@hmharness/evolution 0.14.25 → 0.14.27

package/dist/candidates.d.ts

@@ -1,4 +1,4 @@
-import type { BenchCase } from './bench.ts';
+import type { BenchCase } from './bench.js';
 export type CandidateTarget = 'prompt' | 'skill' | 'context' | 'tool_policy' | 'model_router' | 'workflow' | 'harness';
 export declare const CANDIDATE_TARGETS: CandidateTarget[];
 export interface EvolutionCandidate {

package/dist/cloud-finetune.d.ts

@@ -1,4 +1,4 @@
-import type { DpoPair } from './reward-model.ts';
+import type { DpoPair } from './reward-model.js';
 import type { ProviderConfig } from '@hmharness/kernel';
 /** One provider's fine-tuning dialect. Everything provider-specific lives
  * here; the transport and job flow above it are generic. */

package/dist/evolve.d.ts

@@ -1,5 +1,5 @@
 import { type ProviderConfig } from '@hmharness/kernel';
-import { type BenchCase } from './bench.ts';
+import { type BenchCase } from './bench.js';
 export interface SkillProposal {
     name: string;
     description: string;

package/dist/index.d.ts

@@ -1,26 +1,26 @@
-export * from './memory.ts';
-export * from './ranker.ts';
-export * from './candidates.ts';
-export * from './skillpayload.ts';
-export * from './dataset.ts';
-export * from './readiness.ts';
-export * from './labels.ts';
-export * from './reward-model.ts';
-export * from './judge.ts';
-export * from './sequential.ts';
-export * from './insights.ts';
-export * from './skills.ts';
-export * from './bench.ts';
-export * from './evolve.ts';
-export * from './patches.ts';
-export * from './impact.ts';
-export * from './workflows.ts';
-export * from './knowledge.ts';
-export * from './radar.ts';
-export * from './external.ts';
-export * from './knowledge-graph.ts';
-export * from './router-evolution.ts';
-export * from './context-evolution.ts';
-export * from './rl-governance.ts';
-export * from './policy-optimizer.ts';
-export * from './cloud-finetune.ts';
+export * from './memory.js';
+export * from './ranker.js';
+export * from './candidates.js';
+export * from './skillpayload.js';
+export * from './dataset.js';
+export * from './readiness.js';
+export * from './labels.js';
+export * from './reward-model.js';
+export * from './judge.js';
+export * from './sequential.js';
+export * from './insights.js';
+export * from './skills.js';
+export * from './bench.js';
+export * from './evolve.js';
+export * from './patches.js';
+export * from './impact.js';
+export * from './workflows.js';
+export * from './knowledge.js';
+export * from './radar.js';
+export * from './external.js';
+export * from './knowledge-graph.js';
+export * from './router-evolution.js';
+export * from './context-evolution.js';
+export * from './rl-governance.js';
+export * from './policy-optimizer.js';
+export * from './cloud-finetune.js';

package/dist/insights.d.ts

@@ -16,3 +16,26 @@ export declare function recordInsight(home: string, insight: Insight): Promise<v
 export declare function readInsights(home: string, limit?: number): Promise<Insight[]>;
 /** Summarize recent insights for the system prompt (bounded). */
 export declare function recentInsights(home: string, limit?: number): Promise<string>;
+/** Task-similarity feature set: per-WORD bigrams plus single chars as
+ *  tokens (2026-09-27). Cross-word bigrams that include the space turned
+ *  out to be a noise factory - 'fix' and 'xxx' shared 'x ' and unrelated
+ *  natural sentences scored 0.06. Words are split on non-alphanumerics;
+ *  CJK text has no spaces, so a whole Chinese clause becomes one word and
+ *  keeps every character bigram. Pure. */
+export declare function bigrams(s: string): Set<string>;
+/** Jaccard similarity of two bigram sets. Pure. */
+export declare function jaccard(a: Set<string>, b: Set<string>): number;
+/** Noise floor: below this similarity an insight is unrelated to the task.
+ *  With per-word bigrams the highest unrelated natural-language pairs score
+ *  ~0.04 (stopword overlap: 'the' contributes th/he); related tasks land at
+ *  0.5+. The floor sits between the two, with margin on both sides. */
+export declare const INSIGHT_SIM_FLOOR = 0.05;
+/**
+ * The K most RELEVANT insights for this task (recency as tiebreak), falling
+ * back to the K most recent when nothing clears the noise floor - so the
+ * prompt always carries some experience, exactly like before.
+ */
+export declare function retrieveInsights(home: string, task: string, opts?: {
+    topK?: number;
+    pool?: number;
+}): Promise<string>;

package/dist/insights.js

@@ -78,3 +78,88 @@ export async function recentInsights(home, limit = 5) {
         return '';
     }
 }
+/* ---------------- experience retrieval (2026-09-27, plan A) ----------------
+ * The injection used to be "last 5 insights" - pure recency. After 300
+ * tasks, a question similar to task #50 got five UNRELATED recent lessons
+ * instead of that one. Industry direction (Agent KB / Agent Workflow
+ * Memory, verified 2026-09-27): retrieve relevant experience per task.
+ * Zero-dependency red line -> character-bigram Jaccard (works well for CJK
+ * too), a small recency tiebreak, graceful fallback to recency order. */
+/** Task-similarity feature set: per-WORD bigrams plus single chars as
+ *  tokens (2026-09-27). Cross-word bigrams that include the space turned
+ *  out to be a noise factory - 'fix' and 'xxx' shared 'x ' and unrelated
+ *  natural sentences scored 0.06. Words are split on non-alphanumerics;
+ *  CJK text has no spaces, so a whole Chinese clause becomes one word and
+ *  keeps every character bigram. Pure. */
+export function bigrams(s) {
+    const words = s.toLowerCase().split(/[^a-z0-9\u4e00-\u9fff]+/).filter(Boolean);
+    const out = new Set();
+    for (const w of words) {
+        if (w.length === 1) {
+            out.add(w);
+            continue;
+        }
+        for (let i = 0; i < w.length - 1; i++)
+            out.add(w.slice(i, i + 2));
+    }
+    return out;
+}
+/** Jaccard similarity of two bigram sets. Pure. */
+export function jaccard(a, b) {
+    if (!a.size || !b.size)
+        return 0;
+    let inter = 0;
+    for (const g of a)
+        if (b.has(g))
+            inter++;
+    return inter / (a.size + b.size - inter);
+}
+function formatInsight(i) {
+    return `- [${i.outcome}] ${i.task.slice(0, 60)} (turns ${i.turns}, tools ${i.toolsUsed.join(',') || 'none'})`;
+}
+/** Noise floor: below this similarity an insight is unrelated to the task.
+ *  With per-word bigrams the highest unrelated natural-language pairs score
+ *  ~0.04 (stopword overlap: 'the' contributes th/he); related tasks land at
+ *  0.5+. The floor sits between the two, with margin on both sides. */
+export const INSIGHT_SIM_FLOOR = 0.05;
+/**
+ * The K most RELEVANT insights for this task (recency as tiebreak), falling
+ * back to the K most recent when nothing clears the noise floor - so the
+ * prompt always carries some experience, exactly like before.
+ */
+export async function retrieveInsights(home, task, opts = {}) {
+    const topK = opts.topK ?? 5;
+    const pool = opts.pool ?? 500;
+    let rows = [];
+    try {
+        const text = await readFile(join(home, 'insights', 'insights.jsonl'), 'utf8');
+        rows = text.trim().split('\n').filter(Boolean)
+            .map((l) => { try {
+            return JSON.parse(l);
+        }
+        catch {
+            return null;
+        } })
+            .filter((x) => x !== null)
+            .slice(-pool);
+    }
+    catch {
+        return ''; // no archive at all - same behaviour as recentInsights
+    }
+    if (rows.length === 0)
+        return '';
+    const q = bigrams(task);
+    if (q.size === 0)
+        return rows.slice(-topK).map(formatInsight).join('\n');
+    // newest first within the pool, so idx IS the recency rank
+    const newestFirst = [...rows].reverse();
+    const scored = newestFirst
+        .map((r, idx) => ({ r, s: jaccard(q, bigrams(r.task)) + (1 - idx / (newestFirst.length + 1)) * 1e-4 }))
+        .filter((x) => x.s > INSIGHT_SIM_FLOOR)
+        .sort((a, b) => b.s - a.s)
+        .slice(0, topK)
+        .map((x) => x.r);
+    if (scored.length === 0)
+        return rows.slice(-topK).map(formatInsight).join('\n');
+    return scored.map(formatInsight).join('\n');
+}

package/dist/knowledge.d.ts

@@ -1,5 +1,5 @@
 import { type ProviderConfig } from '@hmharness/kernel';
-import { type SkillProposal } from './evolve.ts';
+import { type SkillProposal } from './evolve.js';
 declare function fetchPage(url: string): Promise<string | null>;
 /** One refresh cycle: fetch -> compare -> maybe one knowledge draft.
  *  Returns a short human-readable summary (or 'offline' etc.). */

package/dist/reward-model.d.ts

@@ -1,4 +1,4 @@
-import type { HumanLabel } from './labels.ts';
+import type { HumanLabel } from './labels.js';
 export type { HumanLabel };
 export interface RewardFeatures {
     /** outcome bucket: 2=ok, 1=turn-budget, 0=error */

package/dist/workflows.d.ts

@@ -10,7 +10,7 @@
  * a workflow is just a skill with a shape.
  */
 import { type ProviderConfig } from '@hmharness/kernel';
-import { type SkillProposal } from './evolve.ts';
+import { type SkillProposal } from './evolve.js';
 export interface WorkflowCluster {
     key: string;
     tasks: string[];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/evolution",
-  "version": "0.14.25",
+  "version": "0.14.27",
   "description": "hmharness evolution subsystem: persistent memory, insight capture, skill library, and the bench that gives evolution its fitness signal. First-class kernel citizen, not a plugin.",
   "type": "module",
   "main": "dist/index.js",
@@ -15,7 +15,7 @@
     "build": "tsc -p tsconfig.build.json"
   },
   "dependencies": {
-    "@hmharness/kernel": "0.9.4"
+    "@hmharness/kernel": "*"
   },
   "files": [
     "dist"