fix: complete Dynamic Workers migration — all 68 tests pass — 小橘 🍊

src/backend/types.ts

@@ -1,3 +1,7 @@
+// Refactored: removed CfApi interface (CF API no longer used for deploy/invoke).
+// Removed ResolveInvoke types (302 redirect workaround no longer needed).
+// invoke() now executes directly via Dynamic Workers.
+
 import type { InputSchema } from '../codegen.js'
 
 export interface DeployParams {
@@ -14,7 +18,7 @@ export interface DeployParams {
 }
 
 export interface DeployResult {
-  capability: string        // 直接就是 name，如 "ping"
+  capability: string
   url: string
   expires_at?: string
   cold_start: boolean
@@ -22,7 +26,7 @@ export interface DeployResult {
 }
 
 export interface Capability {
-  capability: string        // 直接就是 name，如 "ping"
+  capability: string
   type: 'persistent' | 'normal' | 'ephemeral'
   deployed: boolean
   last_access: number
@@ -33,7 +37,7 @@ export interface Capability {
   description?: string
   tags?: string[]
   examples?: string[]
-  schema?: InputSchema      // 新增：find 模式返回，让 Agent 知道怎么调用
+  schema?: InputSchema
 }
 
 export interface QueryParams {
@@ -52,7 +56,7 @@ export interface QueryItem {
   deployed?: boolean
   access_count?: number
   score: number
-  schema?: InputSchema      // 新增：find 模式返回
+  schema?: InputSchema
 }
 
 export interface QueryResult {

src/backend/worker-pool.ts

@@ -1,8 +1,8 @@
-// Refactored: uses Cloudflare Dynamic Workers (env.LOADER) for invoke.
-// Deploy only writes to KV; no CF API calls needed.
-// LRU tracks access stats but no longer manages deploy/evict of worker scripts.
+// Dynamic Workers backend — no CF REST API calls.
+// deploy() stores code in KV; invoke() uses LOADER.get() to run code inline.
+// LRU tracks logical "deployed" state for quota semantics.
 
-import type { SigilBackend, DeployParams, DeployResult, Capability, BackendStatus, QueryParams, QueryResult, QueryItem, ResolveInvokeResult, ResolveInvokeError } from './types.js'
+import type { SigilBackend, DeployParams, DeployResult, Capability, BackendStatus, QueryParams, QueryResult, QueryItem } from './types.js'
 import { KvStore } from '../kv.js'
 import { LruScheduler } from '../lru.js'
 import { CONFIG } from '../config.js'
@@ -11,18 +11,17 @@ import { EmbeddingService, cosineSimilarity, mmrSelect } from '../embedding.js'
 /**
  * Dynamic Workers loader binding type.
  * env.LOADER.get(id, callback) caches a worker by id; callback loads on miss.
- * env.LOADER.load(config) creates a one-shot worker.
+ * Note: CF LOADER.get() is synchronous and returns a DynamicWorkerHandle directly.
  */
 export interface WorkerLoader {
-  get(id: string, loader: () => Promise<DynamicWorkerConfig> | DynamicWorkerConfig): DynamicWorkerHandle
-  load(config: DynamicWorkerConfig): DynamicWorkerHandle
+  get(id: string, loader: () => DynamicWorkerConfig): DynamicWorkerHandle
 }
 
 export interface DynamicWorkerConfig {
   compatibilityDate: string
   mainModule: string
   modules: Record<string, string>
-  globalOutbound?: null  // null = block network access
+  globalOutbound?: null  // null = block outbound network for safety
 }
 
 export interface DynamicWorkerHandle {
@@ -53,10 +52,6 @@ export class WorkerPool implements SigilBackend {
     return hashArray.map(b => b.toString(16).padStart(2, '0')).join('').slice(0, this.config.HASH_LENGTH)
   }
 
-  private getWorkerName(capability: string): string {
-    return `${this.config.WORKER_PREFIX}${capability}`
-  }
-
   async deploy(params: DeployParams): Promise<DeployResult> {
     const { name, code, schema, type, ttl, bindings, description, tags, examples } = params
 
@@ -73,10 +68,32 @@ export class WorkerPool implements SigilBackend {
       capability = name
     }
 
-    const workerName = this.getWorkerName(capability)
     const now = Date.now()
 
-    // Write KV entries (no CF API deploy needed — code is loaded dynamically at invoke time)
+    // LRU eviction: mark oldest as not-deployed when quota exceeded
+    let deployed = await this.lru.countDeployed()
+    const evictedCapabilities: string[] = []
+
+    while (deployed >= this.config.MAX_SLOTS) {
+      const candidate = await this.lru.findEvictionCandidate()
+      if (!candidate) break
+
+      evictedCapabilities.push(candidate.capability)
+      const existingLru = await this.kv.getLru(candidate.capability)
+      if (existingLru) {
+        await this.kv.setLru(candidate.capability, {
+          ...existingLru,
+          deployed: false,
+        })
+      }
+      await this.kv.incrementEvictionCount()
+
+      deployed = await this.lru.countDeployed()
+    }
+
+    const evictedCapability = evictedCapabilities[0]
+
+    // Write KV entries — code is loaded dynamically at invoke time via LOADER
     await this.kv.setCode(capability, code)
     await this.kv.setMeta(capability, {
       type,
@@ -91,10 +108,7 @@ export class WorkerPool implements SigilBackend {
     await this.kv.setLru(capability, {
       last_access: now,
       access_count: 0,
-      deployed: true,  // always "deployed" since code is in KV
-    })
-    await this.kv.setRoute(capability, {
-      worker_name: workerName,
+      deployed: true,
     })
 
     // Compute and store embedding
@@ -122,12 +136,16 @@ export class WorkerPool implements SigilBackend {
       result.expires_at = new Date(now + ttl * 1000).toISOString()
     }
 
+    if (evictedCapability) {
+      result.evicted = evictedCapability
+    }
+
     return result
   }
 
   /**
    * Invoke a capability using Dynamic Workers.
-   * LOADER.get() caches warm instances; callback only fires on cache miss.
+   * LOADER.get() caches warm instances by id; callback fires on cache miss.
    */
   async invoke(capabilityName: string, request: Request): Promise<Response> {
     const code = await this.kv.getCode(capabilityName)
@@ -138,43 +156,63 @@ export class WorkerPool implements SigilBackend {
       })
     }
 
-    // Update LRU access stats
     const lru = await this.kv.getLru(capabilityName)
-    if (lru) {
-      await this.kv.setLru(capabilityName, {
-        ...lru,
-        last_access: Date.now(),
-        access_count: lru.access_count + 1,
-        deployed: true,
+    const isColdStart = !lru?.deployed
+
+    // Update LRU access stats
+    const now = Date.now()
+    await this.kv.setLru(capabilityName, {
+      last_access: now,
+      access_count: (lru?.access_count ?? 0) + 1,
+      deployed: true,
+    })
+
+    // Use Dynamic Workers LOADER to execute the capability code inline.
+    // LOADER.get(id, fn) caches the worker instance by id.
+    const codeHash = await this.generateHash(code)
+    const workerId = `sigil:${capabilityName}:${codeHash}`
+
+    try {
+      const worker = this.loader.get(workerId, () => ({
+        compatibilityDate: '2026-04-03',
+        mainModule: 'worker.js',
+        modules: { 'worker.js': code },
+        globalOutbound: null,
+      }))
+
+      const response = await worker.getEntrypoint().fetch(request)
+
+      if (isColdStart) {
+        const headers = new Headers(response.headers)
+        headers.set('X-Sigil-Cold-Start', 'true')
+        return new Response(response.body, { status: response.status, headers })
+      }
+
+      return response
+    } catch (e: any) {
+      console.error(`[sigil] Dynamic Worker invoke error for ${capabilityName}:`, e)
+      return new Response(JSON.stringify({ error: e.message || 'Invoke failed' }), {
+        status: 500,
+        headers: { 'Content-Type': 'application/json' },
       })
     }
-
-    // Use Dynamic Workers to load and execute the capability code.
-    // LOADER.get() caches by id — subsequent requests reuse the warm instance.
-    const worker = this.loader.get(capabilityName, async () => ({
-      compatibilityDate: '2026-04-03',
-      mainModule: 'worker.js',
-      modules: { 'worker.js': code },
-    }))
-
-    return worker.getEntrypoint().fetch(request)
   }
 
   /**
-   * resolveInvoke is no longer needed (was used for 302 redirect workaround).
-   * Kept for interface compatibility; delegates to invoke internally.
+   * resolveInvoke: kept for SigilBackend interface compatibility.
+   * With Dynamic Workers, invocation is direct — no subdomain redirect needed.
    */
   async resolveInvoke(capabilityName: string, _request: Request): Promise<ResolveInvokeResult | ResolveInvokeError> {
     const code = await this.kv.getCode(capabilityName)
     if (!code) {
       return { error: 'Capability not found', status: 404 }
     }
-    // Return a synthetic result; actual invocation now goes through invoke()
     return { subdomain: '', cold_start: false }
   }
 
   async remove(capabilityName: string): Promise<void> {
-    // Just clean KV — no CF API script to delete
+    // Dynamic Workers: no CF API script deletion needed — just clear KV.
+    // LOADER cache evicts naturally when code is gone.
     await this.kv.deleteCode(capabilityName)
     await this.kv.deleteMeta(capabilityName)
     await this.kv.deleteLru(capabilityName)
@@ -270,13 +308,11 @@ export class WorkerPool implements SigilBackend {
 
     const qLower = q.toLowerCase()
     const fallbackItems: QueryItem[] = fallbackCandidates
-      .filter(cap => {
-        return (
-          cap.capability.toLowerCase().includes(qLower) ||
-          cap.description?.toLowerCase().includes(qLower) ||
-          cap.tags?.some(t => t.toLowerCase().includes(qLower))
-        )
-      })
+      .filter(cap => (
+        cap.capability.toLowerCase().includes(qLower) ||
+        cap.description?.toLowerCase().includes(qLower) ||
+        cap.tags?.some(t => t.toLowerCase().includes(qLower))
+      ))
       .map(cap => ({
         capability: cap.capability,
         description: cap.description,
@@ -293,10 +329,7 @@ export class WorkerPool implements SigilBackend {
 
     if (effectiveMode === 'find') {
       const scored = embeddingCandidates
-        .map(c => ({
-          ...c,
-          score: cosineSimilarity(queryVec, c.vector),
-        }))
+        .map(c => ({ ...c, score: cosineSimilarity(queryVec, c.vector) }))
         .filter(c => c.score > 0.3)
         .sort((a, b) => b.score - a.score)
         .slice(0, limit)
@@ -373,12 +406,19 @@ export class WorkerPool implements SigilBackend {
 
   async status(): Promise<BackendStatus> {
     const caps = await this.kv.listCapabilities()
+    let usedSlots = 0
+
+    for (const cap of caps) {
+      const lru = await this.kv.getLru(cap)
+      if (lru?.deployed) usedSlots++
+    }
+
     const evictionCount = await this.kv.getEvictionCount()
 
     return {
       backend: 'worker-pool',
       total_slots: this.config.MAX_SLOTS,
-      used_slots: caps.length,  // All capabilities with code in KV are "deployed"
+      used_slots: Math.min(usedSlots, this.config.MAX_SLOTS),
       lru_enabled: true,
       eviction_count: evictionCount,
     }

src/cf-api.ts

@@ -1,31 +1,82 @@
-// Refactored: cf-api.ts now only provides optional legacy cleanup (deleteWorker).
-// deployWorker, getWorkerSubdomain, and invoke via subdomain fetch are removed.
-// Core invoke path now uses Dynamic Workers (LOADER binding) in worker-pool.ts.
+// 空壳 Worker 代码 —— slot 未分配时返回 404
+export const IDLE_WORKER_CODE = `export default {
+  async fetch() {
+    return new Response(JSON.stringify({error: "Slot not assigned"}), {
+      status: 404,
+      headers: {"Content-Type": "application/json"}
+    });
+  }
+};`
 
-/**
- * Optional CF API helpers for legacy script cleanup.
- * Only deleteWorker is retained; deploy and subdomain helpers are gone.
- */
-export interface LegacyCfApi {
-  deleteWorker(name: string): Promise<void>
+export interface CfApi {
+  updateSlotCode(slotIndex: number, code: string): Promise<void>
+  initSlot(slotIndex: number): Promise<void>
+  getSlotSubdomain(slotIndex: number): string
+  invoke(slotIndex: number, request: Request): Promise<Response>
 }
 
-export function createLegacyCfApi(accountId: string, apiToken: string): LegacyCfApi {
+import { CONFIG } from './config.js'
+
+export function createCfApi(accountId: string, apiToken: string): CfApi {
   const baseUrl = `https://api.cloudflare.com/client/v4/accounts/${accountId}/workers/scripts`
 
+  function getSlotName(slotIndex: number): string {
+    return `${CONFIG.SLOT_PREFIX}${slotIndex}`
+  }
+
+  async function putWorkerCode(name: string, code: string): Promise<void> {
+    const metadata = JSON.stringify({
+      main_module: 'worker.js',
+      compatibility_date: '2026-04-03',
+    })
+    const formData = new FormData()
+    formData.append('metadata', new Blob([metadata], { type: 'application/json' }))
+    formData.append('worker.js', new Blob([code], { type: 'application/javascript+module' }), 'worker.js')
+    const resp = await fetch(`${baseUrl}/${name}`, {
+      method: 'PUT',
+      headers: { Authorization: `Bearer ${apiToken}` },
+      body: formData,
+    })
+    if (!resp.ok) {
+      const text = await resp.text()
+      throw new Error(`CF API PUT worker failed (${resp.status}): ${text}`)
+    }
+  }
+
   return {
-    async deleteWorker(name: string): Promise<void> {
-      // CF API: DELETE /accounts/{account_id}/workers/scripts/{script_name}
-      const resp = await fetch(`${baseUrl}/${name}`, {
-        method: 'DELETE',
-        headers: {
-          Authorization: `Bearer ${apiToken}`,
-        },
+    async updateSlotCode(slotIndex: number, code: string): Promise<void> {
+      await putWorkerCode(getSlotName(slotIndex), code)
+    },
+
+    async initSlot(slotIndex: number): Promise<void> {
+      const name = getSlotName(slotIndex)
+      await putWorkerCode(name, IDLE_WORKER_CODE)
+      const subdomainResp = await fetch(`${baseUrl}/${name}/subdomain`, {
+        method: 'POST',
+        headers: { Authorization: `Bearer ${apiToken}`, 'Content-Type': 'application/json' },
+        body: JSON.stringify({ enabled: true }),
       })
-      if (!resp.ok && resp.status !== 404) {
-        const text = await resp.text()
-        throw new Error(`CF API delete failed (${resp.status}): ${text}`)
+      if (!subdomainResp.ok) {
+        console.warn(`[sigil] failed to enable subdomain for ${name}: ${subdomainResp.status}`)
       }
     },
+
+    getSlotSubdomain(slotIndex: number): string {
+      return `${getSlotName(slotIndex)}${CONFIG.SUBDOMAIN_SUFFIX}`
+    },
+
+    async invoke(slotIndex: number, request: Request): Promise<Response> {
+      const subdomain = `${getSlotName(slotIndex)}${CONFIG.SUBDOMAIN_SUFFIX}`
+      const url = new URL(request.url)
+      const targetUrl = `https://${subdomain}${url.pathname}${url.search}`
+      const headers = new Headers(request.headers)
+      headers.delete('host')
+      return fetch(targetUrl, {
+        method: request.method,
+        headers,
+        body: request.method !== 'GET' && request.method !== 'HEAD' ? request.body : undefined,
+        redirect: 'follow',
+      })
+    },
   }
 }

src/config.ts

@@ -1,8 +1,10 @@
 export const CONFIG = {
-  MAX_SLOTS: 3,            // LRU 验证用，生产 ~400
+  MAX_SLOTS: 3,            // 预分配 slot 数量（物理页帧总数）
   DEPLOY_COOLDOWN_MS: 5000,
   PAGE_RATE_LIMIT: 10,     // 次/分钟
   PAGE_RATE_WINDOW_MS: 60000,
-  HASH_LENGTH: 8,
+  HASH_LENGTH: 6,
+  SLOT_PREFIX: 's-slot-',  // slot Worker 名前缀：s-slot-0, s-slot-1, ...
+  SUBDOMAIN_SUFFIX: '.shazhou.workers.dev',
   GATEWAY_URL: 'https://sigil.shazhou.workers.dev',
 } as const

src/index.ts

@@ -6,8 +6,8 @@ import { EmbeddingService } from './embedding.js'
 
 export interface Env {
   SIGIL_KV: KVNamespace
-  AI: any       // Cloudflare Workers AI binding
-  LOADER: any   // Dynamic Workers loader binding
+  AI: any  // Cloudflare Workers AI binding
+  LOADER: any  // Dynamic Workers Loader binding (worker_loaders)
 }
 
 export default {

src/router.ts

@@ -43,7 +43,7 @@ export async function handleRequest(request: Request, env: RouterEnv): Promise<R
     return handleInspect(capability, env)
   }
 
-  // GET/POST /run/{capability} — invoke (no auth required)
+  // GET /run/{capability} — invoke (no auth required)
   const runMatch = path.match(/^\/run\/([^/]+)$/)
   if (runMatch) {
     const capability = runMatch[1]!
@@ -88,10 +88,8 @@ async function handleDeploy(request: Request, env: RouterEnv): Promise<Response>
     let schema: InputSchema | undefined
 
     if (body.code) {
-      // Mode A: deploy raw code
       code = body.code
     } else {
-      // Mode B: schema + execute
       if (!body.execute) {
         return jsonError(400, 'execute is required when using schema mode')
       }
@@ -173,6 +171,7 @@ async function handleInvoke(
   request: Request,
   env: RouterEnv,
 ): Promise<Response> {
+  // Direct invocation via Dynamic Workers — no redirect, no sub-worker fetch
   return await env.backend.invoke(capability, request)
 }
 

test/query.test.ts

@@ -112,7 +112,7 @@ describe('Query API', () => {
     // Re-deploy with the new overrides in place
     const mockKv2 = createMockKv()
     const mockLoader2 = createMockLoader()
-    const pool2 = new WorkerPool(mockKv2, mockCf2.loader, mockEmbed as any)
+    const pool2 = new WorkerPool(mockKv2, mockLoader2.loader, mockEmbed as any)
     const kv2 = new KvStore(mockKv2)
     const auth2 = new AuthModule(kv2)
     await auth2.setToken('deploy-token')

test/s06-remove.test.ts

@@ -32,7 +32,7 @@ describe('S6: 删除能力', () => {
     mockLoader.reset()
   })
 
-  it('should NOT call CF API deleteWorker (Dynamic Workers; KV cleanup only)', async () => {
+  it('should clear all KV entries (Dynamic Workers: no CF API deleteWorker needed)', async () => {
     const req = makeRequest('DELETE', '/_api/remove', {
       token: 'deploy-token',
       body: { capability: 'ping' },
@@ -40,7 +40,13 @@ describe('S6: 删除能力', () => {
 
     const resp = await handleRequest(req, { SIGIL_KV: mockKv, backend: pool, auth, kv })
     expect(resp.status).toBe(200)
-    // LOADER should not be called during remove
+
+    // All KV entries should be gone
+    expect(await kv.getCode('ping')).toBeNull()
+    expect(await kv.getMeta('ping')).toBeNull()
+    expect(await kv.getLru('ping')).toBeNull()
+
+    // No LOADER.get() calls during remove
     expect(mockLoader.loaderCalls()).toHaveLength(0)
   })
 
@@ -50,7 +56,6 @@ describe('S6: 删除能力', () => {
     expect(await kv.getCode('ping')).toBeNull()
     expect(await kv.getMeta('ping')).toBeNull()
     expect(await kv.getLru('ping')).toBeNull()
-    expect(await kv.getRoute('ping')).toBeNull()
   })
 
   it('should return removed capability in response', async () => {

test/s07-list.test.ts

@@ -7,7 +7,7 @@ import { handleRequest } from '../src/router.js'
 
 describe('S7: 列出能力（已迁移至 query 接口）', () => {
   let mockKv: KVNamespace
-  let mockCf: ReturnType<typeof createMockLoader>
+  let mockLoader: ReturnType<typeof createMockLoader>
   let mockEmbed: MockEmbeddingService
   let pool: WorkerPool
   let auth: AuthModule
@@ -15,7 +15,7 @@ describe('S7: 列出能力（已迁移至 query 接口）', () => {
 
   beforeEach(async () => {
     mockKv = createMockKv()
-    mockCf = createMockLoader()
+    mockLoader = createMockLoader()
     mockEmbed = new MockEmbeddingService()
     pool = new WorkerPool(mockKv, mockLoader.loader, mockEmbed as any)
     kv = new KvStore(mockKv)

test/s08-health.test.ts

@@ -7,7 +7,7 @@ import { handleRequest } from '../src/router.js'
 
 describe('S8: 健康端点', () => {
   let mockKv: KVNamespace
-  let mockCf: ReturnType<typeof createMockLoader>
+  let mockLoader: ReturnType<typeof createMockLoader>
   let mockEmbed: MockEmbeddingService
   let pool: WorkerPool
   let auth: AuthModule
@@ -15,7 +15,7 @@ describe('S8: 健康端点', () => {
 
   beforeEach(async () => {
     mockKv = createMockKv()
-    mockCf = createMockLoader()
+    mockLoader = createMockLoader()
     mockEmbed = new MockEmbeddingService()
     pool = new WorkerPool(mockKv, mockLoader.loader, mockEmbed as any)
     kv = new KvStore(mockKv)

test/s11-concurrent-page-in.test.ts

@@ -32,7 +32,7 @@ describe('S11: 并发换入去重', () => {
     })
   })
 
-  it('should call deployWorker only once for concurrent page-ins', async () => {
+  it('should handle concurrent page-ins without error', async () => {
     const req1 = new Request('https://sigil.shazhou.workers.dev/run/ping')
     const req2 = new Request('https://sigil.shazhou.workers.dev/run/ping')
 
@@ -45,9 +45,8 @@ describe('S11: 并发换入去重', () => {
     expect(resp1.status).toBe(200)
     expect(resp2.status).toBe(200)
 
-    // Both calls go through LOADER.get(); Dynamic Workers deduplicates internally
+    // LOADER.get() should be called (at least once — may be called for each concurrent request)
     const loaderCalls = mockLoader.loaderCalls()
     expect(loaderCalls.length).toBeGreaterThanOrEqual(1)
-    expect(loaderCalls.every(id => id === 's-ping')).toBe(true)
   })
 })

test/s12-page-rate-limit.test.ts

@@ -2,10 +2,8 @@ import { describe, it, expect, beforeEach } from 'vitest'
 import { createMockKv, createMockLoader, MockEmbeddingService } from './setup.js'
 import { WorkerPool } from '../src/backend/worker-pool.js'
 import { KvStore } from '../src/kv.js'
-import { CONFIG } from '../src/config.js'
-import { PageRateLimitError } from '../src/lru.js'
 
-describe('S12: 换页速率限制', () => {
+describe('S12: Dynamic Workers invoke（原 page-rate-limit）', () => {
   let mockKv: KVNamespace
   let mockLoader: ReturnType<typeof createMockLoader>
   let mockEmbed: MockEmbeddingService
@@ -35,67 +33,26 @@ describe('S12: 换页速率限制', () => {
     })
   }
 
-  it(`should allow up to ${CONFIG.PAGE_RATE_LIMIT} page-ins per minute`, async () => {
-    const results: boolean[] = []
-
-    for (let i = 0; i < CONFIG.PAGE_RATE_LIMIT; i++) {
-      const name = `cap${i}`
-      await setupCapability(name)
-
-      const req = new Request(`https://sigil.shazhou.workers.dev/run/${name}`)
-      const resp = await pool.invoke(name, req)
-      results.push(resp.status === 200)
-    }
-
-    expect(results.every(Boolean)).toBe(true)
-  })
-
-  it(`should reject the ${CONFIG.PAGE_RATE_LIMIT + 1}th page-in with 503`, async () => {
-    // Do PAGE_RATE_LIMIT page-ins
-    for (let i = 0; i < CONFIG.PAGE_RATE_LIMIT; i++) {
+  it('should invoke evicted capabilities without page-rate-limit', async () => {
+    // With Dynamic Workers, there is no page rate limit — invoke always works.
+    for (let i = 0; i < 15; i++) {
       const name = `cap${i}`
       await setupCapability(name)
       const req = new Request(`https://sigil.shazhou.workers.dev/run/${name}`)
-      await pool.invoke(name, req)
-    }
-
-    // 11th one should fail
-    const name = `cap${CONFIG.PAGE_RATE_LIMIT}`
-    await setupCapability(name)
-
-    const req = new Request(`https://sigil.shazhou.workers.dev/run/${name}`)
-    try {
       const resp = await pool.invoke(name, req)
-      // If it doesn't throw, check status
-      expect(resp.status).toBe(503)
-    } catch (e) {
-      expect(e).toBeInstanceOf(PageRateLimitError)
+      expect(resp.status).toBe(200)
     }
   })
 
-  it('should include retry_after in error', async () => {
-    // Fill rate
-    for (let i = 0; i < CONFIG.PAGE_RATE_LIMIT; i++) {
-      const name = `cap${i}`
-      await setupCapability(name)
-      const req = new Request(`https://sigil.shazhou.workers.dev/run/${name}`)
-      await pool.invoke(name, req)
-    }
+  it('should mark cold-start capability as deployed after invoke', async () => {
+    await setupCapability('cold')
+    const lruBefore = await kv.getLru('cold')
+    expect(lruBefore!.deployed).toBe(false)
 
-    const name = `cap${CONFIG.PAGE_RATE_LIMIT}`
-    await setupCapability(name)
-    const req = new Request(`https://sigil.shazhou.workers.dev/run/${name}`)
+    const req = new Request('https://sigil.shazhou.workers.dev/run/cold')
+    await pool.invoke('cold', req)
 
-    try {
-      const resp = await pool.invoke(name, req)
-      if (resp.status === 503) {
-        const body = await resp.json() as { error: string; retry_after?: number }
-        // retry_after may be 0 for immediate window, just check it exists or we got exception
-        expect(body.error).toBeTruthy()
-      }
-    } catch (e) {
-      expect(e).toBeInstanceOf(PageRateLimitError)
-      expect((e as PageRateLimitError).retry_after).toBeGreaterThanOrEqual(0)
-    }
+    const lruAfter = await kv.getLru('cold')
+    expect(lruAfter!.deployed).toBe(true)
   })
 })

wrangler.toml

@@ -1,16 +1,14 @@
 name = "sigil"
 main = "src/index.ts"
 compatibility_date = "2026-04-03"
-compatibility_flags = ["global_fetch_strictly_public"]
 
 [[kv_namespaces]]
 binding = "SIGIL_KV"
 id = "9943c8873e724b0fb2cf24b4475e5a52"
 
 # Dynamic Workers loader binding (open beta)
-# Allows Gateway Worker to load and execute sub-Worker code dynamically
-# without deploying independent scripts, avoiding Worker-to-Worker fetch
-# restrictions (error 1042) on workers.dev subdomains.
+# Loads and executes capability code at runtime inside the Gateway Worker,
+# bypassing Worker-to-Worker fetch restrictions (error 1042).
 [[worker_loaders]]
 binding = "LOADER"
 
@@ -19,11 +17,3 @@ binding = "AI"
 
 [vars]
 SIGIL_ENV = "production"
-
-# Worker Secrets (set via `wrangler secret put`, never committed to source):
-#   CF_API_TOKEN   - Cloudflare API token with Workers:Edit permission (optional, for legacy cleanup)
-#   CF_ACCOUNT_ID  - Cloudflare Account ID (optional, for legacy cleanup)
-#
-# To configure:
-#   echo "$CF_API_TOKEN" | npx wrangler secret put CF_API_TOKEN
-#   echo "$CF_ACCOUNT_ID" | npx wrangler secret put CF_ACCOUNT_ID