We built a 9-item checklist that catches LLM coding agent failures before execution starts

After watching AI coding agents fail repeatedly on the same classes of problems, we identified the root causes. Here's what kills most agent runs before they start: C1 — Incomplete enum handling. Agent references status values that don't exist in the codebase. C2 — Silent null paths. Optional parameters get skipped silently with no documentation. C3 — SSE auth pattern mismatch. Browser EventSource can't send custom headers — agent uses wrong auth. C4 — Unbounded text fields. No truncation on columns that receive full task descriptions or diffs. C5 — Event/DB race condition. SSE event fires before the DB write completes. Frontend queries empty row. C6 — Schema/ORM mismatch. SQL type says nullable, ORM field says required. C7 — Untestable expectations. Test requirements with no implementation path in the spec. C8 — Non-idempotent inserts. Retry logic creates duplicate rows. C9 — Hallucinated imports. Module doesn't exist in the codebase. We now run this as a validation pass after plannin