dsh-decision-layer 0.5.3 → 0.5.4

package/README.en.md

@@ -4,7 +4,7 @@
 
 `dsh-decision-layer` adds structured, explicitly scoped decisions to the DeepSeek Harness (DSH) agent loop, backed by a pluggable adjudication backend. Dangerous-call gating, output self-check, tool narrowing, and loop prevention hook into the agent loop as automatic interventions.
 
-> **0.5.3 is a preview release.** The plugin works, but how much real value the adjudication interventions add is still being explored — every threshold is an unvalidated heuristic default. Treat it as something to observe, not a proven guarantee of impact. See [ROADMAP.md](<ROADMAP.md>) for scope and progress.
+> **0.5.4 is a preview release.** The plugin works, but how much real value the adjudication interventions add is still being explored — every threshold is an unvalidated heuristic default. Treat it as something to observe, not a proven guarantee of impact. See [ROADMAP.md](<ROADMAP.md>) for scope and progress.
 
 ## Installation
 
@@ -49,4 +49,4 @@ Restart `dsh web` and refresh after installing or updating: the host and client
 
 ## Safety boundary and development
 
-Automatic intervention only subtracts: the gate only denies, narrowing only removes tools, and neither grants new permissions. The gate defers to native DSH approval on backend failure rather than allowing; self-check, narrowing, and loop-guard errors skip intervention. The redacted summary sent to the backend carries only tool name and command/path with secret redaction — never file content. Node.js `>=20.11`; `npm test` and `npm run build:client` run local checks. Rebuild and commit `lib/client.js` with any client change. 0.5.3 is a preview: the interventions work, but their real value and thresholds are still being explored — feedback on misjudgments and feel observed in the decision log is welcome.
+Automatic intervention only subtracts: the gate only denies, narrowing only removes tools, and neither grants new permissions. The gate defers to native DSH approval on backend failure rather than allowing; self-check, narrowing, and loop-guard errors skip intervention. The redacted summary sent to the backend carries only tool name and command/path with secret redaction — never file content. Node.js `>=20.11`; `npm test` and `npm run build:client` run local checks. Rebuild and commit `lib/client.js` with any client change. 0.5.4 is a preview: the interventions work, but their real value and thresholds are still being explored — feedback on misjudgments and feel observed in the decision log is welcome.

package/README.md

@@ -4,7 +4,7 @@
 
 DeepSeek Harness（DSH）的结构化裁决插件。它把有明确判断标准的问题交给同协议的裁决后端，供 agent 在需要时使用，并在 agent loop 上接入危险门控、产出自检、工具收窄和防循环等自动介入。
 
-> **0.5.3 是体验版。** 插件功能已可用，但裁决介入能带来多少实际价值仍在探索中——各决策点的阈值都是未经验证的启发式默认值，建议先以观察心态体验，别当成经过验证的效果保证。功能与后续进度以 [ROADMAP.md](<ROADMAP.md>) 为准。
+> **0.5.4 是体验版。** 插件功能已可用，但裁决介入能带来多少实际价值仍在探索中——各决策点的阈值都是未经验证的启发式默认值，建议先以观察心态体验，别当成经过验证的效果保证。功能与后续进度以 [ROADMAP.md](<ROADMAP.md>) 为准。
 
 ## 安装
 
@@ -53,4 +53,4 @@ dsh plugin --profile web add file:/绝对路径/dsh-decision-layer
 
 ## 开发
 
-服务端使用 Node.js `>=20.11`（`.mjs`，无需构建）；客户端用 TypeScript + esbuild 构建。`npm test` 跑本地测试，`npm run build:client` 生成 `lib/client.js`。改动客户端后需重建并连同 `lib/client.js` 一起提交，否则 GitHub 安装拿到的是旧 bundle。0.5.3 为体验版：介入能力已可用，但实际价值与阈值仍在探索，欢迎反馈决策日志观察到的误判与体感。
+服务端使用 Node.js `>=20.11`（`.mjs`，无需构建）；客户端用 TypeScript + esbuild 构建。`npm test` 跑本地测试，`npm run build:client` 生成 `lib/client.js`。改动客户端后需重建并连同 `lib/client.js` 一起提交，否则 GitHub 安装拿到的是旧 bundle。0.5.4 为体验版：介入能力已可用，但实际价值与阈值仍在探索，欢迎反馈决策日志观察到的误判与体感。

package/lib/client.js

@@ -96,6 +96,9 @@ var dictionaries = {
     gatePassThrough: "\u672A\u5E72\u9884\uFF08\u6309\u539F\u8DEF\u5F84\u6267\u884C\uFF09",
     checkResultLow: "\u5F97\u5206\u504F\u4F4E",
     checkResultOk: "\u901A\u8FC7",
+    checkResultLowNoisy: "\u5F97\u5206\u504F\u4F4E\uFF08\u4F4E\u7F6E\u4FE1\uFF09",
+    checkNoiseTag: "\u566A\u58F0\uFF0C\u53EF\u5FFD\u7565",
+    checkNoiseHint: "\u6A21\u578B\u5BF9\u8FD9\u6B21\u6253\u5206\u51E0\u4E4E\u6CA1\u6709\u628A\u63E1\uFF08\u7F6E\u4FE1\u5EA6\u6781\u4F4E\uFF09\uFF0C\u8FD9\u6761\u4F4E\u5206\u4E0D\u8BA1\u5165\u8D28\u91CF\u8D8B\u52BF\u3001\u4E5F\u4E0D\u89E6\u53D1\u4EFB\u4F55\u8865\u5B8C\uFF0C\u89C6\u4F5C\u566A\u58F0\u5FFD\u7565\u5373\u53EF\u3002\u771F\u6B63\u503C\u5F97\u5173\u6CE8\u7684\u662F\u8FDE\u7EED\u591A\u6B21\u9AD8\u7F6E\u4FE1\u7684\u4F4E\u5206\u3002",
     reasonUnavailable: "\u540E\u7AEF\u4E0D\u53EF\u8FBE",
     reasonHttpError: "\u540E\u7AEF\u8FD4\u56DE\u9519\u8BEF",
     reasonInvalidResponse: "\u54CD\u5E94\u683C\u5F0F\u4E0D\u7B26",
@@ -245,6 +248,9 @@ var dictionaries = {
     gatePassThrough: "no intervention (proceeded on its original path)",
     checkResultLow: "low score",
     checkResultOk: "passed",
+    checkResultLowNoisy: "low score (low confidence)",
+    checkNoiseTag: "noise, ignore",
+    checkNoiseHint: "The model had almost no confidence in this score (near-zero confidence). This low score does not count toward the quality trend and triggers no re-prompt \u2014 treat it as noise. What matters is several consecutive high-confidence low scores.",
     reasonUnavailable: "backend unreachable",
     reasonHttpError: "backend returned an error",
     reasonInvalidResponse: "unexpected response shape",
@@ -422,24 +428,33 @@ function renderCheck({ t }, entry) {
       entry.reason ? ` \xB7 ${failDetail(t, entry)}` : ""
     ] })
   ] });
-  const result = entry.outcome === "low" ? t("checkResultLow") : t("checkResultOk");
+  const noisy = entry.outcome === "low" && entry.lowConfidence === true;
+  const result = entry.outcome === "low" ? noisy ? t("checkResultLowNoisy") : t("checkResultLow") : t("checkResultOk");
+  const verdict = /* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", { className: noisy ? "decision-log-muted-verdict" : "decision-log-verdict", children: result });
+  const score = entry.score !== void 0 ? /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("span", { className: "decision-log-detail", title: t("logScoreHint"), children: [
+    " \xB7 ",
+    t("logScore"),
+    " ",
+    entry.score,
+    "/2"
+  ] }) : null;
+  const conf = typeof entry.confidence === "number" ? /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("span", { className: "decision-log-detail", children: [
+    " \xB7 ",
+    t("logConfidence"),
+    " ",
+    entry.confidence
+  ] }) : null;
+  const hint = noisy ? /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("span", { className: "decision-log-detail", title: t("checkNoiseHint"), children: [
+    " \xB7 ",
+    t("checkNoiseTag")
+  ] }) : null;
   return /* @__PURE__ */ (0, import_jsx_runtime.jsxs)(import_jsx_runtime.Fragment, { children: [
     label,
     " \xB7 ",
-    /* @__PURE__ */ (0, import_jsx_runtime.jsx)("span", { className: "decision-log-verdict", children: result }),
-    entry.score !== void 0 ? /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("span", { className: "decision-log-detail", title: t("logScoreHint"), children: [
-      " \xB7 ",
-      t("logScore"),
-      " ",
-      entry.score,
-      "/2"
-    ] }) : null,
-    typeof entry.confidence === "number" ? /* @__PURE__ */ (0, import_jsx_runtime.jsxs)("span", { className: "decision-log-detail", children: [
-      " \xB7 ",
-      t("logConfidence"),
-      " ",
-      entry.confidence
-    ] }) : null
+    verdict,
+    score,
+    conf,
+    hint
   ] });
 }
 function renderNarrow({ t }, entry) {
@@ -538,7 +553,10 @@ function renderComplete({ t }, entry) {
 function outcomeTagOf(entry) {
   if (entry.kind === "narrow") return entry.outcome === "error" ? "error" : "ok";
   if (entry.kind === "complete") return entry.outcome === "error" ? "error" : entry.outcome === "unsatisfied" ? "warn" : "ok";
-  return entry.outcome === "deny" ? "deny" : entry.outcome === "error" ? "error" : entry.outcome === "low" ? "warn" : "ok";
+  if (entry.outcome === "deny") return "deny";
+  if (entry.outcome === "error") return "error";
+  if (entry.outcome === "low") return entry.kind === "check" && entry.lowConfidence === true ? "muted" : "warn";
+  return "ok";
 }
 function renderBody(props, entry) {
   return entry.kind === "gate" ? renderGate(props, entry) : entry.kind === "narrow" ? renderNarrow(props, entry) : entry.kind === "complete" ? renderComplete(props, entry) : renderCheck(props, entry);
@@ -1538,6 +1556,7 @@ var css = `
 .decision-log-item.decision-log-deny{border-left-color:#c0392b}
 .decision-log-item.decision-log-error,.decision-log-item.decision-log-warn{border-left-color:#c98a1e}
 .decision-log-item.decision-log-ok{border-left-color:#1f8a70}
+.decision-log-item.decision-log-muted{border-left-color:var(--dsw-alias-border-l2,#dce1e6)}
 .decision-log-time{color:var(--dsw-alias-label-tertiary,#77818b);font-variant-numeric:tabular-nums;flex:0 0 auto}
 .decision-log-body{color:var(--dsw-alias-label-primary,#202124);flex:1 1 auto;min-width:0}
 .decision-rate{flex:0 0 auto;display:inline-flex;gap:1px;align-self:center}
@@ -1554,6 +1573,11 @@ var css = `
 .decision-log-ok .decision-log-verdict{color:#1f8a70}
 .decision-log-warn .decision-log-verdict{color:#c98a1e}
 .decision-log-detail{color:var(--dsw-alias-label-secondary,#59636e)}
+/* A low-confidence low score is noise: its verdict is shown muted (not the amber
+   warn accent) and a "noise, ignore" tag hangs off the row so it never reads as a
+   real quality dip. */
+.decision-log-muted-verdict{font-weight:600;color:var(--dsw-alias-label-tertiary,#9aa5b1)}
+.decision-log-muted .decision-log-time{opacity:.75}
 
 .decision-status{display:flex;align-items:center;gap:.5rem;margin-top:.7rem;font-size:.82rem}
 .decision-status-dot{width:8px;height:8px;border-radius:50%;flex:0 0 auto;background:#9aa5b1}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-decision-layer",
-  "version": "0.5.3",
+  "version": "0.5.4",
   "description": "A structured decision layer for the DeepSeek Harness agent loop: a danger-call gate, output self-check, tool narrowing, and loop guard backed by a pluggable adjudication model.",
   "keywords": [
     "dsh",

package/service/self-check.mjs

@@ -80,6 +80,10 @@ export function createSelfCheck({ evaluate, sessions, steer, settings, minConfid
       // The trend run advances over every evaluated score; a high-confidence low
       // score is the only thing that grows it. Thresholds flow to session-state.
       const trend = { confident: lowScore && confidence !== undefined && confidence >= trendMinConfidence, trendRun, trendSevereRun };
+      // A low score the model is not confident about (confidence present but below
+      // the trend threshold) does not advance the quality-trend judgment. Flag it
+      // so the decision log can mark it as low-signal noise instead of a real dip.
+      if (lowScore && confidence !== undefined && !trend.confident) detail.lowConfidence = true;
       record(turn, lowScore ? 'low' : 'ok', { score: answer.score, ...detail }, trend);
       if (!lowScore) return { evaluated: true, lowScore: false, steered: false, score: answer.score };
       const confident = confidence !== undefined && confidence >= minConfidence;

package/service/session-state.mjs

@@ -131,6 +131,7 @@ export function createSessionState({ maxSessions = 1024, maxLogEntries = 50, onL
       if (Number.isSafeInteger(entry.unsatisfied) && entry.unsatisfied >= 0) record.unsatisfied = entry.unsatisfied;
       if (Number.isSafeInteger(entry.insufficient) && entry.insufficient >= 0) record.insufficient = entry.insufficient;
       if (entry.steered === true) record.steered = true;
+      if (entry.lowConfidence === true) record.lowConfidence = true;
       if (Array.isArray(entry.tools)) {
         const names = entry.tools.filter(name => typeof name === 'string' && name.trim()).slice(0, 12).map(name => name.trim().slice(0, 64));
         if (names.length > 0) record.tools = names;