CrashCheck · the crash-safety proof an AI coding agent attaches to its fix

Kill it mid-write.
Read what survived.

Tests call the handler once and nothing goes wrong. CrashCheck kills the worker after every store commit, redelivers the same event to a fresh process, and reads what survived.

Run the check Open the evidence

move the kill

One command

One verdict. An exit code CI can block on.

$ uv run nemisis check --base fixture:sqlite-credit-v1/buggy \
    --candidate fixture:sqlite-credit-v1/misleading-green --mode local
NEMISIS CRASHCHECK — LOCAL
execution: COMPLETED
integrity: VALID
verdict: PATCH_FAILED_STILL_REPRODUCES
summary: The candidate replayed evt_1042 to a durable +$50 duplicate effect.
…
hypotheses: 2 run -> selected effect-commit (effect-commit-v1)
…
timeline: $25.00 durable -> SIGKILL -> fresh worker -> $50.00
…
exit code 1

Real stdout of that command. Each … stands for lines left out: the capsule, engine and commit digests; the no-crash control; the capsule, manifest and report paths. Nothing else is elided, and a candidate that fails its boundary worlds is never swept, so there is no sweep: line.

0FIX_PROVEN_FOR_THIS_CAPSULE

Every kill point replayed to exactly once.

1PATCH_FAILED

A crash window still duplicates the effect.

2EVIDENCE_INCOMPLETE

A write CrashCheck could not attribute. No verdict, the reason named.

uv tool install nemisis

Point your agent at it: a Model Context Protocol server (nemisis mcp) and a skill let a coding agent prove its own retry fix, unaided. Add it from a checkout today (claude mcp add nemisis -- uv run --project /path/to/Nemisis nemisis mcp); the published wheel ships it from 0.2.1. The kernel never calls a model.

Proof, labelled

What is proven. What is refused. What is not.

Proven locally: The double-credit hero, five kill worlds per role, the whole suite green on CPython 3.12, 3.13, and 3.14, a committed benchmark. CI counts the tests.
Refused, by design: A write around the store — a table, a header field, a rowid, a file, a permission bit. Two hostile reviews found eleven handlers an earlier engine blessed; fifteen such shapes are now pinned as refusals, each named in the summary.
Red-teamed nightly: Generated handlers graded by an independent oracle: 600 of 600 agreed at the engine that refuses what the nightly itself found (300 per scenario, seed 34820657707); 300 of 300 in its first run.
Not proven: A live model-written patch. LIVE stays BLOCKED until a receipt exists.

Read the proof ledger, the boundary, and the evidence viewer.

The open square is the write that never happened. CrashCheck is how you know.

View on GitHub Read the README