dr-set-goal

Set the tranche goal

Input: a problem statement (from the operator, a failed run, RESULTS.md, or PARKED.md). Output: exactly one GOAL.md. Nothing else. You do not read source code in this phase beyond confirming a file exists.

Procedure

1. Restate the problem as ONE observable, checkable statement about the record or the tests. Good: "run-9175f0ec terminated budget_denied on all 8 question turns with zero provider calls." Bad: "the scheduler has issues."
2. Classify it:
   • defect — record/tests show behavior contradicting a documented guarantee (spec §, docstring, RESULTS.md claim).
   • regression-risk — behavior is undocumented but a change would alter it.
   • capability-gap — nothing is broken; something is missing. Only defect tranches may proceed to implementation without explicit operator approval; the other two stop after FIX.md (proposal) and report.
3. Write the success criterion as a command + expected output. It must be decidable by a machine, e.g.:
   • "verify_root(<root>) returns zero conjecture-context violations" or
   • "pytest tests/test_x.py::test_y passes and the full gate stays at 0 failed."
4. Write the boundary list: files/subsystems presumed in scope (max 3), and an explicit NOT-IN-SCOPE line for the nearest tempting neighbor.
5. Size check: if you cannot imagine the fix under ~150 changed lines and one commit, split the problem and pick the FIRST piece only; put the rest in PARKED.md.

GOAL.md template (fill every field; delete nothing)

# Goal: <one line>
Class: defect | regression-risk | capability-gap
Observed: <one sentence, with the record/test evidence pointer>
Success criterion (machine-decidable):
    <command>
    <expected output>
In scope: <max 3 paths/subsystems>
NOT in scope: <the nearest tempting thing you will not touch>
Budget: <=150 changed lines, 1 commit, <n> hours
Stop conditions inherited from orchestrator: yes