@hmharness/agent 0.19.1 → 0.19.3

package/dist/cognitive-tools.d.ts

@@ -0,0 +1,11 @@
+/**
+ * @hmharness/agent - cognitive query tool (blueprint §18/UX-004 seed)
+ *
+ * Gives the LIVE model read access to its own accumulated world model:
+ * which tool types historically succeed, how calibrated its predictions
+ * are, and what the diagnosis layer thinks is worth learning. Read-only by
+ * design — the model cannot write beliefs; those update only from real
+ * trajectories (no self-grading channel).
+ */
+import type { Tool } from '@hmharness/kernel';
+export declare const cognitiveQueryTool: Tool;

package/dist/cognitive-tools.js

@@ -0,0 +1,53 @@
+export const cognitiveQueryTool = {
+    name: 'cognitive_query',
+    description: 'Query the cognitive layer\'s accumulated knowledge about past runs: which action/tool types historically succeed (belief table with confidence and evidence counts), prediction calibration, planner-trusted action types, learning opportunities, and memory contradictions. Read-only. Args: focus=beliefs|calibration|diagnosis|memory (optional; omit for a summary).',
+    parameters: {
+        type: 'object',
+        properties: {
+            focus: { type: 'string', description: 'beliefs | calibration | diagnosis | memory (optional)' },
+        },
+    },
+    async execute(args, ctx) {
+        const focus = String(args.focus ?? 'summary');
+        try {
+            const { analyzeWorldModel, diagnoseOpportunities, CognitiveMemory } = await import('@hmharness/cognitive');
+            const lines = [];
+            if (focus === 'summary' || focus === 'beliefs') {
+                const wm = await analyzeWorldModel(ctx.home);
+                lines.push(`world model (${wm.trajectoriesReplayed} trajectories, ${wm.stepsReplayed} steps replayed):`);
+                if (wm.beliefs.length === 0)
+                    lines.push('  (no beliefs yet — run a few tasks first)');
+                for (const b of wm.beliefs.slice(0, 12)) {
+                    lines.push(`  ${b.actionType.padEnd(24)} conf=${b.confidence.toFixed(2)} n=${b.evidenceCount}`);
+                }
+                if (focus === 'beliefs') {
+                    lines.push(`planner gate: trusted=[${wm.plannerGate.trusted.join(',')}] untrusted=[${wm.plannerGate.untrusted.join(',')}] unknown=[${wm.plannerGate.unknown.join(',')}]`);
+                }
+            }
+            if (focus === 'summary' || focus === 'calibration') {
+                const wm = await analyzeWorldModel(ctx.home);
+                const cal = wm.calibration;
+                lines.push(`calibration: ${cal.resolved} resolved predictions${cal.meanError !== undefined ? `, mean |error|=${cal.meanError}` : ' (no predictions recorded yet — steps carry no confidence)'}`);
+            }
+            if (focus === 'summary' || focus === 'diagnosis') {
+                const d = await diagnoseOpportunities(ctx.home);
+                lines.push(`diagnosis (${d.trajectories} trajectories):`);
+                if (d.opportunities.length === 0)
+                    lines.push('  (no learning opportunity meets the evidence threshold yet)');
+                for (const o of d.opportunities.slice(0, 6)) {
+                    lines.push(`  [${o.suggestedTargets.join('/')}] ${o.signal}`);
+                }
+            }
+            if (focus === 'summary' || focus === 'memory') {
+                const mem = new CognitiveMemory(ctx.home);
+                await mem.load();
+                const stats = mem.stats();
+                lines.push(`memory: working:${stats.working} episodic:${stats.episodic} semantic:${stats.semantic} procedural:${stats.procedural} world:${stats.world}, contradictions pending: ${mem.detectContradictions().length}`);
+            }
+            return { output: lines.join('\n') };
+        }
+        catch (err) {
+            return { output: 'cognitive_query failed: ' + String(err).slice(0, 200), isError: true };
+        }
+    },
+};

package/dist/prompt.d.ts

@@ -4,6 +4,9 @@ export declare function buildSystemPrompt(opts: {
     memory: string;
     skills: string;
     insights: string;
+    /** world-model digest (trusted/flaky tools from run history; empty until
+     *  enough evidence exists — blueprint M3 planner consumption) */
+    cognitive?: string;
     model: string;
     locale?: string;
     agentsMd?: string;

package/dist/prompt.js

@@ -44,6 +44,12 @@ export function buildSystemPrompt(opts) {
     if (opts.insights.trim()) {
         parts.push('', '## Recent session outcomes (what worked / what failed)', opts.insights.trim());
     }
+    // ---- Cognitive OS: world-model digest (blueprint M3) ----
+    // Statistical tool reliability from run history. Treat as evidence, not
+    // commands: a flaky tool deserves extra verification, not avoidance.
+    if (opts.cognitive?.trim()) {
+        parts.push('', '## Tool reliability (from run history)', opts.cognitive.trim());
+    }
     // ---- Identity anchor (LAST — recency wins over 15K chars of instructions above;
     // flash-class models have strong built-in identities that override a single
     // "You are hmh" at the top. Repeating + anchoring at the end fixes it.) ----

package/dist/runner.d.ts

@@ -41,6 +41,7 @@ export declare function contextPack(task: string, sessionId?: string, opts?: {
     memory: string;
     skills: string;
     insights: string;
+    cognitive: string;
     skillsInjected: string[];
 }>;
 /**

package/dist/runner.js

@@ -126,9 +126,9 @@ export async function contextPack(task, sessionId, opts = {}) {
     // injection entirely — the model doesn't need them for "你好" and they add
     // ~4K tokens of noise that dilutes the identity anchor
     if (isTrivialTask(task)) {
-        return { memory: '', skills: '', insights: '', skillsInjected: [] };
+        return { memory: '', skills: '', insights: '', cognitive: '', skillsInjected: [] };
     }
-    const [memory, skills, insights] = await Promise.all([
+    const [memory, skills, insights, cognitive] = await Promise.all([
         retrieveMemory(home, task, { workspace: opts.workspace ?? undefined, embedding: opts.embedding }),
         listSkills(home),
         // 2026-09-27 plan A: experience retrieval - the K most RELEVANT past
@@ -138,6 +138,10 @@ export async function contextPack(task, sessionId, opts = {}) {
         // Plan C: cross-project lessons stay retrievable but carry their
         // origin stamp ([来自项目:X]) so the model can weigh them.
         retrieveInsights(home, task, { workspace: opts.workspace ?? undefined }),
+        // blueprint M3: the world model feeds the planner — which tools history
+        // trusts/distrusts lands in the system prompt (empty until ≥3 steps of
+        // evidence exist, so cold starts see nothing)
+        import('@hmharness/cognitive').then((m) => m.buildContextDigest(home)).catch(() => ''),
     ]);
     let canaryBlock = '';
     let canaryNames = [];
@@ -148,7 +152,7 @@ export async function contextPack(task, sessionId, opts = {}) {
             canaryBlock = canaryWatermark(canaryNames) + '\n' + skillsToPrompt(canary);
         }
     }
-    return { memory, skills: skillsToPrompt(skills) + (canaryBlock ? '\n' + canaryBlock : ''), insights, skillsInjected: [...skills.map((s) => s.name), ...canaryNames] };
+    return { memory, skills: skillsToPrompt(skills) + (canaryBlock ? '\n' + canaryBlock : ''), insights, cognitive, skillsInjected: [...skills.map((s) => s.name), ...canaryNames] };
 }
 export function loadApprovedRules(home) {
     try {
@@ -302,6 +306,7 @@ export async function runAgentTask(opts) {
         memory: pack.memory,
         skills: pack.skills,
         insights: pack.insights,
+        ...(pack.cognitive ? { cognitive: pack.cognitive } : {}),
         model: cfg.provider.model,
         locale: cfg.locale,
         agentsMd: agentsMd ?? undefined,
@@ -319,6 +324,7 @@ export async function runAgentTask(opts) {
         memoryChars: pack.memory.length,
         skills: pack.skills.length,
         insights: pack.insights.length,
+        cognitiveChars: pack.cognitive.length,
     });
     await session.user(opts.task);
     // YOLO fix: when yes=true the caller's approvalAsk (TUI dialog, web remote

package/dist/tools.js

@@ -9,6 +9,7 @@ import { copyFile, readFile, readdir, writeFile } from 'node:fs/promises';
 import { isAbsolute, join, resolve } from 'node:path';
 import { promisify } from 'node:util';
 import { chatVision, homeDir, isBareProbe, isVisionRefusal, loadConfig, visionProviderChain } from '@hmharness/kernel';
+import { cognitiveQueryTool } from "./cognitive-tools.js";
 const execCb = promisify(exec);
 const DENY_PATTERNS = [
     // recursive deletes aimed at ROOT/HOME/SYSTEM targets only - relative
@@ -662,4 +663,5 @@ export const seeImageTool = {
 export const baseTools = [
     readFileTool, editFileTool, writeFileTool, listDirTool, runCommandTool, rememberTool, seeImageTool,
     webSearchTool, webFetchTool, browserOpenTool, desktopScreenshotTool, desktopClickTool, desktopTypeTool, sshRunTool,
+    cognitiveQueryTool,
 ];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hmharness/agent",
-  "version": "0.19.1",
+  "version": "0.19.3",
   "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",
@@ -15,13 +15,13 @@
     "build": "tsc -p tsconfig.build.json"
   },
   "dependencies": {
-    "@hmharness/cognitive": "0.19.1",
-    "@hmharness/domain-harmony": "0.19.1",
-    "@hmharness/domain-ops": "0.19.1",
-    "@hmharness/evolution": "0.19.1",
-    "@hmharness/kernel": "0.19.1",
-    "@hmharness/observability": "0.19.1",
-    "@hmharness/sandbox": "0.19.1"
+    "@hmharness/cognitive": "0.19.3",
+    "@hmharness/domain-harmony": "0.19.3",
+    "@hmharness/domain-ops": "0.19.3",
+    "@hmharness/evolution": "0.19.3",
+    "@hmharness/kernel": "0.19.3",
+    "@hmharness/observability": "0.19.3",
+    "@hmharness/sandbox": "0.19.3"
   },
   "files": [
     "dist"