Secure Fix Handoff¶

Safe handoff of a security finding to a coding agent for patching.

Pre-conditions¶

• Finding eligible per security-loop.md criteria.
• Owner has authorized automated fix handoff for the repo.
• Authorization scope policy allows code changes for this target.

Allowed patch scopes¶

• Files explicitly named in the finding's Affected component.
• Files those depend on, only if cited in the remediation recommendation.
• New tests in the repo's standard test layout.

Anything outside these is rejected and triggers human escalation.

Required tests before merge recommendation¶

• The newly authored regression test reproducing the finding (red without patch).
• The same test green with patch applied.
• The repo's existing test suite green.
• A re-run of the security test plan in the sandbox green or downgraded severity.

Verifier review requirements¶

• The verifier subagent confirms the patch:
• is within allowed scope,
• reproduces the finding pre-patch,
• resolves the finding post-patch,
• introduces no new high/critical findings.
• The security-checklist skill runs against the patch diff.

Rollback path¶

If a patch:

• degrades functionality,
• introduces a new finding of equal-or-greater severity,
• causes test regressions,
• causes runtime instability,

then revert the patch on the sandboxed branch, mark the finding partial / iterate, and consult the security-loop iteration cap (default 3).

Coding agent constraints¶

• Operates in the bounded edit loop with verifier and tests required.
• May not access user-only secret store.
• May not exfiltrate target code.
• May not alter audit logs.
• Audit events: fix_proposed, fix_applied, fix_reverted per attempt.

Approval gate¶

Final merge to a non-sandboxed branch requires explicit human approval. Until then, all work stays in the sandbox branch.