remediation-loop

Remediation Loop

When a code review or verification step surfaces findings with severity >= HIGH, run a bounded fix-then-re-review cycle instead of leaving the findings unresolved or asking the user to fix manually.

When to Use

• A code-reviewer agent returned HIGH or CRITICAL findings
• A verification step found failing tests or integration issues
• A silent-failure scan identified error-handling gaps
• Any quality gate failed with actionable findings

Protocol

Step 1 — Assess Findings

Classify each finding:

• Actionable: has a clear fix (wrong logic, missing check, broken import)
• Judgment call: requires user input (architecture change, scope decision)

If ALL findings are judgment calls, present them to the user. Do not enter remediation.

Step 2 — Fix (Bounded)

Dispatch a fix agent (or fix directly) scoped to ONLY the actionable findings:

Agent: Fix the following review findings. Do not refactor beyond what is needed.

Findings:
1. [file:line] — description from reviewer
2. [file:line] — description from reviewer

Constraints:
- Fix only what is listed. No drive-by improvements.
- Run existing tests after fixing to verify no regressions.
- Report what you changed and test results.

Step 3 — Re-Review

After the fix, dispatch a fresh reviewer (or re-run verification) scoped to the same areas:

Agent (read-only): Review ONLY the files touched by the fix.
Confirm whether each original finding is resolved.
Report any new issues introduced by the fix.

Step 4 — Decide

Re-review result	Action
All findings resolved, no new issues	Exit loop — mark complete
New issues found (cycle 1)	Run one more fix + re-review cycle (Step 2-3)
New issues found (cycle 2)	Stop. Present remaining issues to user. Do not loop further.
Original findings NOT resolved	Stop. Present to user with evidence of what was tried.

Hard Limits

• Maximum 2 remediation cycles. Never exceed this. If 2 cycles don't resolve it, the user needs to decide.
• No scope creep. Each fix agent is scoped to specific findings. Do not expand scope across cycles.
• Track what was tried. When presenting remaining issues to the user, include: original finding, what fix was attempted, why it didn't resolve.

Anti-Patterns

• Running remediation on LOW/INFO findings (not worth the cost)
• Expanding fix scope beyond the original findings
• Looping more than twice ("if 2 passes didn't fix it, a 3rd won't either")
• Remediating judgment calls without user input
• Skipping the re-review step ("I'm sure the fix is correct")