fix: keep cancelled turns as interrupted instead of merging into the next request
- Build runtime history per turn: a cancelled or failed reply keeps its partial text plus an interruption marker (or a marker-only placeholder), so its user message is never left unanswered and merged into the next request. - Use the same history for send, manual compaction and Main compact validation. - OpenCode: keep a cancelled prompt in the reused native session and prefix the next turn with a one-time synthetic interruption notice; failed prompts are still removed. - Add an opt-in DeepSeek live test covering direct model and OpenCode.
This commit is contained in:
mesalogo
@@ -0,0 +1,147 @@
|
||||
// @vitest-environment node
|
||||
import { mkdtemp, readFile, rm } from 'node:fs/promises'
|
||||
import { tmpdir } from 'node:os'
|
||||
import { join, resolve } from 'node:path'
|
||||
import { parseEnv } from 'node:util'
|
||||
import { afterAll, beforeAll, describe, expect, it } from 'vitest'
|
||||
import { ModelAgentRuntime } from '../src/main/agent/model-runtime'
|
||||
import { OpenCodeRuntime } from '../src/main/agent/opencode-runtime'
|
||||
import type { AgentRuntime, RuntimeEvent } from '../src/main/agent/runtime'
|
||||
import { buildRuntimeHistory } from '../src/shared/runtime-history'
|
||||
|
||||
// Opt-in: GB_CANCEL_LIVE_ENV=.env.deepseek.local npx vitest run tests/cancelled-turn-live.test.ts
|
||||
const envPath = process.env.GB_CANCEL_LIVE_ENV
|
||||
const opencodePath = resolve('.runtime-resources', process.arch, process.platform === 'win32' ? 'opencode.exe' : 'opencode')
|
||||
|
||||
const firstPrompt = 'List 40 European capital cities, one per line, each with a one-sentence description.'
|
||||
const secondPrompt = 'What is 17 + 25? Reply with only the number and nothing else.'
|
||||
|
||||
type Env = { apiKey: string; baseUrl: string; model: string }
|
||||
|
||||
async function loadEnv(): Promise<Env> {
|
||||
const env = parseEnv(await readFile(resolve(envPath!), 'utf8'))
|
||||
return { apiKey: env.DEEPSEEK_API_KEY!, baseUrl: env.DEEPSEEK_BASE_URL!, model: env.DEEPSEEK_MODEL! }
|
||||
}
|
||||
|
||||
/** Runs until some answer text streams, then cancels like the Stop button. */
|
||||
async function runAndCancel(runtime: AgentRuntime, conversationId: string): Promise<string> {
|
||||
const controller = new AbortController()
|
||||
let partial = ''
|
||||
try {
|
||||
for await (const event of runtime.run({
|
||||
requestId: crypto.randomUUID(), conversationId, workMode: 'ask', prompt: firstPrompt
|
||||
}, controller.signal)) {
|
||||
if (event.type === 'text') partial += event.delta
|
||||
if (partial.length >= 60 && !controller.signal.aborted) controller.abort(new Error('用户取消了请求'))
|
||||
}
|
||||
} catch (error) {
|
||||
if (!controller.signal.aborted) throw error
|
||||
}
|
||||
expect(controller.signal.aborted).toBe(true)
|
||||
return partial
|
||||
}
|
||||
|
||||
async function runText(
|
||||
runtime: AgentRuntime,
|
||||
request: Parameters<AgentRuntime['run']>[0]
|
||||
): Promise<string> {
|
||||
let output = ''
|
||||
for await (const event of runtime.run(request, new AbortController().signal) as AsyncGenerator<RuntimeEvent>) {
|
||||
if (event.type === 'text') output += event.delta
|
||||
}
|
||||
return output.trim()
|
||||
}
|
||||
|
||||
function expectOnlyNewAnswer(output: string): void {
|
||||
console.info(`SECOND_TURN_OUTPUT=${JSON.stringify(output)}`)
|
||||
expect(output).toMatch(/\b42\b/u)
|
||||
// Merging the cancelled turn would resume the capital list.
|
||||
expect(output).not.toMatch(/Paris|Berlin|Madrid|Vienna|Rome|London/u)
|
||||
expect(output.length).toBeLessThan(40)
|
||||
}
|
||||
|
||||
describe.skipIf(!envPath)('cancelled turn handling with a real model', () => {
|
||||
let env: Env
|
||||
let workspace: string
|
||||
beforeAll(async () => {
|
||||
env = await loadEnv()
|
||||
workspace = await mkdtemp(join(tmpdir(), 'gb-cancel-live-'))
|
||||
})
|
||||
afterAll(async () => {
|
||||
await rm(workspace, { recursive: true, force: true })
|
||||
})
|
||||
|
||||
it('direct model answers only the new message after a cancelled turn', async () => {
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: env.apiKey, baseUrl: env.baseUrl, model: env.model,
|
||||
protocol: 'openai-chat-completions', authentication: 'api-key'
|
||||
})
|
||||
try {
|
||||
const conversationId = crypto.randomUUID()
|
||||
const partial = await runAndCancel(runtime, conversationId)
|
||||
console.info(`DIRECT_PARTIAL_CHARS=${partial.length}`)
|
||||
// Same stored shape the renderer keeps after Stop.
|
||||
const history = buildRuntimeHistory([
|
||||
{ id: crypto.randomUUID(), role: 'user', content: firstPrompt, state: 'complete' },
|
||||
{ id: crypto.randomUUID(), role: 'assistant', content: partial, state: 'error', terminalStatus: 'cancelled' }
|
||||
])
|
||||
expect(history).toHaveLength(2)
|
||||
expectOnlyNewAnswer(await runText(runtime, {
|
||||
requestId: crypto.randomUUID(), conversationId, workMode: 'ask', prompt: secondPrompt,
|
||||
history: history.map(({ role, content }) => ({ role, content })),
|
||||
historyMessageIds: history.map(({ id }) => id)
|
||||
}))
|
||||
} finally {
|
||||
await runtime.dispose()
|
||||
}
|
||||
}, 180_000)
|
||||
|
||||
it('direct model answers only the new message when the cancelled turn had no output', async () => {
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: env.apiKey, baseUrl: env.baseUrl, model: env.model,
|
||||
protocol: 'openai-chat-completions', authentication: 'api-key'
|
||||
})
|
||||
try {
|
||||
const history = buildRuntimeHistory([
|
||||
{ id: crypto.randomUUID(), role: 'user', content: firstPrompt, state: 'complete' },
|
||||
{ id: crypto.randomUUID(), role: 'assistant', content: '', state: 'error', terminalStatus: 'cancelled' }
|
||||
])
|
||||
expectOnlyNewAnswer(await runText(runtime, {
|
||||
requestId: crypto.randomUUID(), conversationId: crypto.randomUUID(), workMode: 'ask', prompt: secondPrompt,
|
||||
history: history.map(({ role, content }) => ({ role, content })),
|
||||
historyMessageIds: history.map(({ id }) => id)
|
||||
}))
|
||||
} finally {
|
||||
await runtime.dispose()
|
||||
}
|
||||
}, 180_000)
|
||||
|
||||
it('OpenCode reused session answers only the new message after a cancelled turn', async () => {
|
||||
const runtime = new OpenCodeRuntime({
|
||||
embedded: true, binaryPath: '', bundledBinaryPath: opencodePath, configPath: '',
|
||||
defaultWorkspace: workspace,
|
||||
modelProfile: {
|
||||
id: crypto.randomUUID(), name: 'DeepSeek live', baseUrl: env.baseUrl, modelName: env.model,
|
||||
apiKey: env.apiKey, protocol: 'openai-chat-completions', authentication: 'api-key'
|
||||
}
|
||||
})
|
||||
try {
|
||||
const conversationId = crypto.randomUUID()
|
||||
const partial = await runAndCancel(runtime, conversationId)
|
||||
console.info(`OPENCODE_PARTIAL_CHARS=${partial.length}`)
|
||||
expectOnlyNewAnswer(await runText(runtime, {
|
||||
requestId: crypto.randomUUID(), conversationId, workMode: 'ask', prompt: secondPrompt
|
||||
}))
|
||||
// The cancelled request stays in the native transcript, so the model
|
||||
// still knows what was asked before.
|
||||
const recall = await runText(runtime, {
|
||||
requestId: crypto.randomUUID(), conversationId, workMode: 'ask',
|
||||
prompt: 'In one short phrase, what did I ask for in my very first message of this conversation? Do not list anything.'
|
||||
})
|
||||
console.info(`OPENCODE_RECALL=${JSON.stringify(recall)}`)
|
||||
expect(recall).toMatch(/capital/iu)
|
||||
} finally {
|
||||
await runtime.dispose()
|
||||
}
|
||||
}, 300_000)
|
||||
})
|
||||
Reference in New Issue
Block a user