@reinsjs/lowering-fetch 0.4.0 → 0.5.0

package/CHANGELOG.md

@@ -1,5 +1,25 @@
 # @reinsjs/lowering-fetch
 
+## 0.5.0
+
+### Minor Changes
+
+- 0455c28: Deliver `compact`'s request as a directive and ask twice. `SystemNotePayload` gains an optional `directive: true` for notes that ask the model to act; `@reinsjs/lowering-fetch` (all three protocols) and `@reinsjs/lowering-pi` send those as a framed `<system_note>` user message even when the model accepts mid-conversation system messages, and record the landing as `lossy(user-role)`. Measured on a pooled DeepSeek relay, the same request was ignored 4 of 4 times as a mid-conversation system message and followed 5 of 6 times as a framed user note; the official DeepSeek endpoint followed it 6 of 6 either way. Other notes (perception, pins) keep their current landing, and notes already in a log keep theirs, so cached prefixes do not change on upgrade. `compact({ force })` now asks at two lines by default (`at: [0.75, 0.8]`; a single number still works), uses blunter wording on the second request (`CompactRequestInfo.attempt`), skips the turn right after a compaction until the model has acted, and marks requests as directives unless `force.directive` is `false`.
+
+### Patch Changes
+
+- Updated dependencies [0455c28]
+- Updated dependencies [24ba407]
+- Updated dependencies [4a57a7a]
+  - @reinsjs/core@0.5.0
+
+## 0.4.1
+
+### Patch Changes
+
+- Updated dependencies [fb032c1]
+  - @reinsjs/core@0.4.0
+
 ## 0.4.0
 
 ### Minor Changes

package/README.md

@@ -79,7 +79,7 @@ No `node:*`, no `process`, no `Buffer` — the same `dist/index.js` runs on Node
 
 Every landing is declared in `LOSS_MATRIX["openai-chat"]` and asserted by tests — nothing is dropped silently.
 
-- `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`).
+- `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 note marked `directive: true` (an instruction the model must act on, such as `compact`'s request) always goes out as the framed `user` message, on all three protocols.
 - 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: 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`).
@@ -116,6 +116,6 @@ Non-2xx responses throw `HttpError` with `status`, `headers` and the raw body (`
 
 ## Documentation
 
-`docs/技术方案.md` §11, `docs/模块盘点/lowering-fetch.md`, `spikes/f1-chat-live/`, `spikes/f2-anthropic-live/`, `spikes/f3-responses-live/` (live verification against DeepSeek and, through the Cloudflare AI Gateway, OpenAI and Anthropic) and `spikes/edge-runtime-check/` (Cloudflare workerd) — in Chinese, at the repository root.
+`docs/archive/implementation-plans/2026-09-08-技术方案v1.0.md` §11, `docs/系统全景图_模块盘点/lowering-fetch.md`, `spikes/f1-chat-live/`, `spikes/f2-anthropic-live/`, `spikes/f3-responses-live/` (live verification against DeepSeek and, through the Cloudflare AI Gateway, OpenAI and Anthropic) and `spikes/edge-runtime-check/` (Cloudflare workerd) — in Chinese, at the repository root.
 
 MIT © 2026 NewRate Limited.

package/dist/index.d.ts

@@ -63,6 +63,8 @@ type IrItem = {
     kind: "system_note";
     event: Event;
     noteKind: SystemNotePayload["kind"];
+    /** 注入者声明是要模型去做的指令：三条线都以带标签的 user 文本送达，不看 midConversationSystem */
+    directive: boolean;
     text: string;
     escaped: boolean;
     deferred: boolean;
@@ -291,7 +293,7 @@ declare function endpointOf(model: FetchModel): string;
  *   user 侧内容（tool_result、用户消息、摘要）并进同一条 user（F0 A5b 实测同条 user 里 tool_result 后接文本可达）；
  * - **中途 system 摆放**（S1）：不能是首条、必须紧跟 user（含只带 tool_result 的 user）、后接 assistant 或收尾。所以说明
  *   一律攒到"下一条 assistant 之前"或末尾再放出；此刻前一条不是 user（是 assistant 或什么都没有）就退成 `<system_note>`
- *   框住的 user 文本，记 lossy(user-role)。放出位置比时间线晚了一条 user 的，仍算 exact（说明只换位置，见 CLAUDE.md）；
+ *   框住的 user 文本，记 lossy(user-role)。放出位置比时间线晚了一条 user 的，仍算 exact（说明只换位置，见 AGENTS.md）；
  * - **thinking 回放**：带 signature 原样回放（F0 A7b 接受、伪造签名 400）；无签名（流中断）与来自别家的一律 dropped 声明——
  *   不像 pi-ai 那样降成正文，模型的私下推理不该以它"说过的话"出现在历史里；`redacted_thinking` 用 data 原样回放；
  * - **缓存断点**：最多 4 个。我们打三处——system 末块、tools 末项、最后一条 user 末块（厂商按前缀向前找命中，20 块回看窗口）；

package/dist/index.js

@@ -373,7 +373,7 @@ var CHAT_LOSS_MATRIX = {
     lossy(
       "user-role",
       "wrapped in a <system_note> tag and sent with the user role",
-      "the model declares midConversationSystem: false"
+      "the model declares midConversationSystem: false, or the note is a directive the model must act on (SystemNotePayload.directive)"
     )
   ],
   "core.compaction": [
@@ -473,7 +473,7 @@ var ANTHROPIC_LOSS_MATRIX = {
     lossy2(
       "user-role",
       "wrapped in a <system_note> tag and sent with the user role",
-      "the model family does not support mid-conversation system, or the placement rules forbid it (the previous message is an assistant, or the note would be first)"
+      "the model family does not support mid-conversation system, or the placement rules forbid it (the previous message is an assistant, or the note would be first), or the note is a directive the model must act on (SystemNotePayload.directive)"
     )
   ],
   "core.compaction": [
@@ -620,6 +620,7 @@ function eventsToIr(input) {
           kind: "system_note",
           event: e,
           noteKind: e.payload.kind,
+          directive: e.payload.directive === true,
           text: t.text,
           escaped: t.escaped,
           deferred: false
@@ -663,6 +664,7 @@ function framedSystemNote(kind, text) {
 ${text}
 </system_note>`;
 }
+var DIRECTIVE_NOTE = "the note is a directive the model must act on, so it is wrapped in a <system_note> tag and sent with the user role";
 function framedSummary(summary) {
   return `[Summary of earlier conversation]
 ${summary}`;
@@ -846,7 +848,7 @@ function encodeAnthropicRequest(input) {
         break;
       }
       case "system_note": {
-        if (capabilities.midConversationSystem) {
+        if (capabilities.midConversationSystem && !item.directive) {
           pendingNotes.push({ item, moved: false });
         } else {
           openUser().push({ type: "text", text: framedSystemNote(item.noteKind, item.text) });
@@ -855,7 +857,7 @@ function encodeAnthropicRequest(input) {
             item.event,
             "lossy",
             "user-role",
-            "the model does not support mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
+            item.directive ? DIRECTIVE_NOTE : "the model does not support mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
             item.escaped ? ESCAPED_NOTE : void 0,
             item.deferred ? DEFERRED_NOTE : void 0
           );
@@ -1349,7 +1351,7 @@ function encodeChatRequest(input) {
         break;
       }
       case "system_note": {
-        if (capabilities.midConversationSystem) {
+        if (capabilities.midConversationSystem && !item.directive) {
           messages.push({ role: "system", content: item.text });
           land2(
             landings,
@@ -1366,7 +1368,7 @@ function encodeChatRequest(input) {
             item.event,
             "lossy",
             "user-role",
-            "the upstream does not accept mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
+            item.directive ? DIRECTIVE_NOTE : "the upstream does not accept mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
             item.escaped ? ESCAPED_NOTE : void 0,
             item.deferred ? DEFERRED_NOTE : void 0
           );
@@ -2127,7 +2129,7 @@ function encodeResponsesRequest(input) {
         break;
       }
       case "system_note": {
-        if (capabilities.midConversationSystem) {
+        if (capabilities.midConversationSystem && !item.directive) {
           items.push({ role: systemRole, content: [{ type: "input_text", text: item.text }] });
           land3(
             landings,
@@ -2147,7 +2149,7 @@ function encodeResponsesRequest(input) {
             item.event,
             "lossy",
             "user-role",
-            "the host declared that this upstream does not support mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
+            item.directive ? DIRECTIVE_NOTE : "the host declared that this upstream does not support mid-conversation system, so it is wrapped in a <system_note> tag and sent with the user role",
             item.escaped ? ESCAPED_NOTE : void 0,
             item.deferred ? DEFERRED_NOTE : void 0
           );
@@ -2641,7 +2643,7 @@ var RESPONSES_LOSS_MATRIX = {
     lossy3(
       "user-role",
       "wrapped in a <system_note> tag and sent with the user role",
-      "the host declared midConversationSystem: false for an unusual upstream"
+      "the host declared midConversationSystem: false for an unusual upstream, or the note is a directive the model must act on (SystemNotePayload.directive)"
     )
   ],
   "core.compaction": [