Secrets Hygiene in Agent Logs

PROJECT CONTEXT:
My agent's logs, traces, and memory are a goldmine of leaked secrets: API keys in tool output, tokens in prompts, credentials echoed by commands. Observability and security guidance is consistent - secrets must be scrubbed everywhere they could persist.

GOAL:
Implement pipeline-wide secrets hygiene: detection, redaction at every capture point, safe injection of real secrets, and verification that nothing leaks.

REQUIREMENTS:
- A secret detector: known-key patterns plus configured sensitive values and high-entropy heuristics
- Redaction at every persistence point: logs, traces, session records, memory, error reports
- Safe secret injection: real credentials reach tools without entering model context or storage
- Redaction in model-facing content too (the agent rarely needs raw secrets)
- A leak-scan test proving secrets do not reach any stored artifact

CONSTRAINTS:
- Redaction happens at capture time - never store raw then scrub later
- Real secrets flow only through a controlled path, never through the LLM or logs

DEFINITION OF DONE:
- A tool that outputs an API key has it redacted in logs, traces, and model context
- Configured secrets (env values) are redacted wherever they appear
- Tools needing credentials get them via a secure path, not from the model or storage
- A repo/artifact leak scan finds zero real secrets after a secret-heavy session

COMMON FAILURES TO AVOID:
- Secrets in plaintext logs and traces, one breach from disaster
- Redacting logs but leaving secrets in session records or memory files
- Passing real credentials through the model context 'because it is convenient'
- Assuming no leaks without ever scanning stored artifacts