a working miniature of a self-driving development loop
Put a spec document in, and the loop plans, implements, runs the acceptance tests, reads the failures, revises, and repeats until green or until it hits a cap you set. Skills load from files and change the run. A knowledge file grows, but only when a test proves the learning. Every number on this page is computed by the run you trigger; nothing is precomputed. The one honest substitution: no hosted model is called here, so the router's hosted slots visibly decline and a deterministic rule engine fills the implementer slot behind the same interface. That keeps this page free to run and reproducible.
spec intake
S3 contains a contradiction on purpose. The harness must stop and say so, not burn budget.
skills on file
A skill is a markdown file the loop reads. Toggle one and re-run: iteration 1 changes.
model router
Chains are walked for real, per task class. Hosted slots hold no credentials on this public page, so each run shows them declining and landing on the local floor. In your harness, the credentials go here.
budget caps
Units stand in for tokens and spend. Set the iteration cap to 1 and run S2: the loop stops honestly instead of pretending.
knowledge file
Held by your browser between runs on this stateless page; in the real harness it is a file in the repo. Entries carry the tests that proved them.
run console
selftest, recomputed on this call
The harness pattern runs end to end: spec files drive plans, code is generated and executed against acceptance tests, failures are classified and turned into one targeted revision per iteration, caps and stop conditions hold, an unsatisfiable spec is detected and refused, skills change behavior from files, knowledge grows only with test provenance, and the router walks real fallback chains per task class. GET /api/selftest recomputes all of it on every call.
No hosted model runs here, and "a full application, unattended, from one prompt" is not a thing anyone can honestly sell. The implementer slot is played by a deterministic rule engine behind the same router interface. In a live harness that slot is Claude Code or API models with your keys, and the human gates stay load-bearing: acceptance decisions, credentials, anything irreversible. Specs here are synthetic, written for this page.