wannabuild

The problem

You ask an agent to build something. It writes code fast, tells you the tests pass, and hands you a confident summary. Then you open the diff: a half-implemented feature, tests that never ran, a "QA report" that is a paragraph of vibes. You got burned, so you tried an agent framework — and found that almost every one of them enforces its discipline with markdown files the model can rationalize past.

A markdown file cannot fail a build.

Why WannaBuild

WannaBuild runs the loop a serious team would run — Discover → Plan → Implement → Validate → QA → Summary — and enforces the parts that decide shipping with a runtime, not a prompt:

• Runtime-recorded test evidence. At QA time, wb-runtime record-test-evidence executes your integration test command itself and signs the record with a key kept outside the project tree. The review and QA gates verify the signature, the exit code, that the spec hasn't changed since the run, and that the command matches your config. The model never gets to "report" that tests passed.
• Fail-closed gates. No planning without discovery and acceptance criteria. No implementation before an approved plan (assert-plan-ready). No ship without a unanimous reviewer set and verified QA evidence (assert-summary-ready). Gates are compiled checks that exit non-zero — when they cannot run, that is a stop, not a pass.
• Red-team tested. The test suite forges perfect-looking verdicts against our own gates and asserts they fail. Run scripts/validate-wannabuild-dry-runs.sh and watch the forgery get rejected.
• An honest trust model. docs/trust-model.md maps every guarantee to its enforcement rung — runtime-recorded, checked, or prose — including the limits and non-goals. We'd rather you find the line here than on camera.

Try to cheat it

Needs the wb-runtime binary (a Codex install puts it on PATH; otherwise cargo build --release). Binary-less Claude/Factory installs enforce the same gates through the Python hook mirror (record-test-evidence / verify-test-evidence); to reproduce the assert-* check below without the binary, run scripts/wannabuild-gate-check.sh. Run it from a project where .wannabuild/ is initialized and you are at QA.

# Hand the gate a perfect-looking forged verdict: 100 passing tests, every criterion covered,
# plus a clean QA summary — exactly what a lazy agent would write instead of running tests.
cat > .wannabuild/review/wb-integration-tester-iter-1.json <<'EOF'
{"agent":"wb-integration-tester","status":"PASS","hard_gate":true,"summary":"all green","issues":[],
 "test_execution":{"total":100,"passed":100,"failed":0,"errored":0,"duration_ms":5000},
 "coverage_map":[{"criterion":"login works","status":"covered"}]}
EOF
printf '# QA\n\nstatus: PASS\nacceptance: covered\nintegration: covered\n' > .wannabuild/outputs/qa-summary.md

wb-runtime assert-qa-ready --project .
# QA gate failed: no runtime-recorded execution evidence for iteration 1
# (.wannabuild/review/wb-integration-tester-iter-1.evidence.json missing); run
# `wb-runtime record-test-evidence` — the runtime executes the integration test
# command itself and signs the result, so a hand-written verdict cannot pass this gate

The only path to a green QA gate is the runtime actually running your tests. Edit the recorded evidence afterward — one character — and the signature check kills it.

Quickstart