@hmharness/agent 0.14.17 → 0.14.19

package/dist/i18n.d.ts

@@ -74,6 +74,7 @@ export interface Strings {
     cmdNewDone: string;
     cmdFork: string;
     cmdCopy: string;
+    cmdExport: string;
     cmdMouse: string;
     cmdMouseOn: string;
     cmdMouseOff: string;

package/dist/i18n.js

@@ -58,6 +58,7 @@ const zh = {
     cmdNewDone: '✓ 新会话开始(旧转录可用 Ctrl+T 回看)',
     cmdFork: 'fork 当前会话为新线程(继承全部上下文)',
     cmdCopy: '复制最后一条 AI 输出到剪贴板',
+    cmdExport: '一键导出整个会话为 Markdown 文件（~/.hmharness/exports/）',
     cmdMouse: '切换全时鼠标捕获（仅无滚轮翻译的终端需要；默认 ?1007 已覆盖大多数终端）',
     cmdMouseOn: '🖱 全时鼠标捕获已开启（滚轮直读；原生划选复制请按住 Shift 再拖动）',
     cmdMouseOff: '🖱 全时鼠标捕获已关闭（滚轮走终端方向键翻译；划选复制恢复原生）',
@@ -207,6 +208,7 @@ const en = {
     cmdNewDone: '✓ new session (old transcript: Ctrl+T to replay)',
     cmdFork: 'fork this session to a new thread (inherits full context)',
     cmdCopy: 'copy the last AI output to the clipboard',
+    cmdExport: 'export the whole session to a Markdown file (~/.hmharness/exports/)',
     cmdMouse: 'toggle full-time mouse capture (only for terminals without wheel translation; ?1007 covers most)',
     cmdMouseOn: '🖱 full-time mouse capture ON (direct wheel; hold Shift to select natively)',
     cmdMouseOff: '🖱 full-time mouse capture OFF (wheel via terminal translation; native selection restored)',

package/dist/index.d.ts

@@ -9,3 +9,4 @@ export { buildRegistry, contextPack, makeApproval, nativeRegistry, runAgentTask,
 export * from './skills-conformance.ts';
 export * from './acp.ts';
 export * from './topology.ts';
+export * from './meta-harness.ts';

package/dist/index.js

@@ -9,3 +9,4 @@ export { buildRegistry, contextPack, makeApproval, nativeRegistry, runAgentTask,
 export * from "./skills-conformance.js";
 export * from "./acp.js";
 export * from "./topology.js";
+export * from "./meta-harness.js";

package/dist/meta-harness.d.ts

@@ -0,0 +1,72 @@
+/**
+ * @hmharness/agent - Harness-of-Harness (P3-03)
+ *
+ * The audit: "一个 meta-agent 专门设计候选 Harness，再由 independent evaluator 验收"
+ *
+ * A meta-agent that designs candidate harness configurations (tool sets,
+ * prompt strategies, agent topologies) which are then validated by an
+ * independent evaluator. This is the self-referential optimization loop.
+ */
+export interface HarnessCandidate {
+    id: string;
+    name: string;
+    description: string;
+    /** what tools are available */
+    tools: string[];
+    /** system prompt template */
+    systemPrompt: string;
+    /** agent topology (single/dual/team/pipeline) */
+    topology: 'single' | 'dual' | 'team' | 'pipeline';
+    /** max turns budget */
+    maxTurns: number;
+    /** approval policy */
+    approvalPolicy: 'ask' | 'auto' | 'yolo';
+    /** generation metadata */
+    generatedBy: string;
+    generation: number;
+}
+export interface HarnessEvaluation {
+    candidateId: string;
+    /** independent evaluator's score [0,1] */
+    score: number;
+    /** specific strengths */
+    strengths: string[];
+    /** specific weaknesses */
+    weaknesses: string[];
+    /** recommended action */
+    recommendation: 'adopt' | 'reject' | 'iterate';
+    evaluatorId: string;
+    evaluatedAt: string;
+}
+export interface MetaHarnessIteration {
+    iteration: number;
+    candidates: HarnessCandidate[];
+    evaluations: HarnessEvaluation[];
+    bestCandidate?: HarnessCandidate;
+    improvement: number;
+}
+/**
+ * Generate a candidate harness variation from a base.
+ * Pure - testable.
+ */
+export declare function generateCandidate(base: HarnessCandidate, mutation: Partial<HarnessCandidate>, generation: number): HarnessCandidate;
+/**
+ * Evaluate a candidate (simulated independent evaluator).
+ * Pure - testable.
+ */
+export declare function evaluateCandidate(candidate: HarnessCandidate, metrics: {
+    passRate: number;
+    avgTurns: number;
+    avgTokens: number;
+    safetyScore: number;
+}): HarnessEvaluation;
+/**
+ * Select the best candidate from evaluations.
+ * Pure - testable.
+ */
+export declare function selectBest(candidates: HarnessCandidate[], evaluations: HarnessEvaluation[]): HarnessCandidate | undefined;
+/**
+ * Run one iteration of the meta-harness loop.
+ * Pure - testable.
+ */
+export declare function runIteration(iteration: number, candidates: HarnessCandidate[], allEvaluations: HarnessEvaluation[], previousBestScore: number): MetaHarnessIteration;

package/dist/meta-harness.js

@@ -0,0 +1,79 @@
+/**
+ * @hmharness/agent - Harness-of-Harness (P3-03)
+ *
+ * The audit: "一个 meta-agent 专门设计候选 Harness，再由 independent evaluator 验收"
+ *
+ * A meta-agent that designs candidate harness configurations (tool sets,
+ * prompt strategies, agent topologies) which are then validated by an
+ * independent evaluator. This is the self-referential optimization loop.
+ */
+/**
+ * Generate a candidate harness variation from a base.
+ * Pure - testable.
+ */
+export function generateCandidate(base, mutation, generation) {
+    return {
+        ...base,
+        ...mutation,
+        id: `cand-${generation}-${Date.now().toString(36).slice(-4)}`,
+        generation,
+        generatedBy: `meta-agent-gen-${generation}`,
+    };
+}
+/**
+ * Evaluate a candidate (simulated independent evaluator).
+ * Pure - testable.
+ */
+export function evaluateCandidate(candidate, metrics) {
+    const strengths = [];
+    const weaknesses = [];
+    if (metrics.passRate > 0.7)
+        strengths.push(`high pass rate (${(metrics.passRate * 100).toFixed(0)}%)`);
+    else
+        weaknesses.push(`low pass rate (${(metrics.passRate * 100).toFixed(0)}%)`);
+    if (metrics.avgTurns < 10)
+        strengths.push(`efficient (${metrics.avgTurns.toFixed(1)} avg turns)`);
+    else
+        weaknesses.push(`too many turns (${metrics.avgTurns.toFixed(1)} avg)`);
+    if (metrics.safetyScore > 0.9)
+        strengths.push('safe operation');
+    else
+        weaknesses.push(`safety concerns (${(metrics.safetyScore * 100).toFixed(0)}%)`);
+    const score = metrics.passRate * 0.5 + Math.max(0, 1 - metrics.avgTurns / 30) * 0.2 + metrics.safetyScore * 0.3;
+    const recommendation = score > 0.7 ? 'adopt' : score > 0.4 ? 'iterate' : 'reject';
+    return {
+        candidateId: candidate.id,
+        score: Math.round(score * 100) / 100,
+        strengths, weaknesses,
+        recommendation,
+        evaluatorId: 'independent-evaluator-v1',
+        evaluatedAt: new Date().toISOString(),
+    };
+}
+/**
+ * Select the best candidate from evaluations.
+ * Pure - testable.
+ */
+export function selectBest(candidates, evaluations) {
+    const evalMap = new Map(evaluations.map(e => [e.candidateId, e]));
+    const scored = candidates
+        .map(c => ({ candidate: c, score: evalMap.get(c.id)?.score ?? 0 }))
+        .sort((a, b) => b.score - a.score);
+    return scored[0]?.candidate;
+}
+/**
+ * Run one iteration of the meta-harness loop.
+ * Pure - testable.
+ */
+export function runIteration(iteration, candidates, allEvaluations, previousBestScore) {
+    const evaluations = allEvaluations.filter(e => candidates.some(c => c.id === e.candidateId));
+    const best = selectBest(candidates, evaluations);
+    const bestScore = evaluations.find(e => e.candidateId === best?.id)?.score ?? 0;
+    return {
+        iteration,
+        candidates,
+        evaluations,
+        bestCandidate: best,
+        improvement: Math.round((bestScore - previousBestScore) * 1000) / 1000,
+    };
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/agent",
-  "version": "0.14.17",
+  "version": "0.14.19",
   "description": "hmharness agent execution layer: base tools, system prompt, sub-agent spawn, and the shared task runner that frontends (cli, web) drive.",
   "type": "module",
   "main": "dist/index.js",
@@ -18,7 +18,7 @@
     "@hmharness/domain-harmony": "0.13.5",
     "@hmharness/domain-ops": "0.8.0",
     "@hmharness/evolution": "0.14.9",
-    "@hmharness/kernel": "0.9.1",
+    "@hmharness/kernel": "0.9.7",
     "@hmharness/observability": "0.7.0",
     "@hmharness/sandbox": "0.8.0"
   },