diff --git a/packages/node-runtime/src/ai/agent/__tests__/planner.test.ts b/packages/node-runtime/src/ai/agent/__tests__/planner.test.ts
new file mode 100644
index 00000000..d04eb55f
--- /dev/null
+++ b/packages/node-runtime/src/ai/agent/__tests__/planner.test.ts
@@ -0,0 +1,67 @@
+import { describe, it } from 'node:test'
+import assert from 'node:assert/strict'
+import { createAnalysisPlanner } from '../planner'
+
+const baseInput = {
+  userMessage: '分析过去一年群里话题的变化趋势，按季度总结主要变化，并举出证据。',
+  chatType: 'group' as const,
+  locale: 'zh-CN',
+  availableTools: ['search_messages', 'get_time_stats', 'get_member_stats'],
+}
+
+describe('createAnalysisPlanner', () => {
+  it('parses valid JSON and filters suggested tools to the available set', async () => {
+    const planner = createAnalysisPlanner({
+      complete: async () =>
+        JSON.stringify({
+          title: '年度话题趋势分析',
+          intent: 'trend',
+          steps: [
+            {
+              goal: '按季度检索代表性话题',
+              suggestedTools: ['search_messages', 'missing_tool'],
+              evidenceNeeded: '每个季度的代表消息',
+            },
+          ],
+          successCriteria: ['覆盖全年', '引用证据'],
+        }),
+    })
+
+    const plan = await planner(baseInput)
+
+    assert.equal(plan?.route, 'planned_execution')
+    assert.equal(plan?.title, '年度话题趋势分析')
+    assert.deepEqual(plan?.steps[0]?.suggestedTools, ['search_messages'])
+  })
+
+  it('limits steps and success criteria to five items', async () => {
+    const planner = createAnalysisPlanner({
+      complete: async () =>
+        JSON.stringify({
+          title: '复杂分析',
+          intent: 'mixed',
+          steps: Array.from({ length: 7 }, (_, index) => ({
+            goal: `step ${index + 1}`,
+            suggestedTools: ['search_messages'],
+            evidenceNeeded: `evidence ${index + 1}`,
+          })),
+          successCriteria: ['a', 'b', 'c', 'd', 'e', 'f'],
+        }),
+    })
+
+    const plan = await planner(baseInput)
+
+    assert.equal(plan?.steps.length, 5)
+    assert.equal(plan?.successCriteria.length, 5)
+  })
+
+  it('returns null for invalid planner JSON', async () => {
+    const planner = createAnalysisPlanner({
+      complete: async () => '{"title": "broken", "intent": "unknown", "steps": []}',
+    })
+
+    const plan = await planner(baseInput)
+
+    assert.equal(plan, null)
+  })
+})
diff --git a/packages/node-runtime/src/ai/agent/index.ts b/packages/node-runtime/src/ai/agent/index.ts
index 0aa59456..7576d47e 100644
--- a/packages/node-runtime/src/ai/agent/index.ts
+++ b/packages/node-runtime/src/ai/agent/index.ts
@@ -3,3 +3,12 @@ export { runAgentCore } from './core'
 export { DEFAULT_MAX_TOOL_ROUNDS } from './constants'
 export { createLlmRouteDecider, decideRequestRoute } from './router'
 export type { LlmRouteDecider, RequestRoute, RouteDecision, RouteDecisionSource, RouterInput } from './routing-types'
+export { createAnalysisPlanner } from './planner'
+export type {
+  AnalysisPlanIntent,
+  AnalysisPlanner,
+  AnalysisPlanStep,
+  AnalysisPlanSummary,
+  PlannerInput,
+  PlanContentBlock,
+} from './planning-types'
diff --git a/packages/node-runtime/src/ai/agent/planner.ts b/packages/node-runtime/src/ai/agent/planner.ts
new file mode 100644
index 00000000..217775e8
--- /dev/null
+++ b/packages/node-runtime/src/ai/agent/planner.ts
@@ -0,0 +1,181 @@
+import {
+  completeSimple,
+  type Api as PiApi,
+  type Model as PiModel,
+  type TextContent as PiTextContent,
+} from '@earendil-works/pi-ai'
+import type { AnalysisPlanIntent, AnalysisPlanSummary, AnalysisPlanner, PlannerInput } from './planning-types'
+
+export type PlannerCompletionResult = string | { text: string }
+
+export interface CreateAnalysisPlannerOptions {
+  piModel?: PiModel<PiApi>
+  apiKey?: string
+  complete?: (prompt: string, signal?: AbortSignal) => Promise<PlannerCompletionResult> | PlannerCompletionResult
+  maxTokens?: number
+  temperature?: number
+}
+
+const PLAN_INTENTS: readonly AnalysisPlanIntent[] = [
+  'summary',
+  'trend',
+  'relationship',
+  'search',
+  'comparison',
+  'mixed',
+]
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return typeof value === 'object' && value !== null && !Array.isArray(value)
+}
+
+function isIntent(value: unknown): value is AnalysisPlanIntent {
+  return typeof value === 'string' && PLAN_INTENTS.includes(value as AnalysisPlanIntent)
+}
+
+function normalizeString(value: unknown): string | null {
+  if (typeof value !== 'string') return null
+  const text = value.trim()
+  return text.length > 0 ? text : null
+}
+
+function extractJsonObject(text: string): string {
+  const fenced = text.match(/```(?:json)?\s*([\s\S]*?)```/i)
+  if (fenced?.[1]) return fenced[1].trim()
+  const start = text.indexOf('{')
+  const end = text.lastIndexOf('}')
+  if (start >= 0 && end > start) return text.slice(start, end + 1)
+  return text.trim()
+}
+
+function parsePlanJson(rawText: string, availableTools: Set<string>): AnalysisPlanSummary | null {
+  try {
+    const parsed = JSON.parse(extractJsonObject(rawText)) as unknown
+    if (!isRecord(parsed)) return null
+    const title = normalizeString(parsed.title)
+    if (!title || !isIntent(parsed.intent) || !Array.isArray(parsed.steps)) return null
+
+    // 中文注释：Planner 输出来自 LLM，必须在进入 Agent guidance 前做结构约束；
+    // 这里裁剪长度并过滤不存在的工具，避免把幻觉工具名注入后续执行上下文。
+    const steps = parsed.steps
+      .filter(isRecord)
+      .map((step) => {
+        const goal = normalizeString(step.goal)
+        const evidenceNeeded = normalizeString(step.evidenceNeeded)
+        if (!goal || !evidenceNeeded) return null
+        const suggestedTools = Array.isArray(step.suggestedTools)
+          ? step.suggestedTools.filter((tool): tool is string => typeof tool === 'string' && availableTools.has(tool))
+          : []
+        return { goal, suggestedTools: Array.from(new Set(suggestedTools)), evidenceNeeded }
+      })
+      .filter((step): step is NonNullable<typeof step> => step !== null)
+      .slice(0, 5)
+
+    if (steps.length === 0) return null
+
+    const successCriteria = Array.isArray(parsed.successCriteria)
+      ? parsed.successCriteria
+          .map(normalizeString)
+          .filter((item): item is string => item !== null)
+          .slice(0, 5)
+      : []
+    if (successCriteria.length === 0) return null
+
+    return {
+      version: 1,
+      title,
+      route: 'planned_execution',
+      intent: parsed.intent,
+      steps,
+      successCriteria,
+    }
+  } catch {
+    return null
+  }
+}
+
+function resultToText(result: PlannerCompletionResult): string {
+  return typeof result === 'string' ? result : result.text
+}
+
+function buildPlannerPrompt(input: PlannerInput): string {
+  const dataSnapshot = input.dataSnapshot
+    ? `${input.dataSnapshot.name}, ${input.dataSnapshot.totalMessages} messages, ${input.dataSnapshot.totalMembers} members`
+    : '(none)'
+
+  return `Create a concise user-visible analysis plan for a ChatLab planned_execution request.
+
+Return strict JSON only:
+{
+  "title": "short title",
+  "intent": "summary|trend|relationship|search|comparison|mixed",
+  "steps": [
+    {"goal": "what to investigate", "suggestedTools": ["tool_name"], "evidenceNeeded": "what evidence is needed"}
+  ],
+  "successCriteria": ["criterion"]
+}
+
+Limits:
+- Max 5 steps.
+- Max 5 successCriteria.
+- suggestedTools must come only from availableTools.
+- Do not reveal hidden chain-of-thought; write a user-readable plan summary.
+- The plan is soft guidance, not a mandatory execution script.
+
+Context:
+- locale: ${input.locale}
+- chatType: ${input.chatType}
+- availableTools: ${input.availableTools.join(', ') || '(none)'}
+- dataSnapshot: ${dataSnapshot}
+- assistantSummary: ${input.assistantSummary ?? '(none)'}
+- skillSummary: ${input.skillSummary ?? '(none)'}
+- recentIntentSummary: ${input.recentIntentSummary ?? '(none)'}
+
+User request:
+${input.userMessage}`
+}
+
+export function createAnalysisPlanner(options: CreateAnalysisPlannerOptions): AnalysisPlanner {
+  return async (input, signal) => {
+    try {
+      const prompt = buildPlannerPrompt(input)
+      const rawResult = options.complete
+        ? await options.complete(prompt, signal)
+        : await completeWithPiAi(prompt, options, signal)
+      return parsePlanJson(resultToText(rawResult), new Set(input.availableTools))
+    } catch {
+      return null
+    }
+  }
+}
+
+async function completeWithPiAi(
+  prompt: string,
+  options: CreateAnalysisPlannerOptions,
+  signal?: AbortSignal
+): Promise<PlannerCompletionResult> {
+  if (!options.piModel || !options.apiKey) {
+    throw new Error('Planner requires piModel and apiKey')
+  }
+
+  const result = await completeSimple(
+    options.piModel,
+    {
+      systemPrompt: 'You are a strict JSON planner for ChatLab analysis. Return JSON only.',
+      messages: [{ role: 'user', content: [{ type: 'text', text: prompt }], timestamp: Date.now() }] as any,
+    },
+    {
+      apiKey: options.apiKey,
+      maxTokens: options.maxTokens ?? 700,
+      temperature: options.temperature ?? 0.2,
+      signal,
+    }
+  )
+
+  return {
+    text: result.content
+      .filter((item): item is PiTextContent => item.type === 'text')
+      .map((item) => item.text)
+      .join(''),
+  }
+}
diff --git a/packages/node-runtime/src/ai/agent/planning-types.ts b/packages/node-runtime/src/ai/agent/planning-types.ts
new file mode 100644
index 00000000..77250ccb
--- /dev/null
+++ b/packages/node-runtime/src/ai/agent/planning-types.ts
@@ -0,0 +1,38 @@
+import type { DataSnapshot } from './prompt-builder'
+
+export type AnalysisPlanIntent = 'summary' | 'trend' | 'relationship' | 'search' | 'comparison' | 'mixed'
+
+export interface AnalysisPlanStep {
+  goal: string
+  suggestedTools: string[]
+  evidenceNeeded: string
+}
+
+export interface AnalysisPlanSummary {
+  version: 1
+  title: string
+  route: 'planned_execution'
+  intent: AnalysisPlanIntent
+  steps: AnalysisPlanStep[]
+  successCriteria: string[]
+}
+
+export interface PlanContentBlock {
+  type: 'plan'
+  version: 1
+  status: 'created' | 'executing' | 'done' | 'skipped'
+  plan: AnalysisPlanSummary
+}
+
+export interface PlannerInput {
+  userMessage: string
+  chatType: 'group' | 'private'
+  locale: string
+  dataSnapshot?: DataSnapshot
+  availableTools: string[]
+  assistantSummary?: string
+  skillSummary?: string
+  recentIntentSummary?: string
+}
+
+export type AnalysisPlanner = (input: PlannerInput, signal?: AbortSignal) => Promise<AnalysisPlanSummary | null>
diff --git a/packages/node-runtime/src/ai/index.ts b/packages/node-runtime/src/ai/index.ts
index d8a47185..2c71233a 100644
--- a/packages/node-runtime/src/ai/index.ts
+++ b/packages/node-runtime/src/ai/index.ts
@@ -99,6 +99,15 @@ export type { PreprocessingPipelineOptions, PreprocessingPipelineResult } from '
 export type { AgentCoreOptions, AgentCoreEvent, AgentCoreResult, AgentTokenUsage, SimpleHistoryMessage } from './agent'
 export { DEFAULT_MAX_TOOL_ROUNDS, createLlmRouteDecider, decideRequestRoute, runAgentCore } from './agent'
 export type { LlmRouteDecider, RequestRoute, RouteDecision, RouteDecisionSource, RouterInput } from './agent'
+export { createAnalysisPlanner } from './agent'
+export type {
+  AnalysisPlanIntent,
+  AnalysisPlanner,
+  AnalysisPlanStep,
+  AnalysisPlanSummary,
+  PlannerInput,
+  PlanContentBlock,
+} from './agent'
 
 // Agent Event Handler
 export { AgentEventHandler, estimateTokensFromText } from './agent/event-handler'
diff --git a/packages/node-runtime/src/index.ts b/packages/node-runtime/src/index.ts
index 36bcd36d..0e522a9d 100644
--- a/packages/node-runtime/src/index.ts
+++ b/packages/node-runtime/src/index.ts
@@ -205,6 +205,15 @@ export type { AIChat, AIMessage, AIMessageRole, ContentBlock, TokenUsageData, AI
 export type { AgentCoreOptions, AgentCoreEvent, AgentCoreResult, AgentTokenUsage, SimpleHistoryMessage } from './ai'
 export { DEFAULT_MAX_TOOL_ROUNDS, createLlmRouteDecider, decideRequestRoute, runAgentCore } from './ai'
 export type { LlmRouteDecider, RequestRoute, RouteDecision, RouteDecisionSource, RouterInput } from './ai'
+export { createAnalysisPlanner } from './ai'
+export type {
+  AnalysisPlanIntent,
+  AnalysisPlanner,
+  AnalysisPlanStep,
+  AnalysisPlanSummary,
+  PlannerInput,
+  PlanContentBlock,
+} from './ai'
 
 // Summary generation
 export {