From a6f8800630439ac081b33967428bf2cd91dd46c7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E8=B0=A2=E6=A0=8B=E6=A2=81?= Date: Tue, 24 Feb 2026 23:47:15 +0800 Subject: [PATCH] =?UTF-8?q?feat:=20=E7=BB=9F=E4=B8=80=E7=AB=AF=E7=82=B9?= =?UTF-8?q?=E8=BF=94=E5=9B=9E=E6=A0=87=E5=87=86=20OpenAI=20Chat=20Completi?= =?UTF-8?q?ons=20=E6=A0=BC=E5=BC=8F?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 为 unified /v1/chat/completions 端点添加双向格式转换： - Codex: Chat Completions 请求转 Responses API 格式，响应转回 Chat Completions（流式+非流式） - Gemini: 原生响应转为标准 OpenAI Chat Completions 格式（流式+非流式） - 工具名 64 字符缩短与逆向恢复，非 function 工具原样透传 - 注入 Codex CLI 系统提示词，与 handleResponses 适配行为一致 --- src/routes/openaiRoutes.js | 12 +- src/routes/unified.js | 517 +++++++++++++++++++++++- src/services/codexToOpenAI.js | 717 +++++++++++++++++++++++++++++++++ src/services/geminiToOpenAI.js | 392 ++++++++++++++++++ 4 files changed, 1625 insertions(+), 13 deletions(-) create mode 100644 src/services/codexToOpenAI.js create mode 100644 src/services/geminiToOpenAI.js diff --git a/src/routes/openaiRoutes.js b/src/routes/openaiRoutes.js index b912a253..15bbb3a3 100644 --- a/src/routes/openaiRoutes.js +++ b/src/routes/openaiRoutes.js @@ -16,6 +16,10 @@ const { updateRateLimitCounters } = require('../utils/rateLimitHelper') const { IncrementalSSEParser } = require('../utils/sseParser') const { getSafeMessage } = require('../utils/errorSanitizer') +// Codex CLI 系统提示词（非 Codex CLI 客户端请求时注入，统一端点也使用） +const CODEX_CLI_INSTRUCTIONS = + "You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.\n\n## General\n\n- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)\n\n## Editing constraints\n\n- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.\n- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like \"Assigns the value to the variable\", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.\n- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).\n- You may be in a dirty git worktree.\n * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.\n * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.\n * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.\n * If the changes are in unrelated files, just ignore them and don't revert them.\n- Do not amend a commit unless explicitly requested to do so.\n- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.\n- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.\n\n## Plan tool\n\nWhen using the planning tool:\n- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).\n- Do not make single-step plans.\n- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.\n\n## Codex CLI harness, sandboxing, and approvals\n\nThe Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.\n\nFilesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:\n- **read-only**: The sandbox only permits reading files.\n- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.\n- **danger-full-access**: No filesystem sandboxing - all commands are permitted.\n\nNetwork sandboxing defines whether network can be accessed without approval. Options for `network_access` are:\n- **restricted**: Requires approval\n- **enabled**: No approval needed\n\nApprovals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` are\n- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe \"read\" commands.\n- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.\n- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)\n- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.\n\nWhen you are running with `approval_policy == on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:\n- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /var)\n- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.\n- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)\n- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval. ALWAYS proceed to use the `with_escalated_permissions` and `justification` parameters - do not message the user before requesting approval for the command.\n- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for\n- (for all of these, you should weigh alternative paths that do not require approval)\n\nWhen `sandbox_mode` is set to read-only, you'll need to request approval for any command that isn't a read.\n\nYou will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.\n\nAlthough they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to \"never\", in which case never ask for approvals.\n\nWhen requesting approval to execute a command that will require escalated privileges:\n - Provide the `with_escalated_permissions` parameter with the boolean value true\n - Include a short, 1 sentence explanation for why you need to enable `with_escalated_permissions` in the justification parameter\n\n## Special user requests\n\n- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.\n- If the user asks for a \"review\", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.\n\n## Frontend tasks\nWhen doing frontend design tasks, avoid collapsing into \"AI slop\" or safe, average-looking layouts.\nAim for interfaces that feel intentional, bold, and a bit surprising.\n- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).\n- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.\n- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.\n- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.\n- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.\n- Ensure the page loads properly on both desktop and mobile\n\nException: If working within an existing website or design system, preserve the established patterns, structure, and visual language.\n\n## Presenting your work and final message\n\nYou are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.\n\n- Default: be very concise; friendly coding teammate tone.\n- Ask only when needed; suggest ideas; mirror the user's style.\n- For substantial work, summarize clearly; follow final‑answer formatting.\n- Skip heavy formatting for simple confirmations.\n- Don't dump large files you've written; reference paths only.\n- No \"save/copy this file\" - User is on the same machine.\n- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.\n- For code changes:\n * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with \"summary\", just jump right in.\n * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.\n * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.\n- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.\n\n### Final answer structure and style guidelines\n\n- Plain text; CLI handles styling. Use structure only when it helps scanability.\n- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.\n- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.\n- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.\n- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.\n- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.\n- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no \"above/below\"; parallel wording.\n- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.\n- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.\n- File References: When referencing files in your response follow the below rules:\n * Use inline code to make file paths clickable.\n * Each reference should have a stand alone path. Even if it's the same file.\n * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.\n * Optionally include line/column (1‑based): :line[:column] or #Lline[Ccolumn] (column defaults to 1).\n * Do not use URIs like file://, vscode://, or https://.\n * Do not provide range of lines\n * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\\repo\\project\\main.rs:12:5\n" + // 创建代理 Agent（使用统一的代理工具） function createProxyAgent(proxy) { return ProxyHelper.createProxyAgent(proxy) @@ -271,8 +275,8 @@ const handleResponses = async (req, res) => { const codexCliPattern = /^(codex_vscode|codex_cli_rs|codex_exec)\/[\d.]+/i const isCodexCLI = codexCliPattern.test(userAgent) - // 如果不是 Codex CLI 请求，则进行适配 - if (!isCodexCLI) { + // 如果不是 Codex CLI 请求且不是来自 unified 端点（已完成格式转换），则进行适配 + if (!isCodexCLI && !req._fromUnifiedEndpoint) { // 移除不需要的请求体字段 const fieldsToRemove = [ 'temperature', @@ -291,8 +295,7 @@ const handleResponses = async (req, res) => { }) // 设置固定的 Codex CLI instructions - req.body.instructions = - "You are Codex, based on GPT-5. You are running as a coding agent in the Codex CLI on a user's computer.\n\n## General\n\n- When searching for text or files, prefer using `rg` or `rg --files` respectively because `rg` is much faster than alternatives like `grep`. (If the `rg` command is not found, then use alternatives.)\n\n## Editing constraints\n\n- Default to ASCII when editing or creating files. Only introduce non-ASCII or other Unicode characters when there is a clear justification and the file already uses them.\n- Add succinct code comments that explain what is going on if code is not self-explanatory. You should not add comments like \"Assigns the value to the variable\", but a brief comment might be useful ahead of a complex code block that the user would otherwise have to spend time parsing out. Usage of these comments should be rare.\n- Try to use apply_patch for single file edits, but it is fine to explore other options to make the edit if it does not work well. Do not use apply_patch for changes that are auto-generated (i.e. generating package.json or running a lint or format command like gofmt) or when scripting is more efficient (such as search and replacing a string across a codebase).\n- You may be in a dirty git worktree.\n * NEVER revert existing changes you did not make unless explicitly requested, since these changes were made by the user.\n * If asked to make a commit or code edits and there are unrelated changes to your work or changes that you didn't make in those files, don't revert those changes.\n * If the changes are in files you've touched recently, you should read carefully and understand how you can work with the changes rather than reverting them.\n * If the changes are in unrelated files, just ignore them and don't revert them.\n- Do not amend a commit unless explicitly requested to do so.\n- While you are working, you might notice unexpected changes that you didn't make. If this happens, STOP IMMEDIATELY and ask the user how they would like to proceed.\n- **NEVER** use destructive commands like `git reset --hard` or `git checkout --` unless specifically requested or approved by the user.\n\n## Plan tool\n\nWhen using the planning tool:\n- Skip using the planning tool for straightforward tasks (roughly the easiest 25%).\n- Do not make single-step plans.\n- When you made a plan, update it after having performed one of the sub-tasks that you shared on the plan.\n\n## Codex CLI harness, sandboxing, and approvals\n\nThe Codex CLI harness supports several different configurations for sandboxing and escalation approvals that the user can choose from.\n\nFilesystem sandboxing defines which files can be read or written. The options for `sandbox_mode` are:\n- **read-only**: The sandbox only permits reading files.\n- **workspace-write**: The sandbox permits reading files, and editing files in `cwd` and `writable_roots`. Editing files in other directories requires approval.\n- **danger-full-access**: No filesystem sandboxing - all commands are permitted.\n\nNetwork sandboxing defines whether network can be accessed without approval. Options for `network_access` are:\n- **restricted**: Requires approval\n- **enabled**: No approval needed\n\nApprovals are your mechanism to get user consent to run shell commands without the sandbox. Possible configuration options for `approval_policy` are\n- **untrusted**: The harness will escalate most commands for user approval, apart from a limited allowlist of safe \"read\" commands.\n- **on-failure**: The harness will allow all commands to run in the sandbox (if enabled), and failures will be escalated to the user for approval to run again without the sandbox.\n- **on-request**: Commands will be run in the sandbox by default, and you can specify in your tool call if you want to escalate a command to run without sandboxing. (Note that this mode is not always available. If it is, you'll see parameters for it in the `shell` command description.)\n- **never**: This is a non-interactive mode where you may NEVER ask the user for approval to run commands. Instead, you must always persist and work around constraints to solve the task for the user. You MUST do your utmost best to finish the task and validate your work before yielding. If this mode is paired with `danger-full-access`, take advantage of it to deliver the best outcome for the user. Further, in this mode, your default testing philosophy is overridden: Even if you don't see local patterns for testing, you may add tests and scripts to validate your work. Just remove them before yielding.\n\nWhen you are running with `approval_policy == on-request`, and sandboxing enabled, here are scenarios where you'll need to request approval:\n- You need to run a command that writes to a directory that requires it (e.g. running tests that write to /var)\n- You need to run a GUI app (e.g., open/xdg-open/osascript) to open browsers or files.\n- You are running sandboxed and need to run a command that requires network access (e.g. installing packages)\n- If you run a command that is important to solving the user's query, but it fails because of sandboxing, rerun the command with approval. ALWAYS proceed to use the `with_escalated_permissions` and `justification` parameters - do not message the user before requesting approval for the command.\n- You are about to take a potentially destructive action such as an `rm` or `git reset` that the user did not explicitly ask for\n- (for all of these, you should weigh alternative paths that do not require approval)\n\nWhen `sandbox_mode` is set to read-only, you'll need to request approval for any command that isn't a read.\n\nYou will be told what filesystem sandboxing, network sandboxing, and approval mode are active in a developer or user message. If you are not told about this, assume that you are running with workspace-write, network sandboxing enabled, and approval on-failure.\n\nAlthough they introduce friction to the user because your work is paused until the user responds, you should leverage them when necessary to accomplish important work. If the completing the task requires escalated permissions, Do not let these settings or the sandbox deter you from attempting to accomplish the user's task unless it is set to \"never\", in which case never ask for approvals.\n\nWhen requesting approval to execute a command that will require escalated privileges:\n - Provide the `with_escalated_permissions` parameter with the boolean value true\n - Include a short, 1 sentence explanation for why you need to enable `with_escalated_permissions` in the justification parameter\n\n## Special user requests\n\n- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as `date`), you should do so.\n- If the user asks for a \"review\", default to a code review mindset: prioritise identifying bugs, risks, behavioural regressions, and missing tests. Findings must be the primary focus of the response - keep summaries or overviews brief and only after enumerating the issues. Present findings first (ordered by severity with file/line references), follow with open questions or assumptions, and offer a change-summary only as a secondary detail. If no findings are discovered, state that explicitly and mention any residual risks or testing gaps.\n\n## Frontend tasks\nWhen doing frontend design tasks, avoid collapsing into \"AI slop\" or safe, average-looking layouts.\nAim for interfaces that feel intentional, bold, and a bit surprising.\n- Typography: Use expressive, purposeful fonts and avoid default stacks (Inter, Roboto, Arial, system).\n- Color & Look: Choose a clear visual direction; define CSS variables; avoid purple-on-white defaults. No purple bias or dark mode bias.\n- Motion: Use a few meaningful animations (page-load, staggered reveals) instead of generic micro-motions.\n- Background: Don't rely on flat, single-color backgrounds; use gradients, shapes, or subtle patterns to build atmosphere.\n- Overall: Avoid boilerplate layouts and interchangeable UI patterns. Vary themes, type families, and visual languages across outputs.\n- Ensure the page loads properly on both desktop and mobile\n\nException: If working within an existing website or design system, preserve the established patterns, structure, and visual language.\n\n## Presenting your work and final message\n\nYou are producing plain text that will later be styled by the CLI. Follow these rules exactly. Formatting should make results easy to scan, but not feel mechanical. Use judgment to decide how much structure adds value.\n\n- Default: be very concise; friendly coding teammate tone.\n- Ask only when needed; suggest ideas; mirror the user's style.\n- For substantial work, summarize clearly; follow final‑answer formatting.\n- Skip heavy formatting for simple confirmations.\n- Don't dump large files you've written; reference paths only.\n- No \"save/copy this file\" - User is on the same machine.\n- Offer logical next steps (tests, commits, build) briefly; add verify steps if you couldn't do something.\n- For code changes:\n * Lead with a quick explanation of the change, and then give more details on the context covering where and why a change was made. Do not start this explanation with \"summary\", just jump right in.\n * If there are natural next steps the user may want to take, suggest them at the end of your response. Do not make suggestions if there are no natural next steps.\n * When suggesting multiple options, use numeric lists for the suggestions so the user can quickly respond with a single number.\n- The user does not command execution outputs. When asked to show the output of a command (e.g. `git show`), relay the important details in your answer or summarize the key lines so the user understands the result.\n\n### Final answer structure and style guidelines\n\n- Plain text; CLI handles styling. Use structure only when it helps scanability.\n- Headers: optional; short Title Case (1-3 words) wrapped in **…**; no blank line before the first bullet; add only if they truly help.\n- Bullets: use - ; merge related points; keep to one line when possible; 4–6 per list ordered by importance; keep phrasing consistent.\n- Monospace: backticks for commands/paths/env vars/code ids and inline examples; use for literal keyword bullets; never combine with **.\n- Code samples or multi-line snippets should be wrapped in fenced code blocks; include an info string as often as possible.\n- Structure: group related bullets; order sections general → specific → supporting; for subsections, start with a bolded keyword bullet, then items; match complexity to the task.\n- Tone: collaborative, concise, factual; present tense, active voice; self‑contained; no \"above/below\"; parallel wording.\n- Don'ts: no nested bullets/hierarchies; no ANSI codes; don't cram unrelated keywords; keep keyword lists short—wrap/reformat if long; avoid naming formatting styles in answers.\n- Adaptation: code explanations → precise, structured with code refs; simple tasks → lead with outcome; big changes → logical walkthrough + rationale + next actions; casual one-offs → plain sentences, no headers/bullets.\n- File References: When referencing files in your response follow the below rules:\n * Use inline code to make file paths clickable.\n * Each reference should have a stand alone path. Even if it's the same file.\n * Accepted: absolute, workspace‑relative, a/ or b/ diff prefixes, or bare filename/suffix.\n * Optionally include line/column (1‑based): :line[:column] or #Lline[Ccolumn] (column defaults to 1).\n * Do not use URIs like file://, vscode://, or https://.\n * Do not provide range of lines\n * Examples: src/app.ts, src/app.ts:42, b/server/index.js#L10, C:\\repo\\project\\main.rs:12:5\n" + req.body.instructions = CODEX_CLI_INSTRUCTIONS logger.info('📝 Non-Codex CLI request detected, applying Codex CLI adaptation') } else { @@ -934,3 +937,4 @@ router.get('/key-info', authenticateApiKey, async (req, res) => { module.exports = router module.exports.handleResponses = handleResponses +module.exports.CODEX_CLI_INSTRUCTIONS = CODEX_CLI_INSTRUCTIONS diff --git a/src/routes/unified.js b/src/routes/unified.js index c1401137..61212ba8 100644 --- a/src/routes/unified.js +++ b/src/routes/unified.js @@ -8,7 +8,10 @@ const { handleStreamGenerateContent: geminiHandleStreamGenerateContent } = require('../handlers/geminiHandlers') const openaiRoutes = require('./openaiRoutes') +const { CODEX_CLI_INSTRUCTIONS } = require('./openaiRoutes') const apiKeyService = require('../services/apiKeyService') +const GeminiToOpenAIConverter = require('../services/geminiToOpenAI') +const CodexToOpenAIConverter = require('../services/codexToOpenAI') const router = express.Router() @@ -71,6 +74,147 @@ async function routeToBackend(req, res, requestedModel) { } }) } + // 响应格式拦截：Codex/Responses → OpenAI Chat Completions + const codexConverter = new CodexToOpenAIConverter() + const originalJson = res.json.bind(res) + + // 流式：patch res.write/res.end 拦截 SSE 事件 + // 与 openaiRoutes 保持一致：stream 缺省时视为流式（stream !== false） + if (req.body.stream !== false) { + const streamState = codexConverter.createStreamState() + const sseBuffer = { data: '' } + const originalWrite = res.write.bind(res) + const originalEnd = res.end.bind(res) + + res.write = function (chunk, encoding, callback) { + if (res.statusCode >= 400) { + return originalWrite(chunk, encoding, callback) + } + + const str = (typeof chunk === 'string' ? chunk : chunk.toString()).replace(/\r\n/g, '\n') + sseBuffer.data += str + + let idx + while ((idx = sseBuffer.data.indexOf('\n\n')) !== -1) { + const event = sseBuffer.data.slice(0, idx) + sseBuffer.data = sseBuffer.data.slice(idx + 2) + + if (!event.trim()) { + continue + } + + const lines = event.split('\n') + for (const line of lines) { + if (line.startsWith('data: ')) { + const jsonStr = line.slice(6) + if (!jsonStr || jsonStr === '[DONE]') { + continue + } + + try { + const eventData = JSON.parse(jsonStr) + if (eventData.error) { + originalWrite(`data: ${jsonStr}\n\n`) + continue + } + const converted = codexConverter.convertStreamChunk( + eventData, + requestedModel, + streamState + ) + for (const c of converted) { + originalWrite(c) + } + } catch (e) { + originalWrite(`data: ${jsonStr}\n\n`) + } + } + } + } + + if (typeof callback === 'function') { + callback() + } + return true + } + + res.end = function (chunk, encoding, callback) { + if (res.statusCode < 400) { + // 处理 res.end(chunk) 传入的最后一块数据 + if (chunk) { + const str = (typeof chunk === 'string' ? chunk : chunk.toString()).replace( + /\r\n/g, + '\n' + ) + sseBuffer.data += str + chunk = undefined + } + + if (sseBuffer.data.trim()) { + const remaining = `${sseBuffer.data}\n\n` + sseBuffer.data = '' + + const lines = remaining.split('\n') + for (const line of lines) { + if (line.startsWith('data: ')) { + const jsonStr = line.slice(6) + if (!jsonStr || jsonStr === '[DONE]') { + continue + } + try { + const eventData = JSON.parse(jsonStr) + if (eventData.error) { + originalWrite(`data: ${jsonStr}\n\n`) + } else { + const converted = codexConverter.convertStreamChunk( + eventData, + requestedModel, + streamState + ) + for (const c of converted) { + originalWrite(c) + } + } + } catch (e) { + originalWrite(`data: ${jsonStr}\n\n`) + } + } + } + } + + originalWrite('data: [DONE]\n\n') + } + return originalEnd(chunk, encoding, callback) + } + } + + // 非流式：patch res.json 拦截 JSON 响应 + // chatgpt.com 后端返回 { type: "response.completed", response: {...} } + // api.openai.com 后端返回标准 Response 对象 { object: "response", status, output, ... } + res.json = function (data) { + if (res.statusCode >= 400) { + return originalJson(data) + } + if (data && (data.type === 'response.completed' || data.object === 'response')) { + try { + return originalJson(codexConverter.convertResponse(data, requestedModel)) + } catch (e) { + logger.debug('Codex response conversion failed, passing through:', e.message) + return originalJson(data) + } + } + return originalJson(data) + } + + // 输入转换：Chat Completions → Responses API 格式 + req.body = codexConverter.buildRequestFromOpenAI(req.body) + // 注入 Codex CLI 系统提示词（与 handleResponses 非 Codex CLI 适配一致） + req.body.instructions = CODEX_CLI_INSTRUCTIONS + req._fromUnifiedEndpoint = true + // 修正请求路径：body 已转为 Responses 格式，路径需与之匹配 + // Express req.path 是只读 getter（派生自 req.url），需改 req.url + req.url = '/v1/responses' + return await openaiRoutes.handleResponses(req, res) } else if (backend === 'gemini') { // Gemini 后端 @@ -84,20 +228,101 @@ async function routeToBackend(req, res, requestedModel) { }) } - // 转换为 Gemini 格式 - const geminiRequest = { - model: requestedModel, - messages: req.body.messages, - temperature: req.body.temperature || 0.7, - max_tokens: req.body.max_tokens || 4096, - stream: req.body.stream || false + // 将 OpenAI Chat Completions 参数转换为 Gemini 原生格式 + // standard 处理器从 req.body 根层解构 contents/generationConfig 等字段 + const geminiRequest = buildGeminiRequestFromOpenAI(req.body) + + // standard 处理器从 req.params.modelName 获取模型名 + req.params = req.params || {} + req.params.modelName = requestedModel + + // 平铺到 req.body 根层（保留 messages/stream 等原始字段给 sessionHelper 计算 hash） + req.body.contents = geminiRequest.contents + req.body.generationConfig = geminiRequest.generationConfig || {} + req.body.safetySettings = geminiRequest.safetySettings + // standard 处理器读取 camelCase: systemInstruction + if (geminiRequest.system_instruction) { + req.body.systemInstruction = geminiRequest.system_instruction + } + if (geminiRequest.tools) { + req.body.tools = geminiRequest.tools + } + if (geminiRequest.toolConfig) { + req.body.toolConfig = geminiRequest.toolConfig } - req.body = geminiRequest + if (req.body.stream) { + // 响应格式拦截：Gemini SSE → OpenAI Chat Completions chunk + const geminiConverter = new GeminiToOpenAIConverter() + const geminiStreamState = geminiConverter.createStreamState() + const geminiOriginalWrite = res.write.bind(res) + const geminiOriginalEnd = res.end.bind(res) + + res.write = function (chunk, encoding, callback) { + if (res.statusCode >= 400) { + return geminiOriginalWrite(chunk, encoding, callback) + } + + const converted = geminiConverter.convertStreamChunk( + chunk, + requestedModel, + geminiStreamState + ) + if (converted) { + return geminiOriginalWrite(converted, encoding, callback) + } + if (typeof callback === 'function') { + callback() + } + return true + } + + res.end = function (chunk, encoding, callback) { + if (res.statusCode < 400) { + // 处理 res.end(chunk) 传入的最后一块数据 + if (chunk) { + const converted = geminiConverter.convertStreamChunk( + chunk, + requestedModel, + geminiStreamState + ) + if (converted) { + geminiOriginalWrite(converted) + } + chunk = undefined + } + // 刷新 converter 内部 buffer 中的残留数据 + if (geminiStreamState.buffer.trim()) { + const remaining = geminiConverter.convertStreamChunk( + '\n\n', + requestedModel, + geminiStreamState + ) + if (remaining) { + geminiOriginalWrite(remaining) + } + } + geminiOriginalWrite('data: [DONE]\n\n') + } + return geminiOriginalEnd(chunk, encoding, callback) + } - if (geminiRequest.stream) { return await geminiHandleStreamGenerateContent(req, res) } else { + // 响应格式拦截：Gemini JSON → OpenAI chat.completion + const geminiConverter = new GeminiToOpenAIConverter() + const geminiOriginalJson = res.json.bind(res) + + res.json = function (data) { + if (res.statusCode >= 400) { + return geminiOriginalJson(data) + } + if (data && (data.candidates || data.response?.candidates)) { + return geminiOriginalJson(geminiConverter.convertResponse(data, requestedModel)) + } + return geminiOriginalJson(data) + } + return await geminiHandleGenerateContent(req, res) } } else { @@ -198,6 +423,280 @@ router.post('/v1/completions', authenticateApiKey, async (req, res) => { } }) +// --- OpenAI Chat Completions → Gemini 原生请求转换（OpenAI → Gemini 格式映射） --- + +function buildGeminiRequestFromOpenAI(body) { + const request = {} + const generationConfig = {} + const messages = body.messages || [] + + // 第一遍：收集 assistant tool_calls 的 id→name 映射（用于 tool response 关联） + const toolCallNames = Object.create(null) + for (const msg of messages) { + if (msg.role === 'assistant' && msg.tool_calls) { + for (const tc of msg.tool_calls) { + if (tc.id && tc.function?.name) { + toolCallNames[tc.id] = tc.function.name + } + } + } + } + + // 第二遍：构建 contents + system_instruction + const systemParts = [] + const contents = [] + + for (const msg of messages) { + if (msg.role === 'system' || msg.role === 'developer') { + const text = extractTextContent(msg.content) + if (text) { + systemParts.push({ text }) + } + } else if (msg.role === 'user') { + const parts = buildContentParts(msg.content) + if (parts.length > 0) { + contents.push({ role: 'user', parts }) + } + } else if (msg.role === 'assistant') { + // 格式映射: assistant 内容保留 text + image（多模态） + const parts = buildContentParts(msg.content) + // tool_calls → functionCall parts + if (msg.tool_calls) { + for (const tc of msg.tool_calls) { + if (tc.function) { + let args + try { + args = JSON.parse(tc.function.arguments || '{}') + } catch { + // parse 失败时尝试保留原始内容 + args = tc.function.arguments ? { _raw: tc.function.arguments } : {} + } + parts.push({ + functionCall: { name: tc.function.name, args } + }) + } + } + } + if (parts.length > 0) { + contents.push({ role: 'model', parts }) + } + } else if (msg.role === 'tool') { + // tool response → functionResponse（Gemini 用 user role） + const name = toolCallNames[msg.tool_call_id] || msg.name || 'unknown' + let responseContent + try { + responseContent = + typeof msg.content === 'string' ? JSON.parse(msg.content) : msg.content || {} + } catch { + responseContent = { result: msg.content } + } + contents.push({ + role: 'user', + parts: [{ functionResponse: { name, response: responseContent } }] + }) + } + } + + if (systemParts.length > 0) { + if (contents.length === 0) { + // Gemini 格式：只有 system 消息时，将其作为 user content（避免 Gemini 拒绝空 contents） + contents.push({ role: 'user', parts: systemParts }) + } else { + request.system_instruction = { parts: systemParts } + } + } + request.contents = contents + + // Generation config + if (body.temperature !== undefined) { + generationConfig.temperature = body.temperature + } + const maxTokens = body.max_completion_tokens || body.max_tokens + if (maxTokens !== undefined) { + generationConfig.maxOutputTokens = maxTokens + } + if (body.top_p !== undefined) { + generationConfig.topP = body.top_p + } + if (body.top_k !== undefined) { + generationConfig.topK = body.top_k + } + if (body.n !== undefined && body.n > 1) { + generationConfig.candidateCount = body.n + } + if (body.stop) { + generationConfig.stopSequences = Array.isArray(body.stop) ? body.stop : [body.stop] + } + + // modalities → responseModalities（text→TEXT, image→IMAGE, audio→AUDIO） + if (body.modalities && Array.isArray(body.modalities)) { + const modalityMap = { text: 'TEXT', image: 'IMAGE', audio: 'AUDIO' } + const mapped = body.modalities.map((m) => modalityMap[m.toLowerCase()]).filter(Boolean) + if (mapped.length > 0) { + generationConfig.responseModalities = mapped + } + } + + // image_config → imageConfig（Gemini 格式：aspect_ratio→aspectRatio, image_size→imageSize） + if (body.image_config) { + const imageConfig = {} + if (body.image_config.aspect_ratio) { + imageConfig.aspectRatio = body.image_config.aspect_ratio + } + if (body.image_config.image_size) { + imageConfig.imageSize = body.image_config.image_size + } + if (Object.keys(imageConfig).length > 0) { + generationConfig.imageConfig = imageConfig + } + } + + // reasoning_effort → thinkingConfig（Gemini 格式） + if (body.reasoning_effort) { + const effort = body.reasoning_effort.toLowerCase() + if (effort === 'none') { + generationConfig.thinkingConfig = { thinkingLevel: 'none', includeThoughts: false } + } else if (effort === 'auto') { + // 格式映射: auto → thinkingBudget:-1 (让模型自行决定) + generationConfig.thinkingConfig = { thinkingBudget: -1, includeThoughts: true } + } else { + generationConfig.thinkingConfig = { thinkingLevel: effort, includeThoughts: true } + } + } + + // response_format → responseMimeType / responseSchema + if (body.response_format) { + if (body.response_format.type === 'json_object') { + generationConfig.responseMimeType = 'application/json' + } else if ( + body.response_format.type === 'json_schema' && + body.response_format.json_schema?.schema + ) { + generationConfig.responseMimeType = 'application/json' + generationConfig.responseSchema = body.response_format.json_schema.schema + } + } + + if (Object.keys(generationConfig).length > 0) { + request.generationConfig = generationConfig + } + + // Tools: OpenAI function → Gemini functionDeclarations（OpenAI → Gemini 格式映射） + if (body.tools && body.tools.length > 0) { + const functionDeclarations = [] + const extraTools = [] + for (const tool of body.tools) { + if (tool.type === 'function' && tool.function) { + const decl = { + name: tool.function.name, + description: tool.function.description || '' + } + if (tool.function.parameters) { + // 格式映射: parameters → parametersJsonSchema, 删除 strict + const schema = { ...tool.function.parameters } + delete schema.strict + decl.parametersJsonSchema = schema + } else { + decl.parametersJsonSchema = { type: 'object', properties: {} } + } + functionDeclarations.push(decl) + } else if ( + tool.type === 'google_search' || + tool.type === 'code_execution' || + tool.type === 'url_context' + ) { + // 非 function 工具透传，snake_case → camelCase（Gemini 原生格式） + const typeMap = { + google_search: 'googleSearch', + code_execution: 'codeExecution', + url_context: 'urlContext' + } + const geminiType = typeMap[tool.type] + extraTools.push({ [geminiType]: tool[tool.type] || {} }) + } + } + const toolsArray = [] + if (functionDeclarations.length > 0) { + toolsArray.push({ functionDeclarations }) + } + toolsArray.push(...extraTools) + if (toolsArray.length > 0) { + request.tools = toolsArray + } + } + + // tool_choice → toolConfig.functionCallingConfig + if (body.tool_choice) { + if (body.tool_choice === 'none') { + request.toolConfig = { functionCallingConfig: { mode: 'NONE' } } + } else if (body.tool_choice === 'auto') { + request.toolConfig = { functionCallingConfig: { mode: 'AUTO' } } + } else if (body.tool_choice === 'required') { + request.toolConfig = { functionCallingConfig: { mode: 'ANY' } } + } else if (typeof body.tool_choice === 'object' && body.tool_choice.function?.name) { + request.toolConfig = { + functionCallingConfig: { + mode: 'ANY', + allowedFunctionNames: [body.tool_choice.function.name] + } + } + } + } + + // 默认安全设置（Gemini 格式：最大化允许，避免不必要的内容拦截） + if (!request.safetySettings) { + request.safetySettings = [ + { category: 'HARM_CATEGORY_HARASSMENT', threshold: 'OFF' }, + { category: 'HARM_CATEGORY_HATE_SPEECH', threshold: 'OFF' }, + { category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT', threshold: 'OFF' }, + { category: 'HARM_CATEGORY_DANGEROUS_CONTENT', threshold: 'OFF' }, + { category: 'HARM_CATEGORY_CIVIC_INTEGRITY', threshold: 'BLOCK_NONE' } + ] + } + + return request +} + +function extractTextContent(content) { + if (typeof content === 'string') { + return content + } + if (Array.isArray(content)) { + return content + .filter((c) => c.type === 'text') + .map((c) => c.text) + .join('') + } + return '' +} + +function buildContentParts(content) { + if (typeof content === 'string') { + return [{ text: content }] + } + if (Array.isArray(content)) { + const parts = [] + for (const item of content) { + if (item.type === 'text') { + parts.push({ text: item.text }) + } else if (item.type === 'image_url' && item.image_url?.url) { + const { url } = item.image_url + if (url.startsWith('data:')) { + const match = url.match(/^data:([^;]+);base64,(.+)$/) + if (match) { + parts.push({ inlineData: { mimeType: match[1], data: match[2] } }) + } + } + } + } + return parts + } + if (!content) { + return [] + } + return [{ text: String(content) }] +} + module.exports = router module.exports.detectBackendFromModel = detectBackendFromModel module.exports.routeToBackend = routeToBackend diff --git a/src/services/codexToOpenAI.js b/src/services/codexToOpenAI.js new file mode 100644 index 00000000..e040c318 --- /dev/null +++ b/src/services/codexToOpenAI.js @@ -0,0 +1,717 @@ +/** + * Codex Responses API → OpenAI Chat Completions 格式转换器 + * 将 Codex/OpenAI Responses API 的 SSE 事件转为标准 chat.completion / chat.completion.chunk 格式 + */ + +class CodexToOpenAIConverter { + constructor() { + // 工具名缩短映射（buildRequestFromOpenAI 填充，响应转换时逆向恢复） + this._reverseToolNameMap = {} + } + + createStreamState() { + return { + responseId: '', + createdAt: 0, + model: '', + functionCallIndex: -1, + hasReceivedArgumentsDelta: false, + hasToolCallAnnounced: false, + roleSent: false + } + } + + /** + * 流式转换: 单个已解析的 SSE 事件 → OpenAI chunk SSE 字符串数组 + * @param {Object} eventData - 已解析的 SSE JSON 对象 + * @param {string} model - 请求模型名 + * @param {Object} state - createStreamState() 的返回值 + * @returns {string[]} "data: {...}\n\n" 字符串数组（可能为空） + */ + convertStreamChunk(eventData, model, state) { + const { type } = eventData + if (!type) { + return [] + } + + switch (type) { + case 'response.created': + return this._handleResponseCreated(eventData, state) + + case 'response.reasoning_summary_text.delta': + return this._emitChunk(state, model, { reasoning_content: eventData.delta }) + + case 'response.reasoning_summary_text.done': + // done 事件仅为结束信号，delta 已通过 .delta 事件发送，不再注入内容 + return [] + + case 'response.output_text.delta': + return this._emitChunk(state, model, { content: eventData.delta }) + + case 'response.output_item.added': + return this._handleOutputItemAdded(eventData, model, state) + + case 'response.function_call_arguments.delta': + return this._handleArgumentsDelta(eventData, model, state) + + case 'response.function_call_arguments.done': + return this._handleArgumentsDone(eventData, model, state) + + case 'response.output_item.done': + return this._handleOutputItemDone(eventData, model, state) + + case 'response.completed': + return this._handleResponseCompleted(eventData, model, state) + + case 'response.failed': + case 'response.incomplete': + return this._handleResponseError(eventData, model, state) + + case 'error': + return this._handleStreamError(eventData, model, state) + + default: + return [] + } + } + + /** + * 非流式转换: Codex 完整响应 → OpenAI chat.completion + * @param {Object} responseData - Codex 响应对象或 response.completed 事件 + * @param {string} model - 请求模型名 + * @returns {Object} + */ + convertResponse(responseData, model) { + // 自动检测：response.completed 事件包装 vs 直接响应对象 + const resp = responseData.type === 'response.completed' ? responseData.response : responseData + + const message = { role: 'assistant', content: null } + const toolCalls = [] + + const output = resp.output || [] + for (const item of output) { + if (item.type === 'reasoning') { + const summaryTexts = (item.summary || []) + .filter((s) => s.type === 'summary_text') + .map((s) => s.text) + if (summaryTexts.length > 0) { + message.reasoning_content = (message.reasoning_content || '') + summaryTexts.join('') + } + } else if (item.type === 'message') { + const contentTexts = (item.content || []) + .filter((c) => c.type === 'output_text') + .map((c) => c.text) + if (contentTexts.length > 0) { + message.content = (message.content || '') + contentTexts.join('') + } + } else if (item.type === 'function_call') { + toolCalls.push({ + id: item.call_id || item.id, + type: 'function', + function: { + name: this._restoreToolName(item.name), + arguments: item.arguments || '{}' + } + }) + } + } + + if (toolCalls.length > 0) { + message.tool_calls = toolCalls + } + + // response.failed → 返回 error 结构（与流式 _handleResponseError 一致） + if (resp.status === 'failed') { + const err = resp.error || {} + return { + error: { + message: err.message || 'Response failed', + type: err.type || 'server_error', + code: err.code || null + } + } + } + + const finishReason = toolCalls.length > 0 ? 'tool_calls' : this._mapResponseStatus(resp) + + const result = { + id: resp.id || `chatcmpl-${Date.now()}`, + object: 'chat.completion', + created: this._parseCreatedAt(resp.created_at), + model: resp.model || model, + choices: [{ index: 0, message, finish_reason: finishReason }] + } + + const usage = this._mapUsage(resp.usage) + if (usage) { + result.usage = usage + } + + return result + } + + // --- 内部方法：流式事件处理 --- + + _handleResponseCreated(eventData, state) { + const resp = eventData.response || {} + state.responseId = resp.id || '' + if (resp.created_at) { + state.createdAt = this._parseCreatedAt(resp.created_at) + } + state.model = resp.model || '' + return [] + } + + _handleOutputItemAdded(eventData, model, state) { + const { item } = eventData + if (!item || item.type !== 'function_call') { + return [] + } + + state.functionCallIndex++ + state.hasReceivedArgumentsDelta = false + state.hasToolCallAnnounced = true + + return this._emitChunk(state, model, { + tool_calls: [ + { + index: state.functionCallIndex, + id: item.call_id || item.id, + type: 'function', + function: { name: this._restoreToolName(item.name), arguments: '' } + } + ] + }) + } + + _handleArgumentsDelta(eventData, model, state) { + state.hasReceivedArgumentsDelta = true + return this._emitChunk(state, model, { + tool_calls: [ + { + index: state.functionCallIndex, + function: { arguments: eventData.delta } + } + ] + }) + } + + _handleArgumentsDone(eventData, model, state) { + // 如果已收到增量 delta，done 不需要再输出 + if (state.hasReceivedArgumentsDelta) { + return [] + } + + // 没有收到 delta，一次性输出完整参数 + return this._emitChunk(state, model, { + tool_calls: [ + { + index: state.functionCallIndex, + function: { arguments: eventData.arguments || '{}' } + } + ] + }) + } + + _handleOutputItemDone(eventData, model, state) { + const { item } = eventData + if (!item || item.type !== 'function_call') { + return [] + } + + // 如果已经通过 output_item.added 通知过，不重复输出 + if (state.hasToolCallAnnounced) { + state.hasToolCallAnnounced = false + return [] + } + + // Fallback：未收到 added 事件，输出完整 tool call + state.functionCallIndex++ + return this._emitChunk(state, model, { + tool_calls: [ + { + index: state.functionCallIndex, + id: item.call_id || item.id, + type: 'function', + function: { + name: this._restoreToolName(item.name), + arguments: item.arguments || '{}' + } + } + ] + }) + } + + _handleResponseCompleted(eventData, model, state) { + const resp = eventData.response || {} + const chunk = this._makeChunk(state, model) + + if (state.functionCallIndex >= 0) { + chunk.choices[0].finish_reason = 'tool_calls' + } else { + chunk.choices[0].finish_reason = this._mapResponseStatus(resp) + } + + const usage = this._mapUsage(resp.usage) + if (usage) { + chunk.usage = usage + } + + return [`data: ${JSON.stringify(chunk)}\n\n`] + } + + _handleResponseError(eventData, model, state) { + const resp = eventData.response || {} + const results = [] + + // response.failed → 转为 error SSE 事件（保留错误语义） + if (resp.status === 'failed') { + const err = resp.error || {} + results.push( + `data: ${JSON.stringify({ + error: { + message: err.message || 'Response failed', + type: err.type || 'server_error', + code: err.code || null + } + })}\n\n` + ) + } + + // response.incomplete 及其他非 failed 状态 → 带 finish_reason 的终止 chunk + if (resp.status !== 'failed') { + const chunk = this._makeChunk(state, model) + if (state.functionCallIndex >= 0) { + chunk.choices[0].finish_reason = 'tool_calls' + } else { + chunk.choices[0].finish_reason = this._mapResponseStatus(resp) + } + const usage = this._mapUsage(resp.usage) + if (usage) { + chunk.usage = usage + } + results.push(`data: ${JSON.stringify(chunk)}\n\n`) + } + + return results + } + + _handleStreamError(eventData) { + // type: "error" → 转为 OpenAI 格式的 error SSE 事件 + const errorObj = { + error: { + message: eventData.message || 'Unknown error', + type: 'server_error', + code: eventData.code || null + } + } + return [`data: ${JSON.stringify(errorObj)}\n\n`] + } + + // --- 工具方法 --- + + _emitChunk(state, model, delta) { + const chunk = this._makeChunk(state, model) + if (!state.roleSent) { + delta.role = 'assistant' + state.roleSent = true + } + chunk.choices[0].delta = delta + return [`data: ${JSON.stringify(chunk)}\n\n`] + } + + _makeChunk(state, model) { + return { + id: state.responseId || `chatcmpl-${Date.now()}`, + object: 'chat.completion.chunk', + created: state.createdAt || Math.floor(Date.now() / 1000), + model: state.model || model, + choices: [{ index: 0, delta: {}, finish_reason: null }] + } + } + + _mapResponseStatus(resp) { + const { status } = resp + if (!status || status === 'completed') { + return 'stop' + } + if (status === 'incomplete') { + const reason = resp.incomplete_details?.reason + if (reason === 'max_output_tokens') { + return 'length' + } + if (reason === 'content_filter') { + return 'content_filter' + } + return 'length' + } + // failed, cancelled, etc. + return 'stop' + } + + _parseCreatedAt(createdAt) { + if (!createdAt) { + return Math.floor(Date.now() / 1000) + } + if (typeof createdAt === 'number') { + return createdAt + } + const ts = Math.floor(new Date(createdAt).getTime() / 1000) + return isNaN(ts) ? Math.floor(Date.now() / 1000) : ts + } + + _mapUsage(usage) { + if (!usage) { + return undefined + } + const result = { + prompt_tokens: usage.input_tokens || 0, + completion_tokens: usage.output_tokens || 0, + total_tokens: usage.total_tokens || 0 + } + if (usage.input_tokens_details?.cached_tokens > 0) { + result.prompt_tokens_details = { cached_tokens: usage.input_tokens_details.cached_tokens } + } + if (usage.output_tokens_details?.reasoning_tokens > 0) { + result.completion_tokens_details = { + reasoning_tokens: usage.output_tokens_details.reasoning_tokens + } + } + return result + } + + // ============================================= + // 请求转换: Chat Completions → Responses API + // ============================================= + // ============================================= + + /** + * 将 OpenAI Chat Completions 请求体转为 Responses API 格式 + * @param {Object} chatBody - Chat Completions 格式请求体 + * @returns {Object} Responses API 格式请求体 + */ + buildRequestFromOpenAI(chatBody) { + const result = {} + + if (chatBody.model) { + result.model = chatBody.model + } + if (chatBody.stream !== undefined) { + result.stream = chatBody.stream + } + + // messages → input（instructions 由调用方设置，此处只转换消息到 input） + const input = [] + + for (const msg of chatBody.messages || []) { + switch (msg.role) { + case 'system': + case 'developer': + input.push({ + type: 'message', + role: 'developer', + content: this._wrapContent(msg.content, 'user') + }) + break + + case 'user': + input.push({ + type: 'message', + role: 'user', + content: this._wrapContent(msg.content, 'user') + }) + break + + case 'assistant': + if (msg.content) { + input.push({ + type: 'message', + role: 'assistant', + content: this._wrapContent(msg.content, 'assistant') + }) + } + if (msg.tool_calls && msg.tool_calls.length > 0) { + for (const tc of msg.tool_calls) { + if (tc.type === 'function') { + input.push({ + type: 'function_call', + call_id: tc.id, + name: this._shortenToolName(tc.function?.name || ''), + arguments: + typeof tc.function?.arguments === 'string' + ? tc.function.arguments + : JSON.stringify(tc.function?.arguments ?? {}) + }) + } + } + } + break + + case 'tool': + input.push({ + type: 'function_call_output', + call_id: msg.tool_call_id, + output: typeof msg.content === 'string' ? msg.content : JSON.stringify(msg.content) + }) + break + } + } + + result.input = input + + // temperature/top_p/max_output_tokens 不透传，与上游 Codex API 行为保持一致 + + // reasoning 配置 + result.reasoning = { + effort: chatBody.reasoning_effort || 'medium', + summary: 'auto' + } + + // 固定值 + result.parallel_tool_calls = true + result.include = ['reasoning.encrypted_content'] + result.store = false + + // 收集所有工具名（tools + assistant.tool_calls），统一构建缩短映射 + const allToolNames = new Set() + if (chatBody.tools) { + for (const t of chatBody.tools) { + if (t.type === 'function' && t.function?.name) { + allToolNames.add(t.function.name) + } + } + } + for (const msg of chatBody.messages || []) { + if (msg.role === 'assistant' && msg.tool_calls) { + for (const tc of msg.tool_calls) { + if (tc.type === 'function' && tc.function?.name) { + allToolNames.add(tc.function.name) + } + } + } + } + if (allToolNames.size > 0) { + this._toolNameMap = this._buildShortNameMap([...allToolNames]) + this._reverseToolNameMap = {} + for (const [orig, short] of Object.entries(this._toolNameMap)) { + if (orig !== short) { + this._reverseToolNameMap[short] = orig + } + } + } + + // tools 展平 + if (chatBody.tools && chatBody.tools.length > 0) { + result.tools = this._convertTools(chatBody.tools) + } + + // tool_choice + if (chatBody.tool_choice !== undefined) { + result.tool_choice = this._convertToolChoice(chatBody.tool_choice) + } + + // response_format → text.format + if (chatBody.response_format) { + const text = this._convertResponseFormat(chatBody.response_format) + if (text && Object.keys(text).length > 0) { + result.text = text + } + } + + // session 字段透传（handleResponses 用于 session hash） + if (chatBody.session_id) { + result.session_id = chatBody.session_id + } + if (chatBody.conversation_id) { + result.conversation_id = chatBody.conversation_id + } + if (chatBody.prompt_cache_key) { + result.prompt_cache_key = chatBody.prompt_cache_key + } + + return result + } + + // --- 请求转换辅助方法 --- + + _extractTextContent(content) { + if (typeof content === 'string') { + return content + } + if (Array.isArray(content)) { + return content + .filter((c) => c.type === 'text') + .map((c) => c.text) + .join('') + } + return String(content || '') + } + + _wrapContent(content, role) { + const textType = role === 'assistant' ? 'output_text' : 'input_text' + if (typeof content === 'string') { + return [{ type: textType, text: content }] + } + if (Array.isArray(content)) { + return content + .map((item) => { + switch (item.type) { + case 'text': + return { type: textType, text: item.text } + case 'image_url': + return { + type: 'input_image', + image_url: item.image_url?.url || item.image_url + } + default: + return item + } + }) + .filter(Boolean) + } + return [{ type: textType, text: String(content || '') }] + } + + _convertTools(tools) { + return tools + .filter((t) => t && t.type) + .map((t) => { + // 非 function 工具（web_search、code_interpreter 等）原样透传 + if (t.type !== 'function') { + return t + } + // function 工具展平：去除 function wrapper + if (!t.function) { + return null + } + const tool = { type: 'function', name: this._shortenToolName(t.function.name) } + if (t.function.description) { + tool.description = t.function.description + } + if (t.function.parameters) { + tool.parameters = t.function.parameters + } + if (t.function.strict !== undefined) { + tool.strict = t.function.strict + } + return tool + }) + .filter(Boolean) + } + + _convertToolChoice(tc) { + if (typeof tc === 'string') { + return tc + } + if (tc && typeof tc === 'object') { + if (tc.type === 'function' && tc.function?.name) { + return { type: 'function', name: this._shortenToolName(tc.function.name) } + } + return tc + } + return tc + } + + _convertResponseFormat(rf) { + if (!rf || !rf.type) { + return {} + } + if (rf.type === 'text') { + return { format: { type: 'text' } } + } + if (rf.type === 'json_schema' && rf.json_schema) { + const format = { type: 'json_schema' } + if (rf.json_schema.name) { + format.name = rf.json_schema.name + } + if (rf.json_schema.strict !== undefined) { + format.strict = rf.json_schema.strict + } + if (rf.json_schema.schema) { + format.schema = rf.json_schema.schema + } + return { format } + } + return {} + } + + /** + * 工具名缩短：优先使用唯一化 map，无 map 时做简单截断 + */ + _shortenToolName(name) { + if (!name) { + return name + } + if (this._toolNameMap && this._toolNameMap[name]) { + return this._toolNameMap[name] + } + const LIMIT = 64 + if (name.length <= LIMIT) { + return name + } + if (name.startsWith('mcp__')) { + const idx = name.lastIndexOf('__') + if (idx > 0) { + const candidate = `mcp__${name.slice(idx + 2)}` + return candidate.length > LIMIT ? candidate.slice(0, LIMIT) : candidate + } + } + return name.slice(0, LIMIT) + } + + /** + * 构建工具名缩短映射（保证唯一） + * 构建唯一缩短名映射，处理碰撞 + * @returns {Object} { originalName: shortName } + */ + _buildShortNameMap(names) { + const LIMIT = 64 + const used = new Set() + const map = {} + + const baseCandidate = (n) => { + if (n.length <= LIMIT) { + return n + } + if (n.startsWith('mcp__')) { + const idx = n.lastIndexOf('__') + if (idx > 0) { + const cand = `mcp__${n.slice(idx + 2)}` + return cand.length > LIMIT ? cand.slice(0, LIMIT) : cand + } + } + return n.slice(0, LIMIT) + } + + const makeUnique = (cand) => { + if (!used.has(cand)) { + return cand + } + const base = cand + for (let i = 1; ; i++) { + const suffix = `_${i}` + const allowed = LIMIT - suffix.length + const tmp = (base.length > allowed ? base.slice(0, allowed) : base) + suffix + if (!used.has(tmp)) { + return tmp + } + } + } + + for (const n of names) { + const short = makeUnique(baseCandidate(n)) + used.add(short) + map[n] = short + } + return map + } + + /** + * 逆向恢复工具名（用于响应转换） + */ + _restoreToolName(shortName) { + return this._reverseToolNameMap[shortName] || shortName + } +} + +module.exports = CodexToOpenAIConverter diff --git a/src/services/geminiToOpenAI.js b/src/services/geminiToOpenAI.js new file mode 100644 index 00000000..82a60eee --- /dev/null +++ b/src/services/geminiToOpenAI.js @@ -0,0 +1,392 @@ +/** + * Gemini 响应格式 → OpenAI Chat Completions 格式转换器 + * 将 Gemini API 的原生响应转为标准 OpenAI chat.completion / chat.completion.chunk 格式 + */ + +class GeminiToOpenAIConverter { + createStreamState() { + return { + buffer: '', + id: `chatcmpl-${Date.now()}`, + created: Math.floor(Date.now() / 1000), + functionIndex: 0, + candidatesWithFunctionCalls: new Set(), + roleSent: false + } + } + + /** + * 流式转换: 拦截 res.write 的 chunk → OpenAI SSE 字符串 + * @param {Buffer|string} rawChunk + * @param {string} model + * @param {Object} state - createStreamState() 返回的状态 + * @returns {string|null} 转换后的 SSE 字符串，null 表示跳过 + */ + convertStreamChunk(rawChunk, model, state) { + const str = (typeof rawChunk === 'string' ? rawChunk : rawChunk.toString()).replace( + /\r\n/g, + '\n' + ) + + // 心跳透传：仅当 buffer 为空时才透传空白 + // buffer 有数据时需要继续处理（空白可能是 SSE 事件的 \n\n 分隔符） + if (!str.trim() && !state.buffer) { + return str + } + + state.buffer += str + let output = '' + + // 按 \n\n 分割完整 SSE 事件 + let idx + while ((idx = state.buffer.indexOf('\n\n')) !== -1) { + const event = state.buffer.slice(0, idx) + state.buffer = state.buffer.slice(idx + 2) + + if (!event.trim()) { + continue + } + + const lines = event.split('\n') + for (const line of lines) { + if (!line.startsWith('data: ')) { + continue + } + + const jsonStr = line.slice(6).trim() + if (!jsonStr) { + continue + } + + // [DONE] 消费（由 res.end patch 统一发送） + if (jsonStr === '[DONE]') { + continue + } + + let geminiData + try { + geminiData = JSON.parse(jsonStr) + } catch (e) { + // 解析失败透传 + output += `data: ${jsonStr}\n\n` + continue + } + + // 错误事件透传 + if (geminiData.error) { + output += `data: ${jsonStr}\n\n` + continue + } + + const chunks = this._convertGeminiChunkToOpenAI(geminiData, model, state) + for (const c of chunks) { + output += `data: ${JSON.stringify(c)}\n\n` + } + } + } + + return output || null + } + + /** + * 非流式转换: Gemini JSON → OpenAI chat.completion + * @param {Object} geminiData - { candidates, usageMetadata, modelVersion, responseId } + * @param {string} model + * @returns {Object} + */ + convertResponse(geminiData, model) { + // 兼容 v1internal 包裹格式 { response: { candidates: [...] } } + const data = geminiData.response || geminiData + const candidates = data.candidates || [] + const choices = candidates.map((candidate, i) => { + const parts = candidate.content?.parts || [] + const textParts = [] + const thoughtParts = [] + const toolCalls = [] + const images = [] + let fnIndex = 0 + + for (const part of parts) { + if ( + part.thoughtSignature && + !part.text && + !part.functionCall && + !part.inlineData && + !part.inline_data + ) { + continue + } + + if (part.functionCall) { + toolCalls.push({ + id: `${part.functionCall.name}-${Date.now()}-${fnIndex}`, + type: 'function', + function: { + name: part.functionCall.name, + arguments: JSON.stringify(part.functionCall.args || {}) + } + }) + fnIndex++ + } else if (part.text !== undefined) { + if (part.thought) { + thoughtParts.push(part.text) + } else { + textParts.push(part.text) + } + } else if (part.inlineData || part.inline_data) { + const inlineData = part.inlineData || part.inline_data + const imgData = inlineData.data + if (imgData) { + const mimeType = inlineData.mimeType || inlineData.mime_type || 'image/png' + images.push({ + type: 'image_url', + index: images.length, + image_url: { url: `data:${mimeType};base64,${imgData}` } + }) + } + } + } + + const message = { role: 'assistant' } + if (textParts.length > 0) { + message.content = textParts.join('') + } else { + message.content = null + } + if (thoughtParts.length > 0) { + message.reasoning_content = thoughtParts.join('') + } + if (toolCalls.length > 0) { + message.tool_calls = toolCalls + } + if (images.length > 0) { + message.images = images + } + + let finishReason = 'stop' + if (toolCalls.length > 0) { + finishReason = 'tool_calls' + } else if (candidate.finishReason) { + finishReason = this._mapFinishReason(candidate.finishReason) + } + + return { + index: candidate.index !== undefined ? candidate.index : i, + message, + finish_reason: finishReason + } + }) + + const result = { + id: data.responseId || `chatcmpl-${Date.now()}`, + object: 'chat.completion', + created: this._parseCreateTime(data.createTime), + model: data.modelVersion || model, + choices + } + + const usage = this._mapUsage(data.usageMetadata) + if (usage) { + result.usage = usage + } + + return result + } + + // --- 内部方法 --- + + _convertGeminiChunkToOpenAI(geminiData, model, state) { + // 兼容 v1internal 包裹格式 { response: { candidates: [...] } } + const data = geminiData.response || geminiData + + // 更新元数据 + if (data.responseId) { + state.id = data.responseId + } + if (data.modelVersion) { + state.model = data.modelVersion + } + if (data.createTime) { + const ts = this._parseCreateTime(data.createTime) + if (ts !== Math.floor(Date.now() / 1000)) { + state.created = ts + } + } + + const candidates = data.candidates || [] + if (candidates.length === 0 && data.usageMetadata) { + // 仅 usage 的最终 chunk + const chunk = this._makeChunk(state, model) + chunk.choices[0].finish_reason = 'stop' + chunk.usage = this._mapUsage(data.usageMetadata) + return [chunk] + } + + const results = [] + for (let i = 0; i < candidates.length; i++) { + const candidate = candidates[i] + const candidateIndex = candidate.index !== undefined ? candidate.index : i + const parts = candidate.content?.parts || [] + + for (const part of parts) { + if ( + part.thoughtSignature && + !part.text && + !part.functionCall && + !part.inlineData && + !part.inline_data + ) { + continue + } + + const chunk = this._makeChunk(state, model) + chunk.choices[0].index = candidateIndex + + if (part.functionCall) { + state.candidatesWithFunctionCalls.add(candidateIndex) + chunk.choices[0].delta = { + tool_calls: [ + { + index: state.functionIndex, + id: `${part.functionCall.name}-${Date.now()}-${state.functionIndex}`, + type: 'function', + function: { + name: part.functionCall.name, + arguments: JSON.stringify(part.functionCall.args || {}) + } + } + ] + } + this._injectRole(state, chunk.choices[0].delta) + state.functionIndex++ + results.push(chunk) + } else if (part.text !== undefined) { + if (part.thought) { + chunk.choices[0].delta = { reasoning_content: part.text } + } else { + chunk.choices[0].delta = { content: part.text } + } + this._injectRole(state, chunk.choices[0].delta) + results.push(chunk) + } else if (part.inlineData || part.inline_data) { + const inlineData = part.inlineData || part.inline_data + const imgData = inlineData.data + if (imgData) { + const mimeType = inlineData.mimeType || inlineData.mime_type || 'image/png' + chunk.choices[0].delta = { + images: [ + { + type: 'image_url', + index: 0, + image_url: { url: `data:${mimeType};base64,${imgData}` } + } + ] + } + this._injectRole(state, chunk.choices[0].delta) + results.push(chunk) + } + } + } + + // finish_reason + if (candidate.finishReason) { + const chunk = this._makeChunk(state, model) + chunk.choices[0].index = candidateIndex + + if (state.candidatesWithFunctionCalls.has(candidateIndex)) { + chunk.choices[0].finish_reason = 'tool_calls' + } else { + chunk.choices[0].finish_reason = this._mapFinishReason(candidate.finishReason) + } + + if (data.usageMetadata) { + chunk.usage = this._mapUsage(data.usageMetadata) + } + + results.push(chunk) + } + } + + return results + } + + _mapFinishReason(geminiReason) { + // 按 Gemini FinishReason 官方枚举完整映射 + const fr = geminiReason.toLowerCase() + if (fr === 'stop') { + return 'stop' + } + if (fr === 'max_tokens') { + return 'length' + } + // 工具调用异常 → stop（调用失败不等于内容过滤） + if ( + fr === 'malformed_function_call' || + fr === 'too_many_tool_calls' || + fr === 'unexpected_tool_call' + ) { + return 'stop' + } + // 内容策略/安全拦截 → content_filter + if ( + fr === 'safety' || + fr === 'recitation' || + fr === 'blocklist' || + fr === 'prohibited_content' || + fr === 'spii' || + fr === 'image_safety' || + fr === 'language' + ) { + return 'content_filter' + } + // FINISH_REASON_UNSPECIFIED, OTHER, 未知值 → stop + return 'stop' + } + + _makeChunk(state, model) { + return { + id: state.id, + object: 'chat.completion.chunk', + created: state.created, + model: state.model || model, + choices: [{ index: 0, delta: {}, finish_reason: null }] + } + } + + _injectRole(state, delta) { + if (!state.roleSent) { + delta.role = 'assistant' + state.roleSent = true + } + } + + _parseCreateTime(createTime) { + if (!createTime) { + return Math.floor(Date.now() / 1000) + } + // Gemini 官方文档：createTime 为 RFC3339 格式字符串 + const ts = Math.floor(new Date(createTime).getTime() / 1000) + return isNaN(ts) ? Math.floor(Date.now() / 1000) : ts + } + + _mapUsage(meta) { + if (!meta) { + return undefined + } + const completionTokens = (meta.candidatesTokenCount || 0) + (meta.thoughtsTokenCount || 0) + const result = { + prompt_tokens: meta.promptTokenCount || 0, + completion_tokens: completionTokens, + total_tokens: meta.totalTokenCount || 0 + } + if (meta.thoughtsTokenCount > 0) { + result.completion_tokens_details = { reasoning_tokens: meta.thoughtsTokenCount } + } + if (meta.cachedContentTokenCount > 0) { + result.prompt_tokens_details = { cached_tokens: meta.cachedContentTokenCount } + } + return result + } +} + +module.exports = GeminiToOpenAIConverter