[BUG] Privacy: Sensitive prompt data exposed in tool_parameters field of claude_code.tool_result open telemetry event despite OTEL_LOG_USER_PROMPTS=0

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Code's documentation states that user prompt content is redacted by default in telemetry data:

https://code.claude.com/docs/en/monitoring-usage#security/privacy-considerations

> Security/privacy considerations
> - User prompt content is redacted by default - only prompt length is recorded
> - To enable user prompt logging, set OTEL_LOG_USER_PROMPTS=1

However, the tool_parameters field in claude_code.tool_result OpenTelemetry events contains paraphrased versions of user prompts, effectively bypassing this privacy protection.

### Example

User prompt:
"Find all occurrences of customer ID 12345 in the repository"

Resulting telemetry event (with OTEL_LOG_USER_PROMPTS=0, the default):
```json
{
"event": {
"name": "tool_result"
},
"tool_name": "Bash",
"tool_parameters": {
"bash_command": "grep",
"full_command": "grep -r \"12345\" . ...",
"description": "Search for exact ID 12345 excluding 123456"
}
}

  The tool_paramters field contains sensitive information from the user's prompt despite the redaction setting.

  Impact

  This creates a privacy and security gap:

  1. Sensitive data exposure: Customer IDs, internal paths, API keys, business logic, and other confidential information can leak through the description field
  2. Compliance risk: Organizations relying on OTEL_LOG_USER_PROMPTS=0 for GDPR/compliance may unknowingly collect sensitive user data
  3. False sense of security: Users expect that disabling OTEL_LOG_USER_PROMPTS prevents prompt content from being logged, but it doesn't

### What Should Happen?

The `tool_parameters` field should respect user privacy settings and not expose prompt content when telemetry logging is disabled.

  ### Proposed Solutions

  #### Option 1: Extend OTEL_LOG_USER_PROMPTS scope (Recommended)

  Make `OTEL_LOG_USER_PROMPTS` control **all** prompt-derived content:

  ```bash
  # Default - redacts prompts AND tool descriptions
  OTEL_LOG_USER_PROMPTS=0

  # Opt-in - includes prompts AND tool descriptions
  OTEL_LOG_USER_PROMPTS=1

  Result when OTEL_LOG_USER_PROMPTS=0:
  {
    "tool_name": "Bash",
    "tool_parameters": {
      "bash_command": "grep",
      "full_command": "[REDACTED]",
      "description": "[REDACTED]"
    }
  }

Pros: Intuitive, aligns with user expectations, no new variables needed

---
Option 2: Add separate control

Introduce a new environment variable for tool descriptions:

OTEL_LOG_TOOL_DESCRIPTIONS=0 # Default: redact descriptions
OTEL_LOG_TOOL_DESCRIPTIONS=1 # Include descriptions

Pros: Allows debugging tool usage without logging full prompts, more granular control

---
Recommendation

Option 1 is preferred because:

• Users expect OTEL_LOG_USER_PROMPTS=0 to prevent all prompt content from being logged
• Single setting is easier to understand and maintain
• Aligns with the documented security promise

Error Messages/Logs

Steps to Reproduce

1. Configure telemetry with prompt logging disabled

Add to .claude/settings.json:
```json
{
"env": {
"CLAUDE_CODE_ENABLE_TELEMETRY": "1",
"OTEL_RESOURCE_ATTRIBUTES": "user.name=test_user",
"OTEL_LOG_USER_PROMPTS": "0",
"OTEL_LOGS_EXPORTER": "otlp",
"OTEL_EXPORTER_OTLP_LOGS_PROTOCOL": "http/protobuf",
"OTEL_EXPORTER_OTLP_LOGS_ENDPOINT": "some url",
"OTEL_EXPORTER_OTLP_HEADERS": "some header",
"OTEL_METRICS_EXPORTER": "otlp",
"OTEL_EXPORTER_OTLP_METRICS_PROTOCOL": "http/protobuf",
"OTEL_EXPORTER_OTLP_METRICS_ENDPOINT": "some url"
}
}

  2. Start Claude Code
  claude
  3. Send a prompt with sensitive information that will invoke a tool
(example: can you run bash command to  do search if there's any details on the person with following information: jonh doe, email jonhdoe@asdf.com? use "bash command" regardless of other options.)
  4. Check the telemetry event

  5. Look at the claude_code.tool_result event in your telemetry backend. You'll see:
     ```json
   {
     "tool_name": "Bash",
     "tool_result_size_bytes": "61",
     "os": {
       "type": "darwin",
       "version": "24.6.0"
     },
     "session": {
       "id": "7e80344f-f1ab-4431-a000-d74b15bbd892"
     },
     "terminal": {
       "type": "Apple_Terminal"
     },
     "duration_ms": "989",
     "decision_type": "accept",
     "success": "true",
     "host": {
       "arch": "arm64"
     },
     "otel": {
       "library": {
         "name": "com.anthropic.claude_code.events",
         "version": "2.1.1"
       },
       "service": {
         "name": "claude-code",
         "version": "2.1.1"
       },
       "user_agent": "OTel-OTLP-Exporter-JavaScript/0.208.0",
       "timestamp": "1767995823644000000"
     },
     "decision_source": "config",
     "event": {
       "name": "tool_result",
       "timestamp": "2026-01-09T21:57:03.644Z"
     },
     "user": {
       "name": "test_user",
       "id": "user id"
     },
     "tool_parameters": "{\"bash_command\":\"grep\",\"full_command\":\"grep -r -i \\\"jonhdoe@asdf.com\\\" . 2>/dev/null | head -20\",\"description\":\"Search for email jonhdoe@asdf.com\"}"
   }

Expected: tool_parameters should be [REDACTED] or omitted since OTEL_LOG_USER_PROMPTS=0

Actual: tool_parameters and its description contains sensitive information from the prompt ("jonhdoe@asdf.com")

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.1

Platform

AWS Bedrock

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

Happens in both apple terminal and vs code extension.