Merge upstream/main into fix/anthropic-schema-format

2026-04-02 15:42:28 +02:00
parent 0c88dea247 1514220ee7
commit 6319df02f0
71 changed files with 6989 additions and 931 deletions
--- a/src/services/api/openaiShim.ts
+++ b/src/services/api/openaiShim.ts
@@ -14,8 +14,16 @@
 *   OPENAI_BASE_URL=http://...        — base URL (default: https://api.openai.com/v1)
 *   OPENAI_MODEL=gpt-4o              — default model override
 *   CODEX_API_KEY / ~/.codex/auth.json — Codex auth for codexplan/codexspark
+ *
+ * GitHub Models (models.github.ai), OpenAI-compatible:
+ *   CLAUDE_CODE_USE_GITHUB=1         — enable GitHub inference (no need for USE_OPENAI)
+ *   GITHUB_TOKEN or GH_TOKEN         — PAT with models access (mapped to Bearer auth)
+ *   OPENAI_MODEL                     — optional; use github:copilot or openai/gpt-4.1 style IDs
 */

+import { APIError } from '@anthropic-ai/sdk'
+import { isEnvTruthy } from '../../utils/envUtils.js'
+import { hydrateGithubModelsTokenFromSecureStorage } from '../../utils/githubModelsCredentials.js'
 import {
  codexStreamToAnthropic,
  collectCodexCompletedResponse,
@@ -26,10 +34,31 @@ import {
  type ShimCreateParams,
 } from './codexShim.js'
 import {
+  isLocalProviderUrl,
  resolveCodexApiCredentials,
  resolveProviderRequest,
 } from './providerConfig.js'
 import { stripIncompatibleSchemaKeywords } from '../../utils/schemaSanitizer.js'
+import { redactSecretValueForDisplay } from '../../utils/providerProfile.js'
+
+const GITHUB_MODELS_DEFAULT_BASE = 'https://models.github.ai/inference'
+const GITHUB_API_VERSION = '2022-11-28'
+const GITHUB_429_MAX_RETRIES = 3
+const GITHUB_429_BASE_DELAY_SEC = 1
+const GITHUB_429_MAX_DELAY_SEC = 32
+
+function isGithubModelsMode(): boolean {
+  return isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)
+}
+
+function formatRetryAfterHint(response: Response): string {
+  const ra = response.headers.get('retry-after')
+  return ra ? ` (Retry-After: ${ra})` : ''
+}
+
+function sleepMs(ms: number): Promise<void> {
+  return new Promise(resolve => setTimeout(resolve, ms))
+}

 // ---------------------------------------------------------------------------
 // Types — minimal subset of Anthropic SDK types we need to produce
@@ -188,7 +217,10 @@ function convertMessages(

        const assistantMsg: OpenAIMessage = {
          role: 'assistant',
-          content: convertContentBlocks(textContent) as string,
+          content: (() => {
+            const c = convertContentBlocks(textContent)
+            return typeof c === 'string' ? c : Array.isArray(c) ? c.map((p: { text?: string }) => p.text ?? '').join('') : ''
+          })(),
        }

        if (toolUses.length > 0) {
@@ -217,7 +249,10 @@ function convertMessages(
      } else {
        result.push({
          role: 'assistant',
-          content: convertContentBlocks(content) as string,
+          content: (() => {
+            const c = convertContentBlocks(content)
+            return typeof c === 'string' ? c : Array.isArray(c) ? c.map((p: { text?: string }) => p.text ?? '').join('') : ''
+          })(),
        })
      }
    }
@@ -296,9 +331,7 @@ function normalizeSchemaForOpenAI(
 function convertTools(
  tools: Array<{ name: string; description?: string; input_schema?: Record<string, unknown> }>,
 ): OpenAITool[] {
-  const isGemini =
-    process.env.CLAUDE_CODE_USE_GEMINI === '1' ||
-    process.env.CLAUDE_CODE_USE_GEMINI === 'true'
+  const isGemini = isEnvTruthy(process.env.CLAUDE_CODE_USE_GEMINI)

  return tools
    .filter(t => t.name !== 'ToolSearchTool') // Not relevant for OpenAI
@@ -595,7 +628,8 @@ async function* openaiStreamToAnthropic(
      if (
        !hasEmittedFinalUsage &&
        chunkUsage &&
-        (chunk.choices?.length ?? 0) === 0
+        (chunk.choices?.length ?? 0) === 0 &&
+        lastStopReason !== null
      ) {
        yield {
          type: 'message_delta',
@@ -633,9 +667,11 @@ class OpenAIShimStream {

 class OpenAIShimMessages {
  private defaultHeaders: Record<string, string>
+  private reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh'

-  constructor(defaultHeaders: Record<string, string>) {
+  constructor(defaultHeaders: Record<string, string>, reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh') {
    this.defaultHeaders = defaultHeaders
+    this.reasoningEffort = reasoningEffort
  }

  create(
@@ -644,9 +680,12 @@ class OpenAIShimMessages {
  ) {
    const self = this

+    let httpResponse: Response | undefined
+
    const promise = (async () => {
-      const request = resolveProviderRequest({ model: params.model })
+      const request = resolveProviderRequest({ model: params.model, reasoningEffortOverride: self.reasoningEffort })
      const response = await self._doRequest(request, params, options)
+      httpResponse = response

      if (params.stream) {
        return new OpenAIShimStream(
@@ -673,8 +712,9 @@ class OpenAIShimMessages {
          const data = await promise
          return {
            data,
-            response: new Response(),
-            request_id: makeMessageId(),
+            response: httpResponse ?? new Response(),
+            request_id:
+              httpResponse?.headers.get('x-request-id') ?? makeMessageId(),
          }
        }

@@ -692,8 +732,11 @@ class OpenAIShimMessages {
        const authHint = credentials.authPath
          ? ` or place a Codex auth.json at ${credentials.authPath}`
          : ''
+        const safeModel =
+          redactSecretValueForDisplay(request.requestedModel, process.env) ??
+          'the requested model'
        throw new Error(
-          `Codex auth is required for ${request.requestedModel}. Set CODEX_API_KEY${authHint}.`,
+          `Codex auth is required for ${safeModel}. Set CODEX_API_KEY${authHint}.`,
        )
      }
      if (!credentials.accountId) {
@@ -752,10 +795,16 @@ class OpenAIShimMessages {
      body.max_completion_tokens = maxCompletionTokensValue
    }

-    if (params.stream) {
+    if (params.stream && !isLocalProviderUrl(request.baseUrl)) {
      body.stream_options = { include_usage: true }
    }

+    const isGithub = isGithubModelsMode()
+    if (isGithub && body.max_completion_tokens !== undefined) {
+      body.max_tokens = body.max_completion_tokens
+      delete body.max_completion_tokens
+    }
+
    if (params.temperature !== undefined) body.temperature = params.temperature
    if (params.top_p !== undefined) body.top_p = params.top_p

@@ -805,6 +854,11 @@ class OpenAIShimMessages {
      }
    }

+    if (isGithub) {
+      headers.Accept = 'application/vnd.github.v3+json'
+      headers['X-GitHub-Api-Version'] = GITHUB_API_VERSION
+    }
+
    // Build the chat completions URL
    // Azure Cognitive Services / Azure OpenAI require a deployment-specific path
    // and an api-version query parameter.
@@ -827,19 +881,50 @@ class OpenAIShimMessages {
      chatCompletionsUrl = `${request.baseUrl}/chat/completions`
    }

-    const response = await fetch(chatCompletionsUrl, {
-      method: 'POST',
+    const fetchInit = {
+      method: 'POST' as const,
      headers,
      body: JSON.stringify(body),
      signal: options?.signal,
-    })
-
-    if (!response.ok) {
-      const errorBody = await response.text().catch(() => 'unknown error')
-      throw new Error(`OpenAI API error ${response.status}: ${errorBody}`)
    }

-    return response
+    const maxAttempts = isGithub ? GITHUB_429_MAX_RETRIES : 1
+    let response: Response | undefined
+    for (let attempt = 0; attempt < maxAttempts; attempt++) {
+      response = await fetch(chatCompletionsUrl, fetchInit)
+      if (response.ok) {
+        return response
+      }
+      if (
+        isGithub &&
+        response.status === 429 &&
+        attempt < maxAttempts - 1
+      ) {
+        await response.text().catch(() => {})
+        const delaySec = Math.min(
+          GITHUB_429_BASE_DELAY_SEC * 2 ** attempt,
+          GITHUB_429_MAX_DELAY_SEC,
+        )
+        await sleepMs(delaySec * 1000)
+        continue
+      }
+      const errorBody = await response.text().catch(() => 'unknown error')
+      const rateHint =
+        isGithub && response.status === 429 ? formatRetryAfterHint(response) : ''
+      let errorResponse: object | undefined
+      try { errorResponse = JSON.parse(errorBody) } catch { /* raw text */ }
+      throw APIError.generate(
+        response.status,
+        errorResponse,
+        `OpenAI API error ${response.status}: ${errorBody}${rateHint}`,
+        response.headers as unknown as Record<string, string>,
+      )
+    }
+
+    throw APIError.generate(
+      500, undefined, 'OpenAI shim: request loop exited unexpectedly',
+      {} as Record<string, string>,
+    )
  }

  private _convertNonStreamingResponse(
@@ -849,7 +934,10 @@ class OpenAIShimMessages {
      choices?: Array<{
        message?: {
          role?: string
-          content?: string | null
+          content?:
+            | string
+            | null
+            | Array<{ type?: string; text?: string }>
          tool_calls?: Array<{
            id: string
            function: { name: string; arguments: string }
@@ -868,8 +956,25 @@ class OpenAIShimMessages {
    const choice = data.choices?.[0]
    const content: Array<Record<string, unknown>> = []

-    if (choice?.message?.content) {
-      content.push({ type: 'text', text: choice.message.content })
+    const rawContent = choice?.message?.content
+    if (typeof rawContent === 'string' && rawContent) {
+      content.push({ type: 'text', text: rawContent })
+    } else if (Array.isArray(rawContent) && rawContent.length > 0) {
+      const parts: string[] = []
+      for (const part of rawContent) {
+        if (
+          part &&
+          typeof part === 'object' &&
+          part.type === 'text' &&
+          typeof part.text === 'string'
+        ) {
+          parts.push(part.text)
+        }
+      }
+      const joined = parts.join('\n')
+      if (joined) {
+        content.push({ type: 'text', text: joined })
+      }
    }

    if (choice?.message?.tool_calls) {
@@ -917,9 +1022,11 @@ class OpenAIShimMessages {

 class OpenAIShimBeta {
  messages: OpenAIShimMessages
+  reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh'

-  constructor(defaultHeaders: Record<string, string>) {
-    this.messages = new OpenAIShimMessages(defaultHeaders)
+  constructor(defaultHeaders: Record<string, string>, reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh') {
+    this.messages = new OpenAIShimMessages(defaultHeaders, reasoningEffort)
+    this.reasoningEffort = reasoningEffort
  }
 }

@@ -927,13 +1034,13 @@ export function createOpenAIShimClient(options: {
  defaultHeaders?: Record<string, string>
  maxRetries?: number
  timeout?: number
+  reasoningEffort?: 'low' | 'medium' | 'high' | 'xhigh'
 }): unknown {
+  hydrateGithubModelsTokenFromSecureStorage()
+
  // When Gemini provider is active, map Gemini env vars to OpenAI-compatible ones
  // so the existing providerConfig.ts infrastructure picks them up correctly.
-  if (
-    process.env.CLAUDE_CODE_USE_GEMINI === '1' ||
-    process.env.CLAUDE_CODE_USE_GEMINI === 'true'
-  ) {
+  if (isEnvTruthy(process.env.CLAUDE_CODE_USE_GEMINI)) {
    process.env.OPENAI_BASE_URL ??=
      process.env.GEMINI_BASE_URL ??
      'https://generativelanguage.googleapis.com/v1beta/openai'
@@ -942,11 +1049,15 @@ export function createOpenAIShimClient(options: {
    if (process.env.GEMINI_MODEL && !process.env.OPENAI_MODEL) {
      process.env.OPENAI_MODEL = process.env.GEMINI_MODEL
    }
+  } else if (isEnvTruthy(process.env.CLAUDE_CODE_USE_GITHUB)) {
+    process.env.OPENAI_BASE_URL ??= GITHUB_MODELS_DEFAULT_BASE
+    process.env.OPENAI_API_KEY ??=
+      process.env.GITHUB_TOKEN ?? process.env.GH_TOKEN ?? ''
  }

  const beta = new OpenAIShimBeta({
    ...(options.defaultHeaders ?? {}),
-  })
+  }, options.reasoningEffort)

  return {
    beta,