Intent is all you need.
One line of intent becomes planned, test-first, gate-verified software on Claude Code, Codex and OpenCode. A human approves the plan; cheap workers build it in isolated worktrees; deterministic hooks decide what is accepted.
aladdin "add CSV export to the report"
Give the agent something to make false.
Do not merely give an agent a goal. Give it a failing proof of what is missing, then make it earn GREEN. A test written after the code proves the code does what the code does. A test written before it proves the code is missing — and "done" becomes a claim a gate can check, not a sentence a model can say.
writes the tests the plan named — and must fail them.
gate-red-verifydeclares the symbols the tests need; no bodies.
gate-scaffold-verifymakes the tests pass, touching only the task's files.
gate-green-verifysymbol uniqueness, duplicates, orphans; then the sprint's whole matrix once. Every planned test must be on main.
Each step is a sub-agent whose stop is intercepted by a hook. The hook can block the stop and hand the reason back. A gate that crashes blocks. A host that cannot block natively is wrapped from outside. A BLOCK is never downgraded to a warning.
Models explore. Gates decide.
| The model may | The gate decides |
|---|---|
plan, split, pick a route (strongest · strong · economical) | whether the plan has tests for every task and targets that exist or are declared |
| write tests, scaffolds, code, reviews | whether RED fails, SCAFFOLD has no bodies, GREEN passes inside its scope |
| say it is done | nothing — done is a passing gate verdict in the event stream |
| spend tokens and minutes | within a box: soft warning, hard stop, breach classified and re-planned, never retried as-is |
The audit tree
Every run leaves an aladdin.event/1 stream. One command walks a commit back to the human who approved it, with no transcript archaeology:
aloop events trace 3e555af
commit 3e555af feat: csv export (built from the task's rationale)
└─ task S1-01-T2 "wire the export" worktree merged --no-ff
└─ attempt 2 green-worker · economical gate-green-verify: PASS
├─ attempt 1 gate-green-verify: BLOCK scope: touched src/report.py outside targets
└─ targets C2 modify src/report.py:4 render reason: render offers the export
└─ plan S1-01 approved human.approval · autonomy: reviewed
└─ intent int_01a14d9f… "add CSV export to the report" received 2026-10-18T06:07Z
Decisions carry a one-line reason, evidence refs and an authority ref. Usage is reported by class (fresh input, cache read, cache write, output, context) and marked estimated where a host does not report it.
Evidence, not claims
Every public number comes from the harness's own stats.md and is reproducible with make e2e-live YES=1: a one-line intent through intake, standards, planning, RED/SCAFFOLD/GREEN per task, structural review and the final gate, to merged and tested code in a real session. The semver intent ran end to end in 11m10s for $0.65 (run 6); a two-story word-frequency tool in 19m42s for $1.56 (run 9), and in 10m58s for $0.76 with no gate block once every dispatch ran on the project's model (run 23), and in 9m54s for $0.70 on today's master (run 27). Every number here is a row in docs/evidence/README.md with the command that reproduces it, and every requirement on the roadmap is tracked with the test that proves it in docs/TRACKING.md.