feat: 增强 auto mode 的易用性 (#312)

* feat: poor 模式降级 yolo 审阅模型

* feat: 为多模块添加 Langfuse tracing 支持

在 web search、agent creation、away summary、token estimation、
skill improvement 等模块中集成 Langfuse trace，并透传至
compact/apiQueryHook/execPromptHook 等调用链。

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: 让 auto mode 记录回主 trace

* fix: reopen auto mode prompt when classifier is unavailable

* fix: 修复 auto mode 情况下, llm 报错导致弹窗也不打开的问题

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: claude-code-best

packages/builtin-tools/src/tools/WebSearchTool/adapters/apiAdapter.ts

@@ -9,6 +9,9 @@ import type {
 } from '@anthropic-ai/sdk/resources/beta/messages/messages.mjs'
 import { getFeatureValue_CACHED_MAY_BE_STALE } from 'src/services/analytics/growthbook.js'
 import { queryModelWithStreaming } from 'src/services/api/claude.js'
+import { createTrace, endTrace, isLangfuseEnabled } from 'src/services/langfuse/index.js'
+import { getSessionId } from 'src/bootstrap/state.js'
+import { getAPIProvider } from 'src/utils/model/providers.js'
 import { createUserMessage } from 'src/utils/messages.js'
 import { getMainLoopModel, getSmallFastModel } from 'src/utils/model/model.js'
 import { jsonParse } from 'src/utils/slowOperations.js'
@@ -38,6 +41,15 @@ export class ApiSearchAdapter implements WebSearchAdapter {
     const toolSchema = makeToolSchema({ allowedDomains, blockedDomains })
 
     const useHaiku = getFeatureValue_CACHED_MAY_BE_STALE('tengu_plum_vx3', false)
+    const model = useHaiku ? getSmallFastModel() : getMainLoopModel()
+    const langfuseTrace = isLangfuseEnabled()
+      ? createTrace({
+          sessionId: getSessionId(),
+          model,
+          provider: getAPIProvider(),
+          name: 'web-search-tool',
+        })
+      : null
 
     const queryStream = queryModelWithStreaming({
       messages: [userMessage],
@@ -58,7 +70,7 @@ export class ApiSearchAdapter implements WebSearchAdapter {
           alwaysAskRules: {},
           isBypassPermissionsModeAvailable: false,
         }),
-        model: useHaiku ? getSmallFastModel() : getMainLoopModel(),
+        model,
         toolChoice: useHaiku ? { type: 'tool' as const, name: 'web_search' } : undefined,
         isNonInteractiveSession: false,
         hasAppendSystemPrompt: false,
@@ -68,6 +80,7 @@ export class ApiSearchAdapter implements WebSearchAdapter {
         mcpTools: [],
         agentId: undefined,
         effortValue: undefined,
+        langfuseTrace,
       },
     })
 
@@ -148,6 +161,8 @@ export class ApiSearchAdapter implements WebSearchAdapter {
       }
     }
 
+    endTrace(langfuseTrace)
+
     // Extract SearchResult[] from content blocks
     return extractSearchResults(allContentBlocks)
   }

src/Tool.ts

@@ -277,6 +277,8 @@ export type ToolUseContext = {
   criticalSystemReminder_EXPERIMENTAL?: string
   /** Langfuse root trace span for this query turn. Passed down to tool execution for observability. */
   langfuseTrace?: LangfuseSpan | null
+  /** Langfuse root trace span for the outer/main agent trace. Used when subagents need to nest observations under the parent agent trace. */
+  langfuseRootTrace?: LangfuseSpan | null
   /** Langfuse batch span wrapping a concurrent tool group. When set, tool observations are nested under it. */
   langfuseBatchSpan?: LangfuseSpan | null
   /** When true, preserve toolUseResult on messages even for subagents.

src/cli/handlers/autoMode.ts

@@ -6,6 +6,7 @@
 import { errorMessage } from '../../utils/errors.js'
 import {
   getMainLoopModel,
+  getSmallFastModel,
   parseUserSpecifiedModel,
 } from '../../utils/model/model.js'
 import {
@@ -14,6 +15,7 @@ import {
   getDefaultExternalAutoModeRules,
 } from '../../utils/permissions/yoloClassifier.js'
 import { getAutoModeConfig } from '../../utils/settings/settings.js'
+import { isPoorModeActive } from '../../commands/poor/poorMode.js'
 import { sideQuery } from '../../utils/sideQuery.js'
 import { jsonStringify } from '../../utils/slowOperations.js'
 
@@ -90,7 +92,9 @@ export async function autoModeCritiqueHandler(options: {
 
   const model = options.model
     ? parseUserSpecifiedModel(options.model)
-    : getMainLoopModel()
+    : isPoorModeActive()
+      ? getSmallFastModel()
+      : getMainLoopModel()
 
   const defaults = getDefaultExternalAutoModeRules()
   const classifierPrompt = buildDefaultExternalSystemPrompt()

src/components/agents/generateAgent.ts

@@ -14,6 +14,9 @@ import {
   type AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS,
   logEvent,
 } from '../../services/analytics/index.js'
+import { createTrace, endTrace, isLangfuseEnabled } from '../../services/langfuse/index.js'
+import { getSessionId } from '../../bootstrap/state.js'
+import { getAPIProvider } from '../../utils/model/providers.js'
 import { jsonParse } from '../../utils/slowOperations.js'
 import { asSystemPrompt } from '../../utils/systemPromptType.js'
 
@@ -146,6 +149,15 @@ export async function generateAgent(
     ? AGENT_CREATION_SYSTEM_PROMPT + AGENT_MEMORY_INSTRUCTIONS
     : AGENT_CREATION_SYSTEM_PROMPT
 
+  const langfuseTrace = isLangfuseEnabled()
+    ? createTrace({
+        sessionId: getSessionId(),
+        model,
+        provider: getAPIProvider(),
+        name: 'agent-creation',
+      })
+    : null
+
   const response = await queryModelWithoutStreaming({
     messages: normalizeMessagesForAPI(messagesWithContext),
     systemPrompt: asSystemPrompt([systemPrompt]),
@@ -161,9 +173,12 @@ export async function generateAgent(
       hasAppendSystemPrompt: false,
       querySource: 'agent_creation',
       mcpTools: [],
+      langfuseTrace,
     },
   })
 
+  endTrace(langfuseTrace)
+
   const textBlocks = (Array.isArray(response.message.content) ? response.message.content : []).filter(
     (block): block is ContentBlock & { type: 'text' } => block.type === 'text',
   )

src/query.ts

@@ -235,6 +235,9 @@ export async function* query(
   // When called as a sub-agent, langfuseTrace is already set by runAgent()
   // — reuse it instead of creating an independent trace.
   const ownsTrace = !params.toolUseContext.langfuseTrace
+  logForDebugging(
+    `[query] ownsTrace=${ownsTrace} incoming langfuseTrace=${params.toolUseContext.langfuseTrace ? 'present' : 'null/undefined'} isLangfuseEnabled=${isLangfuseEnabled()}`,
+  )
   const langfuseTrace = params.toolUseContext.langfuseTrace
     ?? (isLangfuseEnabled()
       ? createTrace({

src/services/awaySummary.ts

@@ -10,6 +10,9 @@ import { getSmallFastModel } from '../utils/model/model.js'
 import { asSystemPrompt } from '../utils/systemPromptType.js'
 import { getResolvedLanguage } from '../utils/language.js'
 import { queryModelWithoutStreaming } from './api/claude.js'
+import { createTrace, endTrace, isLangfuseEnabled } from './langfuse/index.js'
+import { getSessionId } from '../bootstrap/state.js'
+import { getAPIProvider } from '../utils/model/providers.js'
 import { getSessionMemoryContent } from './SessionMemory/sessionMemoryUtils.js'
 
 // Recap only needs recent context — truncate to avoid "prompt too long" on
@@ -42,6 +45,16 @@ export async function generateAwaySummary(
     return null
   }
 
+  const model = getSmallFastModel()
+  const langfuseTrace = isLangfuseEnabled()
+    ? createTrace({
+        sessionId: getSessionId(),
+        model,
+        provider: getAPIProvider(),
+        name: 'away-summary',
+      })
+    : null
+
   try {
     const memory = await getSessionMemoryContent()
     const recent = messages.slice(-RECENT_MESSAGE_WINDOW)
@@ -54,29 +67,33 @@ export async function generateAwaySummary(
       signal,
       options: {
         getToolPermissionContext: async () => getEmptyToolPermissionContext(),
-        model: getSmallFastModel(),
+        model,
         toolChoice: undefined,
         isNonInteractiveSession: false,
         hasAppendSystemPrompt: false,
         agents: [],
         querySource: 'away_summary',
         mcpTools: [],
         skipCacheWrite: true,
+        langfuseTrace,
       },
     })
 
     if (response.isApiErrorMessage) {
       logForDebugging(
         `[awaySummary] API error: ${getAssistantMessageText(response)}`,
       )
+      endTrace(langfuseTrace, undefined, 'error')
       return null
     }
+    endTrace(langfuseTrace)
     return getAssistantMessageText(response)
   } catch (err) {
     if (err instanceof APIUserAbortError || signal.aborted) {
       return null
     }
     logForDebugging(`[awaySummary] generation failed: ${err}`)
+    endTrace(langfuseTrace, undefined, 'error')
     return null
   }
 }

src/services/compact/compact.ts

@@ -1326,6 +1326,7 @@ async function streamCompactSummary({
           agents: context.options.agentDefinitions.activeAgents,
           mcpTools: [],
           effortValue: appState.effortValue,
+          langfuseTrace: context.langfuseTrace,
         },
       })
       const streamIter = streamingGen[Symbol.asyncIterator]()

src/services/tokenEstimation.ts

@@ -25,6 +25,8 @@ import { jsonStringify } from '../utils/slowOperations.js'
 import { isToolReferenceBlock } from '../utils/toolSearch.js'
 import { getAPIMetadata, getExtraBodyParams } from './api/claude.js'
 import { getAnthropicClient } from './api/client.js'
+import { createTrace, endTrace, isLangfuseEnabled, recordLLMObservation } from './langfuse/index.js'
+import { getSessionId } from '../bootstrap/state.js'
 import { withTokenCountVCR } from './vcr.js'
 
 // Minimal values for token counting with thinking enabled
@@ -309,6 +311,15 @@ export async function countTokensViaHaikuFallback(
       : betas
 
   // biome-ignore lint/plugin: token counting needs specialized parameters (thinking, betas) that sideQuery doesn't support
+  const apiStart = Date.now()
+  const langfuseTrace = isLangfuseEnabled()
+    ? createTrace({
+        sessionId: getSessionId(),
+        model: normalizeModelStringForAPI(model),
+        provider: getAPIProvider(),
+        name: 'token-estimation',
+      })
+    : null
   const response = await anthropic.beta.messages.create({
     model: normalizeModelStringForAPI(model),
     max_tokens: containsThinking ? TOKEN_COUNT_MAX_TOKENS : 1,
@@ -331,6 +342,22 @@ export async function countTokensViaHaikuFallback(
   const cacheCreationTokens = usage.cache_creation_input_tokens || 0
   const cacheReadTokens = usage.cache_read_input_tokens || 0
 
+  recordLLMObservation(langfuseTrace, {
+    model: normalizeModelStringForAPI(model),
+    provider: getAPIProvider(),
+    input: messagesToSend,
+    output: response.content,
+    usage: {
+      input_tokens: inputTokens,
+      output_tokens: usage.output_tokens,
+      cache_creation_input_tokens: cacheCreationTokens || undefined,
+      cache_read_input_tokens: cacheReadTokens || undefined,
+    },
+    startTime: new Date(apiStart),
+    endTime: new Date(),
+  })
+  endTrace(langfuseTrace)
+
   return inputTokens + cacheCreationTokens + cacheReadTokens
 }
 

src/utils/__tests__/messages.test.ts

@@ -457,9 +457,14 @@ describe("buildClassifierUnavailableMessage", () => {
     expect(msg).toContain("classifier-v1");
     expect(msg).toContain("unavailable");
   });
+
+  test("tells the model to wait and retry later", () => {
+    const msg = buildClassifierUnavailableMessage("Bash", "classifier-v1");
+    expect(msg).toContain("Wait briefly and then try this action again.");
+    expect(msg).toContain("come back to it later");
+  });
 });
 
-// ─── normalizeMessages ──────────────────────────────────────────────────
 
 describe("normalizeMessages", () => {
   test("splits multi-block assistant message into individual messages", () => {

src/utils/forkedAgent.ts

@@ -374,6 +374,10 @@ export function createSubagentContext(
         }
 
   return {
+    // Preserve the parent Langfuse trace separately so nested side queries
+    // like auto_mode can attach to the main agent trace instead of the
+    // subagent's own trace.
+    langfuseRootTrace: parentContext.langfuseTrace,
     // Mutable state - cloned by default to maintain isolation
     // Clone overrides.readFileState if provided, otherwise clone from parent
     readFileState: cloneFileStateCache(