diff --git a/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.test.ts b/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.test.ts
new file mode 100644
index 000000000..f36b8c4ad
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.test.ts
@@ -0,0 +1,92 @@
+import { describe, it, expect } from 'vitest';
+import { parsePlannerResponse } from './parser';
+
+const TOOLS = new Set(['create_task', 'log_drink']);
+
+describe('parsePlannerResponse', () => {
+	it('parses a valid fenced json block', () => {
+		const text = `\`\`\`json
+{
+  "summary": "Plan für heute",
+  "steps": [
+    { "summary": "Task anlegen", "toolName": "create_task", "params": { "title": "Foo" }, "rationale": "weil wichtig" }
+  ]
+}
+\`\`\``;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(true);
+		if (!r.ok) return;
+		expect(r.value.summary).toBe('Plan für heute');
+		expect(r.value.steps).toHaveLength(1);
+		expect(r.value.steps[0].toolName).toBe('create_task');
+		expect(r.value.steps[0].params).toEqual({ title: 'Foo' });
+	});
+
+	it('accepts a bare JSON object without fence', () => {
+		const text = `{ "summary": "x", "steps": [
+			{ "summary": "log", "toolName": "log_drink", "params": {}, "rationale": "Routine" }
+		]}`;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(true);
+	});
+
+	it('rejects when no JSON block found', () => {
+		const r = parsePlannerResponse('just prose no JSON here', TOOLS);
+		expect(r.ok).toBe(false);
+	});
+
+	it('rejects invalid JSON inside the fence', () => {
+		const r = parsePlannerResponse('```json\n{not: valid}\n```', TOOLS);
+		expect(r.ok).toBe(false);
+	});
+
+	it('rejects when steps is missing or not an array', () => {
+		const r = parsePlannerResponse('```json\n{"summary":"x"}\n```', TOOLS);
+		expect(r.ok).toBe(false);
+	});
+
+	it('rejects steps referencing unknown tools', () => {
+		const text = `\`\`\`json
+{ "summary": "", "steps": [{ "toolName": "nuke_database", "params": {}, "rationale": "why not" }] }
+\`\`\``;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(false);
+		if (r.ok) return;
+		expect(r.reason).toContain('nuke_database');
+	});
+
+	it('rejects steps missing rationale', () => {
+		const text = `\`\`\`json
+{ "summary": "", "steps": [{ "toolName": "create_task", "params": { "title": "x" } }] }
+\`\`\``;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(false);
+		if (r.ok) return;
+		expect(r.reason).toContain('rationale');
+	});
+
+	it('tolerates missing summary / step summary by defaulting to empty', () => {
+		const text = `\`\`\`json
+{
+  "steps": [
+    { "toolName": "create_task", "params": {}, "rationale": "need one" }
+  ]
+}
+\`\`\``;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(true);
+		if (!r.ok) return;
+		expect(r.value.summary).toBe('');
+		expect(r.value.steps[0].summary).toBe('');
+	});
+
+	it('accepts an empty steps array (no-op iteration)', () => {
+		const text = `\`\`\`json
+{ "summary": "nothing to do today", "steps": [] }
+\`\`\``;
+		const r = parsePlannerResponse(text, TOOLS);
+		expect(r.ok).toBe(true);
+		if (!r.ok) return;
+		expect(r.value.steps).toHaveLength(0);
+	});
+});
diff --git a/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.ts b/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.ts
new file mode 100644
index 000000000..5226bee6a
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/data/ai/missions/planner/parser.ts
@@ -0,0 +1,99 @@
+/**
+ * Parser for Planner LLM output.
+ *
+ * Strict: we only accept the fenced `json` block the system prompt
+ * prescribes, validate shape, and surface errors so the Runner can
+ * record them on the iteration instead of silently producing a bad plan.
+ */
+
+import type { AiPlanOutput, PlannedStep } from './types';
+
+export type ParseResult =
+	| { readonly ok: true; readonly value: AiPlanOutput }
+	| { readonly ok: false; readonly reason: string; readonly raw?: string };
+
+/** Extract and validate the Planner JSON from the LLM's raw response. */
+export function parsePlannerResponse(text: string, knownToolNames: Set<string>): ParseResult {
+	const block = extractJsonBlock(text);
+	if (!block) return { ok: false, reason: 'no JSON block found', raw: text };
+
+	let parsed: unknown;
+	try {
+		parsed = JSON.parse(block);
+	} catch (err) {
+		return {
+			ok: false,
+			reason: `JSON parse failed: ${err instanceof Error ? err.message : String(err)}`,
+			raw: block,
+		};
+	}
+
+	if (typeof parsed !== 'object' || parsed === null) {
+		return { ok: false, reason: 'top-level value is not an object', raw: block };
+	}
+
+	const obj = parsed as Record<string, unknown>;
+	const summary = typeof obj.summary === 'string' ? obj.summary : '';
+	const rawSteps = obj.steps;
+	if (!Array.isArray(rawSteps)) {
+		return { ok: false, reason: '`steps` must be an array', raw: block };
+	}
+
+	const steps: PlannedStep[] = [];
+	for (let i = 0; i < rawSteps.length; i++) {
+		const step = rawSteps[i];
+		const validation = validateStep(step, knownToolNames, i);
+		if (!validation.ok) {
+			return { ok: false, reason: validation.reason, raw: block };
+		}
+		steps.push(validation.value);
+	}
+
+	return { ok: true, value: { summary, steps } };
+}
+
+function extractJsonBlock(text: string): string | null {
+	// Prefer a ```json fenced block
+	const fenced = /```(?:json)?\s*\n?([\s\S]*?)\n?```/;
+	const m = text.match(fenced);
+	if (m) return m[1].trim();
+	// Fallback: the whole response might BE a JSON object
+	const trimmed = text.trim();
+	if (trimmed.startsWith('{') && trimmed.endsWith('}')) return trimmed;
+	return null;
+}
+
+function validateStep(
+	raw: unknown,
+	knownToolNames: Set<string>,
+	index: number
+): { ok: true; value: PlannedStep } | { ok: false; reason: string } {
+	if (typeof raw !== 'object' || raw === null) {
+		return { ok: false, reason: `step[${index}] is not an object` };
+	}
+	const obj = raw as Record<string, unknown>;
+	const toolName = obj.toolName;
+	if (typeof toolName !== 'string' || toolName.length === 0) {
+		return { ok: false, reason: `step[${index}].toolName missing or not a string` };
+	}
+	if (!knownToolNames.has(toolName)) {
+		return {
+			ok: false,
+			reason: `step[${index}].toolName "${toolName}" is not in the allowed tool set`,
+		};
+	}
+	const summary = typeof obj.summary === 'string' ? obj.summary : '';
+	const rationale = typeof obj.rationale === 'string' ? obj.rationale : '';
+	if (rationale.length === 0) {
+		return { ok: false, reason: `step[${index}].rationale is required (user will see this)` };
+	}
+	const params =
+		typeof obj.params === 'object' && obj.params !== null
+			? (obj.params as Record<string, unknown>)
+			: {};
+
+	return {
+		ok: true,
+		value: { summary, toolName, rationale, params },
+	};
+}
diff --git a/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.test.ts b/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.test.ts
new file mode 100644
index 000000000..e2ca62749
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.test.ts
@@ -0,0 +1,127 @@
+import { describe, it, expect } from 'vitest';
+import { buildPlannerPrompt } from './prompt';
+import type { AiPlanInput } from './types';
+import type { Mission } from '../types';
+
+function baseMission(overrides: Partial<Mission> = {}): Mission {
+	return {
+		id: 'm-1',
+		createdAt: '2026-04-14T10:00:00.000Z',
+		updatedAt: '2026-04-14T10:00:00.000Z',
+		title: 'Weekly review',
+		conceptMarkdown: '# Concept\nDo a thing.',
+		objective: 'Review progress each Monday',
+		inputs: [],
+		cadence: { kind: 'weekly', dayOfWeek: 1, atHour: 9 },
+		state: 'active',
+		iterations: [],
+		...overrides,
+	};
+}
+
+describe('buildPlannerPrompt', () => {
+	it('emits system + user messages with mission title and objective', () => {
+		const input: AiPlanInput = {
+			mission: baseMission(),
+			resolvedInputs: [],
+			availableTools: [],
+		};
+		const { system, user } = buildPlannerPrompt(input);
+		expect(user).toContain('Weekly review');
+		expect(user).toContain('Review progress each Monday');
+		expect(system).toContain('JSON');
+		expect(system).toContain('rationale');
+	});
+
+	it('lists available tools with their params in the system prompt', () => {
+		const input: AiPlanInput = {
+			mission: baseMission(),
+			resolvedInputs: [],
+			availableTools: [
+				{
+					name: 'create_task',
+					module: 'todo',
+					description: 'Creates a task',
+					parameters: [
+						{ name: 'title', type: 'string', required: true, description: 'Task title' },
+						{
+							name: 'priority',
+							type: 'string',
+							required: false,
+							description: 'prio',
+							enum: ['low', 'high'],
+						},
+					],
+				},
+			],
+		};
+		const { system } = buildPlannerPrompt(input);
+		expect(system).toContain('create_task');
+		expect(system).toContain('title');
+		expect(system).toContain('(required)');
+		expect(system).toContain('[low|high]');
+	});
+
+	it('injects resolved input content into the user prompt', () => {
+		const input: AiPlanInput = {
+			mission: baseMission({
+				inputs: [{ module: 'notes', table: 'notes', id: 'n-1' }],
+			}),
+			resolvedInputs: [
+				{ id: 'n-1', module: 'notes', table: 'notes', title: 'Strategy', content: 'Be bold.' },
+			],
+			availableTools: [],
+		};
+		const { user } = buildPlannerPrompt(input);
+		expect(user).toContain('Strategy');
+		expect(user).toContain('Be bold.');
+	});
+
+	it('includes user feedback from the most recent iteration', () => {
+		const input: AiPlanInput = {
+			mission: baseMission({
+				iterations: [
+					{
+						id: 'it-1',
+						startedAt: '2026-04-07T09:00:00.000Z',
+						finishedAt: '2026-04-07T09:01:00.000Z',
+						plan: [
+							{
+								id: 's-1',
+								summary: 'Old step',
+								intent: { kind: 'toolCall', toolName: 'create_task', params: {} },
+								status: 'rejected',
+							},
+						],
+						userFeedback: 'Zu aggressiv — bitte zurücknehmen',
+						overallStatus: 'rejected',
+					},
+				],
+			}),
+			resolvedInputs: [],
+			availableTools: [],
+		};
+		const { user } = buildPlannerPrompt(input);
+		expect(user).toContain('Zu aggressiv');
+		expect(user).toContain('[rejected]');
+	});
+
+	it('truncates iteration history to the last 3', () => {
+		const many = Array.from({ length: 10 }, (_, i) => ({
+			id: `it-${i}`,
+			startedAt: `2026-04-${String(i + 1).padStart(2, '0')}T10:00:00.000Z`,
+			plan: [],
+			overallStatus: 'approved' as const,
+			userFeedback: `feedback-${i}`,
+		}));
+		const { user } = buildPlannerPrompt({
+			mission: baseMission({ iterations: many }),
+			resolvedInputs: [],
+			availableTools: [],
+		});
+		// Only the last three iterations (7, 8, 9) should be present
+		expect(user).toContain('feedback-9');
+		expect(user).toContain('feedback-7');
+		expect(user).not.toContain('feedback-5');
+	});
+});
diff --git a/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.ts b/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.ts
new file mode 100644
index 000000000..5e5ae54d0
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/data/ai/missions/planner/prompt.ts
@@ -0,0 +1,112 @@
+/**
+ * Prompt builder for the Mission Planner.
+ *
+ * Produces a system + user message pair the LLM backend can consume. The
+ * grammar we ask the model to return is deliberately small (JSON in a
+ * fenced block, one shape) — the parser is strict, and we'd rather the
+ * LLM mess up in a detectable way than produce half-valid output.
+ *
+ * Keeps everything the Planner sees inside the prompt, nothing ambient —
+ * makes the task reproducible and unit-testable without a live LLM.
+ */
+
+import type { AiPlanInput } from './types';
+
+export interface PlannerMessages {
+	readonly system: string;
+	readonly user: string;
+}
+
+export function buildPlannerPrompt(input: AiPlanInput): PlannerMessages {
+	return {
+		system: buildSystemPrompt(input),
+		user: buildUserPrompt(input),
+	};
+}
+
+function buildSystemPrompt(input: AiPlanInput): string {
+	const toolBlock = input.availableTools
+		.map((t) => {
+			const params = t.parameters
+				.map((p) => {
+					const req = p.required ? ' (required)' : '';
+					const enumeration = p.enum ? ` [${p.enum.join('|')}]` : '';
+					return `    - ${p.name}: ${p.type}${enumeration}${req} — ${p.description}`;
+				})
+				.join('\n');
+			return `  • ${t.name} (${t.module}) — ${t.description}\n${params || '    (no parameters)'}`;
+		})
+		.join('\n');
+
+	return `Du bist eine KI, die im Auftrag des Nutzers an einer langlebigen Mission arbeitet.
+
+Dein Job: aus dem aktuellen Mission-Kontext einen kurzen, konkreten Plan ableiten — 1 bis 5 Schritte, jeder ein Tool-Aufruf auf Nutzerdaten. Jeder Schritt MUSS eine Begründung haben (rationale), die der Nutzer in der Review-UI sieht.
+
+Wichtige Regeln:
+1. Nutze NUR Tools aus der Liste unten. Unbekannte Tools → Plan invalide.
+2. Jeder Step wird als Proposal gestaged — der Nutzer approved oder rejected. Du schreibst nie direkt.
+3. Berücksichtige das Feedback aus vorherigen Iterationen (unten im User-Prompt). Wenn ein Vorschlag rejected wurde, wiederhole ihn nicht ohne Änderung.
+4. Antworte AUSSCHLIESSLICH mit einem JSON-Block in folgendem Format, keine Prosa davor/danach:
+
+\`\`\`json
+{
+  "summary": "Ein Satz was du in dieser Iteration tust.",
+  "steps": [
+    {
+      "summary": "Kurzer Schritt-Titel",
+      "toolName": "create_task",
+      "params": { "title": "…" },
+      "rationale": "Warum genau jetzt, auf Basis welchen Inputs."
+    }
+  ]
+}
+\`\`\`
+
+Verfügbare Tools:
+${toolBlock || '  (keine Tools verfügbar — gib leeren steps zurück)'}`;
+}
+
+function buildUserPrompt(input: AiPlanInput): string {
+	const { mission, resolvedInputs } = input;
+
+	const inputsBlock =
+		resolvedInputs.length === 0
+			? '_(keine verlinkten Inputs)_'
+			: resolvedInputs
+					.map((r) => {
+						const header = `### ${r.module}/${r.table}: ${r.title ?? r.id}`;
+						return `${header}\n${r.content}`;
+					})
+					.join('\n\n');
+
+	const iterationHistory =
+		mission.iterations.length === 0
+			? '_(erste Iteration)_'
+			: mission.iterations
+					.slice(-3) // only the last 3 to keep the prompt tight
+					.map((it) => {
+						const steps = it.plan.map((s) => `  - [${s.status}] ${s.summary}`).join('\n');
+						const feedback = it.userFeedback ? `\n  Nutzer-Feedback: ${it.userFeedback}` : '';
+						const summary = it.summary ? `\n  Summary: ${it.summary}` : '';
+						return `**${it.startedAt}** (${it.overallStatus}):${summary}\n${steps}${feedback}`;
+					})
+					.join('\n\n');
+
+	return `# Mission: ${mission.title}
+
+## Konzept
+${mission.conceptMarkdown || '_(leer)_'}
+
+## Konkretes Ziel
+${mission.objective}
+
+## Verlinkte Inputs
+${inputsBlock}
+
+## Letzte Iterationen (max. 3)
+${iterationHistory}
+
+---
+
+Erzeuge jetzt einen Plan für die nächste Iteration.`;
+}
diff --git a/apps/mana/apps/web/src/lib/data/ai/missions/planner/types.ts b/apps/mana/apps/web/src/lib/data/ai/missions/planner/types.ts
new file mode 100644
index 000000000..d6822de2c
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/data/ai/missions/planner/types.ts
@@ -0,0 +1,56 @@
+/**
+ * Planner contract — shared input / output shape used by prompt builder,
+ * parser, and the `aiPlanTask` LLM wrapper.
+ */
+
+import type { Mission } from '../types';
+
+/** Resolved content for one {@link MissionInputRef}. */
+export interface ResolvedInput {
+	readonly id: string;
+	readonly module: string;
+	readonly table: string;
+	readonly title?: string;
+	readonly content: string;
+}
+
+/** Tool definition the Planner is allowed to propose. */
+export interface AvailableTool {
+	readonly name: string;
+	readonly module: string;
+	readonly description: string;
+	readonly parameters: ReadonlyArray<{
+		readonly name: string;
+		readonly type: string;
+		readonly required: boolean;
+		readonly description: string;
+		readonly enum?: readonly string[];
+	}>;
+}
+
+export interface AiPlanInput {
+	readonly mission: Mission;
+	/** Content of every MissionInputRef, pre-fetched by the caller (Runner). */
+	readonly resolvedInputs: readonly ResolvedInput[];
+	/** Tools the policy has whitelisted for AI proposals in this run. */
+	readonly availableTools: readonly AvailableTool[];
+}
+
+/**
+ * One step in the Planner's output. The Runner turns each step into a
+ * Proposal by calling `executeTool(toolName, params, aiActor)` — the
+ * policy routes it to `propose`, and the Proposal carries the rationale.
+ */
+export interface PlannedStep {
+	readonly summary: string;
+	readonly toolName: string;
+	readonly params: Record<string, unknown>;
+	readonly rationale: string;
+}
+
+export interface AiPlanOutput {
+	/** 1–N steps the AI wants to take this iteration. May be empty (no-op run). */
+	readonly steps: readonly PlannedStep[];
+	/** The AI's one-line summary of the plan, stored on `MissionIteration.summary`. */
+	readonly summary: string;
+}
diff --git a/apps/mana/apps/web/src/lib/llm-tasks/ai-plan.ts b/apps/mana/apps/web/src/lib/llm-tasks/ai-plan.ts
new file mode 100644
index 000000000..e493183da
--- /dev/null
+++ b/apps/mana/apps/web/src/lib/llm-tasks/ai-plan.ts
@@ -0,0 +1,62 @@
+/**
+ * aiPlanTask — LLM task for the Mission Planner.
+ *
+ * Takes a Mission plus pre-resolved inputs + available tools, asks the
+ * configured LLM backend for a structured plan, parses it, and returns
+ * typed steps the Runner turns into Proposals.
+ *
+ * Routing:
+ * - `minTier: 'browser'` — the Planner runs entirely on the device by
+ *   default. Users can override to mana-server / cloud in settings for
+ *   more capable reasoning on long missions.
+ * - `contentClass: 'personal'` — the prompt contains the user's notes and
+ *   goals. If any linked input is from a strictly-sensitive module
+ *   (journal, dreams, finance), the Runner is responsible for narrowing
+ *   to `'sensitive'` on the request so cloud is refused.
+ *
+ * Error path: the parser returns a structured `ParseResult`. If parsing
+ * fails, the task still returns — with `steps: []` and a summary
+ * explaining why — so the Runner can record a failed iteration without
+ * throwing through the whole mission queue.
+ */
+
+import type { LlmBackend, LlmTask } from '@mana/shared-llm';
+import { buildPlannerPrompt } from '$lib/data/ai/missions/planner/prompt';
+import { parsePlannerResponse } from '$lib/data/ai/missions/planner/parser';
+import type { AiPlanInput, AiPlanOutput } from '$lib/data/ai/missions/planner/types';
+
+export type { AiPlanInput, AiPlanOutput } from '$lib/data/ai/missions/planner/types';
+
+export const aiPlanTask: LlmTask<AiPlanInput, AiPlanOutput> = {
+	name: 'ai.plan',
+	minTier: 'browser',
+	contentClass: 'personal',
+	requires: { streaming: false },
+	displayLabel: 'AI Mission Planner',
+
+	async runLlm(input: AiPlanInput, backend: LlmBackend): Promise<AiPlanOutput> {
+		const { system, user } = buildPlannerPrompt(input);
+
+		const result = await backend.generate({
+			taskName: 'ai.plan',
+			contentClass: 'personal',
+			messages: [
+				{ role: 'system', content: system },
+				{ role: 'user', content: user },
+			],
+			temperature: 0.3,
+			maxTokens: 1024,
+		});
+
+		const knownToolNames = new Set(input.availableTools.map((t) => t.name));
+		const parsed = parsePlannerResponse(result.content, knownToolNames);
+
+		if (!parsed.ok) {
+			return {
+				steps: [],
+				summary: `Plan konnte nicht erzeugt werden: ${parsed.reason}`,
+			};
+		}
+		return parsed.value;
+	},
+};