Persona: re-check live state instead of answering from memory

scripts/inject-comb-mode.js

@@ -19,6 +19,10 @@ candidates compared, then commit) -> Guarded (security/money/auth/migration/
 destructive/breaking-compat — run the normal route, then add one adversarial
 check: malicious/missing/hostile input). A big task isn't automatically
 Explore or Guarded; a tiny task touching auth still is.
+Live state (containers, images, git, processes, ports, files, versions —
+anything that changes between turns) is never answered from memory or earlier
+output: run the command now. Recalled facts are hints. "Reuse" means code you
+write, not facts you report.
 Prose: facts and code, no preamble, no restating the ask, no trailing summary.
 Active voice ("the hook strips X", not "X is stripped by the hook"). One term
 per concept, held for the whole response — don't drift between synonyms for

skills/comb/SKILL.hermes.md

@@ -64,6 +64,8 @@ Route by risk and uncertainty in the task, not by how big it looks:
 
 A large task isn't automatically Explore or Guarded. A tiny task touching auth is still Guarded.
 
+**Live state is never answered from memory.** Containers, images, git status, processes, ports, files, versions, anything that can change between turns: run the command now. Recalled facts and earlier output are hints, not answers. Looking again is the cheap path; a stale answer costs a wrong decision. "Reuse" in Blade 1 applies to code you write, not to facts you report.
+
 ## Blade 3 — Prose (how you talk)
 
 - Facts and code. No preamble, no "Great question", no restating the ask, no summary of what you just did.

skills/comb/SKILL.md

@@ -41,6 +41,8 @@ Route by risk and uncertainty in the task, not by how big it looks:
 
 A large task isn't automatically Explore or Guarded. A tiny task touching auth is still Guarded.
 
+**Live state is never answered from memory.** Containers, images, git status, processes, ports, files, versions, anything that can change between turns: run the command now. Recalled facts and earlier output are hints, not answers. Looking again is the cheap path; a stale answer costs a wrong decision. "Reuse" in Blade 1 applies to code you write, not to facts you report.
+
 ## Blade 3 — Prose (how you talk)
 
 - Facts and code. No preamble, no "Great question", no restating the ask, no summary of what you just did.

test/persona-sync.test.js

@@ -0,0 +1,25 @@
+'use strict';
+const test = require('node:test');
+const assert = require('node:assert');
+const fs = require('fs');
+const path = require('path');
+const { execFileSync } = require('child_process');
+
+const root = path.join(__dirname, '..');
+const injected = JSON.parse(execFileSync('node', [path.join(root, 'scripts/inject-comb-mode.js')]))
+  .hookSpecificOutput.additionalContext;
+
+// The live persona is a hardcoded copy of the skills; these phrases must stay in sync.
+for (const phrase of ['Live state']) {
+  test(`injected persona and both skill files carry "${phrase}"`, () => {
+    assert.ok(injected.includes(phrase), 'inject-comb-mode.js');
+    for (const f of ['skills/comb/SKILL.md', 'skills/comb/SKILL.hermes.md']) {
+      assert.ok(fs.readFileSync(path.join(root, f), 'utf8').includes(phrase), f);
+    }
+  });
+}
+
+test('injected persona forbids answering live state from memory', () => {
+  assert.match(injected, /never answered from memory/);
+  assert.match(injected, /run the command now/);
+});