magpie-gemini-web 0.3.0 → 0.3.2

package/README.en.md

@@ -37,8 +37,9 @@ magpie plugin options magpie-gemini-web '{"allowTools":true,"allowImages":true}'
 | `allowTools` | `false` | Allows tool calls (experimental). When off, requests with tools get a 400 |
 | `allowImages` | `false` | Allows image input |
 | `baseURL` | `http://127.0.0.1:8083/v1` | Default backend URL |
-| `contextLimit` / `outputLimit` | `32768` / `8192` | Context and output limits reported to clients |
+| `contextLimit` / `outputLimit` | `1000000` / `8192` | Context and output limits reported to clients |
 | `requestTimeoutMs` / `modelTimeoutMs` | `240000` / `10000` | Timeouts in ms for requests and for fetching the model list |
+| `stripCitations` | `true` | Removes the citation markers Gemini adds to replies, `[cite: 1]` and `[cite_start]`, tool call arguments included. Set it to `false` while Gemini edits code that contains such text itself |
 
 ## Thinking effort
 

package/README.md

@@ -37,8 +37,9 @@ magpie plugin options magpie-gemini-web '{"allowTools":true,"allowImages":true}'
 | `allowTools` | `false` | 允许工具调用（实验性）。关闭时带工具的请求会返回 400 |
 | `allowImages` | `false` | 允许图片输入 |
 | `baseURL` | `http://127.0.0.1:8083/v1` | 默认后端地址 |
-| `contextLimit` / `outputLimit` | `32768` / `8192` | 向客户端声明的上下文 / 输出上限 |
+| `contextLimit` / `outputLimit` | `1000000` / `8192` | 向客户端声明的上下文 / 输出上限 |
 | `requestTimeoutMs` / `modelTimeoutMs` | `240000` / `10000` | 请求 / 拉取模型列表的超时（毫秒） |
+| `stripCitations` | `true` | 去掉 Gemini 在回复里加的引用标记 `[cite: 1]`、`[cite_start]`（包括工具调用参数）。让 Gemini 编辑本身含有这类文本的代码时，可以设为 `false` |
 
 ## 思考档位
 

package/lib/backend.mjs

@@ -1,3 +1,5 @@
+import { citeStream, citeJSON } from './cite.mjs';
+
 export const DEFAULT_BASE = 'http://127.0.0.1:8083/v1';
 
 const MEDIA_SPECIAL = /(?:^|[-_])(image|music|canvas|video|research)(?:[-_]|$)/i;
@@ -23,7 +25,8 @@ const KNOWN_OPTIONS = new Set([
   'contextLimit',
   'outputLimit',
   'requestTimeoutMs',
-  'modelTimeoutMs'
+  'modelTimeoutMs',
+  'stripCitations'
 ]);
 
 function makeOpenAIError(status, message, extraHeaders = {}) {
@@ -127,7 +130,7 @@ export function settings(options = {}) {
     allowImages = options.allowImages;
   }
 
-  let contextLimit = 32768;
+  let contextLimit = 1000000;
   if (options.contextLimit !== undefined) {
     if (!Number.isInteger(options.contextLimit) || options.contextLimit <= 0) {
       throw new Error('Invalid option contextLimit: must be a positive integer');
@@ -163,6 +166,16 @@ export function settings(options = {}) {
     modelTimeoutMs = options.modelTimeoutMs;
   }
 
+  // Gemini's citation markers come out of replies unless this is false: text that means them,
+  // as code about them does, would lose them too.
+  let stripCitations = true;
+  if (options.stripCitations !== undefined) {
+    if (typeof options.stripCitations !== 'boolean') {
+      throw new Error('Invalid option stripCitations: must be a boolean');
+    }
+    stripCitations = options.stripCitations;
+  }
+
   return {
     baseURL,
     allowTools,
@@ -170,7 +183,8 @@ export function settings(options = {}) {
     contextLimit,
     outputLimit,
     requestTimeoutMs,
-    modelTimeoutMs
+    modelTimeoutMs,
+    stripCitations
   };
 }
 
@@ -500,7 +514,15 @@ export async function forward(input, init, getAuth, opts) {
     respHeaders.set('X-Magpie-Sign-In', 'expired');
   }
 
-  return new Response(upstreamRes.body, {
+  // Gemini's citation markers come out of a chat reply, streamed or whole (see cite.mjs).
+  let respBody = upstreamRes.body;
+  if (opts.stripCitations && incomingUrl === chatUrl && upstreamRes.ok && respBody) {
+    const type = upstreamRes.headers.get('content-type') || '';
+    if (type.includes('text/event-stream')) respBody = citeStream(respBody);
+    else if (type.includes('json')) respBody = citeJSON(await upstreamRes.text());
+  }
+
+  return new Response(respBody, {
     status: upstreamRes.status,
     statusText: upstreamRes.statusText,
     headers: respHeaders

package/lib/cite.mjs

@@ -0,0 +1,166 @@
+// Gemini's web app marks the sentences it grounds in a source with [cite_start] … [cite: 1, 2],
+// which its page turns into citation chips. The backend hands the raw text on, so an agent's
+// replies, and the files its tools write, fill with them; the context an agent sends counts as
+// the source. They are taken out of every chat completion here, streamed or whole.
+
+const MARKER = /[ \t]*\[cite:\s*\d+(?:\s*,\s*\d+)*\s*\]|\[cite_(?:start|end)\]/g;
+// The end of a piece that may be the start of a marker the next piece finishes, with the blanks
+// before it, which go with a [cite: n].
+const PARTIAL = /[ \t]*(?:\[(?:c(?:i(?:t(?:e(?:_(?:s(?:t(?:a(?:r(?:t)?)?)?)?|e(?:n(?:d)?)?)?|:[\d,\s]*)?)?)?)?)?)?$/;
+
+export function stripCites(text) {
+  return typeof text === 'string' ? text.replace(MARKER, '') : text;
+}
+
+// CiteFilter strips one streamed text: push gives what can go now, holding back an end that may
+// be a marker's start; flush gives what it held.
+export class CiteFilter {
+  held = '';
+
+  push(piece) {
+    const text = stripCites(this.held + piece);
+    const tail = PARTIAL.exec(text)[0];
+    this.held = tail;
+    return text.slice(0, text.length - tail.length);
+  }
+
+  flush() {
+    const rest = this.held;
+    this.held = '';
+    return rest;
+  }
+}
+
+// The streamed texts of a chat completion chunk: each choice's content and reasoning, and
+// each tool call's arguments.
+function* texts(chunk) {
+  for (const choice of Array.isArray(chunk?.choices) ? chunk.choices : []) {
+    const d = choice?.delta;
+    if (!d || typeof d !== 'object') continue;
+    const at = choice.index ?? 0;
+    for (const field of ['content', 'reasoning_content', 'reasoning']) {
+      if (typeof d[field] === 'string') yield { key: `${at}:${field}`, get: () => d[field], set: v => { d[field] = v; } };
+    }
+    for (const call of Array.isArray(d.tool_calls) ? d.tool_calls : []) {
+      const fn = call?.function;
+      if (typeof fn?.arguments === 'string') {
+        yield { key: `${at}:tool:${call.index ?? 0}`, get: () => fn.arguments, set: v => { fn.arguments = v; } };
+      }
+    }
+  }
+}
+
+// heldChunk is a chunk carrying what the filters still hold, for the end of a stream, made
+// after last, the stream's last chunk.
+function heldChunk(filters, last) {
+  const choices = new Map();
+  for (const [key, f] of filters) {
+    const rest = f.flush();
+    if (!rest) continue;
+    const [at, kind, call] = key.split(':');
+    const index = Number(at);
+    if (!choices.has(index)) choices.set(index, { index, delta: {} });
+    const delta = choices.get(index).delta;
+    if (kind === 'tool') (delta.tool_calls ??= []).push({ index: Number(call), function: { arguments: rest } });
+    else delta[kind] = rest;
+  }
+  if (!choices.size) return null;
+  const { id, object, created, model } = last || {};
+  return { id, object, created, model, choices: [...choices.values()] };
+}
+
+// mergeHeld adds what the filters hold to chunk, in the choices it names (others get their own).
+function mergeHeld(chunk, filters) {
+  const held = heldChunk(filters, chunk);
+  for (const h of held?.choices ?? []) {
+    const c = chunk.choices.find(x => (x?.index ?? 0) === h.index);
+    if (!c) {
+      chunk.choices.push(h);
+      continue;
+    }
+    const d = (c.delta ??= {});
+    for (const [field, rest] of Object.entries(h.delta)) {
+      if (field !== 'tool_calls') {
+        d[field] = (typeof d[field] === 'string' ? d[field] : '') + rest;
+        continue;
+      }
+      for (const call of rest) {
+        const same = (d.tool_calls ??= []).find(x => (x?.index ?? 0) === call.index);
+        if (same) (same.function ??= {}).arguments = (same.function.arguments ?? '') + call.function.arguments;
+        else d.tool_calls.push(call);
+      }
+    }
+  }
+}
+
+// citeStream strips the markers from an SSE stream of chat completion chunks. An event that
+// isn't a chunk goes on as it came; what the filters hold goes out before the stream ends.
+export function citeStream(body) {
+  const filters = new Map();
+  const decoder = new TextDecoder();
+  const encoder = new TextEncoder();
+  let buffer = '';
+  let last = null;
+
+  const event = data => encoder.encode(`data: ${JSON.stringify(data)}\n\n`);
+  const flushAll = controller => {
+    const chunk = heldChunk(filters, last);
+    if (chunk) controller.enqueue(event(chunk));
+  };
+
+  const handle = (raw, controller) => {
+    const data = /^data: ?(.*)$/s.exec(raw)?.[1];
+    if (data === undefined) return controller.enqueue(encoder.encode(raw + '\n\n'));
+    if (data.trim() === '[DONE]') {
+      flushAll(controller);
+      return controller.enqueue(encoder.encode(raw + '\n\n'));
+    }
+    let chunk;
+    try {
+      chunk = JSON.parse(data);
+    } catch {
+      return controller.enqueue(encoder.encode(raw + '\n\n'));
+    }
+    last = chunk;
+    for (const t of texts(chunk)) {
+      if (!filters.has(t.key)) filters.set(t.key, new CiteFilter());
+      t.set(filters.get(t.key).push(t.get()));
+    }
+    // a client may take the chunk that finishes a choice for its end: what is held goes in it
+    if (Array.isArray(chunk.choices) && chunk.choices.some(c => c?.finish_reason)) mergeHeld(chunk, filters);
+    controller.enqueue(event(chunk));
+  };
+
+  return body.pipeThrough(new TransformStream({
+    transform(piece, controller) {
+      buffer += decoder.decode(piece, { stream: true });
+      const events = buffer.split(/\r?\n\r?\n/);
+      buffer = events.pop();
+      for (const raw of events) if (raw) handle(raw, controller);
+    },
+    flush(controller) {
+      buffer += decoder.decode();
+      if (buffer.trim()) handle(buffer.replace(/\s+$/, ''), controller);
+      flushAll(controller);
+    }
+  }));
+}
+
+// citeJSON strips the markers from a whole chat completion.
+export function citeJSON(text) {
+  let body;
+  try {
+    body = JSON.parse(text);
+  } catch {
+    return text;
+  }
+  for (const choice of Array.isArray(body?.choices) ? body.choices : []) {
+    const m = choice?.message;
+    if (!m || typeof m !== 'object') continue;
+    for (const field of ['content', 'reasoning_content', 'reasoning']) m[field] = stripCites(m[field]);
+    for (const call of Array.isArray(m.tool_calls) ? m.tool_calls : []) {
+      if (call?.function) call.function.arguments = stripCites(call.function.arguments);
+    }
+  }
+  return JSON.stringify(body);
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "magpie-gemini-web",
-  "version": "0.3.0",
+  "version": "0.3.2",
   "description": "Gemini Web provider for Magpie, backed by gemini-web2api-go",
   "type": "module",
   "main": "./index.mjs",