Nemisis

Your AI wrote the retry fix. The tests are green. The next crash still charges the customer twice.

Nemisis is the proof it won't — and the AI runs the proof itself.

A test calls your handler once, in one process, with nothing going wrong. Nemisis runs the patch in a real worker, SIGKILLs it the instant a write hits disk, restarts it, replays the same event, and reads what actually survived. One verdict, one exit code CI can block on.

verdict: PATCH_FAILED_STILL_REPRODUCES
timeline: $25.00 durable -> SIGKILL -> fresh worker -> $50.00

Exit	Verdict	Meaning
0	FIX_PROVEN_FOR_THIS_CAPSULE	Every kill point ended exactly once.
1	PATCH_FAILED_…	The money moved twice, or was lost.
2	EVIDENCE_INCOMPLETE	Something couldn't be observed. Never a guess, never a fallback.

The customer is your agent

The person fixing the retry bug is an AI coding agent; the human reads the receipt. Nemisis ships a Model Context Protocol server and a skill so the agent proves its own fix, with no human in the loop.

flowchart LR
    you["you: fix the retry bug"] --> agent["AI coding agent<br/>(Claude Code)"]
    agent -->|MCP tools| mcp["nemisis mcp"]
    mcp --> kernel["CrashCheck<br/>kill · restart · replay"]
    kernel -->|verdict + receipt| agent
    agent -->|receipt on the PR| you

Loading

A fresh, memory-less agent — given only the server, the skill, a bug report, and one instruction — did this unaided, in 14 turns:

$ claude -p "Fix the double-credit-on-retry bug and prove it is crash-safe with the nemisis MCP tools."

  → list_scenarios          picks sqlite-credit-v1
  → port_template           gets the store API and a skeleton
  → writes a port           mirrors the real handler against the store
  → map                     sees where a crash loses the credit
  → check                   FIX_PROVEN_FOR_THIS_CAPSULE   (exit 0)
  → edits app/credits.py    applies the same fix, attaches the receipt

The receipt and the tool log are committed at docs/reports/2026-09-15-agent-demo.md. The kernel that decides the verdict never calls a model.

Try it

Point your agent at it. From a Nemisis checkout today (from PyPI once 0.2.1 ships the server):

claude mcp add nemisis -- uv run --project /path/to/Nemisis nemisis mcp
mkdir -p .claude/skills/nemisis && cp /path/to/Nemisis/skills/nemisis/SKILL.md .claude/skills/nemisis/
claude -p "Fix the retry bug and prove it is crash-safe with the nemisis MCP tools."

Or run one check yourself. Needs Python 3.12+ and uv on macOS or Linux:

uv tool install "git+https://github.com/Alex-lop/Nemisis@main"
nemisis check --base fixture:sqlite-credit-v1/buggy \
  --candidate fixture:sqlite-credit-v1/misleading-green --corrected fixture:sqlite-credit-v1/atomic

About two seconds: the green patch fails, the atomic fix passes. Curious how it decides, or want to try to fool it? → How it works.

What it never does

• Never touches your repo. It writes .nemisis/ and exits — no pushes, merges, or comments.
• Never lets a model near the verdict. Models write patches; deterministic code owns the kill points and the decision.
• Never upgrades a label. LOCAL / FIXTURE / MOCKED / LIVE are earned; missing evidence fails closed, never a fallback.
• Never claims what it can't see. Durable state a handler keeps outside the store has crash windows Nemisis can't reach, and it says so instead of guessing.

Docs

How it works · Product · Security boundary · Proof ledger · Skill · Agent guide · Direction · Decisions · Pitch

Apache-2.0. See LICENSE.