diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3a0cd16..c5fd6be 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -2,6 +2,19 @@
 
 ## Unreleased
 
+## 2.12.2 — 2026-06-20
+
+- **Agent-awareness in the served MCP instructions.** The `PROTOCOL` string the
+  MCP server hands every connected agent now states *when to reach for Keyoku*
+  (multi-step goals with a verifiable end state — migrations, "make X
+  production-ready", get-CI-green, deploys, refactors with a clear
+  done-condition; not one-shot edits or pure Q&A) and the **discipline** that
+  makes the loop work: behavioral (not file-presence) criteria, criteria
+  immutable after `goal_create`, `goal_record` every action honestly *including
+  failures* (they become pitfalls in future suggestions), never fudge an
+  env-blocked criterion, always honor the autonomy level. Also nudges agents to
+  query existing knowledge/goals/workflows first so learned work gets reused.
+
 ## 2.12.1 — 2026-06-18
 
 - **Bounded the audit log.** `audit.jsonl` now self-caps at ~1 MB (keeps the most
diff --git a/package-lock.json b/package-lock.json
index 2ec2b71..fd406c2 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -1,12 +1,12 @@
 {
-  "name": "@keyoku/harness",
-  "version": "0.1.0",
+  "name": "keyoku",
+  "version": "2.12.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
-      "name": "@keyoku/harness",
-      "version": "0.1.0",
+      "name": "keyoku",
+      "version": "2.12.2",
       "license": "MIT",
       "dependencies": {
         "@anthropic-ai/sdk": "^0.104.1",
@@ -16,7 +16,7 @@
         "zod": "^3.25.0"
       },
       "bin": {
-        "keyoku-harness": "dist/index.js"
+        "keyoku": "dist/index.js"
       },
       "devDependencies": {
         "@types/jmespath": "^0.15.2",
diff --git a/package.json b/package.json
index 15b2de5..c29ebe7 100644
--- a/package.json
+++ b/package.json
@@ -1,13 +1,15 @@
 {
   "name": "keyoku",
-  "version": "2.12.1",
+  "version": "2.12.2",
   "description": "The harness with muscle memory — watches what you do in Claude Code, Cursor, or Codex, learns your patterns, and turns them into reusable MCP-native workflows.",
   "type": "module",
   "bin": {
     "keyoku": "dist/index.js"
   },
   "main": "dist/index.js",
-  "exports": { ".": "./dist/index.js" },
+  "exports": {
+    ".": "./dist/index.js"
+  },
   "files": [
     "dist/*.js",
     "README.md",
diff --git a/src/guidance.ts b/src/guidance.ts
index 0744481..c6bd0ff 100644
--- a/src/guidance.ts
+++ b/src/guidance.ts
@@ -12,6 +12,14 @@ import type {
 export const PROTOCOL = `Keyoku is a convergence harness: it turns goals with machine-checkable
 success criteria into a loop you (the agent) drive until the goal is reached.
 
+REACH FOR KEYOKU whenever a task is a multi-step goal with a verifiable end
+state — a migration, "make X production-ready", "get CI green", a deploy, a
+refactor with a clear done-condition — especially work that spans many actions
+or survives context resets, or that you want to become a reusable, learned
+workflow. Not for one-shot edits or pure Q&A. The signal: you could write a
+command/HTTP/probe that checks "is it done yet?". Before starting one, query
+existing knowledge + goals + workflows so you reuse what was already learned.
+
 The protocol:
 1. goal_create — declare a goal. Success criteria are probes (shell command,
    HTTP request, or MCP connector tool call) plus assertions over their
@@ -45,7 +53,15 @@ harness_learn mines traces + observations into reusable patterns (SLM-powered
 when GEMINI_API_KEY / ANTHROPIC_API_KEY is configured, heuristic otherwise).
 Relevant patterns and previously learned workflows appear in goal_assess
 guidance — prefer them; they encode how this user does things. Everything
-consequential lands in the audit trail (audit_list).`;
+consequential lands in the audit trail (audit_list).
+
+Discipline that makes the loop work: make criteria BEHAVIORAL ("does the thing
+actually work?") not mere file-presence; criteria are IMMUTABLE after
+goal_create — get them right, and make a NEW goal to change them; call
+goal_record for EVERY action honestly, including failures (failures teach the
+harness what to avoid and become pitfalls in future suggestions); never fudge a
+criterion that's blocked by the environment — report it honestly and leave the
+goal unconverged; and always honor the goal's autonomy level.`;
 
 function describeExpected(evaluation: CriterionEvaluation): string {
   const { op, value, path } = evaluation.expected;