@vincemakes/kiso-provider-openai 0.1.28 → 0.1.30

package/README.md

@@ -12,7 +12,7 @@ overview.
 
 `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.
+for the request-prefix (D area) 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

package/dist/index.d.ts

@@ -16,13 +16,13 @@
 import OpenAI from "openai";
 import type { ChatCompletionChunk } from "openai/resources/chat/completions";
 import type { Adapter } from "@vincemakes/kiso-core";
-/** Config accepted by the high-level factory (七: the provider owns its SDK). */
+/** Config accepted by the high-level factory (round 7: the provider owns its SDK). */
 export interface OpenAICompatProviderConfig {
     readonly apiKey?: string;
     readonly baseUrl?: string;
 }
 /**
- * High-level factory (七): builds the adapter FROM CONFIG, owning the SDK
+ * High-level factory (round 7): builds the adapter FROM CONFIG, owning the SDK
  * inside this package. Consumers (and the runtime's lazy provider path)
  * import ONLY @vincemakes/kiso-provider-openai — the SDK stays a private dependency
  * of this package, so nested installs resolve it next to here, never

package/dist/index.js

@@ -18,7 +18,7 @@ import { join } from "node:path";
 import OpenAI from "openai";
 import { mapApiError } from "@vincemakes/kiso-core";
 /**
- * High-level factory (七): builds the adapter FROM CONFIG, owning the SDK
+ * High-level factory (round 7): builds the adapter FROM CONFIG, owning the SDK
  * inside this package. Consumers (and the runtime's lazy provider path)
  * import ONLY @vincemakes/kiso-provider-openai — the SDK stays a private dependency
  * of this package, so nested installs resolve it next to here, never
@@ -104,7 +104,7 @@ export function createOpenAICompatAdapter(client) {
                     for (const tc of delta?.tool_calls ?? []) {
                         const buffered = pending.get(tc.index);
                         if (!buffered) {
-                            // 六: no fallback id is adopted yet — the id must
+                            // round 6: no fallback id is adopted yet — the id must
                             // arrive from the provider to become the identity.
                             pending.set(tc.index, {
                                 index: tc.index,
@@ -122,7 +122,7 @@ export function createOpenAICompatAdapter(client) {
                         if (tc.function?.name)
                             call.name = tc.function.name;
                         if (tc.id) {
-                            // 九: the FIRST non-empty id is the call's identity,
+                            // round 9: the FIRST non-empty id is the call's identity,
                             // forever. A DIFFERENT id later is a protocol
                             // violation — a structured error, never a silent
                             // switch (start/delta/end must share one identity).
@@ -136,7 +136,7 @@ export function createOpenAICompatAdapter(client) {
                             call.id = tc.id;
                             // The identity is now known: emit the start, then
                             // flush the argument deltas that arrived before it
-                            // under the SAME id (六: start → delta → end all
+                            // under the SAME id (round 6: start → delta → end all
                             // share one identity, never the fallback).
                             if (!call.emittedStart) {
                                 call.emittedStart = true;
@@ -174,7 +174,7 @@ export function createOpenAICompatAdapter(client) {
                     }
                     if (chunk.usage) {
                         usageSent = true;
-                        // 六: REAL cached-token data is read from the provider's
+                        // round 6: REAL cached-token data is read from the provider's
                         // prompt_tokens_details — an absent value is null, NEVER
                         // faked as a zero-cache turn. OpenAI does not report a
                         // cache write; null is the honest answer.
@@ -210,7 +210,7 @@ export function createOpenAICompatAdapter(client) {
                 catch {
                     input = null; // never a silent repair
                 }
-                // 六: a call whose id NEVER arrived adopts the index fallback
+                // round 6: a call whose id NEVER arrived adopts the index fallback
                 // here — no start/delta was emitted under any other identity,
                 // so this end is the call's first and only identity.
                 yield {
@@ -293,7 +293,7 @@ function toOpenAIContent(content) {
         ? { type: "text", text: block.text }
         : {
             type: "image_url",
-            // 六: a base64 block becomes a REAL data URL —
+            // round 6: a base64 block becomes a REAL data URL —
             // `data:<media>;base64,<data>` — never an empty string URL.
             // URL-sourced blocks pass the provider URL through.
             image_url: {
@@ -304,7 +304,7 @@ function toOpenAIContent(content) {
         });
 }
 /**
- * 六: OpenAI tool results accept TEXT ONLY — an image block is converted to
+ * round 6: OpenAI tool results accept TEXT ONLY — an image block is converted to
  * an explicit, honest text note (what kind of image was omitted and why),
  * never silently dropped.
  */
@@ -325,20 +325,20 @@ function toOpenAIMessages(messages, systemPrompt) {
     if (systemPrompt !== undefined) {
         out.push({ role: "system", content: systemPrompt });
     }
-    // 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)。
+    // the merge round (0.1.23) C7 revision: `reasoning_content`'s PRESENCE follows the whole projection's
+    // monotone state — any reasoning in the projection → every assistant message
+    // carries it (its own reasoning, or ""); otherwise none do. Rationale: the D-area request-level
+    // byte stability. The old implementation keyed on the "current round" (the ergonomics batch C7); once a round boundary passed,
+    // the old rounds' assistant-message fields were stripped and the serialization rewritten — requests N and N+1's common
+    // prefix broke at the old message; the provider's prefix cache broke at every round boundary (the fresh-mystery
+    // empirical proof: in a 14-request session both cache breaks sat at round boundaries; after the fix the same session had 0 breaks,
+    // per-request cached 82-98%). Monotonicity guarantees the field's presence flips at most once per session
+    // (when the first thinking appears, usually in the first round — before it there are almost no old assistant
+    // messages); after it, never flips for the session's life; real OpenAI never produces reasoning → the field
+    // never appears, and its request path is byte-for-byte the old behavior. Old-round messages carrying reasoning
+    // pass their reasoning back — DeepSeek's officially recommended cache-stable shape; old content hitting the prefix
+    // cache bills at 0.1× only — the "repassing wastes tokens" premise does not hold under cache economics.
+    // the "" field is accepted on old rounds too (real API verification: 200 + 2560 cached tokens).
     const hasReasoning = messages.some((m) => m.role === "assistant" && m.reasoning !== undefined);
     for (const msg of messages) {
         if (msg.role === "user") {
@@ -372,7 +372,7 @@ function toOpenAIMessages(messages, systemPrompt) {
         else {
             // tool messages accept text only — images are converted to an
             // EXPLICIT text note (toOpenAIToolResultContent), never dropped
-            // (六).
+            // (round 6).
             out.push({
                 role: "tool",
                 tool_call_id: msg.callId,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vincemakes/kiso-provider-openai",
-  "version": "0.1.28",
+  "version": "0.1.30",
   "description": "kiso OpenAI-compatible adapter — OpenAI + the compat family (GLM, Kimi, DeepSeek, OpenRouter) via base_url swap, reasoning dialects digested.",
   "type": "module",
   "license": "MIT",
@@ -21,7 +21,7 @@
     "test": "vitest run"
   },
   "dependencies": {
-    "@vincemakes/kiso-core": "0.1.28",
+    "@vincemakes/kiso-core": "0.1.29",
     "openai": "^7.3.0"
   },
   "devDependencies": {