refactor

Refactor $ARGUMENTS. A targeted refactor includes the performance pass below — it is on by default, not a separate command.

Guardrails

• Spawn a worker and explicitly select your tool's mid tier. State the tier on the spawn — do not omit it and rely on a default. An omitted tier inherits the parent's tier, which is not the same thing as the balanced one. Pick the judgment-capable tier that is cheaper and faster than your top reasoning tier. Not the cheapest/fastest tier: on judgment work it measurably degrades (misclassification rates several times higher). Choose by tier, not by a vendor model name copied from this file — names drift, and this command ships to several tools. Fall back to running inline if your tool has no subagent mechanism.
• Escalate, never assume. Anything you cannot decide, cannot verify, or that this spec does not cover → stop and report it to the orchestrator (the main session). Never improvise, never widen scope, never fix a side issue you noticed along the way.
• The worker does the work itself — no delegation. The fixer must not spawn subagents of its own. Every edit it reports, and every test run it cites, has to be one it made or ran with its own tool calls: a relayed "I fixed it and the suite is green" from a sub-worker is hearsay, and this command's whole output is the claim that a change landed and the tests still pass. A fix that delegates its work is a report about a report.
• The HITL gates below belong to the orchestrator, not the worker. A subagent cannot hold a conversation with the user, so it cannot run a gate that ends in stop and ask. When one trips — a failing test, a crossed public API boundary, a change bigger than the bullet asked for — the worker stops there and hands the situation back, with the options and its reasoning but no choice made. The orchestrator asks. A worker that picks revert / patch / update-test on the user's behalf has answered a question it was never allowed to ask.
• Edit only what a surviving bullet names. Ledger mode's scope is the bullets that survive revalidation, one change per bullet — not the neighbouring code, not the formatting, not a second finding noticed on the way past. Anything else goes back to the orchestrator to become a new bullet.
• Prove the blast radius with two checks, because neither sees what the other does. git status --porcelain at exit must list only files a surviving bullet named — that is this command's scope guarantee, and unlike /branch-review it is not expected to be empty. It cannot police the memory directory: .amp/ is normally gitignored, so porcelain stays empty whether you deleted a fixed bullet, wrote nothing, or overwrote MEMORY.md. So also take md5sum .amp/remember/* before you start and again before you report, and show the comparison: only fix-ledger.md may differ. last-review.md in particular is /branch-review's to write — a fixer that touches it forges the gate that judges its own work.