failfirst (Python)

Acceptance criteria that must be proven to fail on the unfixed code before they may count as passing. A check that has never failed carries no information.

pip install failfirst
failfirst baseline        # before the work; writes .failfirst.lock.json
failfirst verify          # after; passes only if every criterion failed first, passes now, and is unchanged

Library:

from failfirst import load_criteria, baseline, verify, format_report

criteria = load_criteria("failfirst.yaml")
report, code = baseline(criteria, ".failfirst.lock.json")
# ... do the work ...
report, code = verify(criteria, ".failfirst.lock.json")
print(format_report(report))

Exit codes: 0 pass, 1 fail, 2 bad input, 3 judgment criteria awaiting human review.

Full documentation, the language-neutral spec and the conformance vectors: see the repository root (spec/SPEC.md). MIT licensed.

Agent use

failfirst describe --json (Anthropic / OpenAI / MCP tool definitions), failfirst mcp (MCP stdio server), --json on every command, stable FF_E_* error codes. See agent/AGENTS.md and agent/llms.txt in this package.