtractAutoOnlyPrompt,
   buildExtractCombinedPrompt,
@@ -367,6 +369,11 @@ export function initExtractMemories(): void {
 
     const canUseTool = createAutoMemCanUseTool(memoryDir)
     const cacheSafeParams = createCacheSafeParams(context)
+    const smallModel = getSmallFastModel()
+    const modelOverride =
+      getAPIProvider() === 'costrict'
+        ? { options: { ...cacheSafeParams.toolUseContext.options, mainLoopModel: smallModel, model: smallModel } }
+        : {}
 
     // Only run extraction every N eligible turns (tengu_bramble_lintel, default 1).
     // Trailing extractions (from stashed contexts) skip this check since they
@@ -415,12 +422,9 @@ export function initExtractMemories(): void {
         canUseTool,
         querySource: 'extract_memories',
         forkLabel: 'extract_memories',
-        // The extractMemories subagent does not need to record to transcript.
-        // Doing so can create race conditions with the main thread.
         skipTranscript: true,
-        // Well-behaved extractions complete in 2-4 turns (read → write).
-        // A hard cap prevents verification rabbit-holes from burning turns.
         maxTurns: 5,
+        overrides: modelOverride,
       })
 
       // Advance the cursor only after a successful run. If the agent errors

src/utils/model/model.ts

@@ -36,8 +36,8 @@ export type ModelSetting = ModelName | ModelAlias | null
 
 export function getSmallFastModel(): ModelName {
   const provider = getAPIProvider()
-  // CoStrict has no haiku alias — use the main loop model to stay on the same provider
+  // CoStrict has no haiku alias — resolve to cheapest available model by creditConsumption
-  if (provider === 'costrict') return getMainLoopModel()
+  if (provider === 'costrict') return getDefaultHaikuModel()
   // Provider-specific small fast model
   if (provider === 'openai' && process.env.OPENAI_SMALL_FAST_MODEL) {
     return process.env.OPENAI_SMALL_FAST_MODEL

src/utils/sideQuery.ts

@@ -38,6 +38,7 @@ import { getOpenAIClient } from '../services/api/openai/client.js'
 import { resolveOpenAIModel } from '@ant/model-provider'
 import { createCoStrictFetch } from '../costrict/provider/fetch.js'
 import { resolveCoStrictModel } from '../costrict/provider/modelMapping.js'
+import { getCachedCoStrictModels } from '../costrict/provider/models.js'
 import { getCoStrictBaseURL } from '../costrict/provider/auth.js'
 import { loadCoStrictCredentials } from '../costrict/provider/credentials.js'
 import { getProxyFetchOptions } from './proxy.js'
@@ -157,7 +158,14 @@ export async function sideQuery(opts: SideQueryOptions): Promise<BetaMessage> {
       fetch: createCoStrictFetch() as unknown as typeof fetch,
       defaultHeaders: { 'User-Agent': getUserAgent() },
     })
-    return sideQueryOpenAICompat(opts, client, resolveCoStrictModel(getMainLoopModel()), 'CoStrict')
+    const costrictModel = resolveCoStrictModel(opts.model)
+    const cachedModels = getCachedCoStrictModels()
+    const isKnown =
+      cachedModels.length === 0 || cachedModels.some(m => m.id === costrictModel)
+    const effectiveModel = isKnown
+      ? costrictModel
+      : resolveCoStrictModel(getMainLoopModel())
+    return sideQueryOpenAICompat(opts, client, effectiveModel, 'CoStrict')
   }
 
   const {