feat: optional intercepting proxy for agent traffic (#50)

[itscooleric]

Summary

mitmproxy-based HTTP(S) interceptor for deep agent traffic analysis.

• intercept-proxy.py — mitmproxy addon that logs request/response metadata to JSONL
• intercept-start.sh — starts the proxy, exports HTTP(S)_PROXY env vars
• Auto-starts from claude-entrypoint.sh when CLIDE_INTERCEPT=1
• Secret redaction on API keys, tokens, auth headers
• Optional body capture with CLIDE_INTERCEPT_BODIES=1
• mitmproxy 11.1.3 added to Dockerfile (in pyenv)

Example output

{"event": "intercept_request", "method": "POST", "url": "https://api.anthropic.com/v1/messages",
 "host": "api.anthropic.com", "status_code": 200, "request_size": 4521, "response_size": 1893,
 "duration_ms": 2341, "request_headers": {"authorization": "[REDACTED:Bearer...]"}}

Observability stack now

Feature	Env var	Output
Session logging	(always on)	events.jsonl + conversation.jsonl
Token/cost tracking	(always on)	session_end event
Egress audit	CLIDE_EGRESS_AUDIT=1	egress.jsonl
Intercept proxy	CLIDE_INTERCEPT=1	intercept.jsonl

Closes #50

Test plan

• CI checks pass
• Build with mitmproxy doesn't bloat image excessively
• Claude session works with proxy enabled
• intercept.jsonl captures API calls with redacted secrets

🤖 Generated with Claude Code