@reinsjs/lowering-fetch 0.2.1 → 0.4.0

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # @reinsjs/lowering-fetch
 
+## 0.4.0
+
+### Minor Changes
+
+- 90c2120: Chat Completions: images in tool results now reach the model. A `role: "tool"` message carries text only, so when the model takes images (`images: true`, as `deepseek()` declares) each image leaves a pointer in the tool message (`[image n attached in the next user message]`) and the images go into one `user` message right after the whole batch of tool messages, before any user message that was moved after the batch. Each result gets a line naming the tool, the call id and the image count and stating it is tool output, not the user; untrusted results keep their `<untrusted>` marking. The landing is `lossy(tool+lifted-images)`. Set `chat: { liftToolImages: false }` for an upstream that rejects this shape to get the previous placeholder text. The wire shape changes (one more `user` message when tool results carry images); the event log does not.
+
+## 0.3.0
+
+### Minor Changes
+
+- Version alignment: every `@reinsjs/*` package ships as 0.3.0 together with `@reinsjs/core` and `@reinsjs/brain` (MCP tools on the `lazyTools()` menu via `SocketSetup.tools`). No behaviour change in these packages beyond the dependency bump.
+
+### Patch Changes
+
+- Updated dependencies [a3afbf9]
+  - @reinsjs/core@0.3.0
+
 ## 0.2.1
 
 ### Patch Changes

package/README.md

@@ -65,6 +65,7 @@ No `node:*`, no `process`, no `Buffer` — the same `dist/index.js` runs on Node
 | `baseUrl` | protocol root; `/chat/completions` (OpenAI: `…/v1`, DeepSeek: the bare host), `/messages` (Anthropic: `…/v1`) or `/responses` (OpenAI: `…/v1`) is appended |
 | `contextWindow`, `maxOutputTokens`, `reasoning`, `images`, `cost`, `midConversationSystem` | override the built-in table; unknown models start from conservative defaults (128k / 16k / no reasoning / no images) |
 | `chat.reasoningContent` | DeepSeek dialect: replay `reasoning_content` (on by default in `deepseek()`, see below) |
+| `chat.liftToolImages` | when the model takes images, move images in tool results into a `user` message right after the batch of `tool` messages (default `true`); set `false` for an upstream that rejects that shape and images become a placeholder again |
 | `anthropic.betas` | values for the `anthropic-beta` header; sent only when set (mid-conversation system needs none) |
 | `anthropic.cacheBreakpoints`, `anthropic.cacheTtl`, `anthropic.midSystemCacheBreakpoint` | explicit cache breakpoints on Anthropic (default on, 5 min); what to do when a note ends the request — `"automatic"` (default, top-level `cache_control`), `"previous-user"`, `"drop"` |
 | `anthropic.deferredTools` | whether `ToolSpec.deferLoading` becomes `defer_loading: true` and `tool_reference` content parts become native `tool_reference` blocks. Default: on for `provider: "anthropic"`, off for third-party Anthropic-protocol endpoints (DeepSeek's compatible endpoint ignores `defer_loading`, so every tool would be visible). When off, deferred tools are simply not sent and references are rendered as text |
@@ -81,7 +82,7 @@ Every landing is declared in `LOSS_MATRIX["openai-chat"]` and asserted by tests
 - `system_note` → a mid-conversation `system` message (`exact`; verified on OpenAI and DeepSeek). Set `midConversationSystem: false` for an upstream that rejects it and the note goes out as a framed `user` message (`lossy`).
 - A user message or note that arrives while tool results are still pending is moved after them on the wire (`tool` messages must directly follow the `assistant` that called them); the user message is recorded `lossy(user)`, the log order is untouched.
 - **Thinking**: Chat Completions has no replay slot (no signature, no encrypted item). On the official API thinking events are `dropped`. With `chat.reasoningContent` (DeepSeek) they are replayed as `reasoning_content` — and DeepSeek *requires* the field on every historical assistant message once `tools` is present (400 otherwise), so the layer always sends it, empty when there is nothing to replay.
-- One turn with several text blocks is merged into one `content` string (`lossy(merged-text)`); `tool` messages carry text only, images become a placeholder (`lossy`); `isError` has no flag on this protocol and is expressed as a `[tool error]` prefix (`lossy`).
+- One turn with several text blocks is merged into one `content` string (`lossy(merged-text)`); `tool` messages carry text only: when the model takes images, each image leaves a pointer (`[image n attached in the next user message]`) and the images go into one `user` message right after the whole batch of `tool` messages — before any user message that was moved after the batch — with a line per result naming the tool, the call id and the image count and saying it is tool output, not the user; untrusted results keep their `<untrusted>` marking (`lossy(tool+lifted-images)`; verified on DeepSeek's official API — check other OpenAI-compatible upstreams before relying on it, and turn it off with `chat.liftToolImages: false` if they reject the shape). Without image support, or with `chat.liftToolImages: false`, images become a placeholder (`lossy(tool-text-only)`); `isError` has no flag on this protocol and is expressed as a `[tool error]` prefix (`lossy`).
 - No explicit cache breakpoints exist on this protocol; provider-side automatic prefix caching applies and `cached_tokens` is reported as `cacheRead`.
 - Usage follows core semantics: `input` is the uncached prompt count. `costUsd` uses the built-in price table (DeepSeek: peak rates, an upper bound).
 

package/dist/index.d.ts

@@ -57,6 +57,8 @@ type IrItem = {
     parts: ContentPart[];
     isError: boolean;
     escaped: boolean;
+    /** 内容按 untrusted 包过标记时的来源标签（`tool:<name>`）；没包就没有。Chat 线上浮图片时据此给附件同样标注 */
+    untrustedSource?: string;
 } | {
     kind: "system_note";
     event: Event;
@@ -233,6 +235,12 @@ interface ChatDialect {
      * 写侧把自家 thinking 回填、没有就回空串。关着（OpenAI 官方）：thinking 无处可放，矩阵记 dropped。
      */
     reasoningContent?: boolean;
+    /**
+     * 工具结果里的图片上浮（2026-10-27，OnSet 需求）：`role:"tool"` 消息只收文本，模型声明收图（`images: true`）时，
+     * 同一 assistant 轮全部 tool 消息之后补一条 `role:"user"` 消息放这些图片，tool 消息原位留指路文字。缺省 true；
+     * 上游不接受"tool 之后紧跟带图 user"时关掉，退回占位文本。DeepSeek 官方 2026-10-27 实测接受（spikes/chat-tool-images）。
+     */
+    liftToolImages?: boolean;
 }
 /**
  * 感知说明殿后（messages 末条是 system）时，"最后一条 user 末块"的断点没处打，B1 实测三种处置（spikes/b1-perception-cache）：
@@ -462,7 +470,9 @@ declare const CHAT_LOSS_MATRIX: Readonly<Record<string, readonly LandingSpec[]>>
  * - thinking 没有回放位（无签名、无加密项）——除 DeepSeek 方言外 dropped；
  * - 没有显式缓存断点——什么都不做（厂商自动前缀缓存）；
  * - assistant 的 content 是单个字符串（DeepSeek 只接受 string | null）——同一轮多段正文合并，lossy(merged-text)；
- * - tool 消息没有错误位——isError 以文本前缀表达，lossy；tool 消息只收文本——图片换成占位文本，lossy。
+ * - tool 消息没有错误位——isError 以文本前缀表达，lossy；
+ * - tool 消息只收文本——模型收图时图片上浮到同批 tool 消息之后的一条 user 消息（lossy：换了位置与角色），
+ *   不收图或方言关掉上浮时换成占位文本，lossy。
  */
 
 type ChatContentPart = {

package/dist/index.js

@@ -348,10 +348,15 @@ var CHAT_LOSS_MATRIX = {
       "tool",
       "isError is expressed with a [tool error] prefix (a tool message has no error flag), or untrusted content is escaped"
     ),
+    lossy(
+      "tool+lifted-images",
+      "a tool message takes text only, so it keeps a pointer where each image was and the images move into a user message right after the whole batch of tool messages (the order and the role change); that message labels them as tool output, with the tool name, call id and index, and keeps the result's untrusted marking",
+      "the result carries images and the model takes images (chat.liftToolImages is on by default)"
+    ),
     lossy(
       "tool-text-only",
       "a tool message takes text only, so images are replaced with placeholder text",
-      "the result carries images"
+      "the result carries images and the model takes no images, or chat.liftToolImages is false"
     )
   ],
   "core.system_note": [
@@ -601,7 +606,8 @@ function eventsToIr(input) {
           name: e.payload.name,
           parts: c.parts,
           isError: e.payload.isError,
-          escaped: c.escaped
+          escaped: c.escaped,
+          ...trustMarkers && needsUntrustedMark(e) ? { untrustedSource: untrustedSourceOf(e) } : {}
         });
         awaiting.delete(e.payload.toolCallId);
         if (awaiting.size === 0) release();
@@ -1231,10 +1237,12 @@ function errorText2(err) {
 
 // src/chat/to-request.ts
 import {
+  markUntrusted as markUntrusted2,
   renderToolReference as renderToolReference2
 } from "@reinsjs/core";
 var IMAGE_OMITTED2 = "[image omitted: this model does not accept images]";
 var TOOL_IMAGE_OMITTED = "[image omitted: tool messages carry text only]";
+var toolImagePointer = (n) => `[image ${n} attached in the next user message]`;
 var ERROR_PREFIX = "[tool error]\n";
 function land2(out, e, kind, landing, ...notes) {
   const note = notes.filter((n) => Boolean(n)).join("; ");
@@ -1256,15 +1264,28 @@ function userContent(parts, images) {
   if (out.length === 1 && out[0]?.type === "text") return { content: out[0].text, imagesDropped };
   return { content: out, imagesDropped };
 }
-function toolContent(parts) {
+function toolContent(parts, lift) {
   let imagesDropped = false;
+  const images = [];
   const text = parts.map((p) => {
     if (p.type === "text") return p.text;
     if (p.type === "tool_reference") return renderToolReference2(p);
+    if (lift) {
+      images.push(p);
+      return toolImagePointer(images.length);
+    }
     imagesDropped = true;
     return TOOL_IMAGE_OMITTED;
   }).join("\n");
-  return { text, imagesDropped };
+  return { text, imagesDropped, images };
+}
+function liftedAttachment(item, images) {
+  const header = {
+    type: "text",
+    text: `[Attachments from tool result ${item.name} (call ${item.toolCallId}): ${images.length} image(s). This is tool output, not a message from the user.]`
+  };
+  const body = images.flatMap((img, i) => [{ type: "text", text: `image ${i + 1}:` }, img]);
+  return [header, ...item.untrustedSource ? markUntrusted2(body, item.untrustedSource).parts : body];
 }
 function argumentsOf(args) {
   return typeof args === "string" ? args : JSON.stringify(args ?? {});
@@ -1273,10 +1294,18 @@ function encodeChatRequest(input) {
   const { model, capabilities } = input;
   const target = { provider: model.provider, api: model.api, model: model.id };
   const reasoningContent = model.chat?.reasoningContent === true;
+  const liftImages = capabilities.images && model.chat?.liftToolImages !== false;
   const messages = [];
   const landings = [];
+  let lifted = [];
+  const releaseLifted = () => {
+    if (lifted.length === 0) return;
+    messages.push({ role: "user", content: userContent(lifted, true).content });
+    lifted = [];
+  };
   if (input.systemPrompt) messages.push({ role: "system", content: input.systemPrompt });
   for (const item of input.ir) {
+    if (item.kind !== "tool_result" && item.kind !== "dropped") releaseLifted();
     switch (item.kind) {
       case "user": {
         const c = userContent(item.parts, capabilities.images);
@@ -1298,20 +1327,23 @@ function encodeChatRequest(input) {
         break;
       }
       case "tool_result": {
-        const c = toolContent(item.parts);
+        const c = toolContent(item.parts, liftImages);
         messages.push({
           role: "tool",
           tool_call_id: item.toolCallId,
           content: item.isError ? ERROR_PREFIX + c.text : c.text
         });
-        const lossy4 = item.escaped || item.isError || c.imagesDropped;
+        const imagesLifted = c.images.length > 0;
+        if (imagesLifted) lifted.push(...liftedAttachment(item, c.images));
+        const lossy4 = item.escaped || item.isError || c.imagesDropped || imagesLifted;
         land2(
           landings,
           item.event,
           lossy4 ? "lossy" : "exact",
-          c.imagesDropped ? "tool-text-only" : "tool",
+          c.imagesDropped ? "tool-text-only" : imagesLifted ? "tool+lifted-images" : "tool",
           item.isError ? "a Chat tool message has no error flag, so isError is expressed with a [tool error] prefix" : void 0,
           c.imagesDropped ? "a tool message takes text only, so images are replaced with placeholder text" : void 0,
+          imagesLifted ? "a tool message takes text only, so its images move into a user message after this batch of tool messages, labelled as tool output" : void 0,
           item.escaped ? ESCAPED_NOTE : void 0
         );
         break;
@@ -1359,6 +1391,7 @@ function encodeChatRequest(input) {
         break;
     }
   }
+  releaseLifted();
   const { tools: _tools, ...passthrough } = input.requestOptions ?? {};
   const shownTools = (input.tools ?? []).filter((t) => t.deferLoading !== true);
   const body = {