plan.txt triggers Azure OpenAI content filter

[intelljames]

Problem

The default plan agent prompt in packages/opencode/src/session/prompt/plan.txt triggers Azure OpenAI's content safety policy filter.

The specific phrase that triggers the filter is:

Plan mode ACTIVE - you are in READ-ONLY phase

After testing, the combination of ACTIVE. You are in (uppercase ACTIVE followed by "You are in") causes Azure OpenAI to reject the request with:

{"error":{"message":"The response was filtered due to the prompt triggering Azure OpenAI's content management policy.","type":"invalid_request_error","code":"content_filter"}}

Test Results

Content	Result
ACTIVE. You are in	❌ content_filter
active. You are in (lowercase)	✅ passes
ACTIVE. You are currently in	✅ passes
Plan mode is active. You are currently in	✅ passes

Suggested Fix

Either:

1. Change the wording in plan.txt from:

   CRITICAL: Plan mode ACTIVE - you are in READ-ONLY phase.
   

   to:

   CRITICAL: Plan mode is active. You are currently in READ-ONLY phase.
   

2. Make the plan prompt configurable via opencode.json:

   {
     "agent": {
       "plan": {
         "prompt": "custom plan prompt..."
       }
     }
   }

Currently, the plan prompt is hardcoded and injected directly in session/prompt.ts, ignoring the agent.plan.prompt config option.

Impact

This prevents users using Azure OpenAI (or similar providers with content safety policies) from using the plan agent.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Plan mode <system-reminder> triggers Azure OpenAI jailbreak detection #16442: Same root cause — plan.txt phrases trigger Azure OpenAI's content filter (jailbreak detection) when using plan mode, with the same suggested fix of rewording the prompt to avoid override/injection-style language.