dsh-agy-link 0.4.66 → 0.4.67

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,5 +1,23 @@
1
1
  # Changelog
2
2
 
3
+ ## 0.4.67 (2026-10-20)
4
+
5
+ ### English
6
+
7
+ - **Fix: Host authorization-review calls lost their output protocol (Fixes #50).**
8
+ - When DSH's `@deepseek-ai/dsh-experimental-auto-review` plugin reviews a tool call it sends the entire rubric in `options.system` (`REVIEW_POLICY`, which demands "Return exactly one JSON object and no other text"). The adapter only forwarded `options.system` when `forwardSystemPrompt` was on — and that flag defaults to `false` — so the model received the data sections alone.
9
+ - The model then answered the way a coding agent does ("Based on a…", preceded by a plan and a self-written test script), and the caller's `JSON.parse` threw `Unexpected token 'B', "Based on a"... is not valid JSON`. Every reviewed tool call was denied with "its body was not executed" instead of being reviewed.
10
+ - The adapter already detects these calls (`isReview`, used to exempt reviews from the auxiliary-call gate). System-prompt forwarding now also fires on that probe, so the policy always reaches the model regardless of the global opt-in. Ordinary turns are untouched: `forwardSystemPrompt` remains `false` by default and non-review calls still drop the system prompt.
11
+ - New tests pin both directions: a review call forwards the policy without the opt-in, and an ordinary turn still sends a bare prompt.
12
+
13
+ ### 中文
14
+
15
+ - **修复:宿主授权复核调用丢失了输出协议(修复 #50)。**
16
+ - DSH 的 `@deepseek-ai/dsh-experimental-auto-review` 插件复核工具调用时,把整份规则放在 `options.system` 里(`REVIEW_POLICY`,要求「只返回一个 JSON 对象,不得有其他文本」)。该适配器仅在 `forwardSystemPrompt` 开启时才转发 `options.system`,而此配置默认 `false`,于是模型只拿到了数据段。
17
+ - 模型随后按编码代理的方式作答(「Based on a…」开头,前面还有计划和自写的测试脚本),调用方的 `JSON.parse` 抛出 `Unexpected token 'B', "Based on a"... is not valid JSON`。结果是每一次被复核的工具调用都以「its body was not executed」被拒绝,而不是被复核。
18
+ - 适配器本来就能识别这类调用(`isReview`,用于把复核豁免出辅助调用闸门)。现在该判定同时触发系统提示转发,策略无论全局开关如何都会送到模型。普通轮次不受影响:`forwardSystemPrompt` 默认仍为 `false`,非复核调用照旧丢弃系统提示。
19
+ - 新增测试固定两个方向:复核调用在未开启开关时也转发策略;普通轮次仍只发送裸提示。
20
+
3
21
  ## 0.4.66 (2026-10-19)
4
22
 
5
23
  ### English
package/dist/index.js CHANGED
@@ -30603,7 +30603,7 @@ var AgyAdapter = class AgyAdapter extends LlmAdapter {
30603
30603
  if (binding.interrupted === true) prompt = "[the previous turn was interrupted before it finished — do not continue or answer that unfinished work; respond only to the newest request below]\n\n" + prompt;
30604
30604
  }
30605
30605
  }
30606
- if (cfg.forwardSystemPrompt) {
30606
+ if (cfg.forwardSystemPrompt || isReview) {
30607
30607
  const sysInstruction = options.system || (() => {
30608
30608
  const sm = messages.find((m) => m.role === "system");
30609
30609
  return sm ? textOf(sm).trim() : "";
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "dsh-agy-link",
3
- "version": "0.4.66",
3
+ "version": "0.4.67",
4
4
  "description": "Google Antigravity (agy CLI) models for DeepSeek Harness — stream Gemini/Claude/GPT-OSS subscriptions into DSH with thinking, tool activity, token usage and in-GUI Google OAuth login.",
5
5
  "type": "module",
6
6
  "license": "MIT",