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>

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'
 
@@ -90,6 +92,8 @@ export async function autoModeCritiqueHandler(options: {
 
   const model = options.model
     ? parseUserSpecifiedModel(options.model)
+    : isPoorModeActive()
+      ? getSmallFastModel()
       : getMainLoopModel()
 
   const defaults = getDefaultExternalAutoModeRules()

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)
@@ -54,7 +67,7 @@ export async function generateAwaySummary(
       signal,
       options: {
         getToolPermissionContext: async () => getEmptyToolPermissionContext(),
-        model: getSmallFastModel(),
+        model,
         toolChoice: undefined,
         isNonInteractiveSession: false,
         hasAppendSystemPrompt: false,
@@ -62,6 +75,7 @@ export async function generateAwaySummary(
         querySource: 'away_summary',
         mcpTools: [],
         skipCacheWrite: true,
+        langfuseTrace,
       },
     })
 
@@ -69,14 +83,17 @@ export async function generateAwaySummary(
       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(

src/utils/hooks/apiQueryHookHelper.ts

@@ -104,6 +104,7 @@ export function createApiQueryHook<TResult>(
           querySource: config.name,
           mcpTools: [],
           agentId: context.toolUseContext.agentId,
+          langfuseTrace: context.toolUseContext.langfuseTrace,
         },
       })
 

src/utils/hooks/execPromptHook.ts

@@ -84,6 +84,7 @@ Your response must be a JSON object matching one of the following schemas:
           querySource: 'hook_prompt',
           mcpTools: [],
           agentId: toolUseContext.agentId,
+          langfuseTrace: toolUseContext.langfuseTrace,
           outputFormat: {
             type: 'json_schema',
             schema: {

src/utils/hooks/skillImprovement.ts

@@ -7,6 +7,9 @@ import {
   logEvent,
 } from '../../services/analytics/index.js'
 import { queryModelWithoutStreaming } from '../../services/api/claude.js'
+import { createTrace, endTrace, isLangfuseEnabled } from '../../services/langfuse/index.js'
+import { getSessionId } from '../../bootstrap/state.js'
+import { getAPIProvider } from '../model/providers.js'
 import { getEmptyToolPermissionContext } from '../../Tool.js'
 import type { Message } from '../../types/message.js'
 import { createAbortController } from '../abortController.js'
@@ -209,6 +212,16 @@ export async function applySkillImprovement(
 
   const updateList = updates.map(u => `- ${u.section}: ${u.change}`).join('\n')
 
+  const model = getSmallFastModel()
+  const langfuseTrace = isLangfuseEnabled()
+    ? createTrace({
+        sessionId: getSessionId(),
+        model,
+        provider: getAPIProvider(),
+        name: 'skill-improvement-apply',
+      })
+    : null
+
   const response = await queryModelWithoutStreaming({
     messages: [
       createUserMessage({
@@ -238,7 +251,7 @@ Rules:
     signal: createAbortController().signal,
     options: {
       getToolPermissionContext: async () => getEmptyToolPermissionContext(),
-      model: getSmallFastModel(),
+      model,
       toolChoice: undefined,
       isNonInteractiveSession: false,
       hasAppendSystemPrompt: false,
@@ -246,9 +259,12 @@ Rules:
       agents: [],
       querySource: 'skill_improvement_apply',
       mcpTools: [],
+      langfuseTrace,
     },
   })
 
+  endTrace(langfuseTrace)
+
   const responseText = extractTextContent(Array.isArray(response.message.content) ? response.message.content : []).trim()
 
   const updatedContent = extractTag(responseText, 'updated_file')

src/utils/permissions/__tests__/permissions.test.ts

@@ -1,153 +1,136 @@
-import { mock, describe, expect, test } from "bun:test";
+import { mock, describe, expect, test } from 'bun:test'
+import { createFileStateCacheWithSizeLimit } from '../../../utils/fileStateCache.js'
+import { createSubagentContext } from '../../../utils/forkedAgent.js'
+import { getEmptyToolPermissionContext } from '../../../Tool.js'
 
-// Mock log.ts to cut the heavy dependency chain
+mock.module('src/utils/log.ts', () => ({
-mock.module("src/utils/log.ts", () => ({
   logError: () => {},
   logToFile: () => {},
-  getLogDisplayTitle: () => "",
+  getLogDisplayTitle: () => '',
   logEvent: () => {},
   logMCPError: () => {},
   logMCPDebug: () => {},
-  dateToFilename: (d: Date) => d.toISOString().replace(/[:.]/g, "-"),
+  dateToFilename: (d: Date) => d.toISOString().replace(/[:.]/g, '-'),
-  getLogFilePath: () => "/tmp/mock-log",
+  getLogFilePath: () => '/tmp/mock-log',
   attachErrorLogSink: () => {},
   getInMemoryErrors: () => [],
   loadErrorLogs: async () => [],
   getErrorLogByIndex: async () => null,
   captureAPIRequest: () => {},
   _resetErrorLogForTesting: () => {},
-}));
+}))
 
 const {
   getDenyRuleForTool,
   getAskRuleForTool,
   getDenyRuleForAgent,
   filterDeniedAgents,
-} = await import("../permissions");
+} = await import('../permissions')
 
-import { getEmptyToolPermissionContext } from "../../../Tool";
+function makeContext(opts: { denyRules?: string[]; askRules?: string[] }) {
-
+  const ctx = getEmptyToolPermissionContext()
-// ─── Helper ─────────────────────────────────────────────────────────────
+  const deny: Record<string, string[]> = {}
-
+  const ask: Record<string, string[]> = {}
-function makeContext(opts: {
+  if (opts.denyRules?.length) deny.localSettings = opts.denyRules
-  denyRules?: string[];
+  if (opts.askRules?.length) ask.localSettings = opts.askRules
-  askRules?: string[];
+  return { ...ctx, alwaysDenyRules: deny, alwaysAskRules: ask } as any
-}) {
-  const ctx = getEmptyToolPermissionContext();
-  const deny: Record<string, string[]> = {};
-  const ask: Record<string, string[]> = {};
-
-  // alwaysDenyRules stores raw rule strings — getDenyRules() calls
-  // permissionRuleValueFromString internally
-  if (opts.denyRules?.length) {
-    deny["localSettings"] = opts.denyRules;
-  }
-  if (opts.askRules?.length) {
-    ask["localSettings"] = opts.askRules;
-  }
-
-  return {
-    ...ctx,
-    alwaysDenyRules: deny,
-    alwaysAskRules: ask,
-  } as any;
 }
 
 function makeTool(name: string, mcpInfo?: { serverName: string; toolName: string }) {
-  return { name, mcpInfo };
+  return { name, mcpInfo }
 }
 
-// ─── getDenyRuleForTool ─────────────────────────────────────────────────
+describe('getDenyRuleForTool', () => {
+  test('returns null when no deny rules', () => {
+    const ctx = makeContext({})
+    expect(getDenyRuleForTool(ctx, makeTool('Bash'))).toBeNull()
+  })
+  test('returns matching deny rule for tool', () => {
+    const ctx = makeContext({ denyRules: ['Bash'] })
+    const result = getDenyRuleForTool(ctx, makeTool('Bash'))
+    expect(result).not.toBeNull()
+    expect(result!.ruleValue.toolName).toBe('Bash')
+  })
+  test('returns null for non-matching tool', () => {
+    const ctx = makeContext({ denyRules: ['Bash'] })
+    expect(getDenyRuleForTool(ctx, makeTool('Read'))).toBeNull()
+  })
+  test('rule with content does not match whole-tool deny', () => {
+    const ctx = makeContext({ denyRules: ['Bash(rm -rf)'] })
+    const result = getDenyRuleForTool(ctx, makeTool('Bash'))
+    expect(result).toBeNull()
+  })
+})
 
-describe("getDenyRuleForTool", () => {
+describe('getAskRuleForTool', () => {
-  test("returns null when no deny rules", () => {
+  test('returns null when no ask rules', () => {
-    const ctx = makeContext({});
+    const ctx = makeContext({})
-    expect(getDenyRuleForTool(ctx, makeTool("Bash"))).toBeNull();
+    expect(getAskRuleForTool(ctx, makeTool('Bash'))).toBeNull()
-  });
+  })
+  test('returns matching ask rule', () => {
+    const ctx = makeContext({ askRules: ['Write'] })
+    const result = getAskRuleForTool(ctx, makeTool('Write'))
+    expect(result).not.toBeNull()
+  })
+  test('returns null for non-matching tool', () => {
+    const ctx = makeContext({ askRules: ['Write'] })
+    expect(getAskRuleForTool(ctx, makeTool('Bash'))).toBeNull()
+  })
+})
 
-  test("returns matching deny rule for tool", () => {
+describe('getDenyRuleForAgent', () => {
-    const ctx = makeContext({ denyRules: ["Bash"] });
+  test('returns null when no deny rules', () => {
-    const result = getDenyRuleForTool(ctx, makeTool("Bash"));
+    const ctx = makeContext({})
-    expect(result).not.toBeNull();
+    expect(getDenyRuleForAgent(ctx, 'Agent', 'Explore')).toBeNull()
-    expect(result!.ruleValue.toolName).toBe("Bash");
+  })
-  });
+  test('returns matching deny rule for agent type', () => {
+    const ctx = makeContext({ denyRules: ['Agent(Explore)'] })
+    const result = getDenyRuleForAgent(ctx, 'Agent', 'Explore')
+    expect(result).not.toBeNull()
+  })
+  test('returns null for non-matching agent type', () => {
+    const ctx = makeContext({ denyRules: ['Agent(Explore)'] })
+    expect(getDenyRuleForAgent(ctx, 'Agent', 'Research')).toBeNull()
+  })
+})
 
-  test("returns null for non-matching tool", () => {
+describe('Langfuse trace propagation', () => {
-    const ctx = makeContext({ denyRules: ["Bash"] });
+  test('subagent context preserves parent trace for nested side queries', () => {
-    expect(getDenyRuleForTool(ctx, makeTool("Read"))).toBeNull();
+    const parentTrace = { id: 'parent-trace' } as never
-  });
+    const parentContext = {
+      ...getEmptyToolPermissionContext(),
+      messages: [],
+      abortController: new AbortController(),
+      readFileState: createFileStateCacheWithSizeLimit(1),
+      getAppState: () => ({ toolPermissionContext: getEmptyToolPermissionContext() }),
+      setAppState: () => {},
+      updateFileHistoryState: () => {},
+      updateAttributionState: () => {},
+      setInProgressToolUseIDs: () => {},
+      setResponseLength: () => {},
+      langfuseTrace: parentTrace,
+    } as never
+    const subagentContext = createSubagentContext(parentContext)
+    expect(subagentContext.langfuseRootTrace).toBe(parentTrace)
+  })
+})
 
-  test("rule with content does not match whole-tool deny", () => {
+describe('filterDeniedAgents', () => {
-    // getDenyRuleForTool uses toolMatchesRule which requires ruleContent === undefined
+  test('returns all agents when no deny rules', () => {
-    // Rules like "Bash(rm -rf)" only match specific invocations, not the entire tool
+    const ctx = makeContext({})
-    const ctx = makeContext({ denyRules: ["Bash(rm -rf)"] });
+    const agents = [{ agentType: 'Explore' }, { agentType: 'Research' }]
-    const result = getDenyRuleForTool(ctx, makeTool("Bash"));
+    expect(filterDeniedAgents(agents, ctx, 'Agent')).toEqual(agents)
-    expect(result).toBeNull();
+  })
-  });
+  test('filters out denied agent type', () => {
-});
+    const ctx = makeContext({ denyRules: ['Agent(Explore)'] })
-
+    const agents = [{ agentType: 'Explore' }, { agentType: 'Research' }]
-// ─── getAskRuleForTool ──────────────────────────────────────────────────
+    const result = filterDeniedAgents(agents, ctx, 'Agent')
-
+    expect(result).toHaveLength(1)
-describe("getAskRuleForTool", () => {
+    expect(result[0]!.agentType).toBe('Research')
-  test("returns null when no ask rules", () => {
+  })
-    const ctx = makeContext({});
+  test('returns empty array when all agents denied', () => {
-    expect(getAskRuleForTool(ctx, makeTool("Bash"))).toBeNull();
+    const ctx = makeContext({ denyRules: ['Agent(Explore)', 'Agent(Research)'] })
-  });
+    const agents = [{ agentType: 'Explore' }, { agentType: 'Research' }]
-
+    expect(filterDeniedAgents(agents, ctx, 'Agent')).toEqual([])
-  test("returns matching ask rule", () => {
+  })
-    const ctx = makeContext({ askRules: ["Write"] });
+})
-    const result = getAskRuleForTool(ctx, makeTool("Write"));
-    expect(result).not.toBeNull();
-  });
-
-  test("returns null for non-matching tool", () => {
-    const ctx = makeContext({ askRules: ["Write"] });
-    expect(getAskRuleForTool(ctx, makeTool("Bash"))).toBeNull();
-  });
-});
-
-// ─── getDenyRuleForAgent ────────────────────────────────────────────────
-
-describe("getDenyRuleForAgent", () => {
-  test("returns null when no deny rules", () => {
-    const ctx = makeContext({});
-    expect(getDenyRuleForAgent(ctx, "Agent", "Explore")).toBeNull();
-  });
-
-  test("returns matching deny rule for agent type", () => {
-    const ctx = makeContext({ denyRules: ["Agent(Explore)"] });
-    const result = getDenyRuleForAgent(ctx, "Agent", "Explore");
-    expect(result).not.toBeNull();
-  });
-
-  test("returns null for non-matching agent type", () => {
-    const ctx = makeContext({ denyRules: ["Agent(Explore)"] });
-    expect(getDenyRuleForAgent(ctx, "Agent", "Research")).toBeNull();
-  });
-});
-
-// ─── filterDeniedAgents ─────────────────────────────────────────────────
-
-describe("filterDeniedAgents", () => {
-  test("returns all agents when no deny rules", () => {
-    const ctx = makeContext({});
-    const agents = [{ agentType: "Explore" }, { agentType: "Research" }];
-    expect(filterDeniedAgents(agents, ctx, "Agent")).toEqual(agents);
-  });
-
-  test("filters out denied agent type", () => {
-    const ctx = makeContext({ denyRules: ["Agent(Explore)"] });
-    const agents = [{ agentType: "Explore" }, { agentType: "Research" }];
-    const result = filterDeniedAgents(agents, ctx, "Agent");
-    expect(result).toHaveLength(1);
-    expect(result[0]!.agentType).toBe("Research");
-  });
-
-  test("returns empty array when all agents denied", () => {
-    const ctx = makeContext({
-      denyRules: ["Agent(Explore)", "Agent(Research)"],
-    });
-    const agents = [{ agentType: "Explore" }, { agentType: "Research" }];
-    expect(filterDeniedAgents(agents, ctx, "Agent")).toEqual([]);
-  });
-});

src/utils/permissions/permissionExplainer.ts

@@ -7,7 +7,8 @@ import { logForDebugging } from '../debug.js'
 import { errorMessage } from '../errors.js'
 import { lazySchema } from '../lazySchema.js'
 import { logError } from '../log.js'
-import { getMainLoopModel } from '../model/model.js'
+import { getMainLoopModel, getSmallFastModel } from '../model/model.js'
+import { isPoorModeActive } from '../../commands/poor/poorMode.js'
 import { sideQuery } from '../sideQuery.js'
 import { jsonStringify } from '../slowOperations.js'
 
@@ -172,7 +173,7 @@ ${conversationContext ? `\nRecent conversation context:\n${conversationContext}`
 
 Explain this command in context.`
 
-    const model = getMainLoopModel()
+    const model = isPoorModeActive() ? getSmallFastModel() : getMainLoopModel()
 
     // Use sideQuery with forced tool choice for guaranteed structured output
     const response = await sideQuery({

src/utils/permissions/permissions.ts

@@ -690,13 +690,16 @@ export const hasPermissionsToUseTool: CanUseToolFn = async (
       setClassifierChecking(toolUseID)
       let classifierResult
       try {
+        logForDebugging(
+          `[auto-mode] classifyYoloAction called with langfuseTrace=${context.langfuseTrace ? `id=${(context.langfuseTrace as unknown as Record<string, unknown>).id ?? 'present'}` : 'null/undefined'}`,
+        )
         classifierResult = await classifyYoloAction(
           context.messages,
           action,
           context.options.tools,
           appState.toolPermissionContext,
           context.abortController.signal,
-          context.langfuseTrace,
+          context.langfuseRootTrace ?? context.langfuseTrace,
         )
       } finally {
         clearClassifierChecking(toolUseID)
@@ -851,6 +854,7 @@ export const hasPermissionsToUseTool: CanUseToolFn = async (
               CLASSIFIER_FAIL_CLOSED_REFRESH_MS,
             )
           ) {
+            if (appState.toolPermissionContext.shouldAvoidPermissionPrompts) {
               logForDebugging(
                 'Auto mode classifier unavailable, denying with retry guidance (fail closed)',
                 { level: 'warn' },
@@ -868,6 +872,23 @@ export const hasPermissionsToUseTool: CanUseToolFn = async (
                 ),
               }
             }
+            logForDebugging(
+              'Auto mode classifier unavailable, falling back to prompting with retry guidance (fail closed)',
+              { level: 'warn' },
+            )
+            return {
+              behavior: 'ask',
+              decisionReason: {
+                type: 'classifier',
+                classifier: 'auto-mode',
+                reason: 'Classifier unavailable',
+              },
+              message: buildClassifierUnavailableMessage(
+                tool.name,
+                classifierResult.model,
+              ),
+            }
+          }
           // Fail open: fall back to normal permission handling
           logForDebugging(
             'Auto mode classifier unavailable, falling back to normal permission handling (fail open)',

src/utils/permissions/yoloClassifier.ts

@@ -28,7 +28,8 @@ import { errorMessage } from '../errors.js'
 import { lazySchema } from '../lazySchema.js'
 import { extractTextContent } from '../messages.js'
 import { resolveAntModel } from '../model/antModels.js'
-import { getMainLoopModel } from '../model/model.js'
+import { getDefaultSonnetModel, getMainLoopModel } from '../model/model.js'
+import { isPoorModeActive } from '../../commands/poor/poorMode.js'
 import { getAutoModeConfig } from '../settings/settings.js'
 import { sideQuery } from '../sideQuery.js'
 import type { LangfuseSpan } from '../../services/langfuse/index.js'
@@ -1350,6 +1351,10 @@ function getClassifierModel(): string {
   if (config?.model) {
     return config.model
   }
+  // Poor mode: downgrade classifier to Sonnet to reduce cost
+  if (isPoorModeActive()) {
+    return getDefaultSonnetModel()
+  }
   return getMainLoopModel()
 }
 

src/utils/sideQuery.ts

@@ -19,6 +19,7 @@ import { createTrace, createChildSpan, endTrace, recordLLMObservation } from '..
 import type { LangfuseSpan } from '../services/langfuse/index.js'
 import { convertMessagesToLangfuse, convertOutputToLangfuse, convertToolsToLangfuse } from '../services/langfuse/convert.js'
 import { getModelBetas, modelSupportsStructuredOutputs } from './betas.js'
+import { logForDebugging } from './debug.js'
 import { errorMessage } from './errors.js'
 import { computeFingerprint } from './fingerprint.js'
 import { getAPIProvider } from './model/providers.js'
@@ -194,14 +195,28 @@ export async function sideQuery(opts: SideQueryOptions): Promise<BetaMessage> {
 
   // When parentSpan is provided, create a child span nested under the
   // main agent trace; otherwise create a standalone root trace.
-  const langfuseTrace = opts.parentSpan
+  const _ps = opts.parentSpan
-    ? createChildSpan(opts.parentSpan, {
+  // eslint-disable-next-line no-constant-condition
+  if (opts.querySource === 'auto_mode') {
+    logForDebugging(
+      `[sideQuery] auto_mode parentSpan=${_ps ? `id=${(_ps as unknown as Record<string, unknown>).id ?? 'present'}` : 'null/undefined'} querySource=${opts.querySource}`,
+    )
+  }
+  // When parentSpan is provided, create a child span nested under the
+  // main agent trace. For auto_mode queries, we must always nest under
+  // a parent span — never create a standalone root trace (agent type),
+  // as auto_mode observations should appear as spans within the parent.
+  // For other query sources without a parent, create a standalone trace.
+  const langfuseTrace = _ps
+    ? createChildSpan(_ps, {
         name: traceName,
         sessionId: getSessionId(),
         model: normalizedModel,
         provider,
         querySource: opts.querySource,
       })
+    : opts.querySource === 'auto_mode'
+      ? null
       : createTrace({
           sessionId: getSessionId(),
           model: normalizedModel,