dsh-approval-review 0.2.1 → 0.2.5

package/README-zh.md

@@ -68,7 +68,7 @@ dsh --profile <profile> --dump-config | grep -A6 'id: approval-review'
 | `reviewer.provider` / `.model` | *(继承)* | 复核路由；不填则继承调用 Agent 自己的路由。会话内可用 `/approval-review model [<provider>/]<id>` 覆盖（**「审批」页签右上角可以直接选**：点开即列出本机配置的模型，候选来自客户端自己的模型目录服务 `modelDirectories`——和 `/model` 选择器、输入框里的模型座位读的是同一份目录。列表由插件自己渲染（原生 `datalist`/`select` 的弹层字号字重无法用 CSS 控制，会显得比页面吵），支持输入过滤、方向键+回车，也可以手打目录里没有的 id）。 |
 | `reviewer.subagentProvider` | `fork` | `mode: subagent` 用的子代理后端（`fork` / `spawn`）。 |
 | `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` | `60000` | 单次复核的硬超时。 |
+| `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」，而是按失败策略走 fail-closed。 |
 | `reviewer.maxTokens` | `1024` | 输出上限。 |
 | `reviewer.temperature` | `0` | 采样温度。 |
 | `reviewer.policyText` | *(内置策略)* | 替换裁决策略正文。 |

package/README.md

@@ -88,7 +88,7 @@ schema defaults.
 | `reviewer.provider` / `.model` | *(inherit)* | Reviewer route; unset inherits the calling agent's own route. |
 | `reviewer.subagentProvider` | `fork` | Subagent backend for `mode: subagent` (`fork` / `spawn`). |
 | `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` | `60000` | Hard deadline for one reviewer call. |
+| `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 fail-closed refusal, not a verdict. |
 | `reviewer.maxTokens` | `1024` | Output cap. |
 | `reviewer.temperature` | `0` | Sampling temperature. |
 | `reviewer.policyText` | *(shipping policy)* | Replaces the ruling policy text. |

package/cordis.patch.yml

@@ -35,7 +35,10 @@
         # Reviewer model, prompt, and size limits. `provider`/`model` unset means
         # the reviewer inherits the calling agent's own route.
         reviewer:
-          timeoutMs: 60000
+          # 120s, not 60s: a reasoning reviewer on a slower route measured 52.5s
+          # here, and a deadline that expires mid-review is not a "no" — it is a
+          # fail-closed refusal under the default `onReviewerFailure: rejected`.
+          timeoutMs: 120000
           maxTokens: 1024
           temperature: 0
           argumentMaxChars: 4000

package/lib/index.js

@@ -2,7 +2,6 @@ import Schema from "@deepseek-ai/schemastery";
 import { z } from "zod";
 import { BlockAssembler, LlmError, createUserMessage } from "@deepseek-ai/dsh-llm";
 import { createHash } from "node:crypto";
-import { assertObjectJsonSchema } from "@deepseek-ai/dsh-tools";
 //#region src/review-types.ts
 /** Every {@link RiskLevel}, least to most dangerous (index is the rank). */
 const RISK_LEVELS = [
@@ -56,7 +55,7 @@ const Config = Schema.object({
 			"glob",
 			"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(6e4).description("Hard deadline for one reviewer call."),
+		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`)."),
 		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."),
@@ -1361,41 +1360,6 @@ var VerdictCache = class {
 };
 //#endregion
 //#region src/subagent-reviewer.ts
-/**
-* The reviewer's requested structured output. An object-rooted schema is the
-* reliable channel: a subagent returns it validated rather than as text this
-* plugin has to salvage.
-*/
-const REVIEWER_OUTPUT_SCHEMA = {
-	type: "object",
-	properties: {
-		decision: {
-			type: "string",
-			enum: [
-				"allow",
-				"deny",
-				"uncertain"
-			]
-		},
-		risk: {
-			type: "string",
-			enum: [
-				"low",
-				"medium",
-				"high",
-				"critical"
-			]
-		},
-		reason: { type: "string" },
-		suggestion: { type: "string" }
-	},
-	required: [
-		"decision",
-		"risk",
-		"reason"
-	],
-	additionalProperties: false
-};
 /** Join text blocks from a child's output, walking nested tool-result blocks. */
 function childText(blocks) {
 	const out = [];
@@ -1440,8 +1404,6 @@ async function runSubagentReviewer(ctx, input) {
 		failure: "cancelled before dispatch",
 		durationMs: 0
 	};
-	const schema = REVIEWER_OUTPUT_SCHEMA;
-	assertObjectJsonSchema(schema);
 	const evidence = buildReviewerUserMessage({
 		toolName: input.evidence.toolName,
 		argumentsText: input.evidence.argumentsText,
@@ -1466,7 +1428,6 @@ async function runSubagentReviewer(ctx, input) {
 			"grep"
 		] },
 		maxDepth: 1,
-		outputSchema: schema,
 		...input.provider === void 0 && input.model === void 0 ? {} : { agentOptions: {
 			...input.provider === void 0 ? {} : { provider: input.provider },
 			...input.model === void 0 ? {} : { model: input.model }
@@ -1555,6 +1516,8 @@ var ReviewRuntime = class {
 	reviewerSessions = /* @__PURE__ */ new Set();
 	/** Latest folded audit state per session, for the live view defaults. */
 	auditStates = /* @__PURE__ */ new WeakMap();
+	/** Sessions whose committed log has already been replayed into that fold. */
+	replayed = /* @__PURE__ */ new WeakSet();
 	/** Reused verdicts for identical actions, when the evidence allows it. */
 	cache;
 	/** Number of verdicts served from the cache since mount. */
@@ -1613,6 +1576,16 @@ var ReviewRuntime = class {
 	*/
 	observeEvent(session, event) {
 		this.sessions.observe(session, event);
+		if (!this.replayed.has(session)) {
+			this.replayed.add(session);
+			let state = initAuditState();
+			for (let seq = 0; seq < session.seq; seq += 1) {
+				const committed = session.eventAt(seq);
+				if (committed !== void 0) state = applyAuditEvent(state, committed, this.config);
+			}
+			this.auditStates.set(session, state);
+			return;
+		}
 		const previous = this.auditStates.get(session) ?? initAuditState();
 		this.auditStates.set(session, applyAuditEvent(previous, event, this.config));
 	}
@@ -1887,7 +1860,7 @@ var ReviewRuntime = class {
 		}
 		this.putRefusal(callId, {
 			marker: formatReviewMarker({
-				reason: verdict?.reason ?? gate.note,
+				reason: clampReason(verdict?.reason ?? (failure === void 0 ? gate.note : `${gate.note} — ${failure}`), this.config.reasonMaxChars),
 				...verdict?.suggestion === void 0 ? {} : { suggestion: verdict.suggestion },
 				...verdict?.risk === void 0 ? {} : { risk: verdict.risk },
 				reviewerRoute: `${route.provider}/${route.model}`,
@@ -2042,6 +2015,21 @@ var ReviewRuntime = class {
 		}
 	}
 };
+/**
+* Bound a reason the plugin is about to publish.
+*
+* The marker rides the tool result, which is model context, so `reasonMaxChars`
+* has to hold here rather than only in the config schema. This is also what
+* keeps a provider's multi-line error digest from swallowing the guidance that
+* follows it.
+* @param reason - the assembled reason.
+* @param max - configured cap.
+* @returns the reason, truncated with an ellipsis when it exceeds the cap.
+*/
+function clampReason(reason, max) {
+	if (reason.length <= max) return reason;
+	return `${reason.slice(0, Math.max(0, max - 1))}…`;
+}
 /** Join the text of a content-block list, walking nested tool-result blocks. */
 function blocksToText(blocks) {
 	const out = [];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-approval-review",
-  "version": "0.2.1",
+  "version": "0.2.5",
   "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",