diff --git a/apps/claw-routing-service/src/modules/cost-budget/constants/budget.constants.ts b/apps/claw-routing-service/src/modules/cost-budget/constants/budget.constants.ts
new file mode 100644
index 00000000..026d8bb2
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/constants/budget.constants.ts
@@ -0,0 +1,9 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+export const BUDGET_DEFAULT_WARN_AT_PERCENT = 80;
+export const BUDGET_DEFAULT_RESET_INTERVAL_DAYS = 30;
+
+export const BUDGET_MESSAGE_KEY_NEAR_LIMIT = 'BUDGET_NEAR_LIMIT';
+export const BUDGET_MESSAGE_KEY_EXCEEDED_FORCED_LOCAL = 'BUDGET_EXCEEDED_FORCED_LOCAL';
+export const BUDGET_MESSAGE_KEY_EXCEEDED_OVERRIDDEN = 'BUDGET_EXCEEDED_OVERRIDDEN';
+export const BUDGET_MESSAGE_KEY_UNKNOWN_COST_BLOCKED = 'BUDGET_UNKNOWN_COST_BLOCKED';
diff --git a/apps/claw-routing-service/src/modules/cost-budget/controllers/cost-budget.controller.ts b/apps/claw-routing-service/src/modules/cost-budget/controllers/cost-budget.controller.ts
new file mode 100644
index 00000000..4a85aa4b
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/controllers/cost-budget.controller.ts
@@ -0,0 +1,45 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { Body, Controller, Get, Param, Patch, Post } from '@nestjs/common';
+
+import { CostBudgetService } from '../services/cost-budget.service';
+
+@Controller('routing/cost-budget')
+export class CostBudgetController {
+  constructor(private readonly service: CostBudgetService) {}
+
+  @Get('me')
+  async getMine(): Promise<unknown> {
+    return this.service.getMine();
+  }
+
+  @Get('me/forecast')
+  async getMineForecast(): Promise<unknown> {
+    return this.service.getMineForecast();
+  }
+
+  @Patch('me')
+  async updateMine(@Body() body: unknown): Promise<unknown> {
+    return this.service.updateMine(body);
+  }
+
+  @Get()
+  async listAll(): Promise<unknown> {
+    return this.service.listAll();
+  }
+
+  @Post()
+  async create(@Body() body: unknown): Promise<unknown> {
+    return this.service.create(body);
+  }
+
+  @Patch(':id')
+  async update(@Param('id') id: string, @Body() body: unknown): Promise<unknown> {
+    return this.service.update(id, body);
+  }
+
+  @Post('check')
+  async check(@Body() body: unknown): Promise<unknown> {
+    return this.service.check(body);
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/cost-budget.module.ts b/apps/claw-routing-service/src/modules/cost-budget/cost-budget.module.ts
new file mode 100644
index 00000000..9975a680
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/cost-budget.module.ts
@@ -0,0 +1,26 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+// NEW module — NOT yet registered in app.module.ts.
+
+import { Module } from '@nestjs/common';
+
+import { CostBudgetController } from './controllers/cost-budget.controller';
+import { BudgetGateManager } from './managers/budget-gate.manager';
+import { BudgetResetManager } from './managers/budget-reset.manager';
+import { BudgetWarningManager } from './managers/budget-warning.manager';
+import { SpendTrackerManager } from './managers/spend-tracker.manager';
+import { UserCostBudgetRepository } from './repositories/user-cost-budget.repository';
+import { CostBudgetService } from './services/cost-budget.service';
+
+@Module({
+  controllers: [CostBudgetController],
+  providers: [
+    CostBudgetService,
+    BudgetGateManager,
+    SpendTrackerManager,
+    BudgetWarningManager,
+    BudgetResetManager,
+    UserCostBudgetRepository,
+  ],
+  exports: [CostBudgetService, BudgetGateManager, SpendTrackerManager],
+})
+export class CostBudgetModule {}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/dto/check-budget.dto.ts b/apps/claw-routing-service/src/modules/cost-budget/dto/check-budget.dto.ts
new file mode 100644
index 00000000..852d46f0
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/dto/check-budget.dto.ts
@@ -0,0 +1,11 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { z } from 'zod';
+
+export const checkBudgetSchema = z.object({
+  userId: z.string().min(1).max(200),
+  orgId: z.string().max(200).optional(),
+  estimatedCostUsd: z.number().nonnegative().max(10_000),
+});
+
+export type CheckBudgetDto = z.infer<typeof checkBudgetSchema>;
diff --git a/apps/claw-routing-service/src/modules/cost-budget/dto/create-budget.dto.ts b/apps/claw-routing-service/src/modules/cost-budget/dto/create-budget.dto.ts
new file mode 100644
index 00000000..2b19e744
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/dto/create-budget.dto.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { z } from 'zod';
+
+export const createBudgetSchema = z.object({
+  scope: z.enum(['USER', 'ORG']),
+  ownerId: z.string().min(1).max(200),
+  monthlyCapUsd: z.number().positive().max(1_000_000),
+  warnAtPercent: z.number().int().min(1).max(100).default(80),
+  overrideAllowed: z.boolean().default(false),
+});
+
+export type CreateBudgetDto = z.infer<typeof createBudgetSchema>;
diff --git a/apps/claw-routing-service/src/modules/cost-budget/dto/update-budget.dto.ts b/apps/claw-routing-service/src/modules/cost-budget/dto/update-budget.dto.ts
new file mode 100644
index 00000000..21555947
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/dto/update-budget.dto.ts
@@ -0,0 +1,11 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { z } from 'zod';
+
+export const updateBudgetSchema = z.object({
+  monthlyCapUsd: z.number().positive().max(1_000_000).optional(),
+  warnAtPercent: z.number().int().min(1).max(100).optional(),
+  overrideAllowed: z.boolean().optional(),
+});
+
+export type UpdateBudgetDto = z.infer<typeof updateBudgetSchema>;
diff --git a/apps/claw-routing-service/src/modules/cost-budget/managers/budget-gate.manager.ts b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-gate.manager.ts
new file mode 100644
index 00000000..9cb5fcd5
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-gate.manager.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+// Pre-routing budget check.
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { BudgetCheckInput, BudgetCheckResult } from '../types/budget.types';
+
+@Injectable()
+export class BudgetGateManager {
+  private readonly logger = new Logger(BudgetGateManager.name);
+
+  async check(_input: BudgetCheckInput): Promise<BudgetCheckResult> {
+    this.logger.warn('BudgetGateManager.check: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R4 — BudgetGateManager.check not implemented; see docs/15-ai-context/routing-flagship-streams/05-r4-cost-budget-intelligence.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/managers/budget-reset.manager.ts b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-reset.manager.ts
new file mode 100644
index 00000000..55481619
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-reset.manager.ts
@@ -0,0 +1,16 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+// Monthly cron — resets every UserCostBudget where resetAt <= now.
+
+import { Injectable, Logger } from '@nestjs/common';
+import { Cron, CronExpression } from '@nestjs/schedule';
+
+@Injectable()
+export class BudgetResetManager {
+  private readonly logger = new Logger(BudgetResetManager.name);
+
+  @Cron(CronExpression.EVERY_HOUR)
+  async tick(): Promise<void> {
+    this.logger.warn('BudgetResetManager.tick: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R4 — BudgetResetManager.tick not implemented');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/managers/budget-warning.manager.ts b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-warning.manager.ts
new file mode 100644
index 00000000..7f0808e1
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/managers/budget-warning.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class BudgetWarningManager {
+  private readonly logger = new Logger(BudgetWarningManager.name);
+
+  async checkAndEmitWarning(_budgetId: string, _currentSpendUsd: number, _capUsd: number): Promise<void> {
+    this.logger.warn('BudgetWarningManager.checkAndEmitWarning: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — BudgetWarningManager.checkAndEmitWarning not implemented');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/managers/spend-tracker.manager.ts b/apps/claw-routing-service/src/modules/cost-budget/managers/spend-tracker.manager.ts
new file mode 100644
index 00000000..8e1e3c68
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/managers/spend-tracker.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class SpendTrackerManager {
+  private readonly logger = new Logger(SpendTrackerManager.name);
+
+  async incrementSpend(_userId: string, _orgId: string | undefined, _amountUsd: number): Promise<void> {
+    this.logger.warn('SpendTrackerManager.incrementSpend: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — SpendTrackerManager.incrementSpend not implemented');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/repositories/user-cost-budget.repository.ts b/apps/claw-routing-service/src/modules/cost-budget/repositories/user-cost-budget.repository.ts
new file mode 100644
index 00000000..d0c06196
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/repositories/user-cost-budget.repository.ts
@@ -0,0 +1,30 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+// Pure data access — uses PrismaService once UserCostBudget model is added.
+// Per repository rule: NO throws. Return null/empty until implemented.
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class UserCostBudgetRepository {
+  private readonly logger = new Logger(UserCostBudgetRepository.name);
+
+  async findActiveForOwner(_scope: string, _ownerId: string): Promise<unknown | null> {
+    this.logger.warn('UserCostBudgetRepository.findActiveForOwner: SCAFFOLD-R4 returns null');
+    return null;
+  }
+
+  async create(_input: unknown): Promise<unknown | null> {
+    this.logger.warn('UserCostBudgetRepository.create: SCAFFOLD-R4 returns null');
+    return null;
+  }
+
+  async update(_id: string, _input: unknown): Promise<unknown | null> {
+    this.logger.warn('UserCostBudgetRepository.update: SCAFFOLD-R4 returns null');
+    return null;
+  }
+
+  async findExpired(_now: Date): Promise<unknown[]> {
+    this.logger.warn('UserCostBudgetRepository.findExpired: SCAFFOLD-R4 returns empty');
+    return [];
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/services/cost-budget.service.ts b/apps/claw-routing-service/src/modules/cost-budget/services/cost-budget.service.ts
new file mode 100644
index 00000000..329abe39
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/services/cost-budget.service.ts
@@ -0,0 +1,37 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class CostBudgetService {
+  private readonly logger = new Logger(CostBudgetService.name);
+
+  async getMine(): Promise<unknown> {
+    this.logger.warn('CostBudgetService.getMine: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.getMine not implemented');
+  }
+  async getMineForecast(): Promise<unknown> {
+    this.logger.warn('CostBudgetService.getMineForecast: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.getMineForecast not implemented');
+  }
+  async updateMine(_body: unknown): Promise<unknown> {
+    this.logger.warn('CostBudgetService.updateMine: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.updateMine not implemented');
+  }
+  async listAll(): Promise<unknown> {
+    this.logger.warn('CostBudgetService.listAll: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.listAll not implemented');
+  }
+  async create(_body: unknown): Promise<unknown> {
+    this.logger.warn('CostBudgetService.create: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.create not implemented');
+  }
+  async update(_id: string, _body: unknown): Promise<unknown> {
+    this.logger.warn('CostBudgetService.update: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.update not implemented');
+  }
+  async check(_body: unknown): Promise<unknown> {
+    this.logger.warn('CostBudgetService.check: SCAFFOLD-R4 only');
+    throw new Error('SCAFFOLD-R4 — CostBudgetService.check not implemented');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/cost-budget/types/budget.types.ts b/apps/claw-routing-service/src/modules/cost-budget/types/budget.types.ts
new file mode 100644
index 00000000..6c71032c
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/cost-budget/types/budget.types.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.4 (05-r4-cost-budget-intelligence)
+
+export type BudgetScope = 'USER' | 'ORG';
+export type BudgetStatus = 'OK' | 'WARN' | 'EXCEEDED';
+
+export type BudgetCheckInput = {
+  userId: string;
+  orgId?: string;
+  estimatedCostUsd: number;
+};
+
+export type BudgetCheckResult = {
+  status: BudgetStatus;
+  remainingUsd: number;
+  percentOfCap: number;
+  overrideAllowed: boolean;
+  blockingScope?: BudgetScope;
+};
diff --git a/apps/claw-routing-service/src/modules/intelligence/constants/intelligence.constants.ts b/apps/claw-routing-service/src/modules/intelligence/constants/intelligence.constants.ts
new file mode 100644
index 00000000..0848214d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/constants/intelligence.constants.ts
@@ -0,0 +1,10 @@
+// SCAFFOLD: stream R.8 (09-r8-advanced-intelligence)
+
+export const PROMPT_LENGTH_HEADROOM_FACTOR = 1.2;
+export const LATENCY_CIRCUIT_ROLLING_WINDOW_MS = 5 * 60 * 1000;
+export const LATENCY_CIRCUIT_MIN_SAMPLES = 20;
+export const MID_STREAM_KILL_GRACE_MS = 500;
+export const CONSENSUS_MIN_MODELS = 3;
+export const CONSENSUS_AGREEMENT_THRESHOLD = 0.7;
+export const COST_QUALITY_SLIDER_DEFAULT = 0.5;
+export const FINE_TUNE_BIAS_WEIGHT_MAX = 0.3;
diff --git a/apps/claw-routing-service/src/modules/intelligence/intelligence.module.ts b/apps/claw-routing-service/src/modules/intelligence/intelligence.module.ts
new file mode 100644
index 00000000..87218660
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/intelligence.module.ts
@@ -0,0 +1,40 @@
+// SCAFFOLD: stream R.8 (09-r8-advanced-intelligence)
+// Umbrella module for 9 independent sub-features. NOT yet registered.
+
+import { Module } from '@nestjs/common';
+
+import { ConsensusModeManager } from './managers/consensus-mode.manager';
+import { CostQualitySliderManager } from './managers/cost-quality-slider.manager';
+import { EmbeddingRouterManager } from './managers/embedding-router.manager';
+import { FineTunePreferenceManager } from './managers/fine-tune-preference.manager';
+import { LatencyCircuitBreakerManager } from './managers/latency-circuit-breaker.manager';
+import { MidStreamSwitcherManager } from './managers/mid-stream-switcher.manager';
+import { MultiIntentSplitterManager } from './managers/multi-intent-splitter.manager';
+import { PromptLengthGuardManager } from './managers/prompt-length-guard.manager';
+import { RegionRouterManager } from './managers/region-router.manager';
+
+@Module({
+  providers: [
+    PromptLengthGuardManager,
+    LatencyCircuitBreakerManager,
+    MidStreamSwitcherManager,
+    FineTunePreferenceManager,
+    RegionRouterManager,
+    MultiIntentSplitterManager,
+    EmbeddingRouterManager,
+    ConsensusModeManager,
+    CostQualitySliderManager,
+  ],
+  exports: [
+    PromptLengthGuardManager,
+    LatencyCircuitBreakerManager,
+    MidStreamSwitcherManager,
+    FineTunePreferenceManager,
+    RegionRouterManager,
+    MultiIntentSplitterManager,
+    EmbeddingRouterManager,
+    ConsensusModeManager,
+    CostQualitySliderManager,
+  ],
+})
+export class IntelligenceModule {}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/consensus-mode.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/consensus-mode.manager.ts
new file mode 100644
index 00000000..669e2698
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/consensus-mode.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.8 (09-r8-advanced-intelligence) — fire 3 models, score agreement
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class ConsensusModeManager {
+  private readonly logger = new Logger(ConsensusModeManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('ConsensusModeManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.8 — ConsensusModeManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/cost-quality-slider.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/cost-quality-slider.manager.ts
new file mode 100644
index 00000000..0d148022
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/cost-quality-slider.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.9 (09-r8-advanced-intelligence) — apply user cost/quality slider to scoring weight
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class CostQualitySliderManager {
+  private readonly logger = new Logger(CostQualitySliderManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('CostQualitySliderManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.9 — CostQualitySliderManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/embedding-router.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/embedding-router.manager.ts
new file mode 100644
index 00000000..c9934c3d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/embedding-router.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.7 (09-r8-advanced-intelligence) — pick embedding-specific model
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class EmbeddingRouterManager {
+  private readonly logger = new Logger(EmbeddingRouterManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('EmbeddingRouterManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.7 — EmbeddingRouterManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/fine-tune-preference.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/fine-tune-preference.manager.ts
new file mode 100644
index 00000000..136c49a6
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/fine-tune-preference.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.4 (09-r8-advanced-intelligence) — boost user fine-tunes for their domain
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class FineTunePreferenceManager {
+  private readonly logger = new Logger(FineTunePreferenceManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('FineTunePreferenceManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.4 — FineTunePreferenceManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/latency-circuit-breaker.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/latency-circuit-breaker.manager.ts
new file mode 100644
index 00000000..e663dbc8
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/latency-circuit-breaker.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.2 (09-r8-advanced-intelligence) — open circuit when p95 latency exceeds threshold
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class LatencyCircuitBreakerManager {
+  private readonly logger = new Logger(LatencyCircuitBreakerManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('LatencyCircuitBreakerManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.2 — LatencyCircuitBreakerManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/mid-stream-switcher.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/mid-stream-switcher.manager.ts
new file mode 100644
index 00000000..e56691ef
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/mid-stream-switcher.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.3 (09-r8-advanced-intelligence) — kill slow first-chunk + reroute mid-stream
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class MidStreamSwitcherManager {
+  private readonly logger = new Logger(MidStreamSwitcherManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('MidStreamSwitcherManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.3 — MidStreamSwitcherManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/multi-intent-splitter.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/multi-intent-splitter.manager.ts
new file mode 100644
index 00000000..64fd4bce
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/multi-intent-splitter.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.6 (09-r8-advanced-intelligence) — split N intents into N parallel calls
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class MultiIntentSplitterManager {
+  private readonly logger = new Logger(MultiIntentSplitterManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('MultiIntentSplitterManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.6 — MultiIntentSplitterManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/prompt-length-guard.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/prompt-length-guard.manager.ts
new file mode 100644
index 00000000..dc055a2d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/prompt-length-guard.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.1 (09-r8-advanced-intelligence) — filter candidates with insufficient context window
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class PromptLengthGuardManager {
+  private readonly logger = new Logger(PromptLengthGuardManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('PromptLengthGuardManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.1 — PromptLengthGuardManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/managers/region-router.manager.ts b/apps/claw-routing-service/src/modules/intelligence/managers/region-router.manager.ts
new file mode 100644
index 00000000..c398ab8a
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/managers/region-router.manager.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.8.9.5 (09-r8-advanced-intelligence) — pick regional endpoint
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class RegionRouterManager {
+  private readonly logger = new Logger(RegionRouterManager.name);
+
+  apply(_input: unknown): unknown {
+    this.logger.warn('RegionRouterManager.apply: SCAFFOLD only');
+    throw new Error('SCAFFOLD-R8.9.5 — RegionRouterManager.apply not implemented; see docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md');
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/intelligence/types/intelligence.types.ts b/apps/claw-routing-service/src/modules/intelligence/types/intelligence.types.ts
new file mode 100644
index 00000000..243695e3
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/intelligence/types/intelligence.types.ts
@@ -0,0 +1,23 @@
+// SCAFFOLD: stream R.8 (09-r8-advanced-intelligence)
+
+export type CostQualityWeight = {
+  qualityWeight: number;
+  costWeight: number;
+};
+
+export type CircuitBreakerTrigger = 'FAILURE_RATE' | 'LATENCY_P95';
+
+export type RegionPreferenceScope = 'USER' | 'ORG' | 'GLOBAL';
+
+export type MultiIntent = {
+  intent: string;
+  confidence: number;
+  routedProvider?: string;
+  routedModel?: string;
+};
+
+export type ConsensusResult = {
+  models: { provider: string; model: string; answer: string; confidence: number }[];
+  agreementScore: number;
+  winnerIndex: number;
+};
diff --git a/apps/claw-routing-service/src/modules/language-detection/constants/language-codes.constants.ts b/apps/claw-routing-service/src/modules/language-detection/constants/language-codes.constants.ts
new file mode 100644
index 00000000..c4e0fc3d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/constants/language-codes.constants.ts
@@ -0,0 +1,8 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+export const SUPPORTED_LOCALES = ['en', 'ar', 'de', 'es', 'fr', 'it', 'pt', 'ru'] as const;
+export const SUPPORTED_LOCALES_SET = new Set<string>(SUPPORTED_LOCALES);
+
+export const LANGUAGE_CONFIDENCE_THRESHOLD = 0.6;
+export const LANGUAGE_FALLBACK = 'en';
+export const LANGUAGE_DETECTION_MIN_CHARS = 3;
diff --git a/apps/claw-routing-service/src/modules/language-detection/constants/language-rtl.constants.ts b/apps/claw-routing-service/src/modules/language-detection/constants/language-rtl.constants.ts
new file mode 100644
index 00000000..1dd590df
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/constants/language-rtl.constants.ts
@@ -0,0 +1,7 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+export const RTL_LANGUAGES = new Set<string>(['ar', 'he', 'fa', 'ur', 'ckb', 'sd', 'yi']);
+
+export function isRtlLanguage(lang: string): boolean {
+  return RTL_LANGUAGES.has(lang.toLowerCase());
+}
diff --git a/apps/claw-routing-service/src/modules/language-detection/controllers/language-detection.controller.ts b/apps/claw-routing-service/src/modules/language-detection/controllers/language-detection.controller.ts
new file mode 100644
index 00000000..f30a9bdd
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/controllers/language-detection.controller.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+import { Body, Controller, Post } from '@nestjs/common';
+
+import { LanguageDetectionService } from '../services/language-detection.service';
+
+@Controller('routing/detect-language')
+export class LanguageDetectionController {
+  constructor(private readonly service: LanguageDetectionService) {}
+
+  @Post()
+  async detect(@Body() body: unknown): Promise<unknown> {
+    return this.service.detect(body);
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/language-detection/dto/detect-language.dto.ts b/apps/claw-routing-service/src/modules/language-detection/dto/detect-language.dto.ts
new file mode 100644
index 00000000..23ed3018
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/dto/detect-language.dto.ts
@@ -0,0 +1,9 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+import { z } from 'zod';
+
+export const detectLanguageSchema = z.object({
+  message: z.string().min(0).max(200_000),
+});
+
+export type DetectLanguageDto = z.infer<typeof detectLanguageSchema>;
diff --git a/apps/claw-routing-service/src/modules/language-detection/language-detection.module.ts b/apps/claw-routing-service/src/modules/language-detection/language-detection.module.ts
new file mode 100644
index 00000000..817f9d15
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/language-detection.module.ts
@@ -0,0 +1,16 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+// NEW module — NOT yet registered.
+
+import { Module } from '@nestjs/common';
+
+import { LanguageDetectionController } from './controllers/language-detection.controller';
+import { CodeMixedDetectorManager } from './managers/code-mixed-detector.manager';
+import { LanguageClassifierManager } from './managers/language-classifier.manager';
+import { LanguageDetectionService } from './services/language-detection.service';
+
+@Module({
+  controllers: [LanguageDetectionController],
+  providers: [LanguageDetectionService, LanguageClassifierManager, CodeMixedDetectorManager],
+  exports: [LanguageDetectionService],
+})
+export class LanguageDetectionModule {}
diff --git a/apps/claw-routing-service/src/modules/language-detection/managers/code-mixed-detector.manager.ts b/apps/claw-routing-service/src/modules/language-detection/managers/code-mixed-detector.manager.ts
new file mode 100644
index 00000000..9684b1e2
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/managers/code-mixed-detector.manager.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class CodeMixedDetectorManager {
+  private readonly logger = new Logger(CodeMixedDetectorManager.name);
+
+  detect(_message: string, _primaryLang: string): { isCodeMixed: boolean; secondaryLang?: string } {
+    this.logger.warn('CodeMixedDetectorManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R7 — CodeMixedDetectorManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/language-detection/managers/language-classifier.manager.ts b/apps/claw-routing-service/src/modules/language-detection/managers/language-classifier.manager.ts
new file mode 100644
index 00000000..fda68d23
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/managers/language-classifier.manager.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { LanguageDetectionResult } from '../types/language-detection.types';
+
+@Injectable()
+export class LanguageClassifierManager {
+  private readonly logger = new Logger(LanguageClassifierManager.name);
+
+  classify(_message: string): LanguageDetectionResult {
+    this.logger.warn('LanguageClassifierManager.classify: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R7 — LanguageClassifierManager.classify not implemented; see docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/language-detection/services/language-detection.service.ts b/apps/claw-routing-service/src/modules/language-detection/services/language-detection.service.ts
new file mode 100644
index 00000000..6422e1f9
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/services/language-detection.service.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class LanguageDetectionService {
+  private readonly logger = new Logger(LanguageDetectionService.name);
+
+  async detect(_body: unknown): Promise<unknown> {
+    this.logger.warn('LanguageDetectionService.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R7 — LanguageDetectionService.detect not implemented; see docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/language-detection/types/language-detection.types.ts b/apps/claw-routing-service/src/modules/language-detection/types/language-detection.types.ts
new file mode 100644
index 00000000..36e7b84e
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/types/language-detection.types.ts
@@ -0,0 +1,12 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+
+export type LanguageDetectionResult = {
+  detectedLanguage: string;       // ISO-639-1, e.g. 'en', 'ar', 'es'
+  languageConfidence: number;     // 0..1
+  isCodeMixed: boolean;
+  secondaryLanguage?: string;
+  secondaryConfidence?: number;
+  isRtl: boolean;
+};
+
+export type LanguageStrength = Record<string, number>;  // { en: 1.0, ar: 0.92, ... }
diff --git a/apps/claw-routing-service/src/modules/language-detection/utilities/language-strength-resolver.utility.ts b/apps/claw-routing-service/src/modules/language-detection/utilities/language-strength-resolver.utility.ts
new file mode 100644
index 00000000..f923254e
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/language-detection/utilities/language-strength-resolver.utility.ts
@@ -0,0 +1,14 @@
+// SCAFFOLD: stream R.7 (08-r7-i18n-non-english)
+// Reads RouterModelRegistry.languageStrengthJson and applies a score boost
+// to candidates with high strength in the detected language.
+
+import type { LanguageStrength } from '../types/language-detection.types';
+
+export function resolveLanguageStrength(
+  _registry: { languageStrengthJson?: LanguageStrength | null },
+  _detectedLanguage: string,
+): number {
+  throw new Error(
+    'SCAFFOLD-R7 — resolveLanguageStrength not implemented; see docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md',
+  );
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/audio-mime.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/audio-mime.constants.ts
new file mode 100644
index 00000000..56bd354d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/audio-mime.constants.ts
@@ -0,0 +1,14 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const AUDIO_MIME_TYPES = [
+  'audio/mpeg',
+  'audio/mp4',
+  'audio/x-m4a',
+  'audio/wav',
+  'audio/x-wav',
+  'audio/ogg',
+  'audio/flac',
+  'audio/webm',
+];
+
+export const AUDIO_EXTENSIONS = ['.mp3', '.m4a', '.wav', '.ogg', '.flac', '.opus'];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/embedding-keywords.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/embedding-keywords.constants.ts
new file mode 100644
index 00000000..29f7104b
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/embedding-keywords.constants.ts
@@ -0,0 +1,14 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const EMBEDDING_VERBS = [
+  'embed',
+  'vectorize',
+  'create embedding',
+  'generate embedding',
+  'store in vector db',
+  'retrieve similar',
+  'find similar',
+  'semantic search',
+  'nearest neighbor',
+  'index this',
+];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/image-mime.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/image-mime.constants.ts
new file mode 100644
index 00000000..3e87c472
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/image-mime.constants.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const IMAGE_MIME_TYPES = [
+  'image/png',
+  'image/jpeg',
+  'image/gif',
+  'image/webp',
+  'image/bmp',
+  'image/tiff',
+  'image/svg+xml',
+];
+
+export const IMAGE_EXTENSIONS = ['.png', '.jpg', '.jpeg', '.gif', '.webp', '.bmp', '.tif', '.tiff', '.svg'];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/pdf-mime.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/pdf-mime.constants.ts
new file mode 100644
index 00000000..56aeb5d9
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/pdf-mime.constants.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const PDF_MIME_TYPES = ['application/pdf'];
+export const PDF_EXTENSIONS = ['.pdf'];
+export const PDF_SUMMARIZE_VERBS = [
+  'summarize',
+  'summary',
+  'explain',
+  'extract',
+  'q&a',
+  'tl;dr',
+  'tldr',
+  'key points',
+  'main ideas',
+  'what does this pdf',
+  'read this pdf',
+];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/spreadsheet-mime.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/spreadsheet-mime.constants.ts
new file mode 100644
index 00000000..1cb29a0b
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/spreadsheet-mime.constants.ts
@@ -0,0 +1,10 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const SPREADSHEET_MIME_TYPES = [
+  'application/vnd.ms-excel',
+  'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet',
+  'application/vnd.oasis.opendocument.spreadsheet',
+  'text/csv',
+];
+
+export const SPREADSHEET_EXTENSIONS = ['.xlsx', '.xls', '.csv', '.ods'];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/tool-calling-keywords.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/tool-calling-keywords.constants.ts
new file mode 100644
index 00000000..157360ba
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/tool-calling-keywords.constants.ts
@@ -0,0 +1,20 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const TOOL_CALLING_VERBS = [
+  'book',
+  'schedule',
+  'send email',
+  'send a message',
+  'call function',
+  'invoke',
+  'run command',
+  'execute',
+  'create ticket',
+  'open pr',
+  'open pull request',
+  'create issue',
+  'post to slack',
+  'add to calendar',
+  'set reminder',
+  'create meeting',
+];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/video-mime.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/video-mime.constants.ts
new file mode 100644
index 00000000..a28be5a8
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/video-mime.constants.ts
@@ -0,0 +1,13 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const VIDEO_MIME_TYPES = [
+  'video/mp4',
+  'video/quicktime',
+  'video/webm',
+  'video/x-msvideo',
+  'video/x-matroska',
+  'video/mpeg',
+  'video/ogg',
+];
+
+export const VIDEO_EXTENSIONS = ['.mp4', '.mov', '.webm', '.avi', '.mkv', '.mpeg', '.mpg', '.ogv'];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/web-url.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/web-url.constants.ts
new file mode 100644
index 00000000..7d3ea11c
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/web-url.constants.ts
@@ -0,0 +1,16 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const HTTP_URL_REGEX = /https?:\/\/[^\s<>"']+/giu;
+export const UNSAFE_URL_SCHEMES = ['javascript:', 'data:', 'file:', 'vbscript:'];
+
+export const WEB_SUMMARIZE_VERBS = [
+  'summarize',
+  'summary',
+  'read',
+  'fetch',
+  'scrape',
+  'what does',
+  'explain this article',
+  'tldr',
+  'tl;dr',
+];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/constants/youtube-url.constants.ts b/apps/claw-routing-service/src/modules/modality-detection/constants/youtube-url.constants.ts
new file mode 100644
index 00000000..5f2fff9f
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/constants/youtube-url.constants.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+export const YOUTUBE_HOST_REGEX = /^https?:\/\/(www\.)?(youtube\.com|youtu\.be|m\.youtube\.com)\//iu;
+export const YOUTUBE_WATCH_REGEX = /https?:\/\/(?:www\.|m\.)?youtube\.com\/watch\?v=[\w-]+/giu;
+export const YOUTUBE_SHORT_REGEX = /https?:\/\/youtu\.be\/[\w-]+/giu;
+export const YOUTUBE_SHORTS_REGEX = /https?:\/\/(?:www\.)?youtube\.com\/shorts\/[\w-]+/giu;
+export const YOUTUBE_PLAYLIST_REGEX = /https?:\/\/(?:www\.)?youtube\.com\/playlist\?list=[\w-]+/giu;
+
+export const YOUTUBE_SUMMARIZE_VERBS = [
+  'summarize',
+  'summary',
+  'tl;dr',
+  'transcript',
+  'what does this video say',
+  'explain this video',
+  'key points',
+  'highlights',
+];
diff --git a/apps/claw-routing-service/src/modules/modality-detection/controllers/modality-detection.controller.ts b/apps/claw-routing-service/src/modules/modality-detection/controllers/modality-detection.controller.ts
new file mode 100644
index 00000000..c7ca9a8f
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/controllers/modality-detection.controller.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Body, Controller, Post } from '@nestjs/common';
+
+import { detectModalitySchema, type DetectModalityDto } from '../dto/detect-modality.dto';
+import { ModalityDetectionService } from '../services/modality-detection.service';
+import type { ModalityDetectionResult } from '../types/modality-detection.types';
+
+@Controller('routing/detect-modality')
+export class ModalityDetectionController {
+  constructor(private readonly service: ModalityDetectionService) {}
+
+  @Post()
+  async detect(@Body() body: unknown): Promise<ModalityDetectionResult> {
+    const parsed: DetectModalityDto = detectModalitySchema.parse(body);
+    return this.service.detect(parsed);
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/dto/detect-modality.dto.ts b/apps/claw-routing-service/src/modules/modality-detection/dto/detect-modality.dto.ts
new file mode 100644
index 00000000..9149348b
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/dto/detect-modality.dto.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { z } from 'zod';
+
+const attachmentMetaSchema = z.object({
+  fileId: z.string().min(1).max(200),
+  filename: z.string().min(1).max(500),
+  mimeType: z.string().min(1).max(200),
+  sizeBytes: z.number().int().nonnegative().max(5_000_000_000),
+});
+
+export const detectModalitySchema = z.object({
+  message: z.string().min(0).max(200_000),
+  attachments: z.array(attachmentMetaSchema).max(50).default([]),
+  clientStreamingExpected: z.boolean().optional(),
+});
+
+export type DetectModalityDto = z.infer<typeof detectModalitySchema>;
diff --git a/apps/claw-routing-service/src/modules/modality-detection/managers/attachment-intent.manager.ts b/apps/claw-routing-service/src/modules/modality-detection/managers/attachment-intent.manager.ts
new file mode 100644
index 00000000..c7844841
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/managers/attachment-intent.manager.ts
@@ -0,0 +1,20 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type {
+  AttachmentIntentResult,
+  AttachmentMeta,
+} from '../types/modality-detection.types';
+
+@Injectable()
+export class AttachmentIntentManager {
+  private readonly logger = new Logger(AttachmentIntentManager.name);
+
+  detect(_message: string, _attachments: AttachmentMeta[]): AttachmentIntentResult {
+    this.logger.warn('AttachmentIntentManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — AttachmentIntentManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/managers/embedding-intent.manager.ts b/apps/claw-routing-service/src/modules/modality-detection/managers/embedding-intent.manager.ts
new file mode 100644
index 00000000..902ed55c
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/managers/embedding-intent.manager.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { EmbeddingIntentResult } from '../types/modality-detection.types';
+
+@Injectable()
+export class EmbeddingIntentManager {
+  private readonly logger = new Logger(EmbeddingIntentManager.name);
+
+  detect(_message: string): EmbeddingIntentResult {
+    this.logger.warn('EmbeddingIntentManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — EmbeddingIntentManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/managers/streaming-intent.manager.ts b/apps/claw-routing-service/src/modules/modality-detection/managers/streaming-intent.manager.ts
new file mode 100644
index 00000000..a2ecccf3
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/managers/streaming-intent.manager.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { StreamingIntentResult } from '../types/modality-detection.types';
+
+@Injectable()
+export class StreamingIntentManager {
+  private readonly logger = new Logger(StreamingIntentManager.name);
+
+  detect(_clientStreamingExpected: boolean | undefined): StreamingIntentResult {
+    this.logger.warn('StreamingIntentManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — StreamingIntentManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/managers/tool-calling-intent.manager.ts b/apps/claw-routing-service/src/modules/modality-detection/managers/tool-calling-intent.manager.ts
new file mode 100644
index 00000000..545bc5a9
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/managers/tool-calling-intent.manager.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { ToolCallingIntentResult } from '../types/modality-detection.types';
+
+@Injectable()
+export class ToolCallingIntentManager {
+  private readonly logger = new Logger(ToolCallingIntentManager.name);
+
+  detect(_message: string): ToolCallingIntentResult {
+    this.logger.warn('ToolCallingIntentManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — ToolCallingIntentManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/managers/url-intent.manager.ts b/apps/claw-routing-service/src/modules/modality-detection/managers/url-intent.manager.ts
new file mode 100644
index 00000000..ea4f2cb9
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/managers/url-intent.manager.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { UrlIntentResult } from '../types/modality-detection.types';
+
+@Injectable()
+export class UrlIntentManager {
+  private readonly logger = new Logger(UrlIntentManager.name);
+
+  detect(_message: string): UrlIntentResult {
+    this.logger.warn('UrlIntentManager.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — UrlIntentManager.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/modality-detection.module.ts b/apps/claw-routing-service/src/modules/modality-detection/modality-detection.module.ts
new file mode 100644
index 00000000..d674f068
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/modality-detection.module.ts
@@ -0,0 +1,26 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+// NEW module — NOT yet imported by app.module.ts. Register only after activation.
+
+import { Module } from '@nestjs/common';
+
+import { ModalityDetectionController } from './controllers/modality-detection.controller';
+import { AttachmentIntentManager } from './managers/attachment-intent.manager';
+import { EmbeddingIntentManager } from './managers/embedding-intent.manager';
+import { StreamingIntentManager } from './managers/streaming-intent.manager';
+import { ToolCallingIntentManager } from './managers/tool-calling-intent.manager';
+import { UrlIntentManager } from './managers/url-intent.manager';
+import { ModalityDetectionService } from './services/modality-detection.service';
+
+@Module({
+  controllers: [ModalityDetectionController],
+  providers: [
+    ModalityDetectionService,
+    UrlIntentManager,
+    AttachmentIntentManager,
+    ToolCallingIntentManager,
+    StreamingIntentManager,
+    EmbeddingIntentManager,
+  ],
+  exports: [ModalityDetectionService],
+})
+export class ModalityDetectionModule {}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/services/modality-detection.service.ts b/apps/claw-routing-service/src/modules/modality-detection/services/modality-detection.service.ts
new file mode 100644
index 00000000..3963f97e
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/services/modality-detection.service.ts
@@ -0,0 +1,39 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import { AttachmentIntentManager } from '../managers/attachment-intent.manager';
+import { EmbeddingIntentManager } from '../managers/embedding-intent.manager';
+import { StreamingIntentManager } from '../managers/streaming-intent.manager';
+import { ToolCallingIntentManager } from '../managers/tool-calling-intent.manager';
+import { UrlIntentManager } from '../managers/url-intent.manager';
+import type { DetectModalityDto } from '../dto/detect-modality.dto';
+import type { ModalityDetectionResult } from '../types/modality-detection.types';
+
+@Injectable()
+export class ModalityDetectionService {
+  private readonly logger = new Logger(ModalityDetectionService.name);
+
+  constructor(
+    private readonly urlIntent: UrlIntentManager,
+    private readonly attachmentIntent: AttachmentIntentManager,
+    private readonly toolCallingIntent: ToolCallingIntentManager,
+    private readonly streamingIntent: StreamingIntentManager,
+    private readonly embeddingIntent: EmbeddingIntentManager,
+  ) {}
+
+  async detect(_input: DetectModalityDto): Promise<ModalityDetectionResult> {
+    // Reference each injected manager so the unused-private-check passes
+    // until the real implementation wires them up.
+    void this.urlIntent;
+    void this.attachmentIntent;
+    void this.toolCallingIntent;
+    void this.streamingIntent;
+    void this.embeddingIntent;
+
+    this.logger.warn('ModalityDetectionService.detect: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R2 — ModalityDetectionService.detect not implemented; see docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/modality-detection/types/modality-detection.types.ts b/apps/claw-routing-service/src/modules/modality-detection/types/modality-detection.types.ts
new file mode 100644
index 00000000..79deac57
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/types/modality-detection.types.ts
@@ -0,0 +1,52 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import type { ModalityKind, WorkflowKind } from '../../../generated/prisma';
+
+export type AttachmentMeta = {
+  fileId: string;
+  filename: string;
+  mimeType: string;
+  sizeBytes: number;
+};
+
+export type DetectedUrl = {
+  url: string;
+  kind: 'youtube' | 'web';
+};
+
+export type ModalityDetectionResult = {
+  detectedModalities: ModalityKind[];
+  workflowHint?: WorkflowKind;
+  workflowConfidence: number;
+  reasonTags: string[];
+  fileMetadata: AttachmentMeta[];
+  urlMetadata: DetectedUrl[];
+  fallback?: { reason: string; tag: string };
+};
+
+// Per-manager partial result types — extracted to avoid inline string-literal
+// unions inside Pick<>, which the routing-service eslint config bans in
+// logic files.
+export type UrlIntentResult = Pick<
+  ModalityDetectionResult,
+  'detectedModalities' | 'urlMetadata' | 'workflowHint' | 'reasonTags'
+>;
+
+export type AttachmentIntentResult = Pick<
+  ModalityDetectionResult,
+  'detectedModalities' | 'fileMetadata' | 'workflowHint' | 'reasonTags'
+>;
+
+export type ToolCallingIntentResult = {
+  needsToolCalling: boolean;
+  reasonTag?: string;
+};
+
+export type StreamingIntentResult = {
+  needsStreaming: boolean;
+};
+
+export type EmbeddingIntentResult = {
+  isEmbeddingTask: boolean;
+  reasonTag?: string;
+};
diff --git a/apps/claw-routing-service/src/modules/modality-detection/utilities/extension-detector.utility.ts b/apps/claw-routing-service/src/modules/modality-detection/utilities/extension-detector.utility.ts
new file mode 100644
index 00000000..862dd9be
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/modality-detection/utilities/extension-detector.utility.ts
@@ -0,0 +1,22 @@
+// SCAFFOLD: stream R.2 (03-r2-multimodal-intent-detection)
+
+import { extname } from 'node:path';
+
+export function getExtension(filename: string): string {
+  return extname(filename).toLowerCase();
+}
+
+export function matchesAny(filename: string, extensions: readonly string[]): boolean {
+  const ext = getExtension(filename);
+  return extensions.includes(ext);
+}
+
+export function mimeOrExtensionMatches(
+  mimeType: string,
+  filename: string,
+  mimeTypes: readonly string[],
+  extensions: readonly string[],
+): boolean {
+  if (mimeTypes.includes(mimeType.toLowerCase())) return true;
+  return matchesAny(filename, extensions);
+}
diff --git a/apps/claw-routing-service/src/modules/playground/controllers/playground.controller.ts b/apps/claw-routing-service/src/modules/playground/controllers/playground.controller.ts
new file mode 100644
index 00000000..50a4dc9b
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/playground/controllers/playground.controller.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.5 (06-r5-operator-playground)
+
+import { Body, Controller, Post } from '@nestjs/common';
+
+import { PlaygroundService } from '../services/playground.service';
+
+@Controller('routing/playground')
+export class PlaygroundController {
+  constructor(private readonly service: PlaygroundService) {}
+
+  @Post('evaluate')
+  async evaluate(@Body() body: unknown): Promise<unknown> {
+    return this.service.evaluate(body);
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/playground/dto/playground-evaluate.dto.ts b/apps/claw-routing-service/src/modules/playground/dto/playground-evaluate.dto.ts
new file mode 100644
index 00000000..4911fc79
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/playground/dto/playground-evaluate.dto.ts
@@ -0,0 +1,25 @@
+// SCAFFOLD: stream R.5 (06-r5-operator-playground)
+
+import { z } from 'zod';
+
+export const playgroundEvaluateSchema = z.object({
+  message: z.string().min(1).max(200_000),
+  attachments: z
+    .array(
+      z.object({
+        fileId: z.string().min(1).max(200),
+        filename: z.string().min(1).max(500),
+        mimeType: z.string().min(1).max(200),
+        sizeBytes: z.number().int().nonnegative().max(5_000_000_000),
+      }),
+    )
+    .max(50)
+    .default([]),
+  userMode: z
+    .enum(['AUTO', 'MANUAL_MODEL', 'LOCAL_ONLY', 'PRIVACY_FIRST', 'LOW_LATENCY', 'HIGH_REASONING', 'COST_SAVER'])
+    .default('AUTO'),
+  compareWithV2: z.boolean().default(true),
+  compareWithOllamaRouter: z.boolean().default(true),
+});
+
+export type PlaygroundEvaluateDto = z.infer<typeof playgroundEvaluateSchema>;
diff --git a/apps/claw-routing-service/src/modules/playground/playground.module.ts b/apps/claw-routing-service/src/modules/playground/playground.module.ts
new file mode 100644
index 00000000..773a624a
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/playground/playground.module.ts
@@ -0,0 +1,14 @@
+// SCAFFOLD: stream R.5 (06-r5-operator-playground)
+// NEW module — NOT yet registered.
+
+import { Module } from '@nestjs/common';
+
+import { PlaygroundController } from './controllers/playground.controller';
+import { PlaygroundService } from './services/playground.service';
+
+@Module({
+  controllers: [PlaygroundController],
+  providers: [PlaygroundService],
+  exports: [PlaygroundService],
+})
+export class PlaygroundModule {}
diff --git a/apps/claw-routing-service/src/modules/playground/services/playground.service.ts b/apps/claw-routing-service/src/modules/playground/services/playground.service.ts
new file mode 100644
index 00000000..56c189b4
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/playground/services/playground.service.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.5 (06-r5-operator-playground)
+
+import { Injectable, Logger } from '@nestjs/common';
+
+@Injectable()
+export class PlaygroundService {
+  private readonly logger = new Logger(PlaygroundService.name);
+
+  async evaluate(_body: unknown): Promise<unknown> {
+    this.logger.warn('PlaygroundService.evaluate: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R5 — PlaygroundService.evaluate not implemented; see docs/15-ai-context/routing-flagship-streams/06-r5-operator-playground.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/playground/types/playground.types.ts b/apps/claw-routing-service/src/modules/playground/types/playground.types.ts
new file mode 100644
index 00000000..9665caa1
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/playground/types/playground.types.ts
@@ -0,0 +1,25 @@
+// SCAFFOLD: stream R.5 (06-r5-operator-playground)
+
+export type ScoreBreakdown = {
+  provider: string;
+  model: string;
+  scores: {
+    quality: number;
+    cost: number;
+    latency: number;
+    privacy: number;
+    familiarity: number;
+  };
+  totalScore: number;
+  wasChosen: boolean;
+};
+
+export type PlaygroundEvaluationResult = {
+  v1Decision: unknown;
+  v2Decision?: unknown;
+  ollamaRouterDecision?: unknown;
+  scoreBreakdown: ScoreBreakdown[];
+  candidateList: unknown[];
+  modalityResult?: unknown;
+  workflowChoice?: unknown;
+};
diff --git a/apps/claw-routing-service/src/modules/route-evaluator/constants/canary.constants.ts b/apps/claw-routing-service/src/modules/route-evaluator/constants/canary.constants.ts
new file mode 100644
index 00000000..f941c56d
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/route-evaluator/constants/canary.constants.ts
@@ -0,0 +1,16 @@
+// SCAFFOLD: stream R.1/R.3 (02-r1r3-v2-evaluator-canary)
+
+export const CANARY_HASH_SALT = 'routing-v2-canary';
+export const CANARY_ROLLING_WINDOW_REQUESTS = 100;
+export const CANARY_ROLLING_WINDOW_MS = 5 * 60 * 1000;
+export const CANARY_WARMUP_GRACE_REQUESTS = 50;
+
+export const CANARY_FALLBACK_REASON_INVALID_SCHEMA = 'invalid_schema';
+export const CANARY_FALLBACK_REASON_SAFETY = 'safety';
+export const CANARY_FALLBACK_REASON_GUARDRAIL_BREACH = 'guardrail_breach';
+export const CANARY_FALLBACK_REASON_V2_THREW = 'v2_threw';
+
+export const CANARY_GUARDRAIL_METRIC_REGRESSION = 'regression';
+export const CANARY_GUARDRAIL_METRIC_COST_INCREASE = 'cost_increase';
+export const CANARY_GUARDRAIL_METRIC_CONFIDENCE_DROP = 'confidence_drop';
+export const CANARY_GUARDRAIL_METRIC_FAILURE_RATE = 'failure_rate';
diff --git a/apps/claw-routing-service/src/modules/route-evaluator/managers/canary-bucket.manager.ts b/apps/claw-routing-service/src/modules/route-evaluator/managers/canary-bucket.manager.ts
new file mode 100644
index 00000000..f95aa516
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/route-evaluator/managers/canary-bucket.manager.ts
@@ -0,0 +1,20 @@
+// SCAFFOLD: stream R.1/R.3 (02-r1r3-v2-evaluator-canary) — replace with real implementation before activation.
+// See docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md
+//
+// NOT registered in route-evaluator.module.ts — discoverable but inert.
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { CanaryBucketInput, CanaryBucketResult } from '../types/canary.types';
+
+@Injectable()
+export class CanaryBucketManager {
+  private readonly logger = new Logger(CanaryBucketManager.name);
+
+  isV2Bucket(_input: CanaryBucketInput): CanaryBucketResult {
+    this.logger.warn('CanaryBucketManager.isV2Bucket: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R1/R3 — CanaryBucketManager.isV2Bucket not implemented; see docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/route-evaluator/types/canary.types.ts b/apps/claw-routing-service/src/modules/route-evaluator/types/canary.types.ts
new file mode 100644
index 00000000..83fe1ea8
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/route-evaluator/types/canary.types.ts
@@ -0,0 +1,25 @@
+// SCAFFOLD: stream R.1/R.3 (02-r1r3-v2-evaluator-canary)
+
+export type CanaryBucketInput = {
+  userId: string;
+  orgId?: string;
+  canaryPercent: number;
+  allowlist?: string[];
+  denylist?: string[];
+};
+
+export type CanaryBucketResult = {
+  isV2: boolean;
+  bucketReason: 'percent' | 'allowlist' | 'denylist' | 'rollback_switch_active' | 'flag_off';
+};
+
+export type DecisionComparisonInput = {
+  v1Decision: { provider: string; model: string; confidence: number; costClass: string };
+  v2Decision: { provider: string; model: string; confidence: number; costClass: string };
+};
+
+export type DecisionComparisonResult = {
+  sameDecision: boolean;
+  confidenceDelta: number;
+  costClassChange: 'cheaper' | 'same' | 'more_expensive';
+};
diff --git a/apps/claw-routing-service/src/modules/route-evaluator/utilities/canary-hash.utility.ts b/apps/claw-routing-service/src/modules/route-evaluator/utilities/canary-hash.utility.ts
new file mode 100644
index 00000000..d5c48203
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/route-evaluator/utilities/canary-hash.utility.ts
@@ -0,0 +1,22 @@
+// SCAFFOLD: stream R.1/R.3 (02-r1r3-v2-evaluator-canary)
+// Pure stable bucketing — same input always yields same bucket.
+
+import { createHash } from 'node:crypto';
+
+import { CANARY_HASH_SALT } from '../constants/canary.constants';
+
+export function hashToBucket(userId: string, orgId: string | undefined): number {
+  const input = `${userId}|${orgId ?? ''}|${CANARY_HASH_SALT}`;
+  const hex = createHash('sha256').update(input).digest('hex').slice(0, 8);
+  return Number.parseInt(hex, 16) % 100;
+}
+
+export function isInCanaryBucket(
+  userId: string,
+  orgId: string | undefined,
+  canaryPercent: number,
+): boolean {
+  if (canaryPercent <= 0) return false;
+  if (canaryPercent >= 100) return true;
+  return hashToBucket(userId, orgId) < canaryPercent;
+}
diff --git a/apps/claw-routing-service/src/modules/route-evaluator/utilities/decision-comparator.utility.ts b/apps/claw-routing-service/src/modules/route-evaluator/utilities/decision-comparator.utility.ts
new file mode 100644
index 00000000..a6b55ff3
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/route-evaluator/utilities/decision-comparator.utility.ts
@@ -0,0 +1,9 @@
+// SCAFFOLD: stream R.1/R.3 (02-r1r3-v2-evaluator-canary)
+
+import type { DecisionComparisonInput, DecisionComparisonResult } from '../types/canary.types';
+
+export function compareDecisions(_input: DecisionComparisonInput): DecisionComparisonResult {
+  throw new Error(
+    'SCAFFOLD-R1/R3 — compareDecisions not implemented; see docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md',
+  );
+}
diff --git a/apps/claw-routing-service/src/modules/routing/constants/learned-bias.constants.ts b/apps/claw-routing-service/src/modules/routing/constants/learned-bias.constants.ts
new file mode 100644
index 00000000..8e3f4d80
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/constants/learned-bias.constants.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop)
+
+export const LEARNED_BIAS_REASON_THUMBS_UP_HISTORY = 'learned_bias_thumbs_up_history';
+export const LEARNED_BIAS_REASON_THUMBS_DOWN_HISTORY = 'learned_bias_thumbs_down_history';
+export const LEARNED_BIAS_REASON_TOPIC_PROFILE = 'learned_bias_topic_profile';
+export const LEARNED_BIAS_REASON_SAMPLE_TOO_SMALL = 'learned_bias_sample_too_small';
+export const LEARNED_BIAS_REASON_NO_DATA = 'learned_bias_no_data';
+export const LEARNED_BIAS_REASON_PRIVACY_OVERRIDE = 'learned_bias_privacy_override_beat_bias';
+export const LEARNED_BIAS_REASON_CAPABILITY_OVERRIDE = 'learned_bias_capability_override_beat_bias';
+export const LEARNED_BIAS_REASON_MANUAL_OVERRIDE = 'learned_bias_manual_user_model_beat_bias';
+export const LEARNED_BIAS_REASON_THREAD_STICKY = 'learned_bias_thread_sticky';
+export const LEARNED_BIAS_REASON_THREAD_STICKY_UNHEALTHY = 'learned_bias_sticky_unhealthy_fallthrough';
+
+export const LEARNED_BIAS_CACHE_TTL_MS = 5 * 60 * 1000;
+export const LEARNED_BIAS_FETCH_LIMIT = 50;
diff --git a/apps/claw-routing-service/src/modules/routing/managers/learned-bias.manager.ts b/apps/claw-routing-service/src/modules/routing/managers/learned-bias.manager.ts
new file mode 100644
index 00000000..c01e6a10
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/managers/learned-bias.manager.ts
@@ -0,0 +1,20 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop) — replace this stub with real implementation before activation.
+// See docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md
+//
+// NOT registered in routing.module.ts — discoverable but inert.
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { BiasInput, BiasOutput } from '../types/learned-bias.types';
+
+@Injectable()
+export class LearnedBiasManager {
+  private readonly logger = new Logger(LearnedBiasManager.name);
+
+  async applyBias(_input: BiasInput): Promise<BiasOutput> {
+    this.logger.warn('LearnedBiasManager.applyBias: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R1 — LearnedBiasManager.applyBias not implemented; see docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/routing/types/calibration.types.ts b/apps/claw-routing-service/src/modules/routing/types/calibration.types.ts
new file mode 100644
index 00000000..1411a8f8
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/types/calibration.types.ts
@@ -0,0 +1,14 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop)
+
+export type CalibrationSignal =
+  | 'EXACT_KEYWORD'
+  | 'VERB_NOUN_COMBO'
+  | 'CATEGORY_KEYWORD'
+  | 'HEURISTIC_FALLBACK'
+  | 'PRIVACY_ENFORCED';
+
+export type CalibrationSample = {
+  signal: CalibrationSignal;
+  wasCorrect: boolean;
+  occurredAt: Date;
+};
diff --git a/apps/claw-routing-service/src/modules/routing/types/learned-bias.types.ts b/apps/claw-routing-service/src/modules/routing/types/learned-bias.types.ts
new file mode 100644
index 00000000..20fd02ca
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/types/learned-bias.types.ts
@@ -0,0 +1,40 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop)
+
+import type { DomainTag, PrivacyClass } from '../../../generated/prisma';
+
+export type CandidateModel = {
+  provider: string;
+  model: string;
+  score: number;
+};
+
+export type BiasInput = {
+  userId: string;
+  domain: DomainTag;
+  taskFamily: string;
+  candidates: CandidateModel[];
+  privacyConstraint: PrivacyClass;
+  threadId?: string;
+};
+
+export type AppliedBiasMetadata = {
+  learnedScoreRowIds: string[];
+  topicProfileRowId?: string;
+  sampleSize: number;
+  biasWeight: number;
+  reasonTag: string;
+};
+
+export type BiasOutput = {
+  candidates: CandidateModel[];
+  appliedBias: AppliedBiasMetadata;
+};
+
+// Per-utility input shape (moved out of learned-bias-applier.utility.ts
+// to satisfy the no-inline-type-alias rule for utility files).
+export type LearnedSignal = {
+  provider: string;
+  model: string;
+  scoreDelta: number;
+  sampleSize: number;
+};
diff --git a/apps/claw-routing-service/src/modules/routing/utilities/confidence-calibrator.utility.ts b/apps/claw-routing-service/src/modules/routing/utilities/confidence-calibrator.utility.ts
new file mode 100644
index 00000000..d10c2ed4
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/utilities/confidence-calibrator.utility.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop)
+// Rolling 30-day hit-rate calibration for routing-signal confidence constants.
+
+import type { CalibrationSample, CalibrationSignal } from '../types/calibration.types';
+
+export function calibrateConfidence(
+  _signal: CalibrationSignal,
+  _samples: CalibrationSample[],
+  _windowDays: number,
+  _fallbackConstant: number,
+): number {
+  throw new Error(
+    'SCAFFOLD-R1 — calibrateConfidence not implemented; see docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md',
+  );
+}
diff --git a/apps/claw-routing-service/src/modules/routing/utilities/learned-bias-applier.utility.ts b/apps/claw-routing-service/src/modules/routing/utilities/learned-bias-applier.utility.ts
new file mode 100644
index 00000000..d07ee3b9
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/routing/utilities/learned-bias-applier.utility.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.1 (01-r1-learning-loop)
+// Pure utility — extracts the bias math out of the manager for unit testability.
+
+import type { CandidateModel, LearnedSignal } from '../types/learned-bias.types';
+
+export function applyLearnedBias(
+  _candidates: CandidateModel[],
+  _signals: LearnedSignal[],
+  _weightMax: number,
+  _minSampleSize: number,
+): CandidateModel[] {
+  throw new Error(
+    'SCAFFOLD-R1 — applyLearnedBias not implemented; see docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md',
+  );
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/constants/workflow-priority.constants.ts b/apps/claw-routing-service/src/modules/workflows/constants/workflow-priority.constants.ts
new file mode 100644
index 00000000..802e55e4
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/constants/workflow-priority.constants.ts
@@ -0,0 +1,25 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator)
+// When multiple workflow handlers match, this ordering decides the winner.
+
+import { WorkflowKind } from '../../../generated/prisma';
+
+export const WORKFLOW_PRIORITY: WorkflowKind[] = [
+  WorkflowKind.JUDGE_PIPELINE,
+  WorkflowKind.COMPARE_ENSEMBLE,
+  WorkflowKind.PDF_EXTRACTION,
+  WorkflowKind.YOUTUBE_TRANSCRIPT,
+  WorkflowKind.VIDEO_ANALYSIS,
+  WorkflowKind.AUDIO_TRANSCRIBE,
+  WorkflowKind.IMAGE_ANALYSIS,
+  WorkflowKind.CODE_REVIEW,
+  WorkflowKind.SEARCH_FIRST,
+  WorkflowKind.EXTRACT_FIRST,
+  WorkflowKind.FILE_GENERATION,
+  WorkflowKind.IMAGE_GENERATION,
+  WorkflowKind.DIRECT_LLM,
+];
+
+export const WORKFLOW_FALLBACK_REASON_DISABLED = 'workflow_disabled_fallthrough';
+export const WORKFLOW_FALLBACK_REASON_DEP_UNHEALTHY = 'workflow_dependency_unhealthy';
+export const WORKFLOW_FALLBACK_REASON_EXTRACTION_FAILED = 'workflow_extraction_failed';
+export const WORKFLOW_FALLBACK_REASON_NO_MATCH = 'workflow_no_match_using_direct_llm';
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/audio-transcribe.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/audio-transcribe.handler.ts
new file mode 100644
index 00000000..41c120a2
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/audio-transcribe.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class AudioTranscribeHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.AUDIO_TRANSCRIBE;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — AudioTranscribeHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — AudioTranscribeHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/code-review.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/code-review.handler.ts
new file mode 100644
index 00000000..b8a33e20
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/code-review.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class CodeReviewHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.CODE_REVIEW;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — CodeReviewHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — CodeReviewHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/compare-ensemble.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/compare-ensemble.handler.ts
new file mode 100644
index 00000000..42f94e5f
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/compare-ensemble.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class CompareEnsembleHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.COMPARE_ENSEMBLE;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — CompareEnsembleHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — CompareEnsembleHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/direct-llm.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/direct-llm.handler.ts
new file mode 100644
index 00000000..49266858
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/direct-llm.handler.ts
@@ -0,0 +1,15 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class DirectLlmHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.DIRECT_LLM;
+  canHandle(_c: WorkflowExecutionContext): boolean { return true; }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — DirectLlmHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0.5; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/extract-first.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/extract-first.handler.ts
new file mode 100644
index 00000000..35747a57
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/extract-first.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class ExtractFirstHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.EXTRACT_FIRST;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — ExtractFirstHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — ExtractFirstHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/file-generation.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/file-generation.handler.ts
new file mode 100644
index 00000000..dca38128
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/file-generation.handler.ts
@@ -0,0 +1,19 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+// NOTE: File generation is wired today in routing.manager.ts (Stage 3 of AUTO pipeline).
+// This handler will replace that direct call when the workflow orchestrator goes live.
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class FileGenerationHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.FILE_GENERATION;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — FileGenerationHandler.canHandle not implemented; delegates to existing file-gen detection');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — FileGenerationHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-analysis.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-analysis.handler.ts
new file mode 100644
index 00000000..2129d5d8
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-analysis.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class ImageAnalysisHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.IMAGE_ANALYSIS;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — ImageAnalysisHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — ImageAnalysisHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-generation.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-generation.handler.ts
new file mode 100644
index 00000000..027f07f6
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/image-generation.handler.ts
@@ -0,0 +1,19 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+// NOTE: Image generation is wired today in routing.manager.ts (Stage 2 of AUTO pipeline).
+// This handler will replace that direct call when the workflow orchestrator goes live.
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class ImageGenerationHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.IMAGE_GENERATION;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — ImageGenerationHandler.canHandle not implemented; delegates to existing ImageDetectionManager');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — ImageGenerationHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/judge-pipeline.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/judge-pipeline.handler.ts
new file mode 100644
index 00000000..b7e2a130
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/judge-pipeline.handler.ts
@@ -0,0 +1,18 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+// HIGHEST priority — used for medical/legal/high-risk domains.
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class JudgePipelineHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.JUDGE_PIPELINE;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — JudgePipelineHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — JudgePipelineHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/pdf-extraction.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/pdf-extraction.handler.ts
new file mode 100644
index 00000000..6f718eab
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/pdf-extraction.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class PdfExtractionHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.PDF_EXTRACTION;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — PdfExtractionHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — PdfExtractionHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/search-first.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/search-first.handler.ts
new file mode 100644
index 00000000..81d49dd6
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/search-first.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class SearchFirstHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.SEARCH_FIRST;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — SearchFirstHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — SearchFirstHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/video-analysis.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/video-analysis.handler.ts
new file mode 100644
index 00000000..5ebd0f08
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/video-analysis.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class VideoAnalysisHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.VIDEO_ANALYSIS;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — VideoAnalysisHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — VideoAnalysisHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/handlers/youtube-transcript.handler.ts b/apps/claw-routing-service/src/modules/workflows/managers/handlers/youtube-transcript.handler.ts
new file mode 100644
index 00000000..5c1cebd1
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/handlers/youtube-transcript.handler.ts
@@ -0,0 +1,17 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator) — handler stub
+import { Injectable } from '@nestjs/common';
+import { WorkflowKind } from '../../../../generated/prisma';
+import type { IWorkflowHandler } from '../../types/workflow-handler.interface';
+import type { WorkflowExecutionContext, WorkflowPlan } from '../../types/workflow-execution.types';
+
+@Injectable()
+export class YoutubeTranscriptHandler implements IWorkflowHandler {
+  readonly kind = WorkflowKind.YOUTUBE_TRANSCRIPT;
+  canHandle(_c: WorkflowExecutionContext): boolean {
+    throw new Error('SCAFFOLD-R3 — YoutubeTranscriptHandler.canHandle not implemented');
+  }
+  plan(_c: WorkflowExecutionContext): WorkflowPlan {
+    throw new Error('SCAFFOLD-R3 — YoutubeTranscriptHandler.plan not implemented');
+  }
+  confidence(_c: WorkflowExecutionContext): number { return 0; }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/managers/workflow-orchestrator.manager.ts b/apps/claw-routing-service/src/modules/workflows/managers/workflow-orchestrator.manager.ts
new file mode 100644
index 00000000..23965eab
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/managers/workflow-orchestrator.manager.ts
@@ -0,0 +1,19 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator)
+// Top-level coordinator — picks the right workflow handler based on context.
+// NOT yet wired into RoutingManager.handleAuto().
+
+import { Injectable, Logger } from '@nestjs/common';
+
+import type { WorkflowChoice, WorkflowExecutionContext } from '../types/workflow-execution.types';
+
+@Injectable()
+export class WorkflowOrchestratorManager {
+  private readonly logger = new Logger(WorkflowOrchestratorManager.name);
+
+  pickWorkflow(_context: WorkflowExecutionContext): WorkflowChoice {
+    this.logger.warn('WorkflowOrchestratorManager.pickWorkflow: SCAFFOLD only');
+    throw new Error(
+      'SCAFFOLD-R3 — WorkflowOrchestratorManager.pickWorkflow not implemented; see docs/15-ai-context/routing-flagship-streams/04-r3-workflow-orchestrator.md',
+    );
+  }
+}
diff --git a/apps/claw-routing-service/src/modules/workflows/types/workflow-execution.types.ts b/apps/claw-routing-service/src/modules/workflows/types/workflow-execution.types.ts
new file mode 100644
index 00000000..19cfca37
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/types/workflow-execution.types.ts
@@ -0,0 +1,55 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator)
+
+import type { WorkflowKind } from '../../../generated/prisma';
+import type {
+  AttachmentMeta,
+  DetectedUrl,
+  ModalityDetectionResult,
+} from '../../modality-detection/types/modality-detection.types';
+
+export type WorkflowExecutionContext = {
+  message: string;
+  threadId?: string;
+  userId: string;
+  modalityResult: ModalityDetectionResult;
+  attachments: AttachmentMeta[];
+  urls: DetectedUrl[];
+};
+
+export type WorkflowStep =
+  | {
+      type: 'llm_call';
+      provider: string;
+      model: string;
+      promptTemplate: string;
+      inputFromStepId?: string;
+    }
+  | {
+      type: 'extract';
+      extractor:
+        | 'pdf'
+        | 'spreadsheet'
+        | 'youtube_transcript'
+        | 'web_scrape'
+        | 'audio_transcribe'
+        | 'video_extract';
+      sourceFileId?: string;
+      sourceUrl?: string;
+    }
+  | { type: 'search'; query: string }
+  | { type: 'judge'; primaryStepId: string; criticStepId: string };
+
+export type WorkflowPlan = {
+  kind: WorkflowKind;
+  steps: WorkflowStep[];
+  estimatedDurationMs: number;
+  estimatedCostUsd: number;
+};
+
+export type WorkflowChoice = {
+  kind: WorkflowKind;
+  confidence: number;
+  plan: WorkflowPlan;
+  reasonTags: string[];
+  fallbackUsed: boolean;
+};
diff --git a/apps/claw-routing-service/src/modules/workflows/types/workflow-handler.interface.ts b/apps/claw-routing-service/src/modules/workflows/types/workflow-handler.interface.ts
new file mode 100644
index 00000000..b24a1121
--- /dev/null
+++ b/apps/claw-routing-service/src/modules/workflows/types/workflow-handler.interface.ts
@@ -0,0 +1,11 @@
+// SCAFFOLD: stream R.3 (04-r3-workflow-orchestrator)
+
+import type { WorkflowKind } from '../../../generated/prisma';
+import type { WorkflowExecutionContext, WorkflowPlan } from './workflow-execution.types';
+
+export interface IWorkflowHandler {
+  readonly kind: WorkflowKind;
+  canHandle(context: WorkflowExecutionContext): boolean;
+  plan(context: WorkflowExecutionContext): WorkflowPlan;
+  confidence(context: WorkflowExecutionContext): number;
+}
diff --git a/docs/15-ai-context/routing-flagship-streams/00-master-plan.md b/docs/15-ai-context/routing-flagship-streams/00-master-plan.md
new file mode 100644
index 00000000..a27c3145
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/00-master-plan.md
@@ -0,0 +1,114 @@
+# 00 — Master Routing Flagship Plan
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/00_MASTER_routing_implementation_flagship.md`
+
+## Mission
+
+Turn the routing audit into an implementation-grade flagship plan that makes ClawAI's router smarter, multimodal, learning-aware, workflow-aware, budget-aware, transparent, and production-grade.
+
+## Implementation roadmap (release slices)
+
+| Release | Streams | Goal | Acceptance |
+|---------|---------|------|------------|
+| **R1** | 01, 11.1, 11.2 | Closing the learning loop + visibility | LearnedScore biases hot path; explanation surfaces in chat; workflow kind on every decision |
+| **R2** | 02, 06, 10 | Trust + safety | v2 canary 5% with rollback; playground UI; 500-prompt regression suite in CI |
+| **R3** | 03, 11.4, 11.5, 11.10 | Multimodal detection | YouTube/PDF/video/audio/spreadsheet/URL/lang detection; modality stored on decision |
+| **R4** | 04 | Workflows go live | JUDGE_PIPELINE, COMPARE_ENSEMBLE, SEARCH_FIRST, EXTRACT_FIRST, PDF_EXTRACTION, YOUTUBE_TRANSCRIPT, CODE_REVIEW all execute |
+| **R5** | 05, 07 | Cost + multi-tenant | Per-user + per-org budgets; org-scoped policies; allow/deny lists |
+| **R6** | 08, 09 | Intelligence + i18n | Language-aware routing; prompt-length filter; mid-stream switch; embedding routing; consensus; cost/quality slider |
+
+## Dependency graph
+
+```
+            ┌────────────────────────┐
+            │ 01 learning loop       │◄────────────────┐
+            └───────┬────────────────┘                 │
+                    │                                  │
+            ┌───────▼────────────────┐                 │
+            │ 02 v2 canary           │                 │
+            └───────┬────────────────┘                 │
+                    │                                  │
+            ┌───────▼────────────────┐    ┌────────────┴───┐
+            │ 04 workflows live      │◄──┤ 03 modality   │
+            └───────┬────────────────┘    └────────────────┘
+                    │
+        ┌───────────┼───────────┐
+        ▼           ▼           ▼
+   ┌────────┐ ┌─────────┐ ┌──────────┐
+   │ 05 cost │ │ 06 playg │ │ 07 fleet │
+   └────────┘ └─────────┘ └──────────┘
+                    │
+            ┌───────▼────────────────┐
+            │ 08 i18n   09 advanced  │
+            └────────────────────────┘
+                    │
+            ┌───────▼────────────────┐
+            │ 10 quality (always-on) │
+            └────────────────────────┘
+```
+
+## Owners / agents needed
+
+| Domain | Owner agent |
+|--------|-------------|
+| Backend manager/service/repository | principal backend architect agent |
+| Prisma migrations + seeders | principal data architect agent |
+| Frontend pages/components/hooks | principal frontend architect agent |
+| Scoring/learning algorithms | principal AI routing scientist agent |
+| Workflow orchestration | principal LLM orchestration architect agent |
+| QA / replay / regression | principal QA agency agent |
+| Docs / runbooks | principal documentation owner agent |
+| Release gates / canary / rollback | principal DevOps agent |
+
+## Migration order (Prisma)
+
+1. **R.1** — no new tables; reads existing `RouterLearnedScore`/`RouterTopicProfile`.
+2. **R.4** — add `UserCostBudget` + index.
+3. **R.6** — add nullable `orgId` to `RoutingPolicy` + composite index `(orgId, isActive, priority)`.
+4. **R.7** — extend `RoutingDecision` with `detectedLanguage` + `languageConfidence` columns.
+5. **R.2/R.3** — extend `RoutingDecision` with `detectedModalities Json` + `selectedWorkflow WorkflowKind` columns.
+6. **R.8** — `RouterRegionPreference`, `UserCostQualitySlider`, `UserFineTunePreference` (one migration per sub-feature).
+
+## Test strategy
+
+Per-stream test plan in each `0X-name.md`. Cross-cutting:
+
+- **Unit:** every manager method (≥80% coverage). Jest + ts-jest.
+- **API:** `qa/test-routing-r<N>-<feature>.sh` per stream. Hits live backend, asserts DB rows + response shape.
+- **Integration:** `qa/test-routing-integration-flagship.sh` — fires 50 prompts through R.1+R.2+R.3+R.4 pipeline, verifies decision lifecycle.
+- **Regression:** `qa/test-routing-regression-500.sh` (Stream 10) — runs the 500-prompt validation set every CI.
+- **Load:** `qa/test-routing-load.sh` (Stream 10) — 100 req/s × 10 min, p95 routing decision < 50 ms.
+- **Playwright:** `apps/claw-frontend/e2e/routing-*.spec.ts` per UI stream (06, 11.3, 11.9).
+
+## Rollback strategy
+
+Every stream activation behind a feature flag in `.env`:
+
+```
+ROUTING_R1_LEARNED_BIAS_ENABLED=false
+ROUTING_R2_MODALITY_DETECTION_ENABLED=false
+ROUTING_R3_WORKFLOWS_ENABLED=false
+ROUTING_R4_COST_BUDGET_ENABLED=false
+ROUTING_R6_MULTI_TENANT_ENABLED=false
+ROUTING_R7_LANGUAGE_DETECTION_ENABLED=false
+ROUTING_R8_<subfeature>_ENABLED=false
+ROUTING_V2_CANARY_PERCENT=0
+ROUTING_V2_PRIMARY_ENABLED=false
+ROUTING_V2_ROLLBACK_SWITCH=true
+```
+
+Flip flag → rollback in seconds. No data migration required to roll back a stream.
+
+## Business positioning (1 sentence)
+
+> "ClawAI isn't another AI wrapper — it's the intelligence layer that picks the right model and workflow for the task, with cost, privacy, latency, modality, domain awareness, learned feedback, and full transparency."
+
+## Blockers identified
+
+| # | Blocker | Stream | Resolution |
+|---|---------|--------|------------|
+| B1 | `chat-service` is the consumer of `/routing/evaluate`; activating R.1 requires no chat changes (it's transparent), but **R.2/R.3 require chat to pass attachment+URL metadata in the routing context** | 02, 03 | Add `RoutingContext.attachments` + `RoutingContext.urls` upstream in chat-service |
+| B2 | `connector-service` must expose `freeTierRemaining` for R.4 free-tier awareness; not implemented today | 05 | New connector field + sync event |
+| B3 | `WorkflowKind.JUDGE_PIPELINE` requires the chat execution layer to support multi-step LLM calls; currently single-call | 04 | Refactor `ChatExecutionManager` to support workflow handoff |
+| B4 | Multi-tenant requires `User.orgId` end-to-end (auth-service); today users have no org | 07 | Add `Organization` + `OrganizationMember` to auth-service first (same model exists in agent-service but isolated) |
+| B5 | Region routing needs Bedrock multi-region connector support | 09.5 | New `ConnectorRegion` table in connector-service |
diff --git a/docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md b/docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md
new file mode 100644
index 00000000..db16bfb3
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/01-r1-learning-loop.md
@@ -0,0 +1,204 @@
+# Stream 01 — R.1 Close the Learning Loop
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/01_R1_close_learning_loop.md`
+
+## Mission
+
+Make the data already collected by `RouterLearnedScore`, `RouterTopicProfile`, `RoutingOutcomeRecord`, and `RoutingFeedbackRecord` actually influence future routing decisions in the v1 hot path.
+
+## Current state (from audit)
+
+- `RouterEducationManager.ingestExecutionOutcome()` writes `RouterModelProfile` + `RouterLearnedScore` rows on `message.completed`.
+- `RouterEducationManager.ingestFeedbackSignal()` writes the same on `message.feedback_set`.
+- `LearningLoopManager` produces `RouterLearnedScore` via `bounded-adjust.utility.ts`.
+- **`RoutingManager.handleAuto()` never reads any of it.** Hot path is keyword-based + Ollama-router only.
+- v2 evaluator (`route-evaluator/`) consumes the learned data but runs in shadow mode.
+
+## Files to add (scaffold included in this branch)
+
+```
+apps/claw-routing-service/src/modules/routing/
+├── managers/
+│   └── learned-bias.manager.ts                       (NEW — new manager)
+├── utilities/
+│   ├── learned-bias-applier.utility.ts               (NEW)
+│   └── confidence-calibrator.utility.ts              (NEW)
+├── types/
+│   ├── learned-bias.types.ts                         (NEW)
+│   └── thread-sticky-route.types.ts                  (NEW)
+└── constants/
+    └── learned-bias.constants.ts                     (NEW)
+```
+
+## Files to modify (NOT scaffolded — needs careful integration)
+
+```
+apps/claw-routing-service/src/modules/routing/managers/routing.manager.ts
+  → handleAuto(): after category detection + before final selection, call LearnedBiasManager.applyBias()
+  → handleAuto(): if thread sticky route enabled + thread has prior decision, prefer same provider+model
+  → buildExplanation(): include learnedBiasReason in reasonTags
+
+apps/claw-routing-service/src/modules/routing/services/routing.service.ts
+  → publish new "routing.learned_bias_applied" event when bias changes a decision
+
+apps/claw-routing-service/src/modules/routing/repositories/routing-decisions.repository.ts
+  → findLastDecisionForThread(threadId): used by sticky route
+
+apps/claw-routing-service/src/app/config/app.config.ts
+  → add zod-validated env block for ROUTING_R1_*
+```
+
+## API contract — LearnedBiasManager
+
+```typescript
+// apps/claw-routing-service/src/modules/routing/managers/learned-bias.manager.ts
+
+export type BiasInput = {
+  userId: string;
+  domain: DomainTag;
+  taskFamily: string;
+  candidates: CandidateModel[];
+  privacyConstraint: PrivacyClass;
+  threadId?: string;
+};
+
+export type BiasOutput = {
+  candidates: CandidateModel[]; // re-scored
+  appliedBias: {
+    learnedScoreRows: string[];   // ids of rows that influenced the decision
+    topicProfileRow?: string;
+    sampleSize: number;
+    biasWeight: number;
+    reasonTag: string;            // 'learned_bias_thumbs_up_history' | 'sample_too_small_no_bias' | etc.
+  };
+};
+
+@Injectable()
+export class LearnedBiasManager {
+  async applyBias(input: BiasInput): Promise<BiasOutput>;
+}
+```
+
+## Bounded-bias algorithm
+
+```
+1. Fetch RouterLearnedScore rows for (userId, domain) — limit 50 most recent.
+2. Fetch RouterTopicProfile row for (userId, taskFamily).
+3. For each candidate model:
+   - Look up score adjustment Δ from learned rows + topic profile.
+   - Clamp Δ to ±ROUTING_R1_LEARNED_BIAS_WEIGHT_MAX (default 0.3).
+   - If sample size < ROUTING_R1_MIN_SAMPLE_SIZE (default 10): use Δ × (sampleSize/min).
+   - Add Δ to candidate.score.
+4. Re-rank candidates.
+5. NEVER promote a candidate that fails privacyConstraint (LOCAL_ONLY/LOCAL_PREFERRED).
+6. NEVER promote a candidate that lacks required capability (vision/tools/streaming).
+7. Return re-ranked list + appliedBias metadata.
+```
+
+## Confidence calibration
+
+```
+1. Rolling window: last ROUTING_R1_CONFIDENCE_CALIBRATION_WINDOW_DAYS (default 30) days.
+2. For each routing signal (EXACT_KEYWORD, VERB_NOUN_COMBO, CATEGORY_KEYWORD, HEURISTIC_FALLBACK, PRIVACY_ENFORCED):
+   - Compute hit rate = correct decisions / total decisions with that signal.
+3. Replace constants with rolling values.
+4. Snapshot weekly into RoutingCalibrationSnapshot.
+5. Compare to baseline; alert if drop > drift threshold (Stream 09 / R.9).
+```
+
+## Per-thread sticky route
+
+```
+If ROUTING_R1_STICKY_THREAD_ROUTE_ENABLED=true:
+  → Find last 5 RoutingDecision rows for context.threadId.
+  → If all 5 picked same (provider, model) AND no privacy/capability change triggered → use same.
+  → Otherwise fall through to AUTO.
+```
+
+## Acceptance criteria
+
+| # | Test                                                                                       | Expected                                                                |
+|---|---------------------------------------------------------------------------------------------|-------------------------------------------------------------------------|
+| 1 | User thumbs-downs gpt-4o-mini for 12 legal questions; thumbs-ups claude-opus-4 for 8       | Next legal question routes to claude-opus-4 even if keyword classifier first picks mini |
+| 2 | Same as above with only 3 thumbs-downs                                                     | Sample too small; bias dampened; reasonTag includes `sample_too_small` |
+| 3 | Privacy-keyword message + learned bias toward cloud model                                  | Local model wins; reasonTag includes `privacy_override_beats_learned`  |
+| 4 | User on MANUAL_MODEL=gpt-4o; learned bias toward claude-sonnet-4                           | gpt-4o wins; learned bias not applied                                  |
+| 5 | Thread has 5 prior decisions all on claude-sonnet-4; new message on same topic             | Sticky route picks claude-sonnet-4; reasonTag includes `thread_sticky` |
+| 6 | Sticky route would pick a now-unhealthy model                                              | Falls through to AUTO; reasonTag includes `sticky_unhealthy_fallthrough`|
+| 7 | RoutingDecision.explanation contains "learned bias adjusted score from X to Y"             | Visible in `/api/v1/routing/decisions/:threadId` response               |
+| 8 | Feature flag off                                                                            | Hot path unchanged; no learned bias applied                            |
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/routing/managers/__tests__/learned-bias.manager.spec.ts
+  - applies positive bias from thumbs-ups history
+  - applies negative bias from thumbs-downs history
+  - clamps bias to ±weightMax
+  - dampens bias when sample size below min
+  - never overrides privacy constraint
+  - never overrides capability requirement
+  - never overrides manual user model
+  - calibration uses rolling 30-day window
+
+apps/claw-routing-service/src/modules/routing/utilities/__tests__/confidence-calibrator.utility.spec.ts
+  - hit rate calculated correctly
+  - default to constants when window empty
+  - signals tracked independently
+
+apps/claw-routing-service/src/modules/routing/managers/__tests__/routing.manager.sticky-route.spec.ts
+  - sticky route picks last decision when 5/5 same
+  - sticky route falls through on unhealthy model
+  - sticky route disabled by flag
+
+qa/test-routing-r1-learning-loop.sh
+  - Live API test: seed 12 fake feedback rows, fire routing request, verify response includes learned bias reason tag
+```
+
+## Observability
+
+New log lines (all `info` level, structured):
+
+```
+routing.learned_bias_applied  candidates=N userId=X domain=Y biasWeight=Z reasonTag=...
+routing.sticky_thread_applied threadId=X provider=Y model=Z
+routing.confidence_calibrated signal=EXACT_KEYWORD windowDays=30 hitRate=0.93
+routing.privacy_override_beat_bias userId=X originalCandidate=Y replacedWith=Z
+```
+
+New RabbitMQ events:
+
+```
+routing.learned_bias_applied   { userId, domain, biasWeight, beforeScore, afterScore }
+routing.confidence_recalibrated { signal, oldValue, newValue, windowDays }
+```
+
+## Rollback
+
+```bash
+echo 'ROUTING_R1_LEARNED_BIAS_ENABLED=false' >> .env
+./scripts/claw.sh restart routing-service
+```
+
+Hot path reverts to keyword-only routing. Existing decisions in DB retain `learnedBiasReason` field (nullable).
+
+## Risks
+
+| # | Risk | Mitigation |
+|---|------|------------|
+| 1 | Learned bias amplifies user mistakes (e.g. user incorrectly thumbs-downs all coding answers) | Cap at ±0.3 weight; require min sample size; visible in explanation |
+| 2 | Bias data leaks via reasonTag in chat UI ("we picked X because you hated Y") | reasonTag goes to admin replay only; user-facing explanation is generic |
+| 3 | Hot path latency increases due to extra DB queries | Cache learned-bias rows per (userId, domain) for 5 min |
+| 4 | First-week of activation: no bias data yet | Falls back to existing keyword routing — safe-by-default |
+| 5 | Sticky route gets stuck on a broken model | Sticky checks isHealthy; sticky disabled when circuit breaker open |
+
+## Implementation slices (PRs)
+
+1. **Slice 1.1:** scaffold types + constants + LearnedBiasManager skeleton (this branch) + unit tests for applier
+2. **Slice 1.2:** wire LearnedBiasManager.applyBias() call into `RoutingManager.handleAuto()` behind `ROUTING_R1_LEARNED_BIAS_ENABLED` flag; integration test
+3. **Slice 1.3:** confidence calibrator + replace constants in routing.constants.ts; weekly snapshot cron
+4. **Slice 1.4:** per-thread sticky route + tests
+5. **Slice 1.5:** explanation surface in `RoutingDecision.explanation` field; surface in `/routing` admin UI
+6. **Slice 1.6:** RabbitMQ events + audit logs
+
+Each slice ships behind the flag — activation = flip flag in `.env`.
diff --git a/docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md b/docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md
new file mode 100644
index 00000000..71eddec1
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/02-r1r3-v2-evaluator-canary.md
@@ -0,0 +1,145 @@
+# Stream 02 — R.1/R.3 Promote v2 Evaluator (Canary)
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/02_R1_R3_promote_v2_evaluator_canary.md`
+
+## Mission
+
+Take the existing route-evaluator-v2 from shadow-only to canary primary for a small percentage of traffic, with guardrails, comparison telemetry, and instant rollback.
+
+## Current state (from audit)
+
+- `route-evaluator.controller.ts` (`/routing/evaluate-v2`) exists with full `RoutingDecisionV2` schema.
+- `evaluate-shadow.controller.ts` runs v2 alongside v1 — zero impact on response.
+- chat-service still consumes `/routing/evaluate` (v1).
+
+## Files to add (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/route-evaluator/
+├── managers/
+│   └── canary-bucket.manager.ts                      (NEW)
+├── utilities/
+│   ├── canary-hash.utility.ts                        (NEW — stable per-user/org bucketing)
+│   └── decision-comparator.utility.ts                (NEW)
+├── types/
+│   └── canary.types.ts                               (NEW)
+└── constants/
+    └── canary.constants.ts                           (NEW)
+```
+
+## Files to modify (NOT scaffolded)
+
+```
+apps/claw-chat-service/src/modules/chat/managers/chat-execution.manager.ts
+  → call /routing/evaluate-v2 when CanaryBucketManager.isV2Bucket(userId, threadId) returns true
+
+apps/claw-routing-service/src/modules/route-evaluator/controllers/route-evaluator.controller.ts
+  → primary mode: POST /evaluate-v2 returns the v2 decision; on safety/regression breach, fall back to v1
+
+apps/claw-routing-service/src/app/config/app.config.ts
+  → ROUTING_V2_PRIMARY_ENABLED, ROUTING_V2_CANARY_PERCENT, ROUTING_V2_ROLLBACK_SWITCH, threshold envs
+```
+
+## Canary bucketing
+
+```typescript
+// SHA-256(userId + orgId + "routing-v2-canary") % 100 < ROUTING_V2_CANARY_PERCENT
+// Stable: same user always lands in same bucket across requests.
+// Org-scoped: all users in an org can be canaried together.
+// Override: ROUTING_V2_CANARY_USER_ALLOWLIST=u1,u2,u3  (force-in for testing)
+//           ROUTING_V2_CANARY_USER_DENYLIST=u4,u5      (force-out, e.g. CEO)
+```
+
+## Guardrails
+
+Comparator runs on every canary decision:
+
+| Metric                     | Threshold env                                      | Default | Action if breached       |
+|----------------------------|----------------------------------------------------|---------|--------------------------|
+| regression %               | `ROUTING_V2_REGRESSION_THRESHOLD_PERCENT`          | 1       | Auto-disable canary      |
+| cost increase %            | `ROUTING_V2_COST_INCREASE_THRESHOLD_PERCENT`       | 10      | Auto-disable canary      |
+| confidence drop            | `ROUTING_V2_CONFIDENCE_DROP_THRESHOLD`             | 0.1     | Warn, log audit          |
+| failure rate %             | `ROUTING_V2_FAILURE_RATE_THRESHOLD_PERCENT`        | 2       | Auto-disable canary      |
+
+Rolling window: 100 requests OR 5 minutes (whichever first).
+
+## Acceptance criteria
+
+| # | Test                                                                                       | Expected                                                                |
+|---|---------------------------------------------------------------------------------------------|-------------------------------------------------------------------------|
+| 1 | `ROUTING_V2_CANARY_PERCENT=5` + 1000 requests                                              | ~50 requests hit v2; same userId always same bucket                    |
+| 2 | v2 returns invalid `RoutingDecisionV2` (schema fail)                                       | Fall back to v1 silently; log warn; counter incremented                |
+| 3 | v2 disagrees with v1 + post-execution outcome is `BAD_REGRESSION`                          | Comparator increments regression count; if rolling rate > threshold, canary disabled automatically |
+| 4 | `ROUTING_V2_ROLLBACK_SWITCH=true` set                                                      | All canary traffic routes through v1; canary effectively disabled       |
+| 5 | Comparison dashboard at `/routing/canary-comparison`                                       | Side-by-side: v1 picked X (cost $0.001), v2 picked Y (cost $0.0008), outcome= QUALITY_WIN |
+| 6 | Promote suspicious case from canary → regression fixture                                   | Same flow as existing replay-lab promotion                              |
+| 7 | Feature flag fully off (`ROUTING_V2_PRIMARY_ENABLED=false`)                                | v2 stays in shadow mode (old behavior); zero impact                    |
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/route-evaluator/managers/__tests__/canary-bucket.manager.spec.ts
+  - bucket hash is stable per user
+  - bucket respects canary percent
+  - allowlist forces inclusion
+  - denylist forces exclusion
+  - org-bucket includes all org members
+
+apps/claw-routing-service/src/modules/route-evaluator/utilities/__tests__/decision-comparator.utility.spec.ts
+  - detects regression when v1 was correct + v2 picked worse model
+  - detects cost increase
+  - detects confidence drop
+  - detects rolling-window breach
+
+apps/claw-chat-service/src/modules/chat/managers/__tests__/chat-execution.manager.canary.spec.ts
+  - chat falls back to v1 when v2 invalid
+  - chat respects canary flag
+
+qa/test-routing-r2-v2-canary.sh
+  - enable canary at 100% in a test env
+  - fire 10 routing requests, assert all go to v2
+  - inject a fake "v2 returns invalid decision" → assert v1 fallback
+  - set ROUTING_V2_ROLLBACK_SWITCH=true → assert all subsequent requests bypass v2
+```
+
+## Observability
+
+```
+routing.canary.bucketed       userId=X bucket=v2|v1 percent=5
+routing.canary.v2_succeeded   userId=X v1Provider=A v2Provider=B sameDecision=false
+routing.canary.v2_fellback    userId=X reason=invalid_schema|safety|breach
+routing.canary.guardrail_tripped metric=regression|cost|confidence|failure rate=X threshold=Y → AUTO-DISABLE
+```
+
+RabbitMQ:
+
+```
+routing.v2.canary_started     { userId, bucket }
+routing.v2.canary_succeeded   { userId, v1Decision, v2Decision }
+routing.v2.canary_fellback    { userId, reason }
+routing.v2.canary_disabled    { metric, rate, threshold }
+```
+
+## Dashboard (admin)
+
+New page `/routing/canary-comparison`:
+- Live counter: canary % active, request count, regression %, cost delta %, failure %
+- Recent disagreements table: v1 chose X / v2 chose Y / outcome / cost delta
+- "Disable canary" big red button (sets `ROUTING_V2_ROLLBACK_SWITCH=true` via admin API)
+
+## Rollback
+
+Two layers:
+1. Flip `ROUTING_V2_ROLLBACK_SWITCH=true` — single request bypass v2 in milliseconds.
+2. Set `ROUTING_V2_CANARY_PERCENT=0` — canary disabled for new requests.
+
+Automatic rollback when guardrail trips. Manual rollback via admin UI button.
+
+## Risks
+
+| # | Risk | Mitigation |
+|---|------|------------|
+| 1 | v2 returns subtly worse decisions that aren't caught by guardrails | Run shadow comparison for 7 days BEFORE enabling primary canary; require ≤0.5% disagreement |
+| 2 | Bucketing leaks state (same user gets different bucket on retry) | Hash is pure stateless function of (userId, orgId); test asserts stability |
+| 3 | Guardrail false-positive trips canary on benign traffic | Rolling-window minimum 100 requests; warmup grace period of 50 requests after enable |
+| 4 | Comparator latency drags hot path | Comparator runs async on `message.completed` event, not inline |
diff --git a/docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md b/docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md
new file mode 100644
index 00000000..c03960a5
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/03-r2-multimodal-intent-detection.md
@@ -0,0 +1,182 @@
+# Stream 03 — R.2 Multimodal Intent Detection
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/03_R2_multimodal_intent_detection.md`
+
+## Mission
+
+Stop pretending the router only sees text. Detect attachments / URLs / file types / streaming/tool-calling needs in the message context and tag the RoutingDecision with one or more `ModalityKind` values + a hinted `WorkflowKind`.
+
+## Detection coverage
+
+| Modality                  | Detector                                                                                                  | WorkflowKind hint        |
+| ------------------------- | --------------------------------------------------------------------------------------------------------- | ------------------------ |
+| `YOUTUBE_INPUT`           | URL regex: `(youtube\.com/watch|youtu\.be/|youtube\.com/shorts/|youtube\.com/playlist)`                  | `YOUTUBE_TRANSCRIPT`     |
+| `PDF_INPUT`               | Attachment MIME `application/pdf` + verb (`summarize|explain|extract|q&a`)                               | `PDF_EXTRACTION`         |
+| `VIDEO_INPUT`             | Attachment MIME `video/*` OR extension `.mp4|.mov|.webm|.avi|.mkv`                                       | `VIDEO_ANALYSIS`         |
+| `AUDIO_INPUT`             | Attachment MIME `audio/*` OR extension `.wav|.mp3|.m4a|.ogg|.flac`                                       | `AUDIO_TRANSCRIBE`       |
+| `SPREADSHEET_INPUT`       | Attachment MIME `application/vnd.ms-excel|application/vnd.openxmlformats-officedocument.spreadsheetml.sheet` OR ext `.xlsx|.xls|.csv|.ods` | `EXTRACT_FIRST` |
+| `WEB_INPUT`               | Non-YouTube URL + verb (`summarize|read|fetch|what does X say`)                                          | `SEARCH_FIRST`           |
+| `IMAGE_INPUT`             | Attachment MIME `image/*` OR extension `.png|.jpg|.jpeg|.gif|.webp|.bmp`                                 | `IMAGE_ANALYSIS`         |
+| `TOOL_CALLING`            | Verb intent: `book|schedule|send email|call function|invoke tool|run command`                            | n/a — filter on capability flag |
+| `STREAMING` (client need) | Request header `Accept: text/event-stream` OR query `?stream=true`                                       | n/a — filter on capability flag |
+| `EMBEDDING`               | Verb: `embed|vectorize|create embedding|store in vector db|retrieve similar`                             | n/a — separate router branch |
+
+## Files to add (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/modality-detection/    (NEW MODULE)
+├── modality-detection.module.ts
+├── controllers/
+│   └── modality-detection.controller.ts                     (POST /routing/detect-modality)
+├── services/
+│   └── modality-detection.service.ts
+├── managers/
+│   ├── url-intent.manager.ts                                (YouTube + Web URL)
+│   ├── attachment-intent.manager.ts                         (PDF/video/audio/spreadsheet/image)
+│   ├── tool-calling-intent.manager.ts
+│   ├── streaming-intent.manager.ts
+│   └── embedding-intent.manager.ts
+├── dto/
+│   └── detect-modality.dto.ts
+├── types/
+│   └── modality-detection.types.ts
+├── constants/
+│   ├── youtube-url.constants.ts
+│   ├── web-url.constants.ts
+│   ├── pdf-mime.constants.ts
+│   ├── video-mime.constants.ts
+│   ├── audio-mime.constants.ts
+│   ├── spreadsheet-mime.constants.ts
+│   ├── image-mime.constants.ts
+│   ├── tool-calling-keywords.constants.ts
+│   └── embedding-keywords.constants.ts
+└── utilities/
+    └── extension-detector.utility.ts
+```
+
+## Detection contract
+
+```typescript
+export type ModalityDetectionInput = {
+  message: string;
+  attachments: AttachmentMeta[];
+  clientStreamingExpected?: boolean;
+};
+
+export type AttachmentMeta = {
+  fileId: string;
+  filename: string;
+  mimeType: string;
+  sizeBytes: number;
+};
+
+export type ModalityDetectionResult = {
+  detectedModalities: ModalityKind[];
+  workflowHint?: WorkflowKind;
+  workflowConfidence: number;   // 0..1
+  reasonTags: string[];
+  fileMetadata: AttachmentMeta[];
+  urlMetadata: { url: string; kind: 'youtube' | 'web' }[];
+  fallback?: { reason: string; tag: string };
+};
+```
+
+## Acceptance criteria
+
+| # | Test                                                                                                          | Expected                                                       |
+|---|----------------------------------------------------------------------------------------------------------------|----------------------------------------------------------------|
+| 1 | Message contains `https://youtube.com/watch?v=abc123` + "summarize"                                          | `detectedModalities=[TEXT, YOUTUBE_INPUT]`, `workflowHint=YOUTUBE_TRANSCRIPT` |
+| 2 | Attachment with mime `application/pdf` + "extract the key points"                                            | `detectedModalities=[TEXT, PDF_INPUT]`, `workflowHint=PDF_EXTRACTION` |
+| 3 | Attachment mime `audio/mpeg`                                                                                 | `[TEXT, AUDIO_INPUT]`, `workflowHint=AUDIO_TRANSCRIBE`         |
+| 4 | Attachment mime `video/mp4`                                                                                  | `[TEXT, VIDEO_INPUT]`, `workflowHint=VIDEO_ANALYSIS`           |
+| 5 | Attachment mime `application/vnd.openxmlformats-officedocument.spreadsheetml.sheet`                          | `[TEXT, SPREADSHEET_INPUT]`, `workflowHint=EXTRACT_FIRST`      |
+| 6 | URL `https://example.com/blog` + "summarize this article"                                                    | `[TEXT, WEB_INPUT]`, `workflowHint=SEARCH_FIRST`               |
+| 7 | Message "book me a meeting tomorrow at 3pm"                                                                  | `[TEXT, TOOL_CALLING]`, no workflowHint                        |
+| 8 | Request with `Accept: text/event-stream`                                                                     | `[TEXT, STREAMING]`, no workflowHint                           |
+| 9 | Message "embed this paragraph and store in vector db"                                                        | `[TEXT, EMBEDDING]`, no workflowHint                           |
+| 10 | Multiple attachments: 1 PDF + 1 audio + 1 image                                                             | `[TEXT, PDF_INPUT, AUDIO_INPUT, IMAGE_INPUT]`; workflowHint=highest-priority (PDF_EXTRACTION) |
+| 11 | Malicious URL: `javascript:alert(1)` or non-http scheme                                                     | Stripped from detection; reasonTag `unsafe_url_ignored`        |
+| 12 | YouTube URL but `ROUTING_R2_YOUTUBE_DETECTION_ENABLED=false`                                                | Falls through; not flagged                                     |
+| 13 | All flags off (`ROUTING_R2_MODALITY_DETECTION_ENABLED=false`)                                              | Endpoint returns empty modalities; hot path uses v1 behavior   |
+| 14 | Non-English prompt with file ("الرجاء تلخيص ملف PDF" + PDF attachment)                                     | Detects `PDF_INPUT` regardless of message language             |
+
+## Endpoint contract
+
+```http
+POST /api/v1/routing/detect-modality
+
+{
+  "message": "Please summarize https://youtu.be/abc123",
+  "attachments": [],
+  "clientStreamingExpected": false
+}
+
+Response 200:
+{
+  "detectedModalities": ["TEXT", "YOUTUBE_INPUT"],
+  "workflowHint": "YOUTUBE_TRANSCRIPT",
+  "workflowConfidence": 0.95,
+  "reasonTags": ["youtube_url_with_summarize_verb"],
+  "fileMetadata": [],
+  "urlMetadata": [{"url": "https://youtu.be/abc123", "kind": "youtube"}]
+}
+```
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/modality-detection/managers/__tests__/url-intent.manager.spec.ts
+  - youtube.com/watch detected
+  - youtu.be detected
+  - youtube.com/shorts detected
+  - playlist detected
+  - non-youtube URL with summarize verb → web_input
+  - javascript: URL stripped
+  - data: URL stripped
+
+apps/claw-routing-service/src/modules/modality-detection/managers/__tests__/attachment-intent.manager.spec.ts
+  - PDF + summarize verb → PDF_EXTRACTION
+  - PDF without verb → PDF_INPUT only (no workflow hint)
+  - audio/mp3 → AUDIO_TRANSCRIBE
+  - video/mp4 → VIDEO_ANALYSIS
+  - .xlsx mime → SPREADSHEET_INPUT
+  - .png mime → IMAGE_INPUT
+  - multiple attachments → multiple modalities
+
+qa/test-routing-r2-modality-detection.sh
+  - Hit POST /routing/detect-modality with each MIME type
+  - Assert response shape
+```
+
+## Wiring with Stream 04 (workflow orchestrator)
+
+After this stream is wired, `RoutingManager.handleAuto()` calls:
+
+```typescript
+const modalityResult = await this.modalityDetectionService.detect({
+  message: context.message,
+  attachments: context.attachments,
+});
+
+if (modalityResult.workflowHint && workflowsEnabled(modalityResult.workflowHint)) {
+  return this.workflowOrchestratorManager.orchestrate({
+    workflow: modalityResult.workflowHint,
+    context,
+    modalityResult,
+  });
+}
+// otherwise fall through to existing AUTO pipeline
+```
+
+## Rollback
+
+`ROUTING_R2_MODALITY_DETECTION_ENABLED=false` → endpoint returns empty; hot path unaffected. Per-modality flags allow partial activation (e.g. enable YouTube detection but not video).
+
+## Risks
+
+| # | Risk | Mitigation |
+|---|------|------------|
+| 1 | False-positive URL detection (e.g. `youtube.com.malicious.io`) | Strict regex anchored at host start; only valid scheme `https?:` |
+| 2 | Attachment metadata not passed by chat-service in `RoutingContext` | **Blocker B1** in master plan — chat-service change required first |
+| 3 | Detection adds latency to hot path | Detector is pure (no I/O); regex + array scans; <1ms total |
+| 4 | Workflow hint suggested but workflow disabled | Hot path falls through to v1 routing; reasonTag `workflow_disabled_fallthrough` |
diff --git a/docs/15-ai-context/routing-flagship-streams/04-r3-workflow-orchestrator.md b/docs/15-ai-context/routing-flagship-streams/04-r3-workflow-orchestrator.md
new file mode 100644
index 00000000..a18499e4
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/04-r3-workflow-orchestrator.md
@@ -0,0 +1,236 @@
+# Stream 04 — R.3 Workflow Orchestrator Goes Live
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/04_R3_workflow_orchestrator_goes_live.md`
+
+## Mission
+
+Make all 13 `WorkflowKind` values mean real executable behavior. Today `RoutingManager.handleAuto()` always returns `DIRECT_LLM` implicitly; this stream lights up the other 12 workflows behind per-workflow feature flags.
+
+## The 13 workflows
+
+| Workflow             | Trigger                                                                       | Status today | Activation flag                                          |
+| -------------------- | ----------------------------------------------------------------------------- | ------------ | -------------------------------------------------------- |
+| `DIRECT_LLM`         | Default — fallthrough when no other workflow matches                          | wired        | always on                                                |
+| `SEARCH_FIRST`       | Time-sensitive verbs (latest, today, current, news, regulation, pricing)      | not wired    | `ROUTING_R3_WORKFLOW_SEARCH_FIRST_ENABLED`               |
+| `EXTRACT_FIRST`      | Attachment with structured-data intent (spreadsheet/CSV/JSON)                 | not wired    | `ROUTING_R3_WORKFLOW_EXTRACT_FIRST_ENABLED`              |
+| `PDF_EXTRACTION`     | `PDF_INPUT` modality + summarize/explain/extract verb                         | not wired    | `ROUTING_R3_WORKFLOW_PDF_EXTRACTION_ENABLED`             |
+| `YOUTUBE_TRANSCRIPT` | `YOUTUBE_INPUT` modality                                                      | not wired    | `ROUTING_R3_WORKFLOW_YOUTUBE_TRANSCRIPT_ENABLED`         |
+| `IMAGE_ANALYSIS`     | `IMAGE_INPUT` modality                                                        | not wired    | `ROUTING_R3_WORKFLOW_IMAGE_ANALYSIS_ENABLED`             |
+| `IMAGE_GENERATION`   | Image generation keywords detected (Stage 2 of v1 pipeline)                  | wired today  | `ROUTING_R3_WORKFLOW_IMAGE_GENERATION_ENABLED` (on)      |
+| `VIDEO_ANALYSIS`     | `VIDEO_INPUT` modality                                                        | not wired    | `ROUTING_R3_WORKFLOW_VIDEO_ANALYSIS_ENABLED`             |
+| `AUDIO_TRANSCRIBE`   | `AUDIO_INPUT` modality                                                        | not wired    | `ROUTING_R3_WORKFLOW_AUDIO_TRANSCRIBE_ENABLED`           |
+| `FILE_GENERATION`    | File generation verb+format detected (Stage 3 of v1 pipeline)                | wired today  | `ROUTING_R3_WORKFLOW_FILE_GENERATION_ENABLED` (on)       |
+| `CODE_REVIEW`        | Code-block in message + `review/refactor/debug` verb                          | not wired    | `ROUTING_R3_WORKFLOW_CODE_REVIEW_ENABLED`                |
+| `COMPARE_ENSEMBLE`   | Explicit "compare" mode OR high-uncertainty (confidence < 0.5)                | not wired    | `ROUTING_R3_WORKFLOW_COMPARE_ENSEMBLE_ENABLED`           |
+| `JUDGE_PIPELINE`     | Medical/Legal/high-risk domain detected                                       | not wired    | `ROUTING_R3_WORKFLOW_JUDGE_PIPELINE_ENABLED`             |
+
+## Per-workflow contract
+
+Every workflow handler implements:
+
+```typescript
+export interface IWorkflowHandler {
+  readonly kind: WorkflowKind;
+
+  /** Validate that this workflow CAN run with the given context (deps healthy, attachments present, etc.) */
+  canHandle(context: WorkflowExecutionContext): boolean;
+
+  /** Return the workflow plan: an ordered list of steps. Each step = a model call OR external service call. */
+  plan(context: WorkflowExecutionContext): WorkflowPlan;
+
+  /** Confidence score for whether this workflow is the right choice for this context (0..1) */
+  confidence(context: WorkflowExecutionContext): number;
+}
+
+export type WorkflowExecutionContext = {
+  message: string;
+  threadId?: string;
+  userId: string;
+  modalityResult: ModalityDetectionResult;
+  routingDecision: RoutingDecision;
+  attachments: AttachmentMeta[];
+  urls: DetectedUrl[];
+};
+
+export type WorkflowPlan = {
+  kind: WorkflowKind;
+  steps: WorkflowStep[];
+  estimatedDurationMs: number;
+  estimatedCostUsd: number;
+};
+
+export type WorkflowStep =
+  | { type: 'llm_call'; provider: string; model: string; promptTemplate: string; inputFromStepId?: string }
+  | { type: 'extract'; extractor: 'pdf' | 'spreadsheet' | 'youtube_transcript' | 'web_scrape' | 'audio_transcribe' | 'video_extract'; sourceFileId?: string; sourceUrl?: string }
+  | { type: 'search'; query: string }
+  | { type: 'judge'; primaryStepId: string; criticStepId: string };
+```
+
+## Files to add (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/workflows/
+├── managers/
+│   ├── workflow-orchestrator.manager.ts                   (NEW — top-level coordinator)
+│   └── handlers/                                          (NEW directory — 13 workflow handlers)
+│       ├── direct-llm.handler.ts
+│       ├── search-first.handler.ts
+│       ├── extract-first.handler.ts
+│       ├── pdf-extraction.handler.ts
+│       ├── youtube-transcript.handler.ts
+│       ├── image-analysis.handler.ts
+│       ├── image-generation.handler.ts
+│       ├── video-analysis.handler.ts
+│       ├── audio-transcribe.handler.ts
+│       ├── file-generation.handler.ts
+│       ├── code-review.handler.ts
+│       ├── compare-ensemble.handler.ts
+│       └── judge-pipeline.handler.ts
+├── types/
+│   ├── workflow-execution.types.ts
+│   ├── workflow-plan.types.ts
+│   └── workflow-handler.interface.ts
+└── constants/
+    └── workflow-priority.constants.ts                      (NEW)
+```
+
+## Workflow priority (when multiple match)
+
+```
+1. JUDGE_PIPELINE   (highest — safety wins)
+2. COMPARE_ENSEMBLE
+3. PDF_EXTRACTION
+4. YOUTUBE_TRANSCRIPT
+5. VIDEO_ANALYSIS
+6. AUDIO_TRANSCRIBE
+7. IMAGE_ANALYSIS
+8. CODE_REVIEW
+9. SEARCH_FIRST
+10. EXTRACT_FIRST
+11. FILE_GENERATION
+12. IMAGE_GENERATION
+13. DIRECT_LLM      (lowest — fallthrough)
+```
+
+## High-priority workflow specs
+
+### `JUDGE_PIPELINE`
+
+For medical/legal/high-risk:
+```
+Step 1: llm_call to PRIMARY (e.g. claude-opus-4)
+Step 2: llm_call to CRITIC (different family, e.g. gpt-4o) — asked "is the primary answer correct? cite issues"
+Step 3: judge — if critic agrees → finalize; if disagrees → escalate to a third "judge" model OR human review queue
+```
+
+Records `judgeOutcome` on `RoutingOutcomeRecord`.
+
+### `COMPARE_ENSEMBLE`
+
+```
+Step 1..N: parallel llm_call to N candidate models
+Step N+1: judge — score the N responses (criteria: completeness, correctness, format adherence)
+Step N+2: return winner + comparison summary
+```
+
+### `SEARCH_FIRST`
+
+```
+Step 1: search → research-service /search/run with query
+Step 2: extract — pull top-K results as context
+Step 3: llm_call with context-augmented prompt
+```
+
+### `PDF_EXTRACTION`
+
+```
+Step 1: extract — pdf via file-service chunking
+Step 2: llm_call with chunks + summarize/extract/qa system prompt
+```
+
+### `YOUTUBE_TRANSCRIPT`
+
+```
+Step 1: extract — youtube_transcript via new extractor service (or research-service extension)
+Step 2: llm_call with transcript chunks + user verb
+```
+
+### `CODE_REVIEW`
+
+```
+Step 1: extract code block(s) from message
+Step 2: llm_call to coding-tuned model (claude-sonnet-4 / LOCAL_CODING) with code-review system prompt
+Step 3 (optional): lint via local tool integration
+Step 4: critic LLM verifies the review's suggestions
+```
+
+## Acceptance criteria
+
+| # | Test                                                                                                     | Expected                                                                                |
+|---|----------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------|
+| 1 | `RoutingDecision` includes `selectedWorkflow` field                                                      | All decisions have non-null workflow (default: DIRECT_LLM)                              |
+| 2 | Medical question + JUDGE_PIPELINE enabled                                                                | Decision has `selectedWorkflow=JUDGE_PIPELINE`, response goes through primary+critic    |
+| 3 | "Compare claude and gpt-4 on this question" + COMPARE_ENSEMBLE enabled                                   | Decision has `COMPARE_ENSEMBLE`, response includes both answers + winner               |
+| 4 | "What's the latest news on the EU AI Act" + SEARCH_FIRST enabled                                         | Decision has `SEARCH_FIRST`, response cites recent web sources                          |
+| 5 | PDF attachment + "summarize" + PDF_EXTRACTION disabled                                                   | Decision falls back to `DIRECT_LLM`, reasonTag `workflow_disabled_fallthrough`         |
+| 6 | Unsupported workflow advertised in UI                                                                    | NOT advertised — UI reads per-flag state                                                |
+| 7 | Workflow fails (e.g. YouTube transcript extractor down)                                                  | Falls back to `DIRECT_LLM` with reasonTag `workflow_extraction_failed`                  |
+
+## Frontend surfacing
+
+Each routing decision in `/routing` recent list shows a workflow badge:
+
+```
+[ JUDGE_PIPELINE ] [ COMPARE_ENSEMBLE ] [ SEARCH_FIRST ] [ DIRECT_LLM ] [ disabled — not yet wired ]
+```
+
+Disabled workflows greyed out with tooltip "Activate ROUTING_R3_WORKFLOW_X_ENABLED in .env".
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/workflows/managers/__tests__/workflow-orchestrator.manager.spec.ts
+  - picks JUDGE_PIPELINE over DIRECT_LLM for medical domain
+  - picks COMPARE_ENSEMBLE for explicit "compare" intent
+  - picks PDF_EXTRACTION when PDF + summarize verb
+  - falls back to DIRECT_LLM when chosen workflow disabled
+  - falls back to DIRECT_LLM when chosen workflow's deps unhealthy
+  - respects priority order when multiple match
+
+apps/claw-routing-service/src/modules/workflows/managers/handlers/__tests__/*.handler.spec.ts
+  - one spec per handler (13 specs)
+  - each verifies canHandle + plan + confidence
+
+qa/test-routing-r3-workflows.sh
+  - fire 13 fixture messages, one per workflow
+  - assert decision.selectedWorkflow matches expected
+```
+
+## Wiring
+
+Once activated, `RoutingManager.handleAuto()`:
+
+```typescript
+async handleAuto(context: RoutingContext): Promise<RoutingDecisionResult> {
+  // ... existing stages 1-4 ...
+
+  const modalityResult = await this.modalityDetectionService.detect({ /* ... */ });
+
+  const workflowChoice = await this.workflowOrchestratorManager.pickWorkflow({
+    context,
+    modalityResult,
+    baseDecision: existingV1Decision,
+  });
+
+  return {
+    ...existingV1Decision,
+    selectedWorkflow: workflowChoice.kind,
+    workflowConfidence: workflowChoice.confidence,
+    workflowPlan: workflowChoice.plan,
+  };
+}
+```
+
+## Rollback
+
+Per-workflow flag — disable one without affecting others. Master flag `ROUTING_R3_WORKFLOWS_ENABLED=false` → all back to DIRECT_LLM-only behavior.
diff --git a/docs/15-ai-context/routing-flagship-streams/05-r4-cost-budget-intelligence.md b/docs/15-ai-context/routing-flagship-streams/05-r4-cost-budget-intelligence.md
new file mode 100644
index 00000000..9512ceb5
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/05-r4-cost-budget-intelligence.md
@@ -0,0 +1,130 @@
+# Stream 05 — R.4 Cost Budget Intelligence
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/05_R4_cost_budget_intelligence.md`
+
+## Mission
+
+Stop letting users spend unlimited cloud-model dollars. Add per-user (and optionally per-org) monthly caps; pre-routing gate that blocks cloud calls when over budget; user-visible warnings; free-tier-remaining awareness; admin + user dashboards.
+
+## Files to add (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/cost-budget/             (NEW MODULE)
+├── cost-budget.module.ts
+├── controllers/
+│   └── cost-budget.controller.ts
+├── services/
+│   └── cost-budget.service.ts
+├── managers/
+│   ├── budget-gate.manager.ts                                  (pre-routing check)
+│   ├── spend-tracker.manager.ts                                (post-execution increment)
+│   ├── budget-warning.manager.ts                               (80% threshold warnings)
+│   └── budget-reset.manager.ts                                 (monthly cron)
+├── repositories/
+│   └── user-cost-budget.repository.ts
+├── dto/
+│   ├── create-budget.dto.ts
+│   ├── update-budget.dto.ts
+│   └── check-budget.dto.ts
+├── types/
+│   └── budget.types.ts
+└── constants/
+    └── budget.constants.ts
+```
+
+## Prisma migration (see PRISMA_FUTURE_MODELS.md)
+
+`UserCostBudget` + `CostBudgetScope` + `CostBudgetStatus` enums. Backfill: empty.
+
+## Pre-routing budget gate
+
+```typescript
+// In RoutingManager.handleAuto() — BEFORE picking cloud provider:
+
+const estimatedCost = estimateRequestCost(context, candidatePrimary);
+const budgetCheck = await this.budgetGateManager.check({
+  userId: context.userId,
+  orgId: context.orgId,
+  estimatedCostUsd: estimatedCost,
+});
+
+if (budgetCheck.status === 'EXCEEDED' && !budgetCheck.overrideAllowed) {
+  // Force local route
+  return this.handleLocalOnly(context);
+}
+
+if (budgetCheck.status === 'WARN') {
+  reasonTags.push('cost_budget_warn');
+}
+```
+
+## Acceptance criteria
+
+| # | Test | Expected |
+|---|------|----------|
+| 1 | User under budget + cloud request | Routes to cloud as normal; spend incremented post-execution |
+| 2 | User at 85% of budget | Routes to cloud + `messageKey: 'BUDGET_NEAR_LIMIT'` in response metadata |
+| 3 | User over budget without override permission | Routes to local; `messageKey: 'BUDGET_EXCEEDED_FORCED_LOCAL'` |
+| 4 | User over budget WITH `overrideAllowed=true` | Routes to cloud + audit log + `messageKey: 'BUDGET_EXCEEDED_OVERRIDDEN'` |
+| 5 | Unknown cost (model not in cost table) | Treat cautiously: assume STANDARD class; if user near budget → block |
+| 6 | Privacy keyword present | Privacy beats budget — always local regardless |
+| 7 | Free tier remaining (when connector exposes it) | Prefer free-tier-remaining provider for cost wins |
+| 8 | Budget reset on `resetAt` date | Cron at midnight: `currentSpendUsd=0`, `status=OK`, `resetAt += 1 month` |
+| 9 | Org budget exceeded but user has personal headroom | Org budget wins (most restrictive) |
+| 10 | Cost dashboard at `/settings/budget` | Shows: monthly cap, spent, trend, projected hit date |
+| 11 | Admin dashboard at `/admin/cost-budgets` | List all users + spend + status; filter by org |
+| 12 | Cost/quality slider (`0=cheapest, 100=best`) | Affects scoring engine weight |
+
+## Endpoint contract
+
+```http
+GET    /api/v1/routing/cost-budget/me            (current user's budget)
+GET    /api/v1/routing/cost-budget/me/forecast   (projected hit date + trend)
+PATCH  /api/v1/routing/cost-budget/me            (user updates personal cap if allowed)
+POST   /api/v1/routing/cost-budget                (admin: create budget for any user/org)
+PATCH  /api/v1/routing/cost-budget/:id            (admin: update)
+GET    /api/v1/routing/cost-budget                (admin: list all with filters)
+POST   /api/v1/routing/cost-budget/check          (internal: pre-routing gate)
+```
+
+## Free-tier awareness (blocked on connector-service work)
+
+Stream blocker B2 in master plan. When connector-service exposes `freeTierRemaining`:
+
+```typescript
+// Score boost for providers with free tier
+if (candidate.freeTierRemaining > 0 && candidate.costClass !== 'FREE') {
+  candidate.score += 0.1;  // small boost when free tier still available
+}
+```
+
+## RabbitMQ events
+
+```
+cost_budget.created                 { budgetId, scope, ownerId, monthlyCapUsd }
+cost_budget.spend_incremented       { budgetId, deltaUsd, totalUsd, percentOfCap }
+cost_budget.warning_threshold_crossed { budgetId, percentOfCap }
+cost_budget.exceeded                { budgetId, forcedLocal: boolean, overrideUsed: boolean }
+cost_budget.reset                   { budgetId, previousSpend, newResetAt }
+```
+
+## Frontend
+
+- `/settings/budget` — user page (own budget)
+- `/admin/cost-budgets` — admin page (all budgets)
+- Warning banner in chat composer when budget at 80%+
+- Locked-state in model selector when over budget without override
+- "Budget exceeded" empty state with "Switch to local model" CTA
+
+## Rollback
+
+`ROUTING_R4_COST_BUDGET_ENABLED=false` → budget gate skipped; all routes proceed as before. DB rows remain (no destructive change).
+
+## Risks
+
+| # | Risk | Mitigation |
+|---|------|------------|
+| 1 | Cost estimate wrong (token count off) | Use generous upper bound; log actual vs estimated for calibration |
+| 2 | Race condition: 2 concurrent requests both pass budget check | Increment is post-execution; minor over-cap acceptable; weekly cleanup script |
+| 3 | User can't access their work near reset | "Override" toggle for "this 24h period" |
+| 4 | Privacy keyword + over-budget conflict | Privacy always wins — never expose data to cloud even if user opts in |
diff --git a/docs/15-ai-context/routing-flagship-streams/06-r5-operator-playground.md b/docs/15-ai-context/routing-flagship-streams/06-r5-operator-playground.md
new file mode 100644
index 00000000..3213eff7
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/06-r5-operator-playground.md
@@ -0,0 +1,180 @@
+# Stream 06 — R.5 Operator Playground + Transparency
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/06_R5_operator_playground_transparency.md`
+
+## Mission
+
+Give operators and end users transparency into routing decisions. Add a "Try a message" playground for operators, surface "Why this model?" explanations in chat, and add the missing admin pages (taxonomy CRUD, circuit-breaker dashboard, category drill-down).
+
+## Surfaces to ship
+
+| Surface | Type | Path | Status today | New file paths |
+|---------|------|------|--------------|----------------|
+| Playground | FE page | `/routing/playground` | missing | `apps/claw-frontend/src/app/(portal)/routing/playground/page.tsx` |
+| "Why this model?" | FE component | inline in chat message | missing | `apps/claw-frontend/src/components/chat/why-this-model.tsx` |
+| Category drill-down | FE filter | `/routing` recent list | missing | extend `use-routing-decisions.ts` + add filter to existing page |
+| Circuit-breaker dashboard | FE page | `/routing/circuit-breakers` | missing (BE exists) | `apps/claw-frontend/src/app/(portal)/routing/circuit-breakers/page.tsx` |
+| Taxonomy admin UI | FE page | `/routing/taxonomy` | missing (BE exists) | `apps/claw-frontend/src/app/(portal)/routing/taxonomy/page.tsx` |
+| Per-user routing history | FE page | `/settings/routing-history` | missing | `apps/claw-frontend/src/app/(portal)/settings/routing-history/page.tsx` |
+| "Rerun decision" button | FE button on recent decision row | existing | missing | extend `routing-decision-row.tsx` |
+| "Save playground case as fixture" | FE button on playground | new | missing | inside playground page |
+| "Compare policies on sample prompt" | FE page | `/routing/policy-compare` | missing | `apps/claw-frontend/src/app/(portal)/routing/policy-compare/page.tsx` |
+
+## Playground API contract
+
+```http
+POST /api/v1/routing/playground/evaluate
+{
+  "message": "Write a Python function to sort a list",
+  "attachments": [],
+  "userMode": "AUTO",
+  "compareWithV2": true,
+  "compareWithOllamaRouter": true
+}
+
+Response 200:
+{
+  "v1Decision": { /* full RoutingDecision */ },
+  "v2Decision": { /* full RoutingDecisionV2 */ },
+  "ollamaRouterDecision": { /* OllamaRouter raw output */ },
+  "scoreBreakdown": [
+    { "provider": "Anthropic", "model": "claude-sonnet-4",
+      "scores": { "quality": 0.92, "cost": 0.6, "latency": 0.75, "privacy": 0.5, "familiarity": 0.8 },
+      "totalScore": 0.74, "wasChosen": true },
+    { "provider": "OpenAI", "model": "gpt-4o-mini",
+      "scores": { "quality": 0.7, "cost": 0.95, "latency": 0.9, "privacy": 0.5, "familiarity": 0.6 },
+      "totalScore": 0.71, "wasChosen": false }
+  ],
+  "candidateList": [...full ranked list...],
+  "modalityResult": {...},
+  "workflowChoice": {...}
+}
+```
+
+## "Why this model?" inline
+
+For every chat message, add a small `ⓘ` button that expands to:
+
+```
+Routed to claude-sonnet-4 (Anthropic) because:
+• Category: Coding (CONFIDENCE_EXACT_KEYWORD: "function", "sort", "list")
+• Workflow: DIRECT_LLM (no special workflow triggered)
+• Learned bias: +0.05 (you positively rated claude-sonnet-4 on coding 8/10 times)
+• Cost class: STANDARD ($0.003 estimated)
+• Privacy: PUBLIC_OK (no PII detected)
+[ Rerun decision ]  [ See full reasoning ]
+```
+
+## Files to add (no scaffold this round — frontend skeletons + BE playground module)
+
+Backend scaffold:
+
+```
+apps/claw-routing-service/src/modules/playground/         (NEW MODULE)
+├── playground.module.ts
+├── controllers/
+│   └── playground.controller.ts                          (POST /routing/playground/evaluate)
+├── services/
+│   └── playground.service.ts
+├── dto/
+│   └── playground-evaluate.dto.ts
+└── types/
+    └── playground.types.ts
+```
+
+Frontend pages (NOT scaffolded — needs i18n + auth wiring):
+
+```
+apps/claw-frontend/src/app/(portal)/routing/playground/page.tsx
+apps/claw-frontend/src/app/(portal)/routing/circuit-breakers/page.tsx
+apps/claw-frontend/src/app/(portal)/routing/taxonomy/page.tsx
+apps/claw-frontend/src/app/(portal)/routing/policy-compare/page.tsx
+apps/claw-frontend/src/app/(portal)/settings/routing-history/page.tsx
+apps/claw-frontend/src/components/chat/why-this-model.tsx
+apps/claw-frontend/src/hooks/routing/use-playground.ts
+apps/claw-frontend/src/hooks/routing/use-circuit-breakers.ts
+apps/claw-frontend/src/hooks/routing/use-taxonomy-admin.ts
+apps/claw-frontend/src/hooks/routing/use-routing-history-mine.ts
+```
+
+## Acceptance criteria
+
+| # | Test | Expected |
+|---|------|----------|
+| 1 | Playground happy path: type message → submit | v1+v2+ollama all return; score breakdown visible |
+| 2 | Playground does NOT actually execute the chat model | only the routing decision; no token spend |
+| 3 | "Why this model?" expands in chat | shows reasonTags + cost estimate + workflow |
+| 4 | Category filter on `/routing` | filtering to "Coding" shows only Coding-routed decisions |
+| 5 | Circuit-breaker dashboard shows current state | OPEN/HALF_OPEN/CLOSED + recent failures + manual reset button |
+| 6 | Taxonomy CRUD | admin can create/edit/delete TaxonomyRole entries via UI |
+| 7 | Per-user routing history at `/settings/routing-history` | user sees their own decisions, with override option |
+| 8 | Rerun decision button | hits `POST /routing/evaluate` with same context, shows new decision side-by-side |
+| 9 | Save playground case as fixture | converts the playground decision into a replay-fixture row |
+| 10 | Dark mode / mobile / RTL | all surfaces pass |
+
+## i18n keys (need real translations in all 8 locales)
+
+```typescript
+{
+  routing: {
+    playground: {
+      title: 'Routing Playground',
+      messageLabel: 'Message',
+      compareV2Label: 'Compare with v2 evaluator',
+      compareOllamaLabel: 'Compare with Ollama router',
+      submitButton: 'Evaluate routing',
+      v1Heading: 'v1 decision (current production)',
+      v2Heading: 'v2 decision (shadow / canary)',
+      ollamaHeading: 'Ollama-router decision',
+      scoreBreakdownHeading: 'Score breakdown',
+      candidateListHeading: 'All candidates',
+      saveAsFixtureButton: 'Save as regression fixture',
+      reasonTagsLabel: 'Reason tags',
+    },
+    whyThisModel: {
+      heading: 'Why this model?',
+      categoryLabel: 'Category',
+      workflowLabel: 'Workflow',
+      learnedBiasLabel: 'Learned bias',
+      costEstimateLabel: 'Cost estimate',
+      privacyLabel: 'Privacy class',
+      rerunButton: 'Rerun decision',
+      fullReasoningButton: 'See full reasoning',
+    },
+    circuitBreakers: {
+      title: 'Circuit Breakers',
+      stateColumn: 'State',
+      providerColumn: 'Provider',
+      failureCountColumn: 'Failures',
+      lastFailureAtColumn: 'Last failure',
+      opensUntilColumn: 'Opens until',
+      manualResetButton: 'Reset manually',
+      states: { CLOSED: 'Closed', OPEN: 'Open', HALF_OPEN: 'Half-open' },
+    },
+    taxonomy: {
+      title: 'Taxonomy Admin',
+      roles: { /* ... */ },
+      domains: { /* ... */ },
+    },
+    routingHistoryMine: {
+      title: 'My Routing History',
+      empty: 'No routing decisions yet',
+      overrideButton: 'Always route X to Y for me',
+    },
+  },
+}
+```
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/playground/services/__tests__/playground.service.spec.ts
+qa/test-routing-r5-playground.sh
+apps/claw-frontend/e2e/routing-playground.spec.ts (Playwright)
+apps/claw-frontend/e2e/routing-circuit-breakers.spec.ts
+apps/claw-frontend/e2e/why-this-model.spec.ts
+```
+
+## Rollback
+
+Per-surface flag: `ROUTING_R5_PLAYGROUND_ENABLED`, `ROUTING_R5_EXPLANATION_IN_CHAT_ENABLED`, etc. Each surface can be hidden via flag without touching code.
diff --git a/docs/15-ai-context/routing-flagship-streams/07-r6-multi-tenant-fleet.md b/docs/15-ai-context/routing-flagship-streams/07-r6-multi-tenant-fleet.md
new file mode 100644
index 00000000..e081aed8
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/07-r6-multi-tenant-fleet.md
@@ -0,0 +1,103 @@
+# Stream 07 — R.6 Multi-Tenant Fleet Routing
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/07_R6_multi_tenant_fleet_routing.md`
+
+## Mission
+
+Add org-scoped routing policies, allowed-provider rules, org budgets, org rate limits, and policy propagation. Today policies are global; this stream lets fleet admins manage routing for their org.
+
+**BLOCKER B4:** Requires `User.orgId` end-to-end in auth-service. `Organization` + `OrganizationMember` already exist in agent-service but are isolated there. Must be lifted into auth-service OR cross-service-shared first.
+
+## Schema changes (see PRISMA_FUTURE_MODELS.md)
+
+- `RoutingPolicy`: add nullable `orgId` column + index `(orgId, isActive, priority)`
+- New `OrgProviderRule` table (ALLOW/DENY per provider per org)
+- New `OrgRateLimit` table (req/min per org)
+- (Stream 05) `UserCostBudget.scope='ORG'` already supports org budgets
+
+## Policy resolution order (most-restrictive wins)
+
+```
+1. Explicit user override (per-thread settings)         → wins immediately
+2. Privacy hard constraints (PRIVACY_KEYWORDS)          → wins (never cloud)
+3. Org DENY rules (org banned this provider)            → block + fallback
+4. Org policy (orgId-scoped, highest priority)          → may override AUTO
+5. Org ALLOW list (if set, restricts choices)           → filter candidates
+6. Global policy (orgId=null, priority-ordered)         → may override AUTO
+7. Safe defaults (existing v1 behavior)
+```
+
+## Files to add / modify
+
+```
+apps/claw-routing-service/src/modules/routing/
+├── managers/
+│   └── policy-resolution.manager.ts                       (NEW — replaces inline applyPolicies)
+├── repositories/
+│   └── org-provider-rule.repository.ts                    (NEW)
+│   └── org-rate-limit.repository.ts                       (NEW)
+├── dto/
+│   ├── create-org-policy.dto.ts                           (NEW — extends create-policy.dto)
+│   ├── create-org-provider-rule.dto.ts                    (NEW)
+│   └── update-org-rate-limit.dto.ts                       (NEW)
+└── types/
+    └── policy-resolution.types.ts                         (NEW)
+```
+
+## Acceptance criteria
+
+| # | Test | Expected |
+|---|------|----------|
+| 1 | Org admin creates policy `forceClaude` for orgX | All orgX users see Anthropic on next request; non-orgX users unaffected |
+| 2 | Org admin adds DENY rule for OpenAI on orgX | All orgX users get fallback away from OpenAI; reasonTag `org_provider_denied` |
+| 3 | Existing global policies still apply | When user has no org, global behavior unchanged |
+| 4 | Org rate limit 60/min exceeded | Routing returns 429 with messageKey `ORG_RATE_LIMIT_EXCEEDED` |
+| 5 | Privacy keyword + org ALLOW=[OpenAI] | Privacy wins → local route; org rule ignored |
+| 6 | Org has both allow and deny | DENY wins; ALLOW further restricts remaining |
+| 7 | User in 2 orgs (future) | Org with most-restrictive policy applies |
+| 8 | Org policy CRUD via UI | Admin UI allows org-admin role to create/edit policies for their org only |
+| 9 | Cross-org isolation | Org admin from orgA cannot see/edit orgB policies |
+| 10 | Policy propagation latency | Within 60s of admin save, all org devices see new policy |
+| 11 | Audit on org policy change | RabbitMQ `routing.org_policy.changed` published |
+
+## Endpoint contract
+
+```http
+POST   /api/v1/routing/policies                  (existing — now accepts orgId in body)
+GET    /api/v1/routing/policies?orgId=X          (existing — filter)
+POST   /api/v1/routing/orgs/:orgId/provider-rules
+GET    /api/v1/routing/orgs/:orgId/provider-rules
+DELETE /api/v1/routing/orgs/:orgId/provider-rules/:id
+PUT    /api/v1/routing/orgs/:orgId/rate-limit
+GET    /api/v1/routing/orgs/:orgId/rate-limit
+```
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/routing/managers/__tests__/policy-resolution.manager.spec.ts
+  - org policy overrides global when both match
+  - org DENY blocks despite ALLOW
+  - privacy keyword beats all
+  - org rate limit returns 429
+  - user with no org uses global only
+  - cross-org isolation in queries
+
+qa/test-routing-r6-multi-tenant.sh
+  - 2 orgs, each with different policies
+  - assert orgA user gets orgA policy, orgB user gets orgB policy
+  - assert rate limit blocks 61st request in 1 minute
+```
+
+## Rollback
+
+`ROUTING_R6_MULTI_TENANT_ENABLED=false` → policy-resolution falls back to existing global-only behavior. Org-tagged policies are still in DB but ignored.
+
+## Risks
+
+| # | Risk | Mitigation |
+|---|------|------------|
+| 1 | Blocker B4: User.orgId not in auth-service | Build this stream after auth-service org work |
+| 2 | Policy resolution complexity creates O(N²) scan | Index `(orgId, isActive, priority)`; cache org policies in Redis 30s |
+| 3 | Org admin in orgA accidentally affects orgB | Strict orgId filter in repository layer; integration test |
+| 4 | Rate limit storage | In-memory sliding window per orgId + Redis backup for distributed |
diff --git a/docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md b/docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md
new file mode 100644
index 00000000..bfd53512
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/08-r7-i18n-non-english.md
@@ -0,0 +1,133 @@
+# Stream 08 — R.7 i18n Non-English Routing
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/08_R7_i18n_non_english_routing.md`
+
+## Mission
+
+Stop treating every message as English. Detect language; route to language-aware models; expand keyword arrays for top non-EN languages; track language confidence on every decision.
+
+## Files to add (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/language-detection/    (NEW MODULE)
+├── language-detection.module.ts
+├── controllers/
+│   └── language-detection.controller.ts                     (POST /routing/detect-language)
+├── services/
+│   └── language-detection.service.ts
+├── managers/
+│   ├── language-classifier.manager.ts                       (uses cld3 or fastext-lite)
+│   └── code-mixed-detector.manager.ts                       (e.g. Arabic + English technical terms)
+├── dto/
+│   └── detect-language.dto.ts
+├── types/
+│   └── language-detection.types.ts
+├── constants/
+│   ├── language-codes.constants.ts                          (ISO-639-1 supported set)
+│   └── language-rtl.constants.ts                            (which languages are RTL)
+└── utilities/
+    └── language-strength-resolver.utility.ts                (reads RouterModelRegistry.languageStrengthJson)
+```
+
+## Schema changes (see PRISMA_FUTURE_MODELS.md)
+
+- `RoutingDecision`: add `detectedLanguage`, `languageConfidence`, `isCodeMixed`
+- `RouterModelRegistry`: add `languageStrengthJson` (per-language quality score)
+
+## Detection pipeline
+
+```
+1. Run cld3 (Compact Language Detector) on message → top-1 language + confidence
+2. If confidence < 0.6 → tag as 'uncertain', use 'en' as fallback
+3. If multi-language (e.g. "كود في Python" Arabic+English) → flag isCodeMixed
+4. Save to RoutingDecision.detectedLanguage + languageConfidence
+5. Pass language to scoring engine → boost candidates with high languageStrength
+6. If RTL language (ar/he/fa/ur) and ROUTING_R7_ARABIC_RTL_HINT_ENABLED → prepend "respond in RTL-friendly format" to system prompt
+```
+
+## Translated keyword arrays
+
+Initial scope: Arabic, Spanish, German (top 3 non-EN per audit user demographics).
+
+```
+apps/claw-routing-service/src/modules/routing/constants/locale-keywords/
+├── ar/coding-keywords.constants.ts             ('كود', 'برمجة', 'وظيفة')
+├── ar/legal-keywords.constants.ts              ('قانون', 'محام', 'عقد')
+├── ar/medical-keywords.constants.ts            ('طب', 'تشخيص', 'دواء')
+├── es/coding-keywords.constants.ts             ('código', 'programación', 'función')
+├── es/legal-keywords.constants.ts              ('legal', 'abogado', 'contrato')
+├── es/medical-keywords.constants.ts            ('médico', 'diagnóstico', 'medicamento')
+├── de/coding-keywords.constants.ts             ('Code', 'Programmierung', 'Funktion')
+├── de/legal-keywords.constants.ts              ('Recht', 'Anwalt', 'Vertrag')
+└── de/medical-keywords.constants.ts            ('Medizin', 'Diagnose', 'Medikament')
+```
+
+Activation: `ROUTING_R7_SUPPORTED_LOCALES=en,ar,es,de` (extend per release).
+
+Roadmap: fr, it, pt, ru, hi, ja, zh-cn (post-stream).
+
+## Acceptance criteria
+
+| # | Test | Expected |
+|---|------|----------|
+| 1 | Arabic message "كود تصنيف صور" | `detectedLanguage=ar`, classified as Coding via AR keywords |
+| 2 | Spanish "Necesito un médico" | `detectedLanguage=es`, classified as Medical via ES keywords |
+| 3 | German "Wie funktioniert dieses Gesetz?" | `detectedLanguage=de`, classified as Legal via DE keywords |
+| 4 | Code-mixed "Write me a 'دالة' in Python" | `isCodeMixed=true`, primary lang detected, both keyword sets scanned |
+| 5 | Arabic + routing prefers `claude-opus-4` (high AR strength) | candidate score boosted; opus wins over gemini-flash |
+| 6 | RTL message → response system prompt includes RTL hint | when flag enabled |
+| 7 | Language confidence < 0.6 | falls through to EN keywords; tag `language_uncertain` |
+| 8 | Unsupported locale (e.g. hi for now) | detected + saved + classifier falls back to EN |
+| 9 | Performance: detection adds <10ms | cld3 is fast; measure p95 |
+
+## Endpoint contract
+
+```http
+POST /api/v1/routing/detect-language
+{ "message": "مرحبا، كيف يمكنني كتابة دالة Python؟" }
+
+200:
+{
+  "detectedLanguage": "ar",
+  "languageConfidence": 0.97,
+  "isCodeMixed": true,
+  "secondaryLanguage": "en",
+  "secondaryConfidence": 0.12,
+  "isRtl": true
+}
+```
+
+## RouterModelRegistry language strength seed
+
+```typescript
+{
+  'claude-opus-4':       { en: 1.0, ar: 0.92, es: 0.95, de: 0.94, fr: 0.94, hi: 0.85, ja: 0.88, zh: 0.85 },
+  'claude-sonnet-4':     { en: 1.0, ar: 0.88, es: 0.93, de: 0.92, fr: 0.92, hi: 0.78, ja: 0.85, zh: 0.82 },
+  'gpt-4o':              { en: 1.0, ar: 0.85, es: 0.95, de: 0.94, fr: 0.94, hi: 0.82, ja: 0.92, zh: 0.92 },
+  'gpt-4o-mini':         { en: 1.0, ar: 0.70, es: 0.90, de: 0.88, fr: 0.88, hi: 0.65, ja: 0.80, zh: 0.78 },
+  'gemini-2.5-flash':    { en: 1.0, ar: 0.78, es: 0.92, de: 0.90, fr: 0.90, hi: 0.80, ja: 0.88, zh: 0.85 },
+  'gemma3:4b':           { en: 0.95, ar: 0.65, es: 0.78, de: 0.75, fr: 0.78 },
+  'qwen3:1.7b':          { en: 0.92, ar: 0.55, es: 0.70, de: 0.68, fr: 0.70, ja: 0.80, zh: 0.95 },
+  'phi4-mini':           { en: 0.95, ar: 0.50, es: 0.72, de: 0.70, fr: 0.72 },
+}
+```
+
+Seeded via `apps/claw-routing-service/prisma/seed/language-strength.seed.ts`.
+
+## Tests
+
+```
+apps/claw-routing-service/src/modules/language-detection/managers/__tests__/language-classifier.manager.spec.ts
+  - 8 languages detected correctly with confidence > 0.8
+  - code-mixed detected
+  - very short message (1 word) returns 'en' fallback
+  - empty message returns null
+
+qa/test-routing-r7-language-detection.sh
+  - 20 prompts in 8 languages, assert detection correct
+  - assert routing biases toward stronger-language models
+```
+
+## Rollback
+
+`ROUTING_R7_LANGUAGE_DETECTION_ENABLED=false` → endpoint returns null; hot path skips detection; `detectedLanguage` left null on new decisions.
diff --git a/docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md b/docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md
new file mode 100644
index 00000000..88bb5bb4
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/09-r8-advanced-intelligence.md
@@ -0,0 +1,148 @@
+# Stream 09 — R.8 Advanced Routing Intelligence
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/09_R8_advanced_routing_intelligence.md`
+
+## Mission
+
+Nine independent sub-features that each make the router smarter. Each ships behind its own flag and can be activated independently.
+
+## Sub-features
+
+### 9.1 — Prompt-length-aware filtering
+
+**Trigger:** routing context arrives with token-count estimate.
+**Behavior:** filter out candidates with `contextWindowTokens < estimatedTokens × 1.2` (20% headroom).
+**Data needed:** existing `RouterModelRegistry.contextWindowTokens` (already there).
+**UI:** in playground score breakdown, show "filtered: insufficient context window".
+**Tests:** 100k-token prompt cannot route to 4k-context model; near-boundary uses cushion.
+**Flag:** `ROUTING_R8_PROMPT_LENGTH_FILTER_ENABLED`.
+**Rollback:** flag off.
+
+### 9.2 — Latency-based circuit breaker
+
+**Trigger:** rolling p95 latency for provider exceeds `ROUTING_R8_LATENCY_CIRCUIT_THRESHOLD_MS` (default 15000).
+**Behavior:** open circuit; same lifecycle as existing failure-based breaker (HALF_OPEN probe → CLOSE on success).
+**Data needed:** existing `RouterCircuitBreaker` table extended with `trigger='LATENCY_P95'` + `latencyThresholdMs` column.
+**UI:** circuit-breaker dashboard (stream R.5) shows trigger reason.
+**Tests:** simulate slow provider → assert circuit opens after threshold breach.
+**Flag:** `ROUTING_R8_LATENCY_CIRCUIT_BREAKER_ENABLED`.
+**Rollback:** flag off; failure-based circuit unchanged.
+
+### 9.3 — Switch model mid-stream
+
+**Trigger:** first SSE chunk takes >`ROUTING_R8_MID_STREAM_FIRST_CHUNK_MAX_MS` (default 5000).
+**Behavior:** kill the in-flight call; immediately reroute to next-best candidate; UI shows "switched to faster model" inline notice.
+**Data needed:** chat-service must support call cancellation mid-stream.
+**Risk:** lose partial output; user sees brief flicker.
+**Flag:** `ROUTING_R8_MID_STREAM_SWITCH_ENABLED`.
+**Rollback:** flag off — stay with original model regardless.
+
+### 9.4 — Fine-tuned model preference per user
+
+**Trigger:** user has a `UserFineTunePreference` row for the detected domain.
+**Behavior:** boost their fine-tune in scoring engine by `weight` × `learnedBiasWeightMax`.
+**Data needed:** new `UserFineTunePreference` table (see PRISMA_FUTURE_MODELS.md 09.4).
+**UI:** `/settings/fine-tunes` — manage own preferences.
+**Tests:** user with fine-tune for "legal" → routes to their fine-tune over claude-opus-4 for legal questions.
+**Flag:** `ROUTING_R8_FINE_TUNE_PREFERENCE_ENABLED`.
+**Rollback:** flag off.
+
+### 9.5 — Region-aware routing
+
+**Trigger:** routing context has `userRegion` (EU/US/APAC) OR org has region preference.
+**Behavior:** prefer regional endpoint (e.g. Bedrock us-east-1 vs eu-west-1) for GDPR compliance + latency win.
+**Data needed:** new `RouterRegionPreference` table + connector-service must expose region endpoints (Blocker B5).
+**Tests:** EU user gets eu-west-1; GDPR-tagged data forces EU even for US-based user.
+**Flag:** `ROUTING_R8_REGION_AWARE_ROUTING_ENABLED`.
+**Rollback:** flag off — default region used.
+
+### 9.6 — Multi-intent splitter
+
+**Trigger:** message has 2+ detected intents (e.g. "code + image", "translate + summarize").
+**Behavior:** split into N parallel routed calls; merge results into a single response.
+**Data needed:** existing `MULTI_INTENT_PRIORITY` constants + chat-service must support parallel-result merging.
+**UI:** message shows N pills, one per sub-intent, with per-sub-result.
+**Tests:** "write code AND a marketing email" → 2 routes; "translate AND summarize" → 2 routes.
+**Flag:** `ROUTING_R8_MULTI_INTENT_SPLITTER_ENABLED`.
+**Rollback:** flag off — picks priority winner only (current behavior).
+
+### 9.7 — Embedding-task routing
+
+**Trigger:** `EMBEDDING` modality detected (Stream R.2).
+**Behavior:** new evaluator branch picks embedding-specific model (text-embedding-3-large / nomic-embed-text / etc.).
+**Data needed:** RouterModelRegistry must include embedding models.
+**Tests:** "embed this paragraph" → text-embedding-3-large; "use local" → nomic-embed-text.
+**Flag:** `ROUTING_R8_EMBEDDING_ROUTING_ENABLED`.
+**Rollback:** flag off — embeddings get classified as text routing (suboptimal but not broken).
+
+### 9.8 — Ensemble consensus mode
+
+**Trigger:** new `RoutingMode.CONSENSUS` OR high-stakes domain (medical/legal) auto-promoted.
+**Behavior:** fire 3 models in parallel; score agreement; return highest-confidence answer + agreement score.
+**Data needed:** chat-service parallel infrastructure (exists for `/chat-messages/parallel`).
+**Tests:** medical question → 3 models agree → high-confidence; 3 disagree → flag as uncertain.
+**Flag:** `ROUTING_R8_CONSENSUS_MODE_ENABLED`.
+**Rollback:** flag off — falls back to JUDGE_PIPELINE or DIRECT_LLM.
+
+### 9.9 — Cost / quality slider
+
+**Trigger:** user has set a slider value 0-100 in `/settings`.
+**Behavior:** scoring engine uses `qualityWeight = slider/100`, `costWeight = 1 - qualityWeight`.
+**Data needed:** new `UserCostQualitySlider` table (see PRISMA_FUTURE_MODELS.md 09.9).
+**UI:** `/settings/routing-preferences` — slider with live preview.
+**Tests:** slider=0 → cheapest; slider=100 → best quality; slider=50 → balanced.
+**Flag:** `ROUTING_R8_COST_QUALITY_SLIDER_ENABLED`.
+**Rollback:** flag off — uses default 0.5 weight.
+
+## Module structure (scaffold included)
+
+```
+apps/claw-routing-service/src/modules/intelligence/        (NEW MODULE — umbrella for 9 sub-features)
+├── intelligence.module.ts
+├── managers/
+│   ├── prompt-length-guard.manager.ts                     (9.1)
+│   ├── latency-circuit-breaker.manager.ts                 (9.2)
+│   ├── mid-stream-switcher.manager.ts                     (9.3)
+│   ├── fine-tune-preference.manager.ts                    (9.4)
+│   ├── region-router.manager.ts                           (9.5)
+│   ├── multi-intent-splitter.manager.ts                   (9.6)
+│   ├── embedding-router.manager.ts                        (9.7)
+│   ├── consensus-mode.manager.ts                          (9.8)
+│   └── cost-quality-slider.manager.ts                     (9.9)
+├── types/
+│   └── intelligence.types.ts
+└── constants/
+    └── intelligence.constants.ts
+```
+
+## Acceptance per sub-feature
+
+| # | Feature | Acceptance |
+|---|---------|------------|
+| 9.1 | Prompt length | 100k-token msg → only 128k+ context candidates |
+| 9.2 | Latency circuit | p95 > 15s for 5 min → circuit OPEN |
+| 9.3 | Mid-stream switch | First chunk > 5s → killed + rerouted |
+| 9.4 | Fine-tune | User with fine-tune for legal → biased toward it |
+| 9.5 | Region | EU user → EU endpoint |
+| 9.6 | Multi-intent | 2 intents → 2 parallel calls |
+| 9.7 | Embedding | Embedding intent → embedding model |
+| 9.8 | Consensus | 3 models fire; agreement scored |
+| 9.9 | Slider | Slider value affects scoring weight |
+
+## Tests + flag matrix
+
+```
+qa/test-routing-r8.1-prompt-length.sh
+qa/test-routing-r8.2-latency-circuit.sh
+qa/test-routing-r8.3-mid-stream-switch.sh
+qa/test-routing-r8.4-fine-tune.sh
+qa/test-routing-r8.5-region.sh
+qa/test-routing-r8.6-multi-intent.sh
+qa/test-routing-r8.7-embedding.sh
+qa/test-routing-r8.8-consensus.sh
+qa/test-routing-r8.9-cost-quality-slider.sh
+```
+
+## Rollback
+
+Each sub-feature has its own flag. Disabling one does NOT affect the others.
diff --git a/docs/15-ai-context/routing-flagship-streams/10-r9-quality-hardening.md b/docs/15-ai-context/routing-flagship-streams/10-r9-quality-hardening.md
new file mode 100644
index 00000000..dba44471
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/10-r9-quality-hardening.md
@@ -0,0 +1,142 @@
+# Stream 10 — R.9 Quality + Reliability Hardening
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/10_R9_quality_reliability_hardening.md`
+
+## Mission
+
+Make the routing service production-grade with real test coverage, regression suites, load tests, drift detection, and release gates.
+
+## Test coverage targets
+
+| Component | Current (est) | Target initial | Target final |
+|-----------|--------------:|---------------:|-------------:|
+| `routing.manager.ts` (1635 LOC) | ~10% | 80% | 92% |
+| `route-evaluator/` v2 | ~30% | 80% | 92% |
+| `scoring/scoring-engine.manager.ts` | ~20% | 80% | 92% |
+| `workflows/workflow-orchestrator.manager.ts` | 0% (scaffold) | 80% (after R.3) | 92% |
+| `classifier/classifier.manager.ts` | ~40% | 85% | 92% |
+| `learning-loop/learning-loop.manager.ts` | ~30% | 80% | 92% |
+| `reliability/circuit-breaker.manager.ts` | ~50% | 85% | 92% |
+| `sync/router-sync.manager.ts` | ~40% | 80% | 92% |
+| `observability/observability.service.ts` | ~30% | 75% | 90% |
+
+## Test artifacts to ship
+
+```
+apps/claw-routing-service/test/
+├── fixtures/
+│   ├── routing-500-prompts.json                    (500-prompt regression set)
+│   ├── multilingual-routing-prompts.json           (top 8 languages, 25 each)
+│   ├── multimodal-routing-prompts.json             (PDF/YouTube/audio/video × 25 each)
+│   ├── replay-confirmed-regressions.json           (promoted suspicious cases)
+│   └── provider-mock-responses.json                (mocks for OpenAI/Anthropic/Gemini/etc.)
+├── load/
+│   ├── routing-decision-load.k6.js                 (k6 script: 100 req/s × 10 min)
+│   ├── routing-decision-spike.k6.js                (spike: 500 req/s × 30s)
+│   └── README.md                                   (how to run, expected p95)
+├── integration/
+│   ├── routing-flagship-end-to-end.spec.ts         (50 prompts through full v2 pipeline)
+│   ├── routing-fallback-chain.spec.ts              (force each provider to fail in turn)
+│   └── routing-canary-comparison.spec.ts           (assert v1 vs v2 outcomes)
+└── regression/
+    ├── routing-regression-runner.ts                (re-runs 500-prompt set; outputs diff)
+    └── routing-drift-detector.ts                   (compares against baseline snapshot)
+```
+
+## QA scripts to add
+
+```
+qa/test-routing-r9-coverage.sh           — npm test:cov; fails if <80%
+qa/test-routing-r9-regression-500.sh     — fires 500 prompts; asserts ≥99% stable outcome
+qa/test-routing-r9-load.sh               — runs k6 load test; asserts p95 < 50ms
+qa/test-routing-r9-drift.sh              — runs drift detector; alerts if accuracy drop > 2%
+qa/test-routing-r9-provider-mock.sh      — runs against mock provider suite (no cloud cost)
+qa/test-routing-r9-migration-rollback.sh — applies + rolls back each migration; asserts no data loss
+```
+
+## CI integration
+
+Add to `.github/workflows/ci.yml`:
+
+```yaml
+jobs:
+  routing-regression:
+    runs-on: ubuntu-latest
+    needs: [build]
+    if: github.event_name == 'pull_request'
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-node@v4
+      - run: cd apps/claw-routing-service && npm run test:cov
+      - run: cd apps/claw-routing-service && bash ../../qa/test-routing-r9-regression-500.sh
+      - name: Block PR on regression
+        run: |
+          REGRESSION_COUNT=$(jq '.bad_regression' apps/claw-routing-service/test/regression/results.json)
+          if [ "$REGRESSION_COUNT" -gt "0" ]; then
+            echo "FAIL: $REGRESSION_COUNT regressions"
+            exit 1
+          fi
+```
+
+## Drift detection
+
+Cron `0 4 * * *` (4 AM daily):
+
+```
+1. Run routing on the 500-prompt baseline fixture set
+2. Compare to last RoutingCalibrationSnapshot
+3. If accuracy delta > 2% → emit `routing.drift.detected` event
+4. Alert via existing audit channel
+5. Generate `claudePrompt` for diagnosis (similar to Replay Lab export)
+```
+
+## Provider mock suite
+
+```
+apps/claw-routing-service/test/mocks/
+├── openai.mock.ts                  (deterministic responses for routing tests)
+├── anthropic.mock.ts
+├── gemini.mock.ts
+├── deepseek.mock.ts
+├── grok.mock.ts
+├── ollama.mock.ts
+└── llamacpp.mock.ts
+```
+
+Used by integration + regression tests so they don't cost real API tokens.
+
+## Release gates
+
+Add `docs/16-quality-engineering/ROUTING_RELEASE_GATE.md`:
+
+```
+RELEASE GATE — routing-service
+
+MUST PASS before any merge to main:
+
+[ ] npm run typecheck     → 0 errors
+[ ] npm run lint          → 0 errors
+[ ] npm run test:cov      → ≥80% on all 4 metrics (statements/branches/functions/lines)
+[ ] qa/test-routing-r9-regression-500.sh → 0 confirmed regressions
+[ ] qa/test-routing-r9-load.sh           → p95 < 50ms
+[ ] qa/test-routing-r9-drift.sh          → no drift alert
+[ ] qa/test-routing-r9-migration-rollback.sh → no data loss
+[ ] Docker logs check                    → 0 UnhandledPromiseRejection
+[ ] CHANGELOG updated
+[ ] Per-stream docs updated for any flag activation
+```
+
+## Acceptance
+
+| # | Test | Expected |
+|---|------|----------|
+| 1 | npm run test:cov in clean repo | ≥80% statements/branches |
+| 2 | Add a routing regression by hand-editing routing.constants.ts | 500-prompt regression suite catches it; PR blocked |
+| 3 | Run k6 load test | p95 routing decision < 50ms at 100 req/s |
+| 4 | Drift detector against frozen baseline | Identifies the change; emits alert |
+| 5 | Add a new migration + rollback test | Both forward + backward succeed without data loss |
+| 6 | Provider mock test (no live cloud) | Tests pass deterministically; runs in CI |
+
+## Rollback
+
+These are quality + CI artifacts, not features. "Rollback" = remove from CI workflow (doesn't affect runtime).
diff --git a/docs/15-ai-context/routing-flagship-streams/11-quick-wins.md b/docs/15-ai-context/routing-flagship-streams/11-quick-wins.md
new file mode 100644
index 00000000..6d87c5a2
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/11-quick-wins.md
@@ -0,0 +1,167 @@
+# Stream 11 — Quick Wins Backlog
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/11_quick_wins_backlog.md`
+
+## Mission
+
+10 single-day tickets that move the needle without requiring full streams. Designed to ship safely without dependencies.
+
+---
+
+## 11.1 — Wire `WorkflowKind` to `RoutingDecision`
+
+**Scope:** add `selectedWorkflow` field to RoutingDecision; for now always set `DIRECT_LLM`. Unlocks future R.3 work + UI badge in stream R.5.
+
+**Files:**
+- `apps/claw-routing-service/prisma/schema.prisma` — add `selectedWorkflow WorkflowKind?`
+- `apps/claw-routing-service/src/modules/routing/managers/routing.manager.ts` — set `DIRECT_LLM` on every decision
+- `apps/claw-frontend/src/types/routing.types.ts` — add field
+- `apps/claw-frontend/src/components/routing/routing-decision-row.tsx` — render badge
+
+**Tests:** existing route tests assert `selectedWorkflow=DIRECT_LLM` on response.
+**Acceptance:** every new RoutingDecision row has non-null `selectedWorkflow`.
+**Rollback:** field is nullable; remove badge from UI; no schema rollback needed.
+**Done checklist:** typecheck + lint + test + visual confirm in `/routing`.
+
+---
+
+## 11.2 — Surface `RoutingDecision.explanation` in chat message header
+
+**Scope:** small `ⓘ` icon next to provider/model badge in chat-message bubble; click expands to existing `explanation` text.
+
+**Files:**
+- `apps/claw-frontend/src/components/chat/message-bubble.tsx` — add icon + popover
+- `apps/claw-frontend/src/hooks/chat/use-message-explanation.ts` — new hook; uses existing decision in props
+- i18n keys in 8 locales for "Why this model?"
+
+**Acceptance:** user clicks ⓘ → sees explanation; closes on outside click.
+**Rollback:** hide icon via flag.
+**Done checklist:** Playwright test for popover open/close + a11y check.
+
+---
+
+## 11.3 — Add `/routing/playground` page (skeleton only)
+
+**Scope:** new page that calls existing `/routing/evaluate` (v1) endpoint with a textarea + submit; renders decision JSON.
+
+**Files:**
+- `apps/claw-frontend/src/app/(portal)/routing/playground/page.tsx`
+- `apps/claw-frontend/src/hooks/routing/use-playground.ts`
+- `apps/claw-frontend/src/components/routing/playground-result-display.tsx`
+- i18n keys
+
+**Acceptance:** operator can paste message → see v1 decision; no chat-model call (no cost).
+**Note:** v2 / Ollama side-by-side from Stream R.5 lands later. This is the skeleton.
+**Rollback:** hide nav entry via flag.
+
+---
+
+## 11.4 — Add YouTube URL detection (regex only, no provider)
+
+**Scope:** detect YouTube URLs in message; set `detectedModality=YOUTUBE_INPUT` on decision; log a TODO line. No actual transcript fetching.
+
+**Files:**
+- `apps/claw-routing-service/src/modules/routing/managers/routing.manager.ts` — single regex check before other detection
+- `apps/claw-routing-service/src/modules/routing/constants/youtube-url.constants.ts` — regex (already scaffolded in R.2)
+
+**Acceptance:** YouTube URL → decision has `detectedModalities` including `YOUTUBE_INPUT`; log line says `youtube_detected_workflow_not_implemented`.
+**Rollback:** remove regex check.
+**Note:** real workflow needs R.2 + R.3.
+
+---
+
+## 11.5 — File-attachment MIME → workflow hint
+
+**Scope:** if request has attachment with `application/pdf` MIME + verb (`summarize|explain`) → set `workflowHint=PDF_EXTRACTION` on decision. No actual PDF extraction wired.
+
+**Files:**
+- `apps/claw-routing-service/src/modules/routing/managers/routing.manager.ts` — single check before AUTO
+- Wire attachments into `RoutingContext` (chat-service change required — see Blocker B1)
+
+**Acceptance:** PDF + summarize → decision has `workflowHint=PDF_EXTRACTION`.
+**Rollback:** remove check.
+**Risk:** Blocker B1 — chat-service must pass attachment metadata in routing context.
+
+---
+
+## 11.6 — Add category filter to `useRoutingDecisions`
+
+**Scope:** add `categoryFilter?: string` query param to `GET /routing/decisions/:threadId`; pipe through to repository.
+
+**Files:**
+- `apps/claw-routing-service/src/modules/routing/controllers/routing.controller.ts` — accept query param
+- `apps/claw-routing-service/src/modules/routing/repositories/routing-decisions.repository.ts` — add WHERE clause
+- `apps/claw-frontend/src/hooks/routing/use-routing-decisions.ts` — pass filter
+- `apps/claw-frontend/src/app/(portal)/routing/page.tsx` — dropdown filter UI
+
+**Acceptance:** filter by "Coding" → only Coding decisions shown.
+**Rollback:** remove filter UI; query param ignored.
+
+---
+
+## 11.7 — Provider+model chart on adaptive-insights
+
+**Scope:** new chart on `/routing/adaptive-insights` showing per-(provider, model) success/failure/latency over 7/30/90-day window.
+
+**Files:**
+- `apps/claw-frontend/src/components/routing/provider-model-chart.tsx` — new component using recharts (already in deps)
+- `apps/claw-frontend/src/app/(portal)/routing/adaptive-insights/page.tsx` — add the chart
+- Backend already exposes this data via existing observability endpoint
+
+**Acceptance:** chart renders with real data; hover shows tooltip.
+**Rollback:** hide chart via flag.
+
+---
+
+## 11.8 — Promote 5 confirmed regressions to test fixtures
+
+**Scope:** operational ticket — pick 5 high-value confirmed regressions from Replay Lab, run "Promote to fixture" on each, commit the generated test code.
+
+**Files:**
+- `apps/claw-routing-service/test/regression/promoted-fixtures.spec.ts` — appended with 5 new test cases
+
+**Acceptance:** 5 new tests in repo; all pass; PR-blocking on regression.
+**Rollback:** revert the commit.
+
+---
+
+## 11.9 — "Rerun decision" button on `/routing` recent list
+
+**Scope:** small refresh icon on each row; clicking re-fires `/routing/evaluate` with the original context; shows old vs new decision in modal.
+
+**Files:**
+- `apps/claw-frontend/src/components/routing/routing-decision-row.tsx` — add button
+- `apps/claw-frontend/src/hooks/routing/use-rerun-decision.ts` — new mutation hook
+- `apps/claw-frontend/src/components/routing/rerun-decision-modal.tsx` — comparison modal
+
+**Acceptance:** click → modal shows old vs new side by side.
+**Rollback:** hide button.
+
+---
+
+## 11.10 — Detect non-English text and tag decision
+
+**Scope:** run cheap character-set detector (no lib needed: count non-ASCII chars / total chars > 0.3 → mark as non-EN); save `detectedLanguage='non-en'` flag on decision. Real language detection comes in R.7.
+
+**Files:**
+- `apps/claw-routing-service/src/modules/routing/utilities/quick-language-detector.utility.ts` — new pure utility
+- `apps/claw-routing-service/src/modules/routing/managers/routing.manager.ts` — call before other detection
+
+**Acceptance:** Arabic/Spanish/German message → decision has `detectedLanguage='non-en'`; baseline data for R.7 work.
+**Rollback:** remove the call.
+
+---
+
+## Activation order (no dependencies)
+
+All 10 are independent; ship in any order. Suggest:
+1. 11.1 (workflow on decision) — unlocks UI badges
+2. 11.6 (category filter) — easy admin win
+3. 11.7 (provider+model chart) — visible improvement
+4. 11.2 ("why this model?") — user-visible
+5. 11.9 (rerun button) — operator win
+6. 11.10 (non-EN tag) — sets up R.7 data
+7. 11.4 (YouTube regex) — sets up R.2 data
+8. 11.3 (playground skeleton) — sets up R.5
+9. 11.8 (promote regressions) — anytime
+10. 11.5 (PDF MIME hint) — blocked on chat-service work
diff --git a/docs/15-ai-context/routing-flagship-streams/12-business-roadmap.md b/docs/15-ai-context/routing-flagship-streams/12-business-roadmap.md
new file mode 100644
index 00000000..b9b10d32
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/12-business-roadmap.md
@@ -0,0 +1,135 @@
+# Stream 12 — Business Positioning + Release Roadmap
+
+**Source prompt:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/12_business_positioning_and_release_roadmap.md`
+
+## Positioning (1-sentence)
+
+> "ClawAI isn't another AI wrapper — it's the intelligence layer that picks the right model and workflow for the task, with cost, privacy, latency, modality, domain awareness, learned feedback, and full transparency."
+
+## Why this matters
+
+Every other "AI platform" today:
+- Hardcodes a single LLM per request
+- Charges per cloud call regardless of need
+- Sends private data to whatever model the user picked, with no privacy enforcement
+- Can't tell PDF from YouTube from spreadsheet — treats every input as text
+- Doesn't learn from feedback — same user keeps getting the same bad answers
+- Has no concept of multi-step workflows (search-first, judge-pipeline, ensemble)
+- Hides routing decisions from operators — debugging "why did it pick X?" is impossible
+
+ClawAI flips all of that.
+
+## Release plan
+
+### Release 1 — Foundation (Streams 11.1 + 11.2 + 01)
+**Goal:** close the learning loop; surface explanations; tag workflow on every decision.
+**Streams:** R.1 + quick wins 11.1, 11.2.
+**Why first:** highest leverage. Data is already collected; the hot path just needs to read it. Visibility unlocks user trust.
+**Demo:** open chat; click ⓘ on a message → see "we picked X because of your thumbs-up history on Y topic". Open `/routing` → see workflow badge.
+**Metrics:**
+- learned-bias activation rate (% of decisions where bias changed the pick)
+- user feedback positive rate (does picking better models actually correlate with positive feedback?)
+
+### Release 2 — Trust + safety (Streams 02 + 06 + 10)
+**Goal:** v2 canary at 5%; playground UI; 500-prompt regression suite blocks bad changes.
+**Streams:** R.1/R.3 canary + R.5 playground + R.9 hardening.
+**Why second:** before exposing more capability, harden the foundation.
+**Demo:** open `/routing/playground` → see v1 + v2 + score breakdown live. Show CI failing on an injected regression.
+**Metrics:**
+- canary regression rate
+- CI block rate (regressions caught before merge)
+- p95 routing decision latency
+
+### Release 3 — Multimodal detection + workflows (Streams 03 + 04 + 11.4 + 11.5 + 11.10)
+**Goal:** detect YouTube/PDF/video/audio/spreadsheet/URL/non-English; SEARCH_FIRST + PDF_EXTRACTION + YOUTUBE_TRANSCRIPT live.
+**Streams:** R.2 + R.3 (subset) + quick wins.
+**Why third:** unlocks new product capabilities (YouTube analysis, PDF Q&A, web search).
+**Demo:** paste YouTube URL → get a summary. Upload PDF + "summarize" → structured extraction.
+**Metrics:**
+- modality detection accuracy
+- per-workflow completion rate
+- workflow fallback rate
+
+### Release 4 — Judge + Compare + Code Review (Stream 04 high-priority workflows)
+**Goal:** JUDGE_PIPELINE for medical/legal; COMPARE_ENSEMBLE for explicit compare; CODE_REVIEW for code blocks.
+**Streams:** R.3 high-priority workflows.
+**Why fourth:** safety-critical workflows; raises quality for high-stakes domains.
+**Demo:** medical question → primary + critic agree → high-confidence answer with judge stamp.
+**Metrics:**
+- judge verified rate
+- ensemble agreement rate
+- code review actionable suggestion rate
+
+### Release 5 — Cost + multi-tenant (Streams 05 + 07)
+**Goal:** per-user budgets; org-scoped policies; org rate limits.
+**Streams:** R.4 + R.6.
+**Why fifth:** enterprise-grade pricing controls.
+**Demo:** user over $10/month budget → routed local automatically; org admin bans OpenAI org-wide.
+**Metrics:**
+- budget enforcement rate
+- org policy propagation latency
+- cost saved per user per month
+
+### Release 6 — Advanced intelligence + i18n (Streams 08 + 09)
+**Goal:** language-aware routing; prompt-length filter; mid-stream switch; embedding routing; consensus; cost/quality slider.
+**Streams:** R.7 + R.8.
+**Why last:** polish + non-English markets.
+**Demo:** Arabic legal question → opus (high AR strength). 100k-token prompt → only 128k+ candidates. Slider at 100 → claude-opus-4; slider at 0 → gemma3:4b local.
+**Metrics:**
+- non-EN routing accuracy
+- mid-stream rescue rate
+- consensus agreement on high-stakes
+
+## Success metrics (cross-release)
+
+| Metric | Today | Target after R6 |
+|--------|------:|---------------:|
+| Route accuracy (vs human-judged "right model") | 99.1% (English keywords only) | 96%+ across 8 languages × 13 workflows |
+| Cost saved per user per month (vs naive routing to gpt-4 always) | unknown | $20+ |
+| Latency saved (vs always-cloud routing) | partial via PRIVACY_FIRST | 30%+ median latency reduction |
+| Judge verified rate (medical/legal) | 0% (no judge) | 92%+ |
+| User feedback positive rate | unknown baseline | +15pp over baseline after R.1 lands |
+| Fallback rescue rate (when primary fails) | wired | 99.5%+ |
+| Workflow completion rate (non-DIRECT_LLM) | 0% (workflows paper-only) | 95%+ |
+| Regression count (per quarter) | unknown | 0 (blocked by CI) |
+| User trust score (survey: "do you understand why X was picked?") | unknown | 75%+ after R.2 |
+| Canary regression rate (v2) | n/a | <1% before promotion to 100% |
+| Cost transparency: % of users who know their monthly spend | 0% | 90%+ after R.5 |
+
+## Sales / marketing collateral implications
+
+This routing platform alone justifies:
+
+| Audience | Pitch |
+|----------|-------|
+| Enterprise CTO | "Stop paying for the wrong models. ClawAI's router enforces your org's policies, budgets, and privacy rules — every call." |
+| Developer | "Stop guessing which model to use. ClawAI's router picks the right one, explains why, and learns from your feedback." |
+| Compliance officer | "Privacy keywords never leave your machine. Org-scoped allow/deny lists enforce data residency. Full audit trail." |
+| Security team | "Circuit breakers + canary deployments + 500-prompt regression suite. We treat routing like production code, not magic." |
+| AI researcher | "Multi-dimensional scoring engine + per-domain learned bias + judge pipeline + ensemble consensus. Open metrics, open replay lab." |
+
+## Public-facing pages to ship alongside this work
+
+- `/landing/routing` — "How ClawAI picks the right model" with live demo
+- `/docs/routing` — public documentation
+- `/blog/routing-deep-dive` — engineering blog (judge pipeline, learning loop)
+- `/case-studies/legal-firm-saves-X-with-routing`
+- `/comparison/clawai-vs-openrouter` — vs raw model gateways
+
+## Long-term differentiation
+
+Even if competitors copy the surface features (workflow kinds, modality detection), the **proprietary learning loop** is the long-term moat:
+
+- Year-over-year, ClawAI's router accumulates per-user / per-org / per-topic preferences
+- New users benefit from cohort priors
+- Cost/quality tradeoffs improve quarterly as model capabilities change
+- Competitors starting fresh need 12+ months of feedback data to match
+
+## Risks to positioning
+
+| Risk | Mitigation |
+|------|------------|
+| Cloud providers ship their own "router" (OpenAI's GPT-4-Turbo auto-routes internally) | ClawAI is multi-cloud + local + workflow-aware. They route within their lineup; we route across all. |
+| Privacy enforcement seen as "feature theater" | Audit logs + open source the privacy keyword list + 3rd-party penetration test public report |
+| "Just use OpenRouter" objection | OpenRouter is a price aggregator. ClawAI is an intelligence layer (workflows, judge, learned bias). Different category. |
+| Bad routing decision goes viral | Surface explanation in every chat message + per-user override in 1 click + CI regression suite prevents recurrence |
diff --git a/docs/15-ai-context/routing-flagship-streams/ENV_ADDITIONS.md b/docs/15-ai-context/routing-flagship-streams/ENV_ADDITIONS.md
new file mode 100644
index 00000000..465483fe
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/ENV_ADDITIONS.md
@@ -0,0 +1,130 @@
+# Environment Variables — Flagship Additions
+
+Block to copy into `.env.example`, `.env`, `scripts/install.sh`, `scripts/install.ps1`. All flags default to `false` / `0` so the production hot path is unchanged until each stream is activated.
+
+```bash
+# =============================================================================
+# Routing Flagship — feature flags (one per stream)
+# =============================================================================
+# All disabled by default. Flip to true / non-zero to activate a stream.
+# Streams have hard dependencies — see docs/15-ai-context/routing-flagship-streams/00-master-plan.md
+
+# R.1 — Close learning loop
+ROUTING_R1_LEARNED_BIAS_ENABLED=false
+ROUTING_R1_LEARNED_BIAS_WEIGHT_MAX=0.3
+ROUTING_R1_MIN_SAMPLE_SIZE=10
+ROUTING_R1_STICKY_THREAD_ROUTE_ENABLED=false
+ROUTING_R1_PREFER_SAME_FAMILY=false
+ROUTING_R1_CONFIDENCE_CALIBRATION_ENABLED=false
+ROUTING_R1_CONFIDENCE_CALIBRATION_WINDOW_DAYS=30
+
+# R.1/R.3 — v2 evaluator canary
+ROUTING_V2_PRIMARY_ENABLED=false
+ROUTING_V2_CANARY_PERCENT=0
+ROUTING_V2_ROLLBACK_SWITCH=true
+ROUTING_V2_REGRESSION_THRESHOLD_PERCENT=1
+ROUTING_V2_COST_INCREASE_THRESHOLD_PERCENT=10
+ROUTING_V2_CONFIDENCE_DROP_THRESHOLD=0.1
+ROUTING_V2_FAILURE_RATE_THRESHOLD_PERCENT=2
+
+# R.2 — Multimodal intent detection
+ROUTING_R2_MODALITY_DETECTION_ENABLED=false
+ROUTING_R2_YOUTUBE_DETECTION_ENABLED=false
+ROUTING_R2_PDF_DETECTION_ENABLED=false
+ROUTING_R2_VIDEO_DETECTION_ENABLED=false
+ROUTING_R2_AUDIO_DETECTION_ENABLED=false
+ROUTING_R2_SPREADSHEET_DETECTION_ENABLED=false
+ROUTING_R2_URL_DETECTION_ENABLED=false
+ROUTING_R2_TOOL_CALLING_FILTER_ENABLED=false
+ROUTING_R2_STREAMING_FILTER_ENABLED=false
+ROUTING_R2_EMBEDDING_ROUTING_ENABLED=false
+
+# R.3 — Workflow orchestrator live workflows
+ROUTING_R3_WORKFLOWS_ENABLED=false
+ROUTING_R3_WORKFLOW_DIRECT_LLM_ENABLED=true              # default workflow — always on
+ROUTING_R3_WORKFLOW_SEARCH_FIRST_ENABLED=false
+ROUTING_R3_WORKFLOW_EXTRACT_FIRST_ENABLED=false
+ROUTING_R3_WORKFLOW_PDF_EXTRACTION_ENABLED=false
+ROUTING_R3_WORKFLOW_YOUTUBE_TRANSCRIPT_ENABLED=false
+ROUTING_R3_WORKFLOW_IMAGE_ANALYSIS_ENABLED=false
+ROUTING_R3_WORKFLOW_IMAGE_GENERATION_ENABLED=true        # already wired
+ROUTING_R3_WORKFLOW_VIDEO_ANALYSIS_ENABLED=false
+ROUTING_R3_WORKFLOW_AUDIO_TRANSCRIBE_ENABLED=false
+ROUTING_R3_WORKFLOW_FILE_GENERATION_ENABLED=true         # already wired
+ROUTING_R3_WORKFLOW_CODE_REVIEW_ENABLED=false
+ROUTING_R3_WORKFLOW_COMPARE_ENSEMBLE_ENABLED=false
+ROUTING_R3_WORKFLOW_JUDGE_PIPELINE_ENABLED=false
+
+# R.4 — Cost budget intelligence
+ROUTING_R4_COST_BUDGET_ENABLED=false
+ROUTING_R4_DEFAULT_USER_MONTHLY_CAP_USD=0                # 0 = no default cap
+ROUTING_R4_DEFAULT_ORG_MONTHLY_CAP_USD=0
+ROUTING_R4_BUDGET_WARN_AT_PERCENT=80
+ROUTING_R4_FORCE_LOCAL_WHEN_OVER_BUDGET=true
+ROUTING_R4_FREE_TIER_AWARENESS_ENABLED=false             # blocked on connector-service work
+
+# R.5 — Operator playground + transparency
+ROUTING_R5_PLAYGROUND_ENABLED=false
+ROUTING_R5_EXPLANATION_IN_CHAT_ENABLED=false
+ROUTING_R5_TAXONOMY_ADMIN_UI_ENABLED=false
+ROUTING_R5_CIRCUIT_BREAKER_DASHBOARD_ENABLED=false
+
+# R.6 — Multi-tenant fleet routing
+ROUTING_R6_MULTI_TENANT_ENABLED=false
+ROUTING_R6_ORG_PROVIDER_RULES_ENABLED=false
+ROUTING_R6_ORG_RATE_LIMIT_ENABLED=false
+
+# R.7 — i18n non-English routing
+ROUTING_R7_LANGUAGE_DETECTION_ENABLED=false
+ROUTING_R7_LANGUAGE_AWARE_MODEL_SELECTION_ENABLED=false
+ROUTING_R7_CODE_MIXED_DETECTION_ENABLED=false
+ROUTING_R7_ARABIC_RTL_HINT_ENABLED=false
+ROUTING_R7_SUPPORTED_LOCALES=en,ar,de,es,fr,it,pt,ru
+
+# R.8 — Advanced routing intelligence (each sub-feature gated)
+ROUTING_R8_PROMPT_LENGTH_FILTER_ENABLED=false
+ROUTING_R8_LATENCY_CIRCUIT_BREAKER_ENABLED=false
+ROUTING_R8_LATENCY_CIRCUIT_THRESHOLD_MS=15000
+ROUTING_R8_MID_STREAM_SWITCH_ENABLED=false
+ROUTING_R8_MID_STREAM_FIRST_CHUNK_MAX_MS=5000
+ROUTING_R8_FINE_TUNE_PREFERENCE_ENABLED=false
+ROUTING_R8_REGION_AWARE_ROUTING_ENABLED=false
+ROUTING_R8_MULTI_INTENT_SPLITTER_ENABLED=false
+ROUTING_R8_EMBEDDING_ROUTING_ENABLED=false
+ROUTING_R8_CONSENSUS_MODE_ENABLED=false
+ROUTING_R8_COST_QUALITY_SLIDER_ENABLED=false
+
+# R.9 — Quality + reliability hardening
+ROUTING_R9_REGRESSION_SUITE_ENABLED=false                # CI flag
+ROUTING_R9_DRIFT_DETECTION_ENABLED=false
+ROUTING_R9_DRIFT_ALERT_THRESHOLD_PERCENT=2
+ROUTING_R9_CALIBRATION_SNAPSHOT_CRON="0 4 * * *"
+```
+
+## Activation order (matches dependency graph)
+
+```
+1. ROUTING_R1_LEARNED_BIAS_ENABLED=true
+2. ROUTING_V2_PRIMARY_ENABLED=true + ROUTING_V2_CANARY_PERCENT=5
+3. ROUTING_R3_WORKFLOWS_ENABLED=true (per-workflow flags individually)
+4. ROUTING_R2_MODALITY_DETECTION_ENABLED=true + per-modality flags
+5. ROUTING_R4_COST_BUDGET_ENABLED=true
+6. ROUTING_R5_* (no production risk — safe to enable in any order)
+7. ROUTING_R7_LANGUAGE_DETECTION_ENABLED=true
+8. ROUTING_R6_MULTI_TENANT_ENABLED=true (after auth-service org work)
+9. ROUTING_R8_* (per sub-feature)
+10. ROUTING_R9_* (CI/ops flags — enable in CI first)
+```
+
+## Rollback
+
+```bash
+# Disable a stream instantly:
+sed -i 's/ROUTING_R1_LEARNED_BIAS_ENABLED=true/ROUTING_R1_LEARNED_BIAS_ENABLED=false/' .env
+./scripts/claw.sh restart routing-service
+# < 10 seconds, no DB migration needed.
+```
+
+## Reading flags
+
+All flags MUST be read via `AppConfig` (Zod-validated). NEVER `process.env.ROUTING_*` directly. See `apps/claw-routing-service/src/app/config/app.config.ts` for the existing pattern.
diff --git a/docs/15-ai-context/routing-flagship-streams/INDEX.md b/docs/15-ai-context/routing-flagship-streams/INDEX.md
new file mode 100644
index 00000000..d6dcb113
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/INDEX.md
@@ -0,0 +1,75 @@
+# Routing Flagship — Scaffolding Index
+
+**Branch:** `feature/routing-flagship-implementation`
+**Source plan:** `plan-prompts/ClawAI_routing_implementation_flagship_pack/`
+**Audit basis:** `docs/15-ai-context/routing-system-audit.md`
+
+**Scope of this branch:** *Scaffolding only* — per-stream module folders, type/constant stubs, doc per stream listing what's required, env-var stubs, future-model markers. No production wiring, no tests, no DB migrations, no nginx changes. Each stream doc is the single source of truth for what an implementing agent needs to do.
+
+## Streams in execution order
+
+| # | Stream | Doc | New modules scaffolded |
+|---|--------|-----|------------------------|
+| 00 | Master plan (meta) | [00-master-plan.md](./00-master-plan.md) | — |
+| 01 | R.1 Close learning loop | [01-r1-learning-loop.md](./01-r1-learning-loop.md) | extension on `routing/` (learned-bias util + manager) |
+| 02 | R.1/R.3 v2 evaluator canary | [02-r1r3-v2-evaluator-canary.md](./02-r1r3-v2-evaluator-canary.md) | extension on `route-evaluator/` (canary-bucket) |
+| 03 | R.2 Multimodal intent detection | [03-r2-multimodal-intent-detection.md](./03-r2-multimodal-intent-detection.md) | NEW module `modality-detection/` |
+| 04 | R.3 Workflow orchestrator goes live | [04-r3-workflow-orchestrator.md](./04-r3-workflow-orchestrator.md) | extension on `workflows/` (13 workflow handler stubs) |
+| 05 | R.4 Cost budget intelligence | [05-r4-cost-budget-intelligence.md](./05-r4-cost-budget-intelligence.md) | NEW module `cost-budget/` |
+| 06 | R.5 Operator playground + transparency | [06-r5-operator-playground.md](./06-r5-operator-playground.md) | NEW module `playground/`, FE pages |
+| 07 | R.6 Multi-tenant fleet routing | [07-r6-multi-tenant-fleet.md](./07-r6-multi-tenant-fleet.md) | extension on `routing/policies` (orgId, propagation) |
+| 08 | R.7 i18n non-English routing | [08-r7-i18n-non-english.md](./08-r7-i18n-non-english.md) | NEW module `language-detection/` |
+| 09 | R.8 Advanced routing intelligence | [09-r8-advanced-intelligence.md](./09-r8-advanced-intelligence.md) | NEW module `intelligence/` (9 sub-features) |
+| 10 | R.9 Quality + reliability hardening | [10-r9-quality-hardening.md](./10-r9-quality-hardening.md) | test scaffolds + QA scripts |
+| 11 | Quick wins backlog | [11-quick-wins.md](./11-quick-wins.md) | 10 single-day tickets |
+| 12 | Business positioning + release roadmap | [12-business-roadmap.md](./12-business-roadmap.md) | docs only |
+
+## Module scaffolds in this branch
+
+Each new module folder contains: `module.ts`, `controllers/`, `services/`, `managers/`, `repositories/`, `dto/`, `types/`, `constants/`, `enums/` (where applicable). All methods are stubs that throw `NotImplementedError('SCAFFOLD-Rx')` so the module is discoverable but inert.
+
+```
+apps/claw-routing-service/src/modules/
+├── modality-detection/   (Stream 03 — R.2)
+├── cost-budget/          (Stream 05 — R.4)
+├── playground/           (Stream 06 — R.5)
+├── language-detection/   (Stream 08 — R.7)
+└── intelligence/         (Stream 09 — R.8)
+```
+
+Extensions on existing modules are documented in their per-stream docs; no skeleton files are added to existing modules to avoid conflicts with the other agent on `main`.
+
+## Future Prisma models (scaffold marker — NOT yet in schema.prisma)
+
+Listed in [PRISMA_FUTURE_MODELS.md](./PRISMA_FUTURE_MODELS.md). Add via migration when implementing each stream.
+
+## New environment variables
+
+Listed in [ENV_ADDITIONS.md](./ENV_ADDITIONS.md). Copy block into `.env.example` + `.env` + `scripts/install.{sh,ps1}` when activating a stream.
+
+## Conventions used in this scaffold
+
+1. Every stub method body: `throw new Error('SCAFFOLD-Rx — not implemented; see docs/15-ai-context/routing-flagship-streams/0X-name.md');`
+2. Every stub file has a top-of-file comment: `// SCAFFOLD: stream Rx (NN-name) — replace this stub with real implementation before activation.`
+3. Modules are NOT registered in `app.module.ts` — they're discoverable but won't load at runtime.
+4. Prisma changes live in `PRISMA_FUTURE_MODELS.md` until activation — schema.prisma is untouched.
+5. i18n keys live in per-stream doc tables, not in locale files yet — adding them now without backend support would leak fake-feature strings to users.
+
+## Activation sequence (recommendation)
+
+Per the audit's leverage analysis, activate in this order:
+
+```
+01-r1-learning-loop          →  the single biggest win
+02-r1r3-v2-evaluator-canary  →  unlocks the rest
+04-r3-workflow-orchestrator  →  needs 01 + 02
+03-r2-multimodal-detection   →  needs 04 for routing the detected intents
+05-r4-cost-budget            →  independent
+06-r5-playground             →  no production risk; can run any time
+08-r7-i18n                   →  user-visible win
+07-r6-multi-tenant           →  blocked on org schema
+09-r8-advanced               →  pick individual sub-features
+10-r9-quality                →  ongoing alongside everything
+```
+
+Quick wins (Stream 11) can ship out-of-order; they're explicitly designed to be safe.
diff --git a/docs/15-ai-context/routing-flagship-streams/PRISMA_FUTURE_MODELS.md b/docs/15-ai-context/routing-flagship-streams/PRISMA_FUTURE_MODELS.md
new file mode 100644
index 00000000..d48abace
--- /dev/null
+++ b/docs/15-ai-context/routing-flagship-streams/PRISMA_FUTURE_MODELS.md
@@ -0,0 +1,243 @@
+# Prisma — Future Models (scaffolded, not yet applied)
+
+These models are scaffolded by the flagship pack but **NOT yet present in `schema.prisma`**. Add via migration when activating the relevant stream.
+
+---
+
+## R.4 — Cost budget (Stream 05)
+
+```prisma
+enum CostBudgetScope {
+  USER
+  ORG
+}
+
+enum CostBudgetStatus {
+  OK
+  WARN
+  EXCEEDED
+}
+
+model UserCostBudget {
+  id                String           @id @default(cuid())
+  scope             CostBudgetScope
+  ownerId           String           @map("owner_id")      // userId or orgId
+  monthlyCapUsd     Decimal          @map("monthly_cap_usd") @db.Decimal(12, 4)
+  currentSpendUsd   Decimal          @default(0) @map("current_spend_usd") @db.Decimal(12, 4)
+  status            CostBudgetStatus @default(OK)
+  warnAtPercent     Int              @default(80) @map("warn_at_percent")
+  overrideAllowed   Boolean          @default(false) @map("override_allowed")
+  resetAt           DateTime         @map("reset_at")
+  lastWarningSentAt DateTime?        @map("last_warning_sent_at")
+  createdAt         DateTime         @default(now()) @map("created_at")
+  updatedAt         DateTime         @updatedAt @map("updated_at")
+
+  @@unique([scope, ownerId])
+  @@index([status])
+  @@index([resetAt])
+  @@map("user_cost_budgets")
+}
+```
+
+**Backfill:** none — empty table on activation. First budget check returns OK if no row exists.
+
+---
+
+## R.6 — Multi-tenant fleet (Stream 07)
+
+```prisma
+// Modify existing RoutingPolicy: add orgId column.
+
+model RoutingPolicy {
+  id          String      @id @default(cuid())
+  name        String
+  routingMode RoutingMode @map("routing_mode")
+  priority    Int         @default(0)
+  isActive    Boolean     @default(true) @map("is_active")
+  config      Json
+  orgId       String?     @map("org_id")  // NEW — null = global policy
+  createdAt   DateTime    @default(now()) @map("created_at")
+  updatedAt   DateTime    @updatedAt @map("updated_at")
+
+  @@index([routingMode])
+  @@index([isActive])
+  @@index([priority])
+  @@index([orgId, isActive, priority])  // NEW
+  @@map("routing_policies")
+}
+
+// NEW table for org-scoped allow/deny lists
+
+enum OrgProviderRuleKind {
+  ALLOW
+  DENY
+}
+
+model OrgProviderRule {
+  id        String              @id @default(cuid())
+  orgId     String              @map("org_id")
+  provider  String
+  kind      OrgProviderRuleKind
+  reason    String?
+  createdAt DateTime            @default(now()) @map("created_at")
+
+  @@unique([orgId, provider])
+  @@index([orgId, kind])
+  @@map("org_provider_rules")
+}
+
+// NEW table for org rate limits
+
+model OrgRateLimit {
+  id                  String   @id @default(cuid())
+  orgId               String   @unique @map("org_id")
+  requestsPerMinute   Int      @map("requests_per_minute")
+  burst               Int      @default(0)
+  createdAt           DateTime @default(now()) @map("created_at")
+  updatedAt           DateTime @updatedAt @map("updated_at")
+
+  @@map("org_rate_limits")
+}
+```
+
+**Backfill:** `UPDATE routing_policies SET org_id = NULL;` (already null on creation — preserves existing policies as global).
+
+---
+
+## R.7 — Language detection (Stream 08)
+
+```prisma
+// Modify existing RoutingDecision: add language columns.
+
+model RoutingDecision {
+  // ... existing fields
+  detectedLanguage    String?  @map("detected_language")     // ISO-639-1 (e.g. 'en', 'ar', 'es')
+  languageConfidence  Decimal? @map("language_confidence") @db.Decimal(5, 4)
+  isCodeMixed         Boolean  @default(false) @map("is_code_mixed")
+  // ... existing fields
+}
+
+// Extend RouterModelRegistry with per-language strength.
+
+model RouterModelRegistry {
+  // ... existing fields
+  languageStrengthJson Json?  @map("language_strength_json")
+  // Shape: { "en": 1.0, "ar": 0.85, "es": 0.9, ... }
+  // ... existing fields
+}
+```
+
+**Backfill:** none. Existing decisions remain `detectedLanguage=NULL`.
+
+---
+
+## R.2/R.3 — Modality + workflow on RoutingDecision (Streams 03, 04)
+
+```prisma
+// Modify existing RoutingDecision: add modality + workflow columns.
+
+model RoutingDecision {
+  // ... existing fields
+  detectedModalities  Json?           @map("detected_modalities")
+  // Shape: ["TEXT", "PDF_INPUT", "YOUTUBE_INPUT"]
+  selectedWorkflow    WorkflowKind?   @map("selected_workflow")
+  workflowConfidence  Decimal?        @map("workflow_confidence") @db.Decimal(5, 4)
+  // ... existing fields
+}
+```
+
+**Backfill:** none.
+
+---
+
+## R.8 — Advanced intelligence (Stream 09 — one sub-feature per migration)
+
+### 09.1 Prompt-length filtering — no new model, uses RouterModelRegistry.contextWindowTokens (already exists)
+
+### 09.2 Latency circuit breaker — extends existing RouterCircuitBreaker
+
+```prisma
+model RouterCircuitBreaker {
+  // ... existing fields
+  trigger              String   @default("FAILURE_RATE")  // 'FAILURE_RATE' | 'LATENCY_P95'
+  latencyThresholdMs   Int?     @map("latency_threshold_ms")
+  // ... existing fields
+}
+```
+
+### 09.4 Fine-tuned model preference
+
+```prisma
+model UserFineTunePreference {
+  id        String   @id @default(cuid())
+  userId    String   @map("user_id")
+  domain    DomainTag
+  provider  String
+  model     String
+  weight    Decimal  @default(0.5) @map("weight") @db.Decimal(3, 2)
+  createdAt DateTime @default(now()) @map("created_at")
+
+  @@unique([userId, domain])
+  @@index([userId])
+  @@map("user_fine_tune_preferences")
+}
+```
+
+### 09.5 Region preference
+
+```prisma
+model RouterRegionPreference {
+  id            String   @id @default(cuid())
+  scope         String   // 'USER' | 'ORG' | 'GLOBAL'
+  ownerId       String?  @map("owner_id")
+  provider      String
+  preferredRegion String @map("preferred_region")
+  reason        String?  // 'GDPR' | 'LATENCY' | 'COST'
+  createdAt     DateTime @default(now()) @map("created_at")
+
+  @@index([scope, ownerId, provider])
+  @@map("router_region_preferences")
+}
+```
+
+### 09.9 Cost/quality slider
+
+```prisma
+model UserCostQualitySlider {
+  userId             String   @id @map("user_id")
+  qualityWeight      Decimal  @default(0.5) @map("quality_weight") @db.Decimal(3, 2)  // 0 = cheapest, 1 = best
+  updatedAt          DateTime @updatedAt @map("updated_at")
+
+  @@map("user_cost_quality_sliders")
+}
+```
+
+---
+
+## Apply order
+
+When activating a stream, run migrations in this order to avoid FK conflicts:
+
+```bash
+# R.4
+npx prisma migrate dev --name r4_add_user_cost_budget
+
+# R.6
+npx prisma migrate dev --name r6_add_org_id_to_routing_policy
+npx prisma migrate dev --name r6_add_org_provider_rules_and_rate_limit
+
+# R.7
+npx prisma migrate dev --name r7_add_language_to_routing_decision
+npx prisma migrate dev --name r7_add_language_strength_to_model_registry
+
+# R.2/R.3
+npx prisma migrate dev --name r2_r3_add_modality_workflow_to_decision
+
+# R.8 (per sub-feature)
+npx prisma migrate dev --name r8_2_extend_circuit_breaker_latency
+npx prisma migrate dev --name r8_4_add_user_fine_tune_preference
+npx prisma migrate dev --name r8_5_add_router_region_preference
+npx prisma migrate dev --name r8_9_add_user_cost_quality_slider
+```
+
+All migrations are **additive** — no destructive changes, all new columns nullable or have defaults.