Anthropic Inspector¶

The Anthropic Inspector is a debugging tool that captures and logs complete request/response flows through Olla's Anthropic translator. Use it to debug model behaviour, compare different models, understand request transformations and troubleshoot integration issues.

Key Features:

• Complete request/response logging in JSONL format
• Multi-stage inspection (Anthropic to OpenAI to Anthropic)
• Session-based organisation for comparing requests
• Configurable redaction for sensitive data
• Built-in analysis capabilities with jq
• Minimal performance impact

Security Warning¶

DO NOT ENABLE IN PRODUCTION

The inspector logs full request/response bodies including user messages, system prompts, tool definitions, model responses, and potentially sensitive data.

Only enable in development and testing environments where you control all data.

When to Use¶

Good use cases:

• Compare Model responses to identical prompts
• Debug Anthropic to OpenAI conversion issues
• Record sessions during feature development
• Track request/response timing per model

Avoid using for:

• Production deployments (security risk, PII data leaks)
• Environments with sensitive user data
• High-throughput systems (adds latency overhead; unoptimised code)

Configuration¶

Enable the inspector in your config.yaml:

translators:
  anthropic:
    enabled: true
    max_message_size: 10485760

    inspector:
      enabled: true
      output_dir: "logs/inspector"
      session_header: "X-Session-ID"

Usage¶

Include the session header in your requests:

curl -X POST http://localhost:40114/olla/anthropic/v1/messages \
  -H "Content-Type: application/json" \
  -H "X-Session-ID: debug-qwen-factorial" \
  -d '{"model":"qwen/qwen3-coder","max_tokens":4096,"messages":[...]}'

Output Format¶

Each session creates a JSONL file with request/response entries:

{"type":"request","ts":"2025-10-21T10:30:00Z","model":"qwen/qwen3-coder","body":{...}}
{"type":"response","ts":"2025-10-21T10:30:03Z","body":{...}}

Analysing Output¶

Extract all requests:

cat logs/inspector/2025-10-21/debug-session.jsonl | jq 'select(.type == "request") | .body'

Extract all responses:

cat logs/inspector/2025-10-21/debug-session.jsonl | jq 'select(.type == "response") | .body.content[0].text'

Compare responses from different models:

jq 'select(.type == "response") | .body.content[0].text' logs/inspector/2025-10-21/compare-qwen.jsonl > qwen.txt
jq 'select(.type == "response") | .body.content[0].text' logs/inspector/2025-10-21/compare-glm.jsonl > glm.txt
diff -u qwen.txt glm.txt

Security Considerations¶

The inspector logs everything. Protect logs:

chmod 700 logs/inspector/
chown olla:olla logs/inspector/

Manual redaction if needed:

jq 'if .type == "request" then .body.messages = [] else . end' session.jsonl > redacted.jsonl