Problem

The compaction summarizer (src/synthorg/engine/compaction/summarizer.py, line ~226-228) extracts msg.content[:100] from every ASSISTANT role message and joins them into a SYSTEM message that is permanently injected into the agent's context. If the LLM response embedded sensitive tool output (credentials, PII, internal file paths) in its content field, those first 100 characters survive compaction and are sent to the LLM indefinitely.

Impact

Expanded by PR #601 -- compaction_callback is now wired into all auto-selected loops (previously only available to statically configured loops), substantially increasing the surface where compaction triggers.

Suggested Fix

In _build_summary, strip or redact known sensitive patterns (secrets-style regexes, file paths) from snippet text before joining, consistent with the sanitization already done in _handle_fatal_error (agent_engine.py lines ~1242-1253).

Requires product-level decision on what constitutes "sensitive content" in the summarizer context.

Context

Discovered during PR #601 review (security-reviewer agent, confidence 85).