plugin-apply.spec.ts 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427
  1. import { createServer } from 'node:http'
  2. import type { IncomingMessage, Server, ServerResponse } from 'node:http'
  3. import { mkdtemp, rm } from 'node:fs/promises'
  4. import { join } from 'node:path'
  5. import { tmpdir } from 'node:os'
  6. import { PassThrough, Writable } from 'node:stream'
  7. import { afterEach, describe, expect, it, vi } from 'vitest'
  8. import { Context } from '@deepseek-ai/cordis'
  9. import Loader from '@deepseek-ai/cordis-plugin-loader'
  10. import * as agentCore from '@deepseek-ai/dsh-agent-spine-demo'
  11. import { LlmAdapter } from '@deepseek-ai/dsh-llm'
  12. import type { GenerateOptions, StreamChunk } from '@deepseek-ai/dsh-llm'
  13. import JsonlSessionPersistence from '@deepseek-ai/dsh-session-persistence-jsonl'
  14. import { defineTool } from '@deepseek-ai/dsh-tools'
  15. import * as jsonrpc from '../src/index.ts'
  16. /**
  17. * Mount the real namespace plugin with in-memory stdio and exit hooks. Covers
  18. * the full transport/server path, response-before-exit shutdown exactly once,
  19. * and bare-fiber disposal without process exit.
  20. */
  21. /** One ordered frame, write completion, or exit observation. */
  22. type WireEvent =
  23. | { kind: 'frame'; frame: Record<string, unknown> }
  24. | { kind: 'write-complete'; ids: (string | number)[] }
  25. | { kind: 'root-disposed' }
  26. | { kind: 'exit'; code: number }
  27. interface ApplyHarness {
  28. ctx: Context
  29. /** The plugin fiber used by the bare-dispose case. */
  30. fiber: Awaited<ReturnType<Context['plugin']>>
  31. /** Frames, write completions, and exits in observation order. */
  32. events: WireEvent[]
  33. outputErrors: Error[]
  34. send(frame: Record<string, unknown>): void
  35. sendRaw(text: string): void
  36. frames(): Record<string, unknown>[]
  37. exits(): number[]
  38. waitForFrame(predicate: (frame: Record<string, unknown>) => boolean, description: string): Promise<Record<string, unknown>>
  39. dispose(): Promise<void>
  40. }
  41. /** Adapter whose route registration is the delayed Loader entry's readiness fact. */
  42. class DelayedAdapter extends LlmAdapter {
  43. async * stream(_options: GenerateOptions): AsyncIterable<StreamChunk> {
  44. throw new Error('not exercised')
  45. }
  46. }
  47. /** Poll asynchronous output for up to five seconds. */
  48. async function waitFor<T>(get: () => T | undefined, description: string): Promise<T> {
  49. const deadline = Date.now() + 5000
  50. for (;;) {
  51. const value = get()
  52. if (value !== undefined) return value
  53. if (Date.now() > deadline) throw new Error(`timed out waiting for ${description}`)
  54. await new Promise(resolve => setTimeout(resolve, 5))
  55. }
  56. }
  57. /** Drain asynchronous work before a negative assertion. */
  58. async function settle(): Promise<void> {
  59. await new Promise(resolve => setTimeout(resolve, 25))
  60. }
  61. /** Mount the real plugin on a minimal harness with in-memory stdio and exit. */
  62. async function mountPlugin(
  63. storageDir: string,
  64. options: {
  65. writeDelayMs?: number
  66. failFlush?: boolean
  67. beforeServer?: (ctx: Context) => Promise<void> | void
  68. toolFilter?: jsonrpc.JsonRpcConfig['toolFilter']
  69. } = {},
  70. ): Promise<ApplyHarness> {
  71. const ctx = new Context()
  72. await ctx.plugin(agentCore, { workspaceContext: false })
  73. await ctx.plugin(JsonlSessionPersistence, { root: storageDir })
  74. await new Promise(resolve => setTimeout(resolve, 50))
  75. await options.beforeServer?.(ctx)
  76. const input = new PassThrough()
  77. const events: WireEvent[] = []
  78. const outputErrors: Error[] = []
  79. let pendingOutput = ''
  80. // Record frame admission separately from write completion so delayed output
  81. // tests the flush barrier.
  82. const output = new Writable({
  83. write(chunk: Buffer, _encoding, callback) {
  84. const ids: (string | number)[] = []
  85. pendingOutput += chunk.toString('utf8')
  86. for (;;) {
  87. const newline = pendingOutput.indexOf('\n')
  88. if (newline < 0) break
  89. const line = pendingOutput.slice(0, newline).trim()
  90. pendingOutput = pendingOutput.slice(newline + 1)
  91. if (line) {
  92. const frame = JSON.parse(line) as Record<string, unknown>
  93. events.push({ kind: 'frame', frame })
  94. if (typeof frame.id === 'string' || typeof frame.id === 'number') ids.push(frame.id)
  95. }
  96. }
  97. const complete = (): void => {
  98. if (options.failFlush === true && chunk.length === 0) {
  99. callback(new Error('flush callback failed'))
  100. return
  101. }
  102. events.push({ kind: 'write-complete', ids })
  103. callback()
  104. }
  105. if ((options.writeDelayMs ?? 0) > 0) setTimeout(complete, options.writeDelayMs)
  106. else complete()
  107. },
  108. })
  109. output.on('error', (error: Error) => { outputErrors.push(error) })
  110. const exit = (code: number): void => { events.push({ kind: 'exit', code }) }
  111. ctx.effect(() => () => { events.push({ kind: 'root-disposed' }) }, 'jsonrpc test root-disposal witness')
  112. const fiber = await ctx.plugin(jsonrpc, {
  113. input,
  114. output,
  115. exit,
  116. ...options.toolFilter === undefined ? {} : { toolFilter: options.toolFilter },
  117. })
  118. const frames = (): Record<string, unknown>[] =>
  119. events.flatMap(event => event.kind === 'frame' ? [event.frame] : [])
  120. return {
  121. ctx,
  122. fiber,
  123. events,
  124. outputErrors,
  125. send: (frame) => { input.write(`${JSON.stringify(frame)}\n`) },
  126. sendRaw: (text) => { input.write(text) },
  127. frames,
  128. exits: () => events.flatMap(event => event.kind === 'exit' ? [event.code] : []),
  129. waitForFrame: (predicate, description) => waitFor(() => frames().find(predicate), description),
  130. dispose: async () => { await ctx.fiber.dispose() },
  131. }
  132. }
  133. const servers: Server[] = []
  134. afterEach(async () => {
  135. await Promise.all(servers.splice(0).map(server => new Promise(resolve => server.close(resolve))))
  136. vi.unstubAllEnvs()
  137. })
  138. /** Keyless SSE endpoint for completing a prompt turn. */
  139. async function mockCompletionServer(): Promise<{ url: string; requests: unknown[] }> {
  140. const requests: unknown[] = []
  141. const server = createServer((request: IncomingMessage, response: ServerResponse) => {
  142. let body = ''
  143. request.on('data', (chunk: Buffer) => { body += chunk.toString('utf8') })
  144. request.on('end', () => {
  145. requests.push(JSON.parse(body))
  146. response.writeHead(200, { 'content-type': 'text/event-stream' })
  147. response.write('data: {"choices":[{"delta":{"role":"assistant","content":null,"reasoning_content":""}}]}\n\n')
  148. response.write('data: {"choices":[{"delta":{"content":"done"}}]}\n\n')
  149. response.write('data: {"choices":[{"delta":{"content":""},"finish_reason":"stop"}],"usage":{"prompt_tokens":3,"completion_tokens":1}}\n\n')
  150. response.write('data: [DONE]\n\n')
  151. response.end()
  152. })
  153. })
  154. servers.push(server)
  155. await new Promise<void>(resolve => server.listen(0, '127.0.0.1', resolve))
  156. const address = server.address()
  157. if (address === null || typeof address === 'string') throw new Error('no port')
  158. return { url: `http://127.0.0.1:${address.port}`, requests }
  159. }
  160. describe('dsh-sdk-jsonrpc-server plugin apply', () => {
  161. it('serves initialize over the injected stdio pair', async () => {
  162. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-init-'))
  163. vi.stubEnv('DEEPSEEK_API_KEY', 'test-key')
  164. const harness = await mountPlugin(storageDir)
  165. try {
  166. harness.send({ jsonrpc: '2.0', id: 'init-1', method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'apply-model' } })
  167. const response = await harness.waitForFrame(frame => frame.id === 'init-1', 'initialize response')
  168. expect(response).toEqual({
  169. jsonrpc: '2.0',
  170. id: 'init-1',
  171. result: { serverInfo: { name: 'deepseek-harness-sdk-runtime', version: '0.0.1' } },
  172. })
  173. expect(harness.exits()).toEqual([])
  174. } finally {
  175. await harness.dispose()
  176. await rm(storageDir, { recursive: true, force: true })
  177. }
  178. })
  179. it('waits for Loader-owned adapter registration before initialize', async () => {
  180. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-readiness-'))
  181. vi.stubEnv('DEEPSEEK_API_KEY', 'test-key')
  182. let markStarted!: () => void
  183. let release!: () => void
  184. const started = new Promise<void>((resolve) => { markStarted = resolve })
  185. const ready = new Promise<void>((resolve) => { release = resolve })
  186. let delayedEntry: Promise<string> | undefined
  187. const harness = await mountPlugin(storageDir, {
  188. beforeServer: async (ctx) => {
  189. await ctx.plugin(Loader)
  190. ctx.loader.builtins['delayed-readiness'] = {
  191. inject: ['llm'],
  192. async apply(entryCtx: Context) {
  193. markStarted()
  194. await ready
  195. entryCtx.llm.registerAdapter(['delayed-private'], new DelayedAdapter())
  196. },
  197. }
  198. delayedEntry = ctx.loader.create({ name: 'cordis:delayed-readiness' })
  199. await started
  200. },
  201. })
  202. try {
  203. const initialize = {
  204. jsonrpc: '2.0',
  205. id: 'init-delayed',
  206. method: 'initialize',
  207. params: { cwd: storageDir, provider: 'delayed-private', model: 'apply-model' },
  208. }
  209. const probe = { jsonrpc: '2.0', id: 'probe-during-delay', method: 'nope/unknown' }
  210. harness.sendRaw(`${JSON.stringify(initialize)}\n${JSON.stringify(probe)}\n`)
  211. // The transport processes independent requests concurrently. Receiving
  212. // this later probe proves the preceding initialize handler has reached
  213. // its Loader wait, without relying on a scheduler delay.
  214. await harness.waitForFrame(frame => frame.id === 'probe-during-delay', 'probe while initialize waits')
  215. expect(harness.frames().some(frame => frame.id === 'init-delayed')).toBe(false)
  216. release()
  217. await delayedEntry
  218. const response = await harness.waitForFrame(frame => frame.id === 'init-delayed', 'initialize response after Loader settlement')
  219. expect(response).toMatchObject({
  220. id: 'init-delayed',
  221. result: { serverInfo: { name: 'deepseek-harness-sdk-runtime' } },
  222. })
  223. expect(harness.ctx.llm.listProviders()).toContainEqual({ id: 'delayed-private', name: 'delayed-private' })
  224. } finally {
  225. release()
  226. await Promise.allSettled(delayedEntry === undefined ? [] : [delayedEntry])
  227. await harness.dispose()
  228. await rm(storageDir, { recursive: true, force: true })
  229. }
  230. })
  231. it('drives a session/prompt turn end-to-end and forwards session notifications as output frames', async () => {
  232. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-prompt-'))
  233. const llmServer = await mockCompletionServer()
  234. vi.stubEnv('DEEPSEEK_API_KEY', 'test-key')
  235. vi.stubEnv('DEEPSEEK_BASE_URL', llmServer.url)
  236. const harness = await mountPlugin(storageDir)
  237. try {
  238. harness.send({ jsonrpc: '2.0', id: 1, method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'dsagent-model' } })
  239. await harness.waitForFrame(frame => frame.id === 1, 'initialize response')
  240. harness.send({
  241. jsonrpc: '2.0',
  242. id: 2,
  243. method: 'session/prompt',
  244. params: { sessionId: 'main', contentBlocks: [{ type: 'text', text: 'fix it' }] },
  245. })
  246. const response = await harness.waitForFrame(frame => frame.id === 2, 'prompt response')
  247. expect((response.result as { messageId?: unknown }).messageId).toBeTypeOf('string')
  248. await harness.waitForFrame(
  249. frame => frame.method === 'session.status'
  250. && (frame.params as { status?: string } | undefined)?.status === 'idle',
  251. 'idle session status',
  252. )
  253. expect(llmServer.requests).toHaveLength(1)
  254. const body = llmServer.requests[0] as { model: string; messages: { role: string }[] }
  255. expect(body.model).toBe('dsagent-model')
  256. expect(body.messages.at(-1)?.role).toBe('user')
  257. // Notifications use the same transport and arrive as id-less frames.
  258. const notifications = harness.frames().filter(frame => frame.id === undefined)
  259. expect(notifications.some(frame => frame.method === 'session.event')).toBe(true)
  260. expect(notifications.findLast(frame => frame.method === 'session.status')).toMatchObject({
  261. jsonrpc: '2.0',
  262. params: { sessionId: 'main', status: 'idle' },
  263. })
  264. } finally {
  265. await harness.dispose()
  266. await rm(storageDir, { recursive: true, force: true })
  267. }
  268. })
  269. it('applies the configured root-agent tool filter through the Loader plugin', async () => {
  270. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-tool-filter-'))
  271. const llmServer = await mockCompletionServer()
  272. vi.stubEnv('DEEPSEEK_API_KEY', 'test-key')
  273. vi.stubEnv('DEEPSEEK_BASE_URL', llmServer.url)
  274. const harness = await mountPlugin(storageDir, {
  275. toolFilter: { allow: ['kept'] },
  276. beforeServer: (ctx) => {
  277. for (const name of ['kept', 'excluded']) {
  278. ctx.tools.register(defineTool({
  279. name,
  280. description: name,
  281. parameters: {},
  282. output: {
  283. schema: { type: 'string' },
  284. render: (_args, value) => [{ type: 'text', text: value }],
  285. },
  286. execute: async () => name,
  287. }))
  288. }
  289. },
  290. })
  291. try {
  292. harness.send({ jsonrpc: '2.0', id: 1, method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'filtered-model' } })
  293. await harness.waitForFrame(frame => frame.id === 1, 'initialize response')
  294. harness.send({
  295. jsonrpc: '2.0',
  296. id: 2,
  297. method: 'session/prompt',
  298. params: { sessionId: 'filtered', contentBlocks: [{ type: 'text', text: 'inspect tools' }] },
  299. })
  300. await harness.waitForFrame(
  301. frame => frame.method === 'session.status'
  302. && (frame.params as { status?: string } | undefined)?.status === 'idle',
  303. 'filtered session idle status',
  304. )
  305. const request = llmServer.requests[0] as { tools?: Array<{ function?: { name?: string } }> }
  306. expect(request.tools?.map(entry => entry.function?.name)).toEqual(['kept'])
  307. } finally {
  308. await harness.dispose()
  309. await rm(storageDir, { recursive: true, force: true })
  310. }
  311. })
  312. it('answers shutdown before exiting 0 exactly once, even against a racing second shutdown', async () => {
  313. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-shutdown-'))
  314. const harness = await mountPlugin(storageDir, { writeDelayMs: 10 })
  315. try {
  316. // One chunk makes the two deferred exit callbacks race.
  317. const first = { jsonrpc: '2.0', id: 'sd-1', method: 'shutdown' }
  318. const second = { jsonrpc: '2.0', id: 'sd-2', method: 'shutdown' }
  319. harness.sendRaw(`${JSON.stringify(first)}\n${JSON.stringify(second)}\n`)
  320. await waitFor(() => harness.exits().length > 0 ? true : undefined, 'exit recorder call')
  321. expect(harness.exits()).toEqual([0])
  322. // Both response writes and the flush barrier complete before exit.
  323. const exitIndex = harness.events.findIndex(event => event.kind === 'exit')
  324. const firstResponse = harness.events.findIndex(event => event.kind === 'frame' && event.frame.id === 'sd-1')
  325. const secondResponse = harness.events.findIndex(event => event.kind === 'frame' && event.frame.id === 'sd-2')
  326. const firstComplete = harness.events.findIndex(event => event.kind === 'write-complete' && event.ids.includes('sd-1'))
  327. const secondComplete = harness.events.findIndex(event => event.kind === 'write-complete' && event.ids.includes('sd-2'))
  328. const flushComplete = harness.events.findIndex(event => event.kind === 'write-complete' && event.ids.length === 0)
  329. const rootDisposed = harness.events.findIndex(event => event.kind === 'root-disposed')
  330. expect(firstResponse).toBeGreaterThanOrEqual(0)
  331. expect(secondResponse).toBeGreaterThanOrEqual(0)
  332. expect(firstComplete).toBeGreaterThan(firstResponse)
  333. expect(secondComplete).toBeGreaterThan(secondResponse)
  334. expect(flushComplete).toBeGreaterThan(firstComplete)
  335. expect(flushComplete).toBeGreaterThan(secondComplete)
  336. expect(rootDisposed).toBeGreaterThan(flushComplete)
  337. expect(exitIndex).toBeGreaterThan(rootDisposed)
  338. await settle()
  339. expect(harness.exits()).toEqual([0])
  340. expect(harness.events.filter(event => event.kind === 'root-disposed')).toHaveLength(1)
  341. const before = harness.frames().length
  342. harness.send({ jsonrpc: '2.0', id: 'after-exit', method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'x' } })
  343. await settle()
  344. expect(harness.frames().length).toBe(before)
  345. } finally {
  346. await harness.dispose()
  347. await rm(storageDir, { recursive: true, force: true })
  348. }
  349. })
  350. it('still disposes and exits once when the flush callback fails', async () => {
  351. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-flush-failure-'))
  352. const harness = await mountPlugin(storageDir, { failFlush: true })
  353. try {
  354. harness.send({ jsonrpc: '2.0', id: 'sd-fail', method: 'shutdown' })
  355. await waitFor(() => harness.exits().length > 0 ? true : undefined, 'exit after flush failure')
  356. await settle()
  357. expect(harness.exits()).toEqual([0])
  358. expect(harness.events.filter(event => event.kind === 'root-disposed')).toHaveLength(1)
  359. expect(harness.outputErrors.map(error => error.message)).toEqual(['flush callback failed'])
  360. const before = harness.frames().length
  361. harness.send({ jsonrpc: '2.0', id: 'after-flush-failure', method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'x' } })
  362. await settle()
  363. expect(harness.frames().length).toBe(before)
  364. } finally {
  365. await harness.dispose()
  366. await rm(storageDir, { recursive: true, force: true })
  367. }
  368. })
  369. it('stops serving on a bare fiber dispose (HMR-style unload) without calling exit', async () => {
  370. const storageDir = await mkdtemp(join(tmpdir(), 'dsh-jsonrpc-apply-dispose-'))
  371. const harness = await mountPlugin(storageDir)
  372. try {
  373. // Prove the handler-rejection path is live before disposal.
  374. harness.send({ jsonrpc: '2.0', id: 'probe-1', method: 'nope/unknown' })
  375. const error = await harness.waitForFrame(frame => frame.id === 'probe-1', 'error response for unknown method')
  376. expect(error.error).toMatchObject({
  377. code: -32603,
  378. message: 'unknown DeepSeek Harness SDK runtime method: nope/unknown',
  379. })
  380. await harness.fiber.dispose()
  381. expect(harness.events.some(event => event.kind === 'root-disposed')).toBe(false)
  382. const before = harness.frames().length
  383. harness.send({ jsonrpc: '2.0', id: 'probe-2', method: 'initialize', params: { cwd: storageDir, provider: 'deepseek-official', model: 'x' } })
  384. await settle()
  385. expect(harness.frames().length).toBe(before)
  386. expect(harness.exits()).toEqual([])
  387. } finally {
  388. await harness.dispose()
  389. await rm(storageDir, { recursive: true, force: true })
  390. }
  391. })
  392. })