HetCreep Resilience Audit
Language: Generate EVERYTHING at runtime in the user's language — questions, answer options, menu labels, recommendations, report narrative. Detect from their messages; never default to English just because this file is English. English is allowed only for technical terms: commands, paths, code identifiers, severity labels (CRITICAL/HIGH/MEDIUM/LOW), and tier names (Light/Standard/Heavy).
Config reads — every config key, always the CASCADE, never the bare project file: ~/.claude/.coalmine.json first, then the project config (own agent dir → other known agent dirs → legacy <gitroot>/.coalmine.json), project wins per key. A bare project read is ABSENT on a machine configured only globally, so it silently yields defaults.
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 |
A denial reaches the WORKER as a visible message and propagates no further — never to a caller, never as a catchable condition. Every row above states a grant or an explicit death; a step that dies says so in the output, never as a false "done"/"skipped"/"clean".
- read denied → refuse before scanning; never a false clean bill.
- write denied → report the change as NOT applied — never claim done.
- network denied/unfetchable →
⚠️ unverified: check [source]. - spawn denied → degrade per Escalation's own capability-lever fallback (never fake parallelism) and say the fan-out did not happen — already discharged there; a row above is only for a spawn this skill does OUTSIDE tier escalation.
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)
Reporting: call ReportFindings when callable — file/line MUST be the defect site, never the enclosing function; an unresolvable line reports your best guess, named imprecise in the wrap-up — never dropped, never faked. Severity prefixed in summary (e.g. [HIGH] …), ranked most-severe first, SUSPECTED as verdict: PLAUSIBLE; chat then carries only the wrap-up line (counts · coverage gaps · overflow past 32 · any imprecise-line findings) + the fix menu, never a restatement of findings. Not callable → the table above, unchanged. An Apply-fixes click = consent to the safe-fix class only — gated the same as this skill's own fix-mode (Hook Context needs an interactive session, per the Hook Context rule below) — composing with (never bypassing) the fix-mode discipline. After any fix round, re-report the same findings with outcome: fixed/skipped/no_change_needed — skipping this leaves the round UNFINISHED.
Escalation — Scope & Model Quality
Tiers are capability targets, not platform commands — resolve each to your host's nearest lever. No lever for one? Degrade gracefully — never fake parallelism you can't do; escalate via model tier + reasoning depth instead.
| Level | Intent | Capability target | Cost |
|---|---|---|---|
| Light | Spot failure-mode check, key paths only | Cheapest model · single agent, no sub-agents. | Low |
| Standard | Balanced FMEA, multi-category coverage | Balanced model · raised reasoning · sub-agents per category only if your platform runs concurrent workers (else single-agent). | Balanced |
| Heavy | Full 8-category FMEA + adversarial verify | Most capable model + largest context · deepest reasoning · max sub-agent fan-out if supported · adversarial cross-check where available. | High |
Per-platform Heavy levers + Heavy-run durability: read references/escalation.md before a Heavy run. No concurrent fan-out on your host → escalate by model + reasoning only.
Agent Context (interactive): score the tier rubric, then call ask_question once with the 3 tiers — the pick marked ✓, score shown, labels localized — and wait for the choice before starting. ask_question = your platform's question tool: Claude Code AskUserQuestion · Cline ask_question · Copilot askQuestions · Gemini CLI ask_user (business-tier product; individual tiers ended 2026-06-18 → Antigravity CLI) · Codex request_user_input · Cursor/Devin Desktop (ex-Windsurf)/Antigravity built-in prompts; none → numbered text menu.
Tier rubric (deterministic): +1 each — ① >20 files or whole-repo/cross-module reach ② >2 of this skill's categories relevant ③ release/security/pre-ship context ④ findings will drive code changes. 0–1 Light · 2–3 Standard · 4 Heavy. Freshness cap: scope already audited ≥Standard this session → cap at Light (re-auditing fresh ground wastes tokens; scope to what changed). Default tier: honor .coalmine.json defaultTier unless the user requests a tier for that run — an explicit request overrides everything.
Hook Context (auto-triggered): auto-Light, no tier question, no sub-agents — report first. Interactive session (a user is present) → follow this skill's own Fix mode section, if it defines one, for what to offer after the report; non-interactive → report-only. Where a Fix mode section exists, never fix without a chosen option.
Entanglement: after the report, if confirmed findings fall in another canary's domain, offer it once via ask_question (one line, max one offer): perf/N+1 → scale-canary · contract/serialization/config → drift-canary · failure-path/retry → resilience-audit · logging/metrics → telemetry-canary · coupling/DI → testability-canary · dependency/CVE → supply-chain-audit · unverified version-sensitive claim → source-grounding · missing/stale rule → gold-standard.
Self error-report: if this skill misbehaves (contradictory instruction, broken procedure, wrong finding class), OFFER to file it at https://github.com/HetCreep/CoalMine/issues/new/choose with a user-reviewed summary — never auto-submit, never include unapproved code or paths.