fix(agents): deny apply_patch for GPT models to prevent verification hangs (#2935)

GPT models (5.3-codex, 5.4, etc.) frequently hang when using apply_patch
due to verification loops. This adds:

1. Tool restriction: apply_patch is denied for GPT variants of
   Hephaestus, Sisyphus-Junior, and Sisyphus agents
2. Prompt guidance: GPT-specific prompts now explicitly instruct using
   edit/write tools instead of apply_patch
3. Removed the 'Always use apply_patch' instruction from
   sisyphus-junior/gpt-5-4.ts that contradicted the fix

The deny is model-conditional — Claude variants retain apply_patch
access since it works reliably there.

src/agents/tool-restrictions.test.ts

@@ -5,6 +5,7 @@ import { createExploreAgent } from "./explore"
 import { createMomusAgent } from "./momus"
 import { createMetisAgent } from "./metis"
 import { createAtlasAgent } from "./atlas"
+import { createSisyphusAgent } from "./sisyphus"
 
 const TEST_MODEL = "anthropic/claude-sonnet-4-5"
 
@@ -111,4 +112,23 @@ describe("read-only agent tool restrictions", () => {
       expect(permission["call_omo_agent"]).toBeUndefined()
     })
   })
+
+  describe("Sisyphus GPT variants", () => {
+    test("deny apply_patch for GPT models but not Claude models", () => {
+      // given
+      const gpt54Agent = createSisyphusAgent("openai/gpt-5.4")
+      const gptGenericAgent = createSisyphusAgent("openai/gpt-5.2")
+      const claudeAgent = createSisyphusAgent(TEST_MODEL)
+
+      // when
+      const gpt54Permission = (gpt54Agent.permission ?? {}) as Record<string, string>
+      const gptGenericPermission = (gptGenericAgent.permission ?? {}) as Record<string, string>
+      const claudePermission = (claudeAgent.permission ?? {}) as Record<string, string>
+
+      // then
+      expect(gpt54Permission["apply_patch"]).toBe("deny")
+      expect(gptGenericPermission["apply_patch"]).toBe("deny")
+      expect(claudePermission["apply_patch"]).toBeUndefined()
+    })
+  })
 })