dsh-approval-review 0.4.0 → 0.4.1

package/README-zh.md

@@ -80,7 +80,7 @@ dsh --profile <profile> --dump-config | grep -A6 'id: approval-review'
 | `reviewer.inspectLocalState` | `true` | 为 direct 模式启用受限本地只读检查。 |
 | `reviewer.tools` | `[read, glob, grep]` | 复核子代理的工具白名单。留空会回退到只读默认，而不是继承父代理的全部工具。 |
 | `reviewer.timeoutMs` | `120000` | 单次复核的硬超时。慢路由 + 推理型复核者实测要 ~50 秒；超时不是「否决」，而是按失败策略处理——出厂设置是转回人工链。 |
-| `reviewer.maxTokens` | `1024` | 输出上限。 |
+| `reviewer.maxTokens` | 模型路由默认值 | 可选输出上限；默认不覆盖适配器或模型配置。 |
 | `reviewer.temperature` | `0` | 采样温度。 |
 | `reviewer.policyText` | *(内置策略)* | 替换裁决策略正文。 |
 | `reviewer.guidance` | *(无)* | 追加在策略之后的部署专属指引。 |

package/README.md

@@ -100,7 +100,7 @@ schema defaults.
 | `reviewer.inspectLocalState` | `true` | Enable the bounded local inspector in direct mode. |
 | `reviewer.tools` | `[read, glob, grep]` | The reviewer child's tool allow-list. An empty list falls back to the read-only default rather than the parent's whole face. |
 | `reviewer.timeoutMs` | `120000` | Hard deadline for one reviewer call. A slow route plus a reasoning reviewer can take ~50s; a deadline that expires mid-review becomes a failure-policy outcome (a delegation to the human by default), not a verdict. |
-| `reviewer.maxTokens` | `1024` | Output cap. |
+| `reviewer.maxTokens` | *(model route default)* | Optional output cap; omitted by default so the adapter/model configuration applies. |
 | `reviewer.temperature` | `0` | Sampling temperature. |
 | `reviewer.policyText` | *(shipping policy)* | Replaces the ruling policy text. |
 | `reviewer.guidance` | *(none)* | Extra deployment guidance appended after the policy. |

package/cordis.patch.yml

@@ -44,7 +44,7 @@
           # here, and a deadline that expires mid-review is not a "no" — it is a
           # human handoff under the default `onReviewerFailure: delegate`.
           timeoutMs: 120000
-          maxTokens: 1024
+          # maxTokens: omit to use the reviewer model route default
           temperature: 0
           argumentMaxChars: 4000
           argumentsBudgetChars: 16000

package/docs/approval-parity.md

@@ -1,4 +1,13 @@
-# 0.4.0 审批策略与验证
+# 审批策略与验证
+
+## 0.4.1 复核续轮修复
+
+- 独立复核器在 `inspect_path` 后保留上一轮的 `source.replayState`，让 pi-ai 适配器恢复原生推理消息及签名，避免续轮丢失服务端要求的推理内容。
+- `reviewer.maxTokens` 改为可选，默认不覆盖模型路由的输出预算；已有显式配置仍生效。未固定为 8192，也未改变推理等级。输出截断、超时与取证次数限制仍按原策略处理。
+- DSH Desktop 2.0.10、DSH 0.1.5-rc.2 的真实 `pi-ai / openai-responses` 路由验证通过：沙箱拒绝后进入复核，独立检查脚本再续轮裁决，排他创建一次性探针并读回成功。该次复核耗时 5655 ms，未出现缺失推理内容错误。这是一次实测结果，不代表所有模型路由均已验证。
+- 回归覆盖原生 replay 元数据续传、默认不传输出上限、显式上限保留，以及截断答案不得被当作有效裁决；发布检查为 14 个测试文件、299 项测试通过，类型检查和构建通过。
+
+## 0.4.0 审批策略
 
 本版合并审批页签语言切换、裁决语言选择和审批策略改动。默认使用独立模型复核，保留 DSH 的原始沙箱边界，减少常规操作的人工审批，同时对高风险操作单独校验授权和范围。
 

package/lib/index.d.ts

@@ -67,8 +67,8 @@ interface ReviewerConfig {
   readonly tools: string[];
   /** Hard deadline for one reviewer call. */
   readonly timeoutMs: number;
-  /** Output-token cap for one reviewer call. */
-  readonly maxTokens: number;
+  /** Optional output-token cap; omission uses the adapter/model route default. */
+  readonly maxTokens?: number;
   /** Sampling temperature; the reviewer should be near-deterministic. */
   readonly temperature: number;
   /** Ruling policy appended to the reviewer prompt (Codex-style policy text). */

package/lib/index.js

@@ -57,7 +57,7 @@ const Config = Schema.object({
 			"grep"
 		]).description("The reviewer child's tool allow-list. An empty list falls back to the read-only default rather than the parent's whole face."),
 		timeoutMs: Schema.number().step(1).min(1e3).default(12e4).description("Hard deadline for one reviewer call. A reasoning reviewer on a slow route can take tens of seconds; a deadline that is too tight turns into a fail-closed refusal (the default failure policy) rather than a verdict."),
-		maxTokens: Schema.number().step(1).min(64).default(1024).description("Output-token cap for one reviewer call (`mode: direct`)."),
+		maxTokens: Schema.number().step(1).min(64).description("Optional output-token cap for direct review; omit to use the adapter/model route default."),
 		temperature: Schema.number().min(0).max(2).default(0).description("Sampling temperature; 0 keeps the reviewer near-deterministic."),
 		policyText: Schema.string().description("Ruling policy appended to the reviewer prompt."),
 		guidance: Schema.string().description("Extra deployment-specific reviewer guidance."),
@@ -1193,8 +1193,8 @@ async function runReviewerCall(ctx, route, system, message, limits) {
 					messages: [...messages],
 					system,
 					temperature: limits.temperature,
-					maxTokens: limits.maxTokens,
 					signal: controller.signal,
+					...limits.maxTokens === void 0 ? {} : { maxTokens: limits.maxTokens },
 					...limits.inspect === void 0 ? {} : { tools: [INSPECTION_TOOL] },
 					...limits.sessionId === void 0 ? {} : { sessionId: limits.sessionId }
 				};
@@ -1204,11 +1204,13 @@ async function runReviewerCall(ctx, route, system, message, limits) {
 				if (calls.length === 0) return assembler;
 				if (assembler.finish.kind === "max-tokens" || assembler.finish.kind === "error" || assembler.finish.kind === "aborted") return assembler;
 				if (limits.inspect === void 0 || inspections + calls.length > 4) throw new Error("read-only investigation budget exhausted");
+				const replayState = assembler.replayState;
 				messages.push(createAssistantMessage({
 					content: blocks,
 					source: {
 						provider: route.provider,
-						model: route.model
+						model: route.model,
+						...replayState === void 0 ? {} : { replayState }
 					}
 				}));
 				for (const call of calls) {
@@ -2117,7 +2119,7 @@ var ReviewRuntime = class {
 			userIntent,
 			...req.reason === void 0 ? {} : { askReason: req.reason }
 		}), {
-			maxTokens: this.config.reviewer.maxTokens,
+			...this.config.reviewer.maxTokens === void 0 ? {} : { maxTokens: this.config.reviewer.maxTokens },
 			temperature: this.config.reviewer.temperature,
 			timeoutMs: this.config.reviewer.timeoutMs,
 			...req.signal === void 0 ? {} : { signal: req.signal },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-approval-review",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "description": "Codex-style agent auto-approval for DeepSeek Harness: an independent reviewer model decides allow/deny on the approval answerer chain, fail-closed, with a per-decision rationale — refusals and allows alike — in a dedicated Approvals tab. · DSH 插件：Codex 风格的 Agent 自动审批，独立 reviewer 模型裁决、fail-closed、每次审批（放行与否决）都留下详细理由，并在独立的「审批」页签中展示。",
   "type": "module",
   "main": "./lib/index.js",