Fix Loop

Rounds 1-3 — resume the original implementer. Send it the open findings verbatim. Its context is intact: it knows the task, the code, and its own choices. If your harness cannot send another message to a live subagent, dispatch a fresh implementer carrying the brief path, the report-file path, and the findings — the report file is the persistent memory either way.

Rounds 4-5 — dispatch a fresh implementer on a more capable model (per SKILL.md’s Model Selection), with the brief path, the report-file path, the open findings, and this framing: “A prior implementer attempted this task [N] times; you own it now. Read the report file for what was tried.” A loop that survives three resumes usually means the implementer cannot see its own problem — fresh eyes and a capability bump in one move.

Every round, either way: the implementer fixes, re-runs the tests covering the amended code, appends its fix report to the same report file, and returns the short contract. Before re-dispatching the reviewer, confirm the fix report contains the covering tests, the command run, and the output; dispatch the re-review once all three are present. Name the covering test files in the fix message — a one-line fix does not need the whole suite.

The re-review is scoped. Run bash scripts/review-package PLAN_FILE FIX_BASE HEAD by its absolute path, from the working tree you’re executing in, where FIX_BASE is the head the previous review saw, and dispatch re-review-prompt.md with the findings list, the brief, the report file, and the printed diff path. The re-reviewer verdicts each finding ADDRESSED or NOT ADDRESSED and flags new breakage in the fix diff only. New Critical/Important breakage in the fix diff joins the open findings list. Out-of-scope observations go to the ledger as deferred minors — they never extend the loop.

After each round, append to the ledger: Task <N>: fix round <R>/5 (<X> addressed, <Y> open — <finding one-liners>; commits <a7>..<b7>)

Never fix findings yourself in the controller session — your context stays clean for coordination, and controller fixes skip review.

The breaker. When round 5’s re-review still leaves findings open, stop dispatching. Adjudicate each open finding yourself — you hold the plan and the cross-task context the reviewer lacks:

• The reviewer is wrong, or the point is contestable: park it — Task <N>: parked — <finding> — Ruling: <why the code stands>. The final review sees both sides.
• Real, but nothing downstream builds on it: park it the same way, with a ruling that says it’s real and deferred.
• Real and load-bearing — a later task builds on it, or it reveals a plan defect: rule on the smallest change that unblocks the dependent work, ledger it as Task <N>: Ruling: <finding> — <what you decided and why>, and carry it into the next task’s dispatch. Parking a structural failure silently lets every dependent task build on it. Stop only when the defect leaves every path forward a guess.

Adjudicate only at the cap. Adjudicating earlier to end a loop is pre-judging with a different name. Every adjudication is a ledger entry — a silent discard is forbidden.