fix(enforce): hole 2 — Task/Agent count as mutating actions

Brain-retro #5 candidate C, hole 2: enforce-classifier-match.mjs's
MUTATING_TOOLS set missed Task/Agent, so delegating mutations via Task()
bypassed the rule. Added Task and Agent to the set; nodeMatches already
handles Task.subagent_type matching.

Regression test asserts Task with matching subagent_type does NOT block
(keeps the existing nodeMatches Task path intact).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tools/enforce-classifier-match.mjs

@@ -28,7 +28,7 @@ import {
 const RULE_KEY = 'classifier-mismatch';
 const CONFIDENCE_THRESHOLD = 0.7;
 
-const MUTATING_TOOLS = new Set(['Edit', 'Write', 'MultiEdit', 'NotebookEdit', 'Bash']);
+const MUTATING_TOOLS = new Set(['Edit', 'Write', 'MultiEdit', 'NotebookEdit', 'Bash', 'Task', 'Agent']);
 
 /** Normalize a node id: strip "superpowers:" / "skill:" prefix; allow #ID. */
 function normalizeNode(s) {

tools/enforce-classifier-match.test.mjs

@@ -103,4 +103,26 @@ describe('enforce-classifier-match / decide', () => {
     });
     expect(r.block).toBe(false);
   });
+
+  it('blocks when Task subagent is spawned without matching recommendation (hole 2)', () => {
+    const r = decide({
+      toolUses: [{ name: 'Task', input: { subagent_type: 'general-purpose', prompt: 'do stuff' } }],
+      recommendation: 'superpowers:writing-plans',
+      confidence: 0.9,
+      assistantText: '',
+      override: null,
+    });
+    expect(r.block).toBe(true);
+  });
+
+  it('does NOT block when Task subagent matches recommendation (regression — Task should count as match when right type)', () => {
+    const r = decide({
+      toolUses: [{ name: 'Task', input: { subagent_type: 'writing-plans', prompt: '...' } }],
+      recommendation: 'writing-plans',
+      confidence: 0.9,
+      assistantText: '',
+      override: null,
+    });
+    expect(r.block).toBe(false);
+  });
 });