atlas-review-handler

Atlas Review Handler — Review Workflow Orchestrator

CRITICAL RULES (LOOP PREVENTION)

1. MAXIMUM 2 REVIEW CYCLES per original task. After 2 cycles, STOP — run CRITICAL CLOSEOUT (RULE/BLOCK/PARK) on every remaining CRITICAL finding. Never silently demote a CRITICAL finding to INFO.
2. Fix tasks ([REVIEW-FIX]) do NOT trigger reviews. The REVIEW-ENFORCER plugin already skips [REVIEW-FIX] markers — but YOU must also not spawn reviews for fix results.
3. Track your cycle count explicitly. Maintain a mental counter: "This is review cycle N of 2."
4. If a review sub-agent runs build/test commands, that sub-agent is malfunctioning. The review-protocol skill explicitly forbids build/test. If you see build output in review results, STOP the cycle — the review skill is not loaded properly.
5. NEVER nest reviews. A review of a review of a fix is an infinite loop. One review per cycle, maximum 2 cycles.

WORKFLOW

When REVIEW-ENFORCER fires (after a task completes):

Cycle 1:

Step 1: Delegate the review

task(
  category="unspecified-low",
  load_skills=["review-protocol"],
  run_in_background=true,
  description="Review [task-name] changes",
  prompt="[REVIEW-TASK] Review the changes made by the previous task. Run git diff to see the changes, analyze them, and return structured findings in CRITICAL/WARNING/INFO format. DO NOT run builds or tests. DO NOT modify files.

IF the task produced or modified data artifacts (pipeline stages, generated datasets/corpora, model outputs, scored results): ALSO read the actual intermediate output files (at least 5 per stage) — raw values, not aggregates — and check for degenerate patterns: empty/whitespace outputs, zero-variance or all-perfect scores across conditions meant to differ, identical outputs across variants, duplicated template clauses, cardinality mismatches at stage boundaries. Report any degenerate pattern as CRITICAL; downstream stages silently mask upstream bugs.")
)