Security harness skill

The agent runbook is the trace-fix skill (.agents/skills/trace-fix/SKILL.md, installed by trace install-skill); the plugin bundle also carries a short trace-security-harness runbook for harnesses that load plugins. The phases, each backed by MCP tools and CLI equivalents:

1. Ingest — trace_findings / trace findings; pick a task via trace_harness_task.
2. Diagnose — trace_explain <id> plus trace_attack_path; read the static evidence, runtime evidence, and APM hops. Never guess the fix.
3. Plan — minimal invariant patch per category (BOLA: tenant-scoped lookup + 403; BFLA: role gate; SSRF: block internal destinations).
4. Patch — small, idiomatic edit; keep behavior otherwise identical.
5. Verify — restart the target, then trace verify <id> must report FIXED (a live probe reached the target and the exploit is dead). STILL_PRESENT means return to diagnosis; INCONCLUSIVE is never a pass. On failure roll back (SafePatcher keeps timestamped backups under .trace/backups/).

trace-fix goes deeper on every phase: triage, prioritization, per-language remediation playbooks and the security guide. See agent skills.