fix: stream interactive model tool rounds
This commit is contained in:
mesalogo
@@ -119,6 +119,19 @@ function createResponsesEventStream(
|
||||
].join('\n')
|
||||
}
|
||||
|
||||
function createSseEventStream(
|
||||
events: ReadonlyArray<Record<string, unknown>>
|
||||
): string {
|
||||
return [
|
||||
...events.flatMap((event) => [
|
||||
`event: ${event.type as string}`,
|
||||
`data: ${JSON.stringify(event)}`,
|
||||
''
|
||||
]),
|
||||
''
|
||||
].join('\n')
|
||||
}
|
||||
|
||||
function createToolProvider(
|
||||
overrides: Partial<ModelToolProviderLike> = {}
|
||||
): ModelToolProviderLike {
|
||||
@@ -1808,6 +1821,227 @@ describe('ModelAgentRuntime', () => {
|
||||
expect(events.at(-1)).toMatchObject({ type: 'done' })
|
||||
})
|
||||
|
||||
it('streams OpenAI Responses text and reasoning through tool rounds', async () => {
|
||||
const streams = [
|
||||
createSseEventStream([
|
||||
{
|
||||
type: 'response.reasoning_summary_text.delta',
|
||||
delta: '先分析。'
|
||||
},
|
||||
{
|
||||
type: 'response.output_text.delta',
|
||||
delta: '准备读取。'
|
||||
},
|
||||
{
|
||||
type: 'response.completed',
|
||||
response: {
|
||||
id: 'resp-stream-tool-1',
|
||||
model: 'gpt-5',
|
||||
status: 'completed',
|
||||
output: [
|
||||
{
|
||||
id: 'reasoning-stream-1',
|
||||
type: 'reasoning',
|
||||
summary: [
|
||||
{ type: 'summary_text', text: '先分析。' }
|
||||
]
|
||||
},
|
||||
{
|
||||
id: 'message-stream-1',
|
||||
type: 'message',
|
||||
role: 'assistant',
|
||||
status: 'completed',
|
||||
content: [
|
||||
{ type: 'output_text', text: '准备读取。' }
|
||||
]
|
||||
},
|
||||
{
|
||||
id: 'function-stream-1',
|
||||
type: 'function_call',
|
||||
call_id: 'call-stream-1',
|
||||
name: 'workspace_read_text',
|
||||
arguments: '{"path":"README.md"}'
|
||||
}
|
||||
],
|
||||
usage: { input_tokens: 12, output_tokens: 4 }
|
||||
}
|
||||
}
|
||||
]),
|
||||
createSseEventStream([
|
||||
{
|
||||
type: 'response.output_text.delta',
|
||||
delta: '读取完成。'
|
||||
},
|
||||
{
|
||||
type: 'response.completed',
|
||||
response: {
|
||||
id: 'resp-stream-tool-2',
|
||||
model: 'gpt-5',
|
||||
status: 'completed',
|
||||
output: [
|
||||
{
|
||||
id: 'message-stream-2',
|
||||
type: 'message',
|
||||
role: 'assistant',
|
||||
status: 'completed',
|
||||
content: [
|
||||
{ type: 'output_text', text: '读取完成。' }
|
||||
]
|
||||
}
|
||||
],
|
||||
usage: { input_tokens: 20, output_tokens: 5 }
|
||||
}
|
||||
}
|
||||
])
|
||||
]
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
new Response(streams.shift(), {
|
||||
status: 200,
|
||||
headers: { 'content-type': 'text/event-stream' }
|
||||
})
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.openai.com/v1',
|
||||
model: 'gpt-5',
|
||||
protocol: 'openai-responses',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const events = []
|
||||
|
||||
for await (const event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed151',
|
||||
conversationId: 'conversation-responses-streaming-tools',
|
||||
prompt: '读取 README',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
events.push(event)
|
||||
}
|
||||
|
||||
for (const [, init] of fetcher.mock.calls) {
|
||||
expect(JSON.parse(init?.body as string)).toMatchObject({
|
||||
stream: true
|
||||
})
|
||||
}
|
||||
expect(
|
||||
events
|
||||
.filter((event) => event.type === 'reasoning')
|
||||
.map((event) => event.delta)
|
||||
).toEqual(['先分析。'])
|
||||
expect(
|
||||
events
|
||||
.filter((event) => event.type === 'text')
|
||||
.map((event) => event.delta)
|
||||
).toEqual(['准备读取。', '读取完成。'])
|
||||
expect(
|
||||
events.findIndex((event) => event.type === 'text')
|
||||
).toBeLessThan(
|
||||
events.findIndex(
|
||||
(event) =>
|
||||
event.type === 'tool' && event.state === 'running'
|
||||
)
|
||||
)
|
||||
const secondBody = JSON.parse(
|
||||
fetcher.mock.calls[1]?.[1]?.body as string
|
||||
) as { input: Array<Record<string, unknown>> }
|
||||
expect(secondBody.input).toEqual([
|
||||
{
|
||||
role: 'user',
|
||||
content: '读取 README'
|
||||
},
|
||||
{
|
||||
id: 'reasoning-stream-1',
|
||||
type: 'reasoning',
|
||||
summary: [{ type: 'summary_text', text: '先分析。' }]
|
||||
},
|
||||
{
|
||||
id: 'message-stream-1',
|
||||
type: 'message',
|
||||
role: 'assistant',
|
||||
status: 'completed',
|
||||
content: [
|
||||
{ type: 'output_text', text: '准备读取。' }
|
||||
]
|
||||
},
|
||||
{
|
||||
id: 'function-stream-1',
|
||||
type: 'function_call',
|
||||
call_id: 'call-stream-1',
|
||||
name: 'workspace_read_text',
|
||||
arguments: '{"path":"README.md"}'
|
||||
},
|
||||
{
|
||||
type: 'function_call_output',
|
||||
call_id: 'call-stream-1',
|
||||
output: [
|
||||
{
|
||||
type: 'input_text',
|
||||
text: 'tool result'
|
||||
}
|
||||
]
|
||||
}
|
||||
])
|
||||
expect(toolProvider.callTool).toHaveBeenCalledOnce()
|
||||
expect(events.at(-1)).toMatchObject({ type: 'done' })
|
||||
})
|
||||
|
||||
it('rejects an incomplete OpenAI Responses tool stream', async () => {
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
new Response(
|
||||
[
|
||||
'event: response.output_text.delta',
|
||||
`data: ${JSON.stringify({
|
||||
type: 'response.output_text.delta',
|
||||
delta: 'partial'
|
||||
})}`,
|
||||
'',
|
||||
'data: [DONE]',
|
||||
'',
|
||||
''
|
||||
].join('\n'),
|
||||
{
|
||||
status: 200,
|
||||
headers: { 'content-type': 'text/event-stream' }
|
||||
}
|
||||
)
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.openai.com/v1',
|
||||
model: 'gpt-5',
|
||||
protocol: 'openai-responses',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const consume = async (): Promise<void> => {
|
||||
for await (const _event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed153',
|
||||
conversationId: 'conversation-responses-incomplete-tools',
|
||||
prompt: '读取 README',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
void _event
|
||||
}
|
||||
}
|
||||
|
||||
await expect(consume()).rejects.toThrow('流式响应意外中断')
|
||||
expect(toolProvider.callTool).not.toHaveBeenCalled()
|
||||
expect(fetcher).toHaveBeenCalledOnce()
|
||||
})
|
||||
|
||||
it('continues OpenAI Responses with function_call_output', async () => {
|
||||
const responses = [
|
||||
{
|
||||
@@ -1902,7 +2136,7 @@ describe('ModelAgentRuntime', () => {
|
||||
) as Record<string, unknown>
|
||||
expect(firstBody).toMatchObject({
|
||||
model: 'gpt-5',
|
||||
stream: false,
|
||||
stream: true,
|
||||
tools: [
|
||||
{
|
||||
type: 'function',
|
||||
@@ -2197,7 +2431,7 @@ describe('ModelAgentRuntime', () => {
|
||||
fetcher.mock.calls[0]?.[1]?.body as string
|
||||
) as Record<string, unknown>
|
||||
expect(firstBody).toMatchObject({
|
||||
stream: false,
|
||||
stream: true,
|
||||
tools: [
|
||||
{
|
||||
name: 'workspace_read_text',
|
||||
@@ -2233,6 +2467,372 @@ describe('ModelAgentRuntime', () => {
|
||||
})
|
||||
})
|
||||
|
||||
it('streams Anthropic text and thinking through tool rounds', async () => {
|
||||
const streams = [
|
||||
createSseEventStream([
|
||||
{
|
||||
type: 'message_start',
|
||||
message: {
|
||||
id: 'message-stream-tool-1',
|
||||
model: 'claude',
|
||||
usage: { input_tokens: 10 }
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 0,
|
||||
content_block: { type: 'thinking', thinking: '' }
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: { type: 'thinking_delta', thinking: '先分析。' }
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: { type: 'signature_delta', signature: 'signed' }
|
||||
},
|
||||
{ type: 'content_block_stop', index: 0 },
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 1,
|
||||
content_block: { type: 'text', text: '' }
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 1,
|
||||
delta: { type: 'text_delta', text: '准备读取。' }
|
||||
},
|
||||
{ type: 'content_block_stop', index: 1 },
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 2,
|
||||
content_block: {
|
||||
type: 'tool_use',
|
||||
id: 'toolu-stream-1',
|
||||
name: 'workspace_read_text',
|
||||
input: {}
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 2,
|
||||
delta: {
|
||||
type: 'input_json_delta',
|
||||
partial_json: '{"path":'
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 2,
|
||||
delta: {
|
||||
type: 'input_json_delta',
|
||||
partial_json: '"notes.md"}'
|
||||
}
|
||||
},
|
||||
{ type: 'content_block_stop', index: 2 },
|
||||
{
|
||||
type: 'message_delta',
|
||||
delta: { stop_reason: 'tool_use' },
|
||||
usage: { output_tokens: 4 }
|
||||
},
|
||||
{ type: 'message_stop' }
|
||||
]),
|
||||
createSseEventStream([
|
||||
{
|
||||
type: 'message_start',
|
||||
message: {
|
||||
id: 'message-stream-tool-2',
|
||||
model: 'claude',
|
||||
usage: { input_tokens: 18 }
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 0,
|
||||
content_block: { type: 'text', text: '' }
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: { type: 'text_delta', text: '读取完成。' }
|
||||
},
|
||||
{ type: 'content_block_stop', index: 0 },
|
||||
{
|
||||
type: 'message_delta',
|
||||
delta: { stop_reason: 'end_turn' },
|
||||
usage: { output_tokens: 5 }
|
||||
},
|
||||
{ type: 'message_stop' }
|
||||
])
|
||||
]
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
new Response(streams.shift(), {
|
||||
status: 200,
|
||||
headers: { 'content-type': 'text/event-stream' }
|
||||
})
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.anthropic.com',
|
||||
model: 'claude',
|
||||
protocol: 'anthropic-messages',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const events = []
|
||||
|
||||
for await (const event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed152',
|
||||
conversationId: 'conversation-anthropic-streaming-tools',
|
||||
prompt: '读取 notes',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
events.push(event)
|
||||
}
|
||||
|
||||
for (const [, init] of fetcher.mock.calls) {
|
||||
expect(JSON.parse(init?.body as string)).toMatchObject({
|
||||
stream: true
|
||||
})
|
||||
}
|
||||
expect(
|
||||
events
|
||||
.filter((event) => event.type === 'reasoning')
|
||||
.map((event) => event.delta)
|
||||
).toEqual(['先分析。'])
|
||||
expect(
|
||||
events
|
||||
.filter((event) => event.type === 'text')
|
||||
.map((event) => event.delta)
|
||||
).toEqual(['准备读取。', '读取完成。'])
|
||||
expect(
|
||||
events.findIndex((event) => event.type === 'text')
|
||||
).toBeLessThan(
|
||||
events.findIndex(
|
||||
(event) =>
|
||||
event.type === 'tool' && event.state === 'running'
|
||||
)
|
||||
)
|
||||
const secondBody = JSON.parse(
|
||||
fetcher.mock.calls[1]?.[1]?.body as string
|
||||
) as { messages: Array<Record<string, unknown>> }
|
||||
expect(secondBody.messages.at(-2)).toEqual({
|
||||
role: 'assistant',
|
||||
content: [
|
||||
{
|
||||
type: 'thinking',
|
||||
thinking: '先分析。',
|
||||
signature: 'signed'
|
||||
},
|
||||
{
|
||||
type: 'text',
|
||||
text: '准备读取。'
|
||||
},
|
||||
{
|
||||
type: 'tool_use',
|
||||
id: 'toolu-stream-1',
|
||||
name: 'workspace_read_text',
|
||||
input: { path: 'notes.md' }
|
||||
}
|
||||
]
|
||||
})
|
||||
expect(secondBody.messages.at(-1)).toEqual({
|
||||
role: 'user',
|
||||
content: [
|
||||
{
|
||||
type: 'tool_result',
|
||||
tool_use_id: 'toolu-stream-1',
|
||||
content: [{ type: 'text', text: 'tool result' }]
|
||||
}
|
||||
]
|
||||
})
|
||||
expect(toolProvider.callTool).toHaveBeenCalledOnce()
|
||||
expect(events.at(-1)).toMatchObject({ type: 'done' })
|
||||
})
|
||||
|
||||
it('rejects malformed streamed Anthropic tool arguments', async () => {
|
||||
const stream = createSseEventStream([
|
||||
{
|
||||
type: 'message_start',
|
||||
message: {
|
||||
id: 'message-invalid-tool-1',
|
||||
model: 'claude',
|
||||
usage: { input_tokens: 8 }
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 0,
|
||||
content_block: {
|
||||
type: 'tool_use',
|
||||
id: 'toolu-invalid-1',
|
||||
name: 'workspace_read_text',
|
||||
input: {}
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: {
|
||||
type: 'input_json_delta',
|
||||
partial_json: '{"path":'
|
||||
}
|
||||
},
|
||||
{ type: 'content_block_stop', index: 0 },
|
||||
{ type: 'message_stop' }
|
||||
])
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
new Response(stream, {
|
||||
status: 200,
|
||||
headers: { 'content-type': 'text/event-stream' }
|
||||
})
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.anthropic.com',
|
||||
model: 'claude',
|
||||
protocol: 'anthropic-messages',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const consume = async (): Promise<void> => {
|
||||
for await (const _event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed154',
|
||||
conversationId: 'conversation-anthropic-invalid-tools',
|
||||
prompt: '读取 notes',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
void _event
|
||||
}
|
||||
}
|
||||
|
||||
await expect(consume()).rejects.toThrow(
|
||||
'模型返回了无效的工具参数 JSON'
|
||||
)
|
||||
expect(toolProvider.callTool).not.toHaveBeenCalled()
|
||||
expect(fetcher).toHaveBeenCalledOnce()
|
||||
})
|
||||
|
||||
it('rejects a truncated streamed Anthropic tool round', async () => {
|
||||
const stream = createSseEventStream([
|
||||
{
|
||||
type: 'message_start',
|
||||
message: {
|
||||
id: 'message-truncated-stream-1',
|
||||
model: 'claude',
|
||||
usage: { input_tokens: 8 }
|
||||
}
|
||||
},
|
||||
{
|
||||
type: 'content_block_start',
|
||||
index: 0,
|
||||
content_block: { type: 'text', text: '' }
|
||||
},
|
||||
{
|
||||
type: 'content_block_delta',
|
||||
index: 0,
|
||||
delta: { type: 'text_delta', text: 'partial' }
|
||||
},
|
||||
{ type: 'content_block_stop', index: 0 },
|
||||
{
|
||||
type: 'message_delta',
|
||||
delta: { stop_reason: 'max_tokens' },
|
||||
usage: { output_tokens: 4 }
|
||||
},
|
||||
{ type: 'message_stop' }
|
||||
])
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
new Response(stream, {
|
||||
status: 200,
|
||||
headers: { 'content-type': 'text/event-stream' }
|
||||
})
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.anthropic.com',
|
||||
model: 'claude',
|
||||
protocol: 'anthropic-messages',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const consume = async (): Promise<void> => {
|
||||
for await (const _event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed155',
|
||||
conversationId: 'conversation-anthropic-truncated-stream',
|
||||
prompt: '读取 notes',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
void _event
|
||||
}
|
||||
}
|
||||
|
||||
await expect(consume()).rejects.toThrow(
|
||||
'Anthropic 返回未完成结果：max_tokens'
|
||||
)
|
||||
expect(toolProvider.callTool).not.toHaveBeenCalled()
|
||||
})
|
||||
|
||||
it('rejects a truncated Anthropic JSON fallback', async () => {
|
||||
const fetcher = vi.fn<typeof fetch>(async () =>
|
||||
Response.json({
|
||||
id: 'message-truncated-json-1',
|
||||
model: 'claude',
|
||||
content: [{ type: 'text', text: 'partial' }],
|
||||
stop_reason: 'model_context_window_exceeded',
|
||||
usage: { input_tokens: 8, output_tokens: 4 }
|
||||
})
|
||||
)
|
||||
const toolProvider = createToolProvider()
|
||||
const runtime = new ModelAgentRuntime({
|
||||
apiKey: 'test-key',
|
||||
baseUrl: 'https://api.anthropic.com',
|
||||
model: 'claude',
|
||||
protocol: 'anthropic-messages',
|
||||
authentication: 'api-key',
|
||||
fetcher,
|
||||
toolProvider
|
||||
})
|
||||
const consume = async (): Promise<void> => {
|
||||
for await (const _event of runtime.run(
|
||||
{
|
||||
requestId: 'a431666e-5ec8-45e6-beb4-654132eed156',
|
||||
conversationId: 'conversation-anthropic-truncated-json',
|
||||
prompt: '读取 notes',
|
||||
workMode: 'execute'
|
||||
},
|
||||
new AbortController().signal,
|
||||
async () => 'once'
|
||||
)) {
|
||||
void _event
|
||||
}
|
||||
}
|
||||
|
||||
await expect(consume()).rejects.toThrow(
|
||||
'Anthropic 返回未完成结果：model_context_window_exceeded'
|
||||
)
|
||||
expect(toolProvider.callTool).not.toHaveBeenCalled()
|
||||
})
|
||||
|
||||
it('synthesizes and pairs a missing Anthropic tool_use id', async () => {
|
||||
const responses = [
|
||||
{
|
||||
|
||||
@@ -799,6 +799,20 @@ function parseModelToolResponse(
|
||||
throw new Error('模型接口返回格式无效')
|
||||
}
|
||||
if (protocol === 'anthropic') {
|
||||
const stopReason = payload.stop_reason
|
||||
if (
|
||||
stopReason !== undefined &&
|
||||
stopReason !== null &&
|
||||
typeof stopReason !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 模型接口返回了无效停止原因')
|
||||
}
|
||||
if (
|
||||
stopReason === 'max_tokens' ||
|
||||
stopReason === 'model_context_window_exceeded'
|
||||
) {
|
||||
throw new Error(`Anthropic 返回未完成结果：${stopReason}`)
|
||||
}
|
||||
if (!Array.isArray(payload.content)) {
|
||||
throw new Error('Anthropic 模型接口未返回 content')
|
||||
}
|
||||
@@ -1124,6 +1138,349 @@ async function* readBoundedSseBlocks(
|
||||
}
|
||||
}
|
||||
|
||||
async function* readOpenAIResponsesToolStream(
|
||||
response: Response,
|
||||
requestId: string
|
||||
): AsyncGenerator<RuntimeEvent, ModelToolResponse, void> {
|
||||
let answer = ''
|
||||
let reasoning = ''
|
||||
const usage: ModelUsageAccumulator = {
|
||||
reported: false
|
||||
}
|
||||
|
||||
for await (const block of readBoundedSseBlocks(response)) {
|
||||
const parsed = parseSseData(block)
|
||||
if (parsed.stopped) {
|
||||
break
|
||||
}
|
||||
if (parsed.event === undefined) {
|
||||
continue
|
||||
}
|
||||
const providerError = getErrorMessage(parsed.event)
|
||||
if (providerError) {
|
||||
throw new Error(providerError)
|
||||
}
|
||||
const event = getRecord(parsed.event)
|
||||
if (!event) {
|
||||
throw new Error('OpenAI Responses 返回了无效流式事件')
|
||||
}
|
||||
if (event.type === 'response.failed') {
|
||||
const failedResponse = getRecord(event.response)
|
||||
throw new Error(
|
||||
getErrorMessage(failedResponse) ?? 'OpenAI Responses 请求失败'
|
||||
)
|
||||
}
|
||||
if (event.type === 'response.incomplete') {
|
||||
const incompleteResponse = getRecord(event.response)
|
||||
const details = getRecord(incompleteResponse?.incomplete_details)
|
||||
const reason =
|
||||
typeof details?.reason === 'string'
|
||||
? `：${details.reason.slice(0, 200)}`
|
||||
: ''
|
||||
throw new Error(`OpenAI Responses 返回未完成结果${reason}`)
|
||||
}
|
||||
applyUsageUpdate(usage, getUsageUpdate(event, 'openai'))
|
||||
|
||||
const reasoningDelta = getOpenAIResponsesReasoningDelta(event)
|
||||
if (reasoningDelta) {
|
||||
reasoning += reasoningDelta
|
||||
yield {
|
||||
requestId,
|
||||
type: 'reasoning',
|
||||
delta: reasoningDelta
|
||||
}
|
||||
}
|
||||
const textDelta = getOpenAIResponsesTextDelta(event)
|
||||
if (textDelta) {
|
||||
answer += textDelta
|
||||
yield {
|
||||
requestId,
|
||||
type: 'text',
|
||||
delta: textDelta
|
||||
}
|
||||
}
|
||||
if (event.type !== 'response.completed') {
|
||||
continue
|
||||
}
|
||||
|
||||
const completedResponse = getRecord(event.response)
|
||||
const result = parseModelToolResponse(
|
||||
completedResponse,
|
||||
'openai-responses'
|
||||
)
|
||||
applyUsageUpdate(usage, result.usage)
|
||||
if (result.reasoning !== reasoning) {
|
||||
if (!result.reasoning.startsWith(reasoning)) {
|
||||
throw new Error(
|
||||
'OpenAI Responses 流式推理与完成结果不一致'
|
||||
)
|
||||
}
|
||||
const remainingReasoning = result.reasoning.slice(reasoning.length)
|
||||
if (remainingReasoning) {
|
||||
reasoning += remainingReasoning
|
||||
yield {
|
||||
requestId,
|
||||
type: 'reasoning',
|
||||
delta: remainingReasoning
|
||||
}
|
||||
}
|
||||
}
|
||||
if (result.text !== answer) {
|
||||
if (!result.text.startsWith(answer)) {
|
||||
throw new Error('OpenAI Responses 流式文本与完成结果不一致')
|
||||
}
|
||||
const remainingText = result.text.slice(answer.length)
|
||||
if (remainingText) {
|
||||
answer += remainingText
|
||||
yield {
|
||||
requestId,
|
||||
type: 'text',
|
||||
delta: remainingText
|
||||
}
|
||||
}
|
||||
}
|
||||
return {
|
||||
...result,
|
||||
text: answer,
|
||||
reasoning,
|
||||
usage,
|
||||
streamed: true
|
||||
}
|
||||
}
|
||||
|
||||
throw new Error('模型接口流式响应意外中断')
|
||||
}
|
||||
|
||||
type AnthropicStreamBlock = {
|
||||
content: Record<string, unknown>
|
||||
initialInput?: unknown
|
||||
kind: 'other' | 'text' | 'thinking' | 'tool_use'
|
||||
open: boolean
|
||||
partialJson: string
|
||||
}
|
||||
|
||||
function getAnthropicStreamIndex(
|
||||
event: Record<string, unknown>
|
||||
): number {
|
||||
if (
|
||||
!Number.isSafeInteger(event.index) ||
|
||||
(event.index as number) < 0
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式内容块序号')
|
||||
}
|
||||
return event.index as number
|
||||
}
|
||||
|
||||
async function* readAnthropicToolStream(
|
||||
response: Response,
|
||||
requestId: string
|
||||
): AsyncGenerator<RuntimeEvent, ModelToolResponse, void> {
|
||||
const blocks = new Map<number, AnthropicStreamBlock>()
|
||||
const usage: ModelUsageAccumulator = {
|
||||
reported: false
|
||||
}
|
||||
let answer = ''
|
||||
let reasoning = ''
|
||||
let stopReason: unknown
|
||||
let toolCallCount = 0
|
||||
|
||||
for await (const block of readBoundedSseBlocks(response)) {
|
||||
const parsed = parseSseData(block)
|
||||
if (parsed.stopped) {
|
||||
break
|
||||
}
|
||||
if (parsed.event === undefined) {
|
||||
continue
|
||||
}
|
||||
const providerError = getErrorMessage(parsed.event)
|
||||
if (providerError) {
|
||||
throw new Error(providerError)
|
||||
}
|
||||
const event = getRecord(parsed.event)
|
||||
if (!event || typeof event.type !== 'string') {
|
||||
throw new Error('Anthropic 返回了无效流式事件')
|
||||
}
|
||||
applyUsageUpdate(usage, getUsageUpdate(event, 'anthropic'))
|
||||
|
||||
if (event.type === 'message_delta') {
|
||||
const delta = getRecord(event.delta)
|
||||
if (delta?.stop_reason !== undefined) {
|
||||
stopReason = delta.stop_reason
|
||||
}
|
||||
}
|
||||
|
||||
if (event.type === 'content_block_start') {
|
||||
const index = getAnthropicStreamIndex(event)
|
||||
const content = getRecord(event.content_block)
|
||||
if (!content || blocks.has(index)) {
|
||||
throw new Error('Anthropic 返回了无效流式内容块')
|
||||
}
|
||||
const next: AnthropicStreamBlock = {
|
||||
content: { ...content },
|
||||
kind: 'other',
|
||||
open: true,
|
||||
partialJson: ''
|
||||
}
|
||||
if (content.type === 'text') {
|
||||
if (
|
||||
content.text !== undefined &&
|
||||
typeof content.text !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式文本块')
|
||||
}
|
||||
const text = typeof content.text === 'string' ? content.text : ''
|
||||
next.kind = 'text'
|
||||
next.content.text = text
|
||||
if (text) {
|
||||
answer += text
|
||||
yield {
|
||||
requestId,
|
||||
type: 'text',
|
||||
delta: text
|
||||
}
|
||||
}
|
||||
} else if (content.type === 'thinking') {
|
||||
if (
|
||||
content.thinking !== undefined &&
|
||||
typeof content.thinking !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式推理块')
|
||||
}
|
||||
const thinking =
|
||||
typeof content.thinking === 'string' ? content.thinking : ''
|
||||
next.kind = 'thinking'
|
||||
next.content.thinking = thinking
|
||||
if (thinking) {
|
||||
reasoning += thinking
|
||||
yield {
|
||||
requestId,
|
||||
type: 'reasoning',
|
||||
delta: thinking
|
||||
}
|
||||
}
|
||||
} else if (content.type === 'tool_use') {
|
||||
toolCallCount += 1
|
||||
if (toolCallCount > maxToolCallsPerRun) {
|
||||
throw new Error('模型单轮工具调用超过安全限制')
|
||||
}
|
||||
const identity = parseToolCallIdentity(content.id, content.name)
|
||||
next.kind = 'tool_use'
|
||||
next.initialInput = content.input
|
||||
next.content.id = identity.id
|
||||
next.content.name = identity.name
|
||||
next.content.input = {}
|
||||
}
|
||||
blocks.set(index, next)
|
||||
continue
|
||||
}
|
||||
|
||||
if (event.type === 'content_block_delta') {
|
||||
const index = getAnthropicStreamIndex(event)
|
||||
const current = blocks.get(index)
|
||||
const delta = getRecord(event.delta)
|
||||
if (!current?.open || !delta || typeof delta.type !== 'string') {
|
||||
throw new Error('Anthropic 返回了无效流式内容增量')
|
||||
}
|
||||
if (delta.type === 'text_delta') {
|
||||
if (current.kind !== 'text' || typeof delta.text !== 'string') {
|
||||
throw new Error('Anthropic 返回了无效流式文本增量')
|
||||
}
|
||||
current.content.text =
|
||||
`${current.content.text as string}${delta.text}`
|
||||
if (delta.text) {
|
||||
answer += delta.text
|
||||
yield {
|
||||
requestId,
|
||||
type: 'text',
|
||||
delta: delta.text
|
||||
}
|
||||
}
|
||||
} else if (delta.type === 'thinking_delta') {
|
||||
if (
|
||||
current.kind !== 'thinking' ||
|
||||
typeof delta.thinking !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式推理增量')
|
||||
}
|
||||
current.content.thinking =
|
||||
`${current.content.thinking as string}${delta.thinking}`
|
||||
if (delta.thinking) {
|
||||
reasoning += delta.thinking
|
||||
yield {
|
||||
requestId,
|
||||
type: 'reasoning',
|
||||
delta: delta.thinking
|
||||
}
|
||||
}
|
||||
} else if (delta.type === 'signature_delta') {
|
||||
if (
|
||||
current.kind !== 'thinking' ||
|
||||
typeof delta.signature !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式签名增量')
|
||||
}
|
||||
current.content.signature =
|
||||
`${typeof current.content.signature === 'string'
|
||||
? current.content.signature
|
||||
: ''}${delta.signature}`
|
||||
} else if (delta.type === 'input_json_delta') {
|
||||
if (
|
||||
current.kind !== 'tool_use' ||
|
||||
typeof delta.partial_json !== 'string'
|
||||
) {
|
||||
throw new Error('Anthropic 返回了无效流式工具参数增量')
|
||||
}
|
||||
current.partialJson += delta.partial_json
|
||||
if (
|
||||
Buffer.byteLength(current.partialJson) >
|
||||
maxToolArgumentBytes
|
||||
) {
|
||||
throw new Error('模型工具参数超过 128KB 安全限制')
|
||||
}
|
||||
}
|
||||
continue
|
||||
}
|
||||
|
||||
if (event.type === 'content_block_stop') {
|
||||
const index = getAnthropicStreamIndex(event)
|
||||
const current = blocks.get(index)
|
||||
if (!current?.open) {
|
||||
throw new Error('Anthropic 返回了无效流式内容结束事件')
|
||||
}
|
||||
current.open = false
|
||||
if (current.kind === 'tool_use') {
|
||||
current.content.input = parseToolArguments(
|
||||
current.partialJson || current.initialInput || {}
|
||||
)
|
||||
}
|
||||
continue
|
||||
}
|
||||
|
||||
if (event.type !== 'message_stop') {
|
||||
continue
|
||||
}
|
||||
if ([...blocks.values()].some((item) => item.open)) {
|
||||
throw new Error('Anthropic 流式响应包含未结束的内容块')
|
||||
}
|
||||
const content = [...blocks.entries()]
|
||||
.sort(([left], [right]) => left - right)
|
||||
.map(([, item]) => item.content)
|
||||
const result = parseModelToolResponse(
|
||||
{ content, stop_reason: stopReason },
|
||||
'anthropic'
|
||||
)
|
||||
return {
|
||||
...result,
|
||||
text: answer,
|
||||
reasoning,
|
||||
usage,
|
||||
streamed: true
|
||||
}
|
||||
}
|
||||
|
||||
throw new Error('模型接口流式响应意外中断')
|
||||
}
|
||||
|
||||
export class ModelAgentRuntime implements AgentRuntime {
|
||||
readonly runtimeId = 'model'
|
||||
readonly requiresToolApproval = false
|
||||
@@ -1784,7 +2141,7 @@ export class ModelAgentRuntime implements AgentRuntime {
|
||||
? {
|
||||
model: this.options.model,
|
||||
max_output_tokens: this.maxOutputTokens,
|
||||
stream: false,
|
||||
stream: true,
|
||||
instructions: system,
|
||||
input: messages,
|
||||
tools: providerTools
|
||||
@@ -1793,7 +2150,7 @@ export class ModelAgentRuntime implements AgentRuntime {
|
||||
? {
|
||||
model: this.options.model,
|
||||
max_tokens: this.maxOutputTokens,
|
||||
stream: false,
|
||||
stream: true,
|
||||
system,
|
||||
messages,
|
||||
tools: providerTools
|
||||
@@ -1843,12 +2200,22 @@ export class ModelAgentRuntime implements AgentRuntime {
|
||||
detail ?? `模型接口请求失败（HTTP ${response.status}）`
|
||||
)
|
||||
}
|
||||
const isEventStream = response.headers
|
||||
.get('content-type')
|
||||
?.toLocaleLowerCase()
|
||||
.includes('text/event-stream') === true
|
||||
if (responses && isEventStream) {
|
||||
return yield* readOpenAIResponsesToolStream(
|
||||
response,
|
||||
requestId
|
||||
)
|
||||
}
|
||||
if (anthropic && isEventStream) {
|
||||
return yield* readAnthropicToolStream(response, requestId)
|
||||
}
|
||||
if (
|
||||
streamOpenAIChat &&
|
||||
response.headers
|
||||
.get('content-type')
|
||||
?.toLocaleLowerCase()
|
||||
.includes('text/event-stream')
|
||||
isEventStream
|
||||
) {
|
||||
const streamedToolCalls = new Map<
|
||||
number,
|
||||
|
||||
Reference in New Issue
Block a user