versun/claude-code
1
0
Fork 0
mirror of https://github.com/claude-code-best/claude-code.git synced 2026-06-22 00:05:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'guunergooner/fix/openai-stop-reason-usage'

Browse Source
This commit is contained in:
claude-code-best
2026-04-10 22:27:16 +08:00
parent b060eabda9 c82f59943c
commit 7088fe3c8b
6 changed files with 865 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

138
src/services/api/openai/index.ts
View File

@@ -30,6 +30,7 @@ import { logForDebugging } from '../../../utils/debug.js'
import { addToTotalSessionCost } from '../../../cost-tracker.js'
import { calculateUSDCost } from '../../../utils/modelCost.js'
import { isEnvTruthy, isEnvDefinedFalsy } from '../../../utils/envUtils.js'
import { getModelMaxOutputTokens } from '../../../utils/context.js'
import type { Options } from '../claude.js'
import { randomUUID } from 'crypto'
import {
@@ -87,16 +88,18 @@ export function buildOpenAIRequestBody(params: {
tools: any[]
toolChoice: any
enableThinking: boolean
maxTokens: number
temperatureOverride?: number
}): ChatCompletionCreateParamsStreaming & {
thinking?: { type: string }
enable_thinking?: boolean
chat_template_kwargs?: { thinking: boolean }
} {
const { model, messages, tools, toolChoice, enableThinking, temperatureOverride } = params
const { model, messages, tools, toolChoice, enableThinking, maxTokens, temperatureOverride } = params
return {
model,
messages,
max_tokens: maxTokens,
...(tools.length > 0 && {
tools,
...(toolChoice && { tool_choice: toolChoice }),
@@ -120,6 +123,56 @@ export function buildOpenAIRequestBody(params: {
}
}
/**
* Assemble the final AssistantMessage (and optional max_tokens error) from
* accumulated stream state. Extracted to avoid duplication between the
* `message_stop` handler and the post-loop safety fallback.
*/
function assembleFinalAssistantOutputs(params: {
partialMessage: any
contentBlocks: Record<number, any>
tools: Tools
agentId: string | undefined
usage: { input_tokens: number; output_tokens: number; cache_creation_input_tokens: number; cache_read_input_tokens: number }
stopReason: string | null
maxTokens: number
}): (AssistantMessage | SystemAPIErrorMessage)[] {
const { partialMessage, contentBlocks, tools, agentId, usage, stopReason, maxTokens } = params
const outputs: (AssistantMessage | SystemAPIErrorMessage)[] = []
const allBlocks = Object.keys(contentBlocks)
.sort((a, b) => Number(a) - Number(b))
.map(k => contentBlocks[Number(k)])
.filter(Boolean)
if (allBlocks.length > 0) {
outputs.push({
message: {
...partialMessage,
content: normalizeContentFromAPI(allBlocks, tools, agentId),
usage,
stop_reason: stopReason,
stop_sequence: null,
},
requestId: undefined,
type: 'assistant',
uuid: randomUUID(),
timestamp: new Date().toISOString(),
} as AssistantMessage)
}
if (stopReason === 'max_tokens') {
outputs.push(createAssistantAPIErrorMessage({
content: `Output truncated: response exceeded the ${maxTokens} token limit. ` +
`Set CLAUDE_CODE_MAX_OUTPUT_TOKENS to override.`,
apiError: 'max_output_tokens',
error: 'max_output_tokens',
}))
}
return outputs
}
/**
* OpenAI-compatible query path. Converts Anthropic-format messages/tools to
* OpenAI format, calls the OpenAI-compatible endpoint, and converts the
@@ -222,7 +275,20 @@ export async function* queryModelOpenAI(
)
}
// 10. Get client and make streaming request
// 10. Compute max_tokens — required by most OpenAI-compatible endpoints.
// Without this the server uses a tiny default, and when
// thinking is enabled the thinking phase consumes the entire budget
// leaving no tokens for the final response.
//
// Use upperLimit (not the slot-cap default) because the Anthropic path's
// slot-reservation cap (CAPPED_DEFAULT_MAX_TOKENS=8k) is paired with an
// auto-retry at 64k in query.ts. The OpenAI path has no such retry, so
// using the capped 8k default would silently truncate responses in
// multi-turn conversations where thinking consumes most of the budget.
const { upperLimit } = getModelMaxOutputTokens(openaiModel)
const maxTokens = options.maxOutputTokensOverride ?? upperLimit
// 11. Get client
const client = getOpenAIClient({
maxRetries: 0,
fetchOverride: options.fetchOverride as unknown as typeof fetch,
@@ -233,13 +299,14 @@ export async function* queryModelOpenAI(
`[OpenAI] Calling model=${openaiModel}, messages=${openaiMessages.length}, tools=${openaiTools.length}, thinking=${enableThinking}`,
)
// 11. Call OpenAI API with streaming
// 12. Call OpenAI API with streaming
const requestBody = buildOpenAIRequestBody({
model: openaiModel,
messages: openaiMessages,
tools: openaiTools,
toolChoice: openaiToolChoice,
enableThinking,
maxTokens,
temperatureOverride: options.temperatureOverride,
})
const stream = await client.chat.completions.create(
@@ -254,6 +321,7 @@ export async function* queryModelOpenAI(
// Accumulate content blocks and usage, same as the Anthropic path in claude.ts
const contentBlocks: Record<number, any> = {}
let partialMessage: any
let stopReason: string | null = null
let usage = {
input_tokens: 0,
output_tokens: 0,
@@ -307,21 +375,7 @@ export async function* queryModelOpenAI(
break
}
case 'content_block_stop': {
const idx = (event as any).index
const block = contentBlocks[idx]
if (!block || !partialMessage) break
const m: AssistantMessage = {
message: {
...partialMessage,
content: normalizeContentFromAPI([block], tools, options.agentId),
},
requestId: undefined,
type: 'assistant',
uuid: randomUUID(),
timestamp: new Date().toISOString(),
}
yield m
// Block accumulation is complete; assembly happens at message_stop.
break
}
case 'message_delta': {
@@ -329,21 +383,33 @@ export async function* queryModelOpenAI(
if (deltaUsage) {
usage = { ...usage, ...deltaUsage }
}
// Update the stop_reason on the last yielded message
// (we don't have a reference here, but the consumer handles this)
if ((event as any).delta?.stop_reason != null) {
stopReason = (event as any).delta.stop_reason
}
break
}
case 'message_stop':
case 'message_stop': {
// Assemble ONE AssistantMessage with ALL content blocks, matching the
// Anthropic SDK path. Real usage (input + output tokens) is available
// here and injected so tokenCountWithEstimation() can read it.
if (partialMessage) {
for (const output of assembleFinalAssistantOutputs({
partialMessage, contentBlocks, tools, agentId: options.agentId,
usage, stopReason, maxTokens,
})) {
yield output
}
// Reset partialMessage so the post-loop safety fallback does not
// yield a second identical AssistantMessage.
partialMessage = null
}
// Track cost and token usage
if (usage.input_tokens + usage.output_tokens > 0) {
const costUSD = calculateUSDCost(openaiModel, usage as any)
addToTotalSessionCost(costUSD, usage as any, options.model)
}
break
}
// Track cost and token usage (matching the Anthropic path in claude.ts)
if (
event.type === 'message_stop' &&
usage.input_tokens + usage.output_tokens > 0
) {
const costUSD = calculateUSDCost(openaiModel, usage as any)
addToTotalSessionCost(costUSD, usage as any, options.model)
}
}
// Also yield as StreamEvent for real-time display (matching Anthropic path)
@@ -353,6 +419,16 @@ export async function* queryModelOpenAI(
...(event.type === 'message_start' ? { ttftMs } : undefined),
} as StreamEvent
}
// Safety: if stream ended without message_stop, assemble and yield whatever we have
if (partialMessage) {
for (const output of assembleFinalAssistantOutputs({
partialMessage, contentBlocks, tools, agentId: options.agentId,
usage, stopReason, maxTokens,
})) {
yield output
}
}
} catch (error) {
const errorMessage = error instanceof Error ? error.message : String(error)
logForDebugging(`[OpenAI] Error: ${errorMessage}`, { level: 'error' })
@@ -362,4 +438,4 @@ export async function* queryModelOpenAI(
error: (error instanceof Error ? error : new Error(String(error))) as unknown as SDKAssistantMessageError,
})
}
}
}