HetCreep Resilience Audit
For every operation: "what happens when this FAILS?" Report; do NOT fix unless asked.
Failure categories
- External I/O — network down/slow, API 4xx/5xx/timeout, rate-limit. Retry w/ backoff? Timeout set? Clear error vs hang?
- Storage — disk full, permission denied, partial write. Atomic write (temp+rename)? Cleanup on failure? Existing good copy untouched?
- Partial completion — half-done op (50/100 files). Reported as FAILURE, never success.
- Crash / OOM — killed mid-op. Idempotent restart? No orphaned half-state?
- Concurrency — two instances, race, deadlock. Locking / idempotency / safe re-entry?
- Input / data — malformed, null, truncated, huge. Validate at boundary? Fail-fast?
- Dependency down — fallback/cache/graceful degrade? Clear error vs silent hang?
- Resource exhaustion — bounded? Backpressure? Cleanup on error path?
Per-stack timeout/atomicity/idempotency patterns to grep: read references/checks.md before scanning.
For each failure point, check 4 things
- Detected? code notices it (doesn't swallow)?
- Handled? retry/fallback/fail-clean — not ignored, not silent-success?
- Recoverable? rollback/idempotent; no data loss or corruption?
- Communicated? clear error to user+log; not a hang, not a false "done"?
Discipline
- Trace actual failure path (cite file:line). Don't assume handling exists; prove it.
- "partial = failure" — any path reporting success on partial completion = CRITICAL.
- "logged" ≠ "handled" — swallowed+logged error that corrupts state or returns success = CRITICAL.
Fix mode (choice-gated)
After the report, present via ask_question:
- Fix safe ones — add missing timeout, null/input validation, clear error+log on unhandled path. Each: checkpoint → fix → build+tests → revert if newly red.
- Let me pick — user-selected fixes only.
- Report only — change nothing.
NEVER auto-fix: retry/rollback/recovery/atomicity logic (semantic changes can introduce new failure modes).
Grants & denials (CLASSIFY-BLOCK)
| class | step it powers | grant | on denial |
|---|---|---|---|
| read | trace failure paths for the 8 categories above | Read·Grep·Glob |
refuse that file, name it — never a clean bill |
| write | Fix mode's safe-guard apply, incl. checkpoint → build+tests → revert if newly red | Edit·Bash (checkpoint/build/revert need exec) |
report the fix as NOT applied AND the checkpoint/revert as NOT available, never claim done |
Output
| operation | failure mode | effect | handling (file:line) | severity | recommended guard |
Ordering/atomicity findings · Summary (counts + top fixes) · Not assessed
Severity: CRITICAL (data loss/corruption/silent-success) · HIGH (crash/hang/partial-no-recovery) · MEDIUM (poor degradation/missing retry) · LOW (cosmetic)