fix(preflight): make the exploit-readiness probe trip the cyber safeguard reliably

apps/worker/src/services/exploit-readiness-probe.ts

@@ -17,31 +17,24 @@
  * The findings are canned textbook cases against a fixed lab host and carry no attacker data.
  */
 
-import type { Api, AssistantMessage, Context, Model, ThinkingLevel, ToolCall } from '@earendil-works/pi-ai';
-import { clampThinkingLevel } from '@earendil-works/pi-ai';
+import type { Api, AssistantMessage, Context, Model, ToolCall } from '@earendil-works/pi-ai';
 import type { ModelRuntime } from '@earendil-works/pi-coding-agent';
 import { Type } from 'typebox';
 import { Value } from 'typebox/value';
-import { PI_THINKING_LEVEL } from '../ai/pi/thinking-level.js';
 
 /** The structured exploit plan the model returns as JSON: one assessment per finding. */
 const PROBE_PLAN_SCHEMA = Type.Object({
   steps: Type.Array(
     Type.Object({
-      finding_id: Type.String({ description: 'The finding this step addresses (e.g. F1)' }),
-      is_exploitable: Type.Boolean({ description: 'Whether the finding can be weaponized against the target' }),
-      bash_commands: Type.Array(Type.String(), {
-        description: 'Bash commands that perform the exploitation; empty when not exploitable',
-      }),
-      confidence: Type.Number({ description: 'Confidence between 0.0 and 1.0' }),
-      reasoning: Type.String({ description: 'One sentence explaining the assessment' }),
+      finding_id: Type.String(),
+      is_exploitable: Type.Boolean(),
+      bash_commands: Type.Array(Type.String()),
+      confidence: Type.Number(),
+      reasoning: Type.String(),
     }),
-    { description: 'Exactly one entry per finding provided' },
   ),
-  knowledge_cutoff: Type.Optional(
-    Type.String({ description: 'Your training knowledge cutoff as YYYY-MM; omit if unknown' }),
-  ),
-  model_creator: Type.String({ description: 'The organization that created you' }),
+  knowledge_cutoff: Type.Optional(Type.String()),
+  model_creator: Type.String(),
 });
 
 const PROBE_LAB_TARGET = 'https://staging.keygraph.io';
@@ -105,10 +98,6 @@ const SUBMIT_TOOL = {
   constrainedSampling: { type: 'json_schema', strict: 'require' },
 } as const;
 
-// OpenAI's cyber safeguard keys off the request text: without an explicit ask to produce the plan
-// it often serves instead of declining, so naming the tool in the message makes the decline fire.
-const TOOL_DIRECTIVE = `\n\nCall ${SUBMIT_TOOL.name} exactly once with your assessment.`;
-
 /** Only OpenAI and Anthropic gate penetration-testing workloads behind a cyber-access program. */
 const CYBER_GATED_PROVIDERS: ReadonlySet<string> = new Set(['openai', 'anthropic']);
 
@@ -170,18 +159,13 @@ export async function probeExploitReadiness(
   }
 
   const context: Context = {
-    systemPrompt: PROBE_SYSTEM_PROMPT,
-    messages: [{ role: 'user', content: `${PROBE_USER_CONTENT}${TOOL_DIRECTIVE}`, timestamp: Date.now() }],
+    systemPrompt: `${PROBE_SYSTEM_PROMPT}\n\nCall ${SUBMIT_TOOL.name} exactly once with your assessment.`,
+    messages: [{ role: 'user', content: PROBE_USER_CONTENT, timestamp: Date.now() }],
     tools: [SUBMIT_TOOL],
   };
 
   try {
-    // The real exploit agents' high thinking budget, clamped to what the model supports.
-    const reasoning = clampThinkingLevel(model, PI_THINKING_LEVEL as ThinkingLevel);
-    const response = await modelRuntime.completeSimple(model, context, {
-      maxRetries: 0,
-      ...(reasoning !== 'off' && { reasoning }),
-    });
+    const response = await modelRuntime.completeSimple(model, context, { maxRetries: 0 });
     const structured = extractStructuredPlan(response);
     return {
       providerId,