Overview Shipped

What it does

You author an acceptance contract — atomic criteria, each with a check hint — and pin it before any code is written. Then the loop runs: pick a lane, brief a coder, verify every citation resolves, commit, and put the result in front of a conformance oracle. The oracle answers per criterion: met, unmet, or divergent — done in a way that conflicts with the intent. Divergent is the verdict that matters, because it is the one a naive fix can hide.

One iteration

guard    refuse to start if the run is over budget
build    a depth-0 builder makes one small increment
commit   git commit the WIP — every accepted increment is revertible
gate     myteams devteam review conformance HEAD --base <base>  → merge-readiness
record   write the measured facts, get back a next action
decide   keep · reset --hard HEAD~1 on regress · escalate once on stall · stop

The stop is decided in Python

Whether to continue is never re-derived by a model. A deterministic recorder compares this iteration's criterion statuses to every prior one and returns a next action the driver executes verbatim: finish_ready, continue, rerun_gate, reset_and_retry, retarget_unjudged, escalate_once, or halt with a reason. The halt ordering is fixed — done, then oscillating, then budget, then the iteration backstop — so two halts that both apply always resolve the same way.

The guards that exist because something went wrong

• Carry-forward semantics — only criteria this iteration targeted trust the fresh verdict; everything else keeps its prior status. Without it the oracle silently forgets what it already proved.
• Anchor checking — a coder can hallucinate an execution claim, but not a diff anchor, so every cited line is checked against the increment.
• A learned backstop — the diminishing-returns inflection is computed from cross-run data per contract size, not guessed at.
• Budget guard before dispatch, so an over-budget run halts before spending rather than after.

Contracts are immutable input

The contract is pinned before the work and never quietly edited mid-run to make a gap disappear. Criteria have to be atomic, and soft criteria are actively discouraged — an observed failure had "log everything" technically satisfying "surface real errors" while reintroducing exactly the suppressed noise the criterion existed to prevent.

Run it

myteams devteam plan code                     # interview → a pinned contract
/myagenticteams:devteam build --contract <path>
myteams devteam build stats                   # the convergence curve