@vincemakes/kiso-provider-openai 0.1.21 → 0.1.24

package/README.md

@@ -7,3 +7,14 @@ connection/timeout/5xx classification.
 
 Requires Node >= 22. See the repository README for the framework
 overview.
+
+## Debug tooling
+
+`KISO_DUMP_REQUESTS=<dir>` writes every outgoing request body to
+`<dir>/req-<pid>-<n>.json` before it is sent — the diagnosis instrument
+for the request-prefix (D 区) contract, kept as a permanent debug sink.
+⚠ The bodies are REAL conversation data (the model may have seen repo
+contents) — never share a dump dir; a dump failure never breaks the
+request. `bench/dumpdiff.py` byte-diffs consecutive dumps and localizes
+the first divergence (healthy = at the older request's last-message end;
+violation = inside an old message).

package/dist/index.js

@@ -13,6 +13,8 @@
  * provider reports usage (not all compat providers do); `stop` reason maps
  * from finish_reason.
  */
+import { mkdirSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
 import OpenAI from "openai";
 import { mapApiError } from "@vincemakes/kiso-core";
 /**
@@ -32,23 +34,28 @@ export function createOpenAICompatProvider(config = {}) {
 export function createOpenAICompatAdapter(client) {
     return {
         async *stream(options) {
+            // The explicit streaming params type keeps the `stream: true`
+            // literal overload — a widened `stream: boolean` would type the
+            // create() call as the NON-streaming variant (TS2504).
+            const body = {
+                model: options.model,
+                messages: toOpenAIMessages(options.messages, options.systemPrompt),
+                stream: true,
+                // D5: request real streaming usage — without this the
+                // provider never sends a usage chunk and we would
+                // report known:false forever.
+                stream_options: { include_usage: true },
+                ...(options.tools?.length ? { tools: options.tools.map(toOpenAITool) } : {}),
+                ...(options.maxTokens !== undefined ? { max_tokens: options.maxTokens } : {}),
+                ...(options.temperature !== undefined ? { temperature: options.temperature } : {}),
+            };
+            maybeDumpRequest(body);
             // Area 6: the stream CREATION is inside the error normalization —
             // a 429/5xx/connection failure before the first byte is a mapped,
             // retryable StructuredError, so the loop's pre-stream retry works.
             let stream;
             try {
-                stream = await client.chat.completions.create({
-                    model: options.model,
-                    messages: toOpenAIMessages(options.messages, options.systemPrompt),
-                    stream: true,
-                    // D5: request real streaming usage — without this the
-                    // provider never sends a usage chunk and we would
-                    // report known:false forever.
-                    stream_options: { include_usage: true },
-                    ...(options.tools?.length ? { tools: options.tools.map(toOpenAITool) } : {}),
-                    ...(options.maxTokens !== undefined ? { max_tokens: options.maxTokens } : {}),
-                    ...(options.temperature !== undefined ? { temperature: options.temperature } : {}),
-                }, 
+                stream = await client.chat.completions.create(body, 
                 // Phase B: cancellation reaches the SDK; the system prompt is a
                 // first-class message, not a dropped option.
                 options.signal !== undefined ? { signal: options.signal } : undefined);
@@ -226,6 +233,30 @@ export function createOpenAICompatAdapter(client) {
         },
     };
 }
+// ── Request dump (debug tool) ───────────────────────────────────────────
+let dumpSeq = 0;
+/**
+ * KISO_DUMP_REQUESTS=<dir> — DEBUG TOOL: writes the FULL request body (the
+ * exact JSON the SDK will send) to `<dir>/req-<pid>-<n>.json` before it is
+ * sent, numbered per request in process order. Built for the fresh-mystery
+ * diagnosis (byte-diff consecutive requests' common prefix); kept as a
+ * permanent debug sink. ⚠ The bodies are REAL conversation data (the model
+ * may have seen repo contents) — never share a dump dir. A dump failure
+ * never breaks the request.
+ */
+function maybeDumpRequest(body) {
+    const dir = process.env.KISO_DUMP_REQUESTS;
+    if (dir === undefined || dir === "")
+        return;
+    dumpSeq += 1;
+    try {
+        mkdirSync(dir, { recursive: true });
+        writeFileSync(join(dir, `req-${process.pid}-${dumpSeq}.json`), JSON.stringify(body));
+    }
+    catch {
+        // debug tool: silent on failure
+    }
+}
 // ── Mapping helpers ────────────────────────────────────────────────────
 /**
  * Exhaustive over the SDK's CLOSED finish_reason union (Area 6): a new SDK
@@ -294,28 +325,22 @@ function toOpenAIMessages(messages, systemPrompt) {
     if (systemPrompt !== undefined) {
         out.push({ role: "system", content: systemPrompt });
     }
-    let lastUser = -1;
-    for (const [i, m] of messages.entries()) {
-        if (m.role === "user")
-            lastUser = i;
-    }
-    // 自举 P1/P2 + 手感批 C7: thinking mode is detected by the presence of
-    // reasoning in the CURRENT turn only (messages after the last user
-    // message). Real OpenAI never emits thinking events, so its requests
-    // never see the field. The current-turn messages carry no source
-    // marker, so the simple judgment is "the current session adapter is
-    // their source" — their reasoning is THIS family's own and may be
-    // echoed. Reasoning in OLDER turns belongs to whatever adapter
-    // produced them (an anthropic-thinking history continued by an openai
-    // adapter is the case this guards): it must NEVER flip the mode —
-    // otherwise this adapter would send `reasoning_content: ""` on a
-    // request it has no business tagging as thinking. In thinking mode,
-    // ONLY the CURRENT turn's assistant messages carry reasoning_content —
-    // their own, or "" when the step produced no thinking (the field must
-    // still be present, or DeepSeek 400s); old turns' CoT is never echoed
-    // (DeepSeek does not need it, and echoing it is token waste).
-    const thinkingMode = messages.slice(lastUser + 1).some((m) => m.role === "assistant" && m.reasoning !== undefined);
-    for (const [i, msg] of messages.entries()) {
+    // 合并轮 (0.1.23) C7 修订: `reasoning_content` 的存在性由整个投影的
+    // 单调状态决定 — 投影里存在任一 reasoning → 每条 assistant 消息都
+    // 携带(自身 reasoning,或 "")；否则一条都不带。理由: D 区请求级
+    // 字节稳定。旧实现按"当前轮"判定(手感批 C7),轮边界一过,旧轮
+    // assistant 消息的字段被剥掉,序列化被改写 — 请求 N 与 N+1 的公共
+    // 前缀断在旧消息处,provider 前缀缓存每轮边界断一次(fresh 之谜
+    // 实证: 14 请求的会话里两次缓存断点都在轮边界;修复后同会话 0 断
+    // 点、逐请求 cached 82-98%)。单调性保证存在性在会话内只翻转一次
+    // (首个 thinking 出现时,通常在首轮 — 此前几乎没有旧 assistant
+    // 消息),之后从生到死不翻转;真 OpenAI 永不产生 reasoning → 字段
+    // 永不出现,其请求路径与旧行为逐字节相同。带 reasoning 的旧轮消息
+    // 回传其推理是 DeepSeek 官方推荐的缓存稳定形态;旧内容命中前缀
+    // 缓存只按 0.1× 计费,"回传是 token 浪费"的前提在缓存经济下不成立。
+    // "" 字段在旧轮同样被接受(真 API 验证: 200 + 2560 cached tokens)。
+    const hasReasoning = messages.some((m) => m.role === "assistant" && m.reasoning !== undefined);
+    for (const msg of messages) {
         if (msg.role === "user") {
             out.push({ role: "user", content: toOpenAIContent(msg.content) });
         }
@@ -337,7 +362,11 @@ function toOpenAIMessages(messages, systemPrompt) {
                 role: "assistant",
                 content: msg.blocks.filter((b) => b.type === "text").map((b) => b.text).join(""),
                 ...(toolCalls.length ? { tool_calls: toolCalls } : {}),
-                ...(thinkingMode && i > lastUser ? { reasoning_content: msg.reasoning ?? "" } : {}),
+                // DeepSeek extension field — absent from the SDK's param type,
+                // added via spread so the literal stays assignable (the union
+                // type carries the extra key). Presence follows hasReasoning
+                // (the monotone rule above — the field never flips mid-history).
+                ...(hasReasoning ? { reasoning_content: msg.reasoning ?? "" } : {}),
             });
         }
         else {

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@vincemakes/kiso-provider-openai",
-  "version": "0.1.21",
-  "description": "kiso OpenAI-compatible adapter — OpenAI + the compat family (GLM, Kimi, DeepSeek, OpenRouter) via base_url swap, reasoning dialects digested.",
+  "version": "0.1.24",
+  "description": "kiso OpenAI-compatible adapter \u2014 OpenAI + the compat family (GLM, Kimi, DeepSeek, OpenRouter) via base_url swap, reasoning dialects digested.",
   "type": "module",
   "license": "MIT",
   "exports": {
@@ -21,7 +21,7 @@
     "test": "vitest run"
   },
   "dependencies": {
-    "@vincemakes/kiso-core": "0.1.21",
+    "@vincemakes/kiso-core": "0.1.24",
     "openai": "^7.3.0"
   },
   "devDependencies": {