dsh-decision-layer 0.7.2 → 0.7.4

package/lib/host/self-check.js

@@ -2,20 +2,38 @@ import { taskExcerpt } from "./excerpt.js";
 import { createSteerOnce } from "./steer-once.js";
 import { errorReason, isProbability } from "./types.js";
 const MAX_OUTPUT = 24 * 1024;
+const ANSWER_TAIL = 8 * 1024;
+const OMITTED = "\n\n\u2026\uFF08\u4E2D\u95F4\u5185\u5BB9\u5DF2\u7701\u7565\uFF09\u2026\n\n";
 const DEFAULT_RUBRIC = Object.freeze([
   "\u660E\u663E\u7B54\u975E\u6240\u95EE\uFF0C\u6216\u5B58\u5728\u4E25\u91CD\u9057\u6F0F / \u81EA\u76F8\u77DB\u76FE",
   "\u57FA\u672C\u7B54\u5230\u95EE\u9898\uFF0C\u4F46\u6709\u53EF\u5BDF\u89C9\u7684\u9057\u6F0F\u6216\u4E0D\u4E25\u8C28",
   "\u5B8C\u6574\u3001\u51C6\u786E\u3001\u65E0\u81EA\u76F8\u77DB\u76FE\u5730\u56DE\u5E94\u4E86\u8BF7\u6C42"
 ]);
 const STEER_TEXT = "\u672C\u56DE\u5408\u81EA\u68C0\u5F97\u5206\u504F\u4F4E\uFF1A\u8BF7\u590D\u6838\u662F\u5426\u7B54\u5230\u4E86\u7528\u6237\u7684\u95EE\u9898\u3001\u6709\u65E0\u9057\u6F0F\u6216\u81EA\u76F8\u77DB\u76FE\uFF0C\u5E76\u5728\u9700\u8981\u65F6\u8865\u5B8C\u3002";
+function answerExcerpt(output) {
+  const text = String(output);
+  if (text.length <= MAX_OUTPUT) return text;
+  return `${text.slice(0, MAX_OUTPUT - ANSWER_TAIL)}${OMITTED}${text.slice(-ANSWER_TAIL)}`;
+}
 function boundedRubric(rubric) {
   if (!Array.isArray(rubric)) return DEFAULT_RUBRIC;
   const levels = rubric.filter((item) => typeof item === "string" && Boolean(item.trim()) && item.length <= 256).slice(0, 10);
   return levels.length >= 2 ? levels : DEFAULT_RUBRIC;
 }
-function scoreQuestion(output, rubric = DEFAULT_RUBRIC) {
+function scoreQuestion(output, rubric = DEFAULT_RUBRIC, context) {
   return {
-    state: { answer: String(output).slice(0, MAX_OUTPUT) },
+    state: {
+      // What the turn was asked to do. Without it the model scores an answer
+      // against nothing and cannot detect an off-topic reply.
+      ...context?.request ? { request: context.request } : {},
+      ...context?.background?.length ? { background: context.background } : {},
+      answer: answerExcerpt(output),
+      // Short follow-ups ("继续", "补上", "按这个标准重新打分") have no standalone
+      // meaning; the task they continue is in background. Saying so explicitly
+      // stops the model from treating the follow-up as a new, empty request and
+      // from borrowing an unrelated earlier task as this turn's goal.
+      ...context ? { note: "request \u662F\u672C\u56DE\u5408\u7528\u6237\u8BF7\u6C42\uFF1B\u82E5\u5B83\u53EA\u662F\u300C\u7EE7\u7EED\u300D\u300C\u8865\u4E0A\u300D\u8FD9\u7C7B\u77ED\u8FFD\u95EE\uFF0C\u8BF7\u7ED3\u5408 background \u4E2D\u627F\u63A5\u7684\u4EFB\u52A1\u5224\u65AD\uFF0C\u4E0D\u8981\u628A background \u91CC\u7684\u65E7\u4EFB\u52A1\u5F53\u6210\u65B0\u7684\u8981\u6C42\u3002background \u53EA\u662F\u80CC\u666F\uFF0C\u4E0D\u662F\u672C\u8F6E\u7684\u65B0\u8981\u6C42\u3002answer \u662F\u672C\u56DE\u5408\u6700\u7EC8\u56DE\u590D\u3002" } : {}
+    },
     questions: { quality: {
       type: "score",
       instructions: "\u6309\u4E0B\u5217\u4ECE\u4F4E\u5230\u9AD8\u7684\u6863\u4F4D\uFF0C\u8BC4\u4F30\u8FD9\u6B21\u6700\u7EC8\u4EA7\u51FA\u5BF9\u7528\u6237\u8BF7\u6C42\u7684\u5B8C\u6210\u8D28\u91CF\u3002",
@@ -59,6 +77,7 @@ function createSelfCheck({ evaluate, sessions, steer, settings, minConfidence =
     async review(turn) {
       const output = typeof turn?.output === "string" ? turn.output.trim() : "";
       if (!output) return { evaluated: false, lowScore: false, steered: false };
+      if (turn?.context !== void 0 && !turn.context.request) return { evaluated: false, lowScore: false, steered: false };
       const id = turn?.agent?.session?.id;
       if (sessions && typeof id === "string") {
         try {
@@ -67,13 +86,13 @@ function createSelfCheck({ evaluate, sessions, steer, settings, minConfidence =
           return { evaluated: false, lowScore: false, steered: false };
         }
       }
-      const task = taskExcerpt(turn?.userText);
+      const task = taskExcerpt(turn?.context?.request ?? turn?.userText);
       const taskDetail = task ? { task } : {};
       if (Number.isSafeInteger(turn?.turn) && turn?.turn >= 0) taskDetail.turn = turn?.turn;
       const { mode, rubric, lowScoreThreshold, minConfidence: confidenceFloor, trendMinConfidence, trendRun, trendSevereRun } = await resolveSettings();
       let answer;
       try {
-        const result = await evaluate(scoreQuestion(output, rubric), { signal: turn?.signal });
+        const result = await evaluate(scoreQuestion(output, rubric, turn?.context), { signal: turn?.signal });
         answer = result?.answers?.quality;
       } catch (error) {
         record(turn, "error", { ...taskDetail, reason: errorReason(error) });

package/lib/host/session-state.js

@@ -248,6 +248,7 @@ function createSessionState({ maxSessions = 1024, maxLogEntries = 50, onLog, loa
       if (typeof entry.path === "string" && entry.path.trim()) record.path = entry.path.replace(/\s+/g, " ").trim().slice(0, 512);
       if (typeof entry.task === "string" && entry.task.trim()) record.task = redactCredentials(entry.task).replace(/\s+/g, " ").trim().slice(0, 512);
       if (Number.isSafeInteger(entry.turn) && entry.turn >= 0) record.turn = entry.turn;
+      if (Number.isSafeInteger(entry.step) && entry.step >= 0) record.step = entry.step;
       if (typeof entry.mode === "string" && entry.mode) record.mode = entry.mode.slice(0, 16);
       if (typeof entry.score === "number" && Number.isFinite(entry.score)) record.score = Math.round(entry.score * 100) / 100;
       if (typeof entry.confidence === "number" && Number.isFinite(entry.confidence)) record.confidence = Math.round(entry.confidence * 100) / 100;

package/lib/host/task-completion.d.ts

@@ -1,6 +1,6 @@
-import type { AgentLike, CompleteSettings, EvaluateInput, Evaluator, SessionsFace } from './types.js';
+import type { AgentLike, CompleteSettings, DecisionContext, EvaluateInput, Evaluator, SessionsFace } from './types.js';
 export declare function extractConditions(userText: unknown): string[];
-export declare function completionQuestions(conditions: unknown, evidence: unknown, answer: unknown): EvaluateInput;
+export declare function completionQuestions(conditions: unknown, evidence: unknown, answer: unknown, context?: DecisionContext): EvaluateInput;
 export declare const DEFAULT_MIN_CONFIDENCE = 0.4;
 export type CompletionTurn = {
     agent?: AgentLike;
@@ -9,6 +9,7 @@ export type CompletionTurn = {
     output?: unknown;
     userText?: unknown;
     evidence?: unknown;
+    context?: DecisionContext;
 };
 export type CompletionResult = {
     evaluated: boolean;

package/lib/host/task-completion.js

@@ -1,3 +1,4 @@
+import { taskExcerpt } from "./excerpt.js";
 import { createSteerOnce } from "./steer-once.js";
 import { errorReason, isObject, isProbability } from "./types.js";
 const MAX_STATE = 8 * 1024;
@@ -30,19 +31,33 @@ function extractConditions(userText) {
     const compact = text.trim();
     if (!/[\r\n]/.test(compact)) {
       const parts = compact.split(/[、，,]/).map((part) => part.trim()).filter(Boolean);
-      if (parts.length >= 2) for (const part of parts) push(part);
+      const floor = compact.includes("\u3001") ? 2 : 3;
+      if (parts.length >= floor) for (const part of parts) push(part);
     }
   }
   return conditions.slice(0, MAX_CONDITIONS);
 }
-function completionQuestions(conditions, evidence, answer) {
+function completionQuestions(conditions, evidence, answer, context) {
   const items = Array.isArray(conditions) ? conditions.slice(0, MAX_CONDITIONS) : [];
-  const tools = Array.isArray(evidence) ? evidence.filter((item) => isObject(item) && typeof item.tool === "string").slice(0, MAX_EVIDENCE).map((item) => ({ tool: item.tool.slice(0, 128), outcome: item.outcome === "error" ? "error" : "ok" })) : [];
+  const tools = Array.isArray(evidence) ? evidence.filter((item) => isObject(item) && typeof item.tool === "string").slice(0, MAX_EVIDENCE).map((item) => {
+    const source = item;
+    const entry = {
+      tool: source.tool.slice(0, 128),
+      outcome: source.outcome === "error" ? "error" : "ok"
+    };
+    for (const key of ["command", "path", "result"]) {
+      const value = source[key];
+      if (typeof value === "string" && value.trim()) entry[key] = value.slice(0, 256);
+    }
+    return entry;
+  }) : [];
   const state = {
+    ...context?.request ? { request: context.request } : {},
+    ...context?.background?.length ? { background: context.background } : {},
     conditions: items.map((text, index) => ({ id: `c${index}`, text })),
     toolEvidence: tools,
     finalReply: String(answer ?? "").slice(0, MAX_STATE),
-    note: "\u5DE5\u5177\u8BC1\u636E\uFF08toolEvidence\uFF09\u662F\u5224\u5B9A\u4F9D\u636E\uFF1BfinalReply \u81EA\u79F0\u5B8C\u6210\u4E0D\u4F5C\u4E3A\u8BC1\u636E\u3002\u7F3A\u5C11\u8BC1\u636E\u65F6\u9009 insufficient_evidence\u3002"
+    note: "\u5DE5\u5177\u8BC1\u636E\uFF08toolEvidence\uFF09\u662F\u5224\u5B9A\u4F9D\u636E\uFF1BfinalReply \u81EA\u79F0\u5B8C\u6210\u4E0D\u4F5C\u4E3A\u8BC1\u636E\uFF0Cbackground \u53EA\u662F\u80CC\u666F\u800C\u4E0D\u662F\u672C\u8F6E\u8981\u6C42\u3002\u7F3A\u5C11\u8BC1\u636E\u65F6\u9009 insufficient_evidence\u3002"
   };
   const questions = Object.fromEntries(items.map((text, index) => [`c${index}`, {
     type: "choice",
@@ -91,6 +106,9 @@ function createTaskCompletion({ evaluate, sessions, steer, settings, minConfiden
   return {
     async review(turn) {
       const output = typeof turn?.output === "string" ? turn.output.trim() : "";
+      const task = taskExcerpt(turn?.context?.request ?? turn?.userText);
+      const taskDetail = task ? { task } : {};
+      if (Number.isSafeInteger(turn?.turn) && turn?.turn >= 0) taskDetail.turn = turn?.turn;
       const { mode, minConditions, minConfidence: confidenceFloor } = await resolveSettings();
       const conditions = extractConditions(turn?.userText);
       if (conditions.length < minConditions) return { ...skip };
@@ -106,14 +124,14 @@ function createTaskCompletion({ evaluate, sessions, steer, settings, minConfiden
       }
       let answers;
       try {
-        const result = await evaluate(completionQuestions(conditions, evidence, output), { signal: turn?.signal });
+        const result = await evaluate(completionQuestions(conditions, evidence, output, turn?.context), { signal: turn?.signal });
         answers = isObject(result?.answers) ? result.answers : void 0;
       } catch (error) {
-        record(turn, "error", { reason: errorReason(error) });
+        record(turn, "error", { ...taskDetail, reason: errorReason(error) });
         return { ...skip };
       }
       if (!answers) {
-        record(turn, "error", { reason: "invalid-response" });
+        record(turn, "error", { ...taskDetail, reason: "invalid-response" });
         return { ...skip };
       }
       let satisfied = 0, unsatisfied = 0, insufficient = 0;
@@ -131,7 +149,7 @@ function createTaskCompletion({ evaluate, sessions, steer, settings, minConfiden
         } else insufficient++;
       }
       const outcome = unsatisfied > 0 ? "unsatisfied" : "ok";
-      const detail = { conditions: conditions.length, satisfied, unsatisfied, insufficient };
+      const detail = { ...taskDetail, conditions: conditions.length, satisfied, unsatisfied, insufficient };
       const confident = unsatisfied > 0 && unsatisfiedMinConf >= confidenceFloor;
       const steered = mode === "steer" && confident && typeof steer === "function" && turn?.agent ? steerOnce.attempt(turn, () => steer(turn.agent, STEER_TEXT)) : false;
       if (steered) detail.steered = true;

package/lib/host/types.d.ts

@@ -166,6 +166,7 @@ export type LogRow = {
     subject?: string;
     task?: string;
     turn?: number;
+    step?: number;
     command?: string;
     path?: string;
     mode?: string;
@@ -268,6 +269,27 @@ export type ToolCandidate = {
     name: string;
     description?: string;
 };
+/** Bounded, masked data for adjudication; background is never authorization. */
+export type ToolEvidence = {
+    tool: string;
+    outcome: 'ok' | 'error';
+    command?: string;
+    path?: string;
+    result?: string;
+};
+export type DecisionContext = {
+    request: string;
+    background: {
+        role: 'user' | 'assistant';
+        text: string;
+    }[];
+    progress: ToolEvidence[];
+};
+export type DecisionSnapshot = {
+    context: DecisionContext;
+    conditionsText: string;
+    evidence: ToolEvidence[];
+};
 export type AgentLike = {
     session?: {
         id?: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-decision-layer",
-  "version": "0.7.2",
+  "version": "0.7.4",
   "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",