Temper

All subagent dispatches use disk-mediated dispatch. See shared/dispatch-convention.md for the full protocol.

The finder angles, the verify gate, the effort tiers, the cap semantics, and the eight-field output schema are NOT redefined here — they live in shared/delve-engine.md. /temper drives that engine: Round 1 dispatches it (bug-angle subset, high effort) to enumerate the tracked set T, and each later round drives it again over the fixed regions to hunt new gating findings (Track A, Step 4).

The severity and verdict vocabularies, and the gating rule T = {CONFIRMED, PLAUSIBLE} × {Critical, Important}, are the contract's — not temper's (I11). See shared/severity-verdict-contract.md; temper consumes that gating rule to build T and defines no scale or verdict of its own.

Like tempering steel after forging — iterative heat-and-quench cycles that set final hardness and elasticity — /temper runs successive fresh-eyes review rounds until the change converges. Round 1 drives the delve-engine fan-out (bug-angle subset) to enumerate a tracked set T of gating findings; each later round re-verifies whether every member of T is resolved against the fixed code and admits any new gating finding the fix introduced. The loop exits when T is fully resolved and no new gating finding entered.

Core principle: Review early, review often. Fresh eyes every round — but the per-round instrument is the engine's parallel finder fan-out + verify gate (plus a per-member re-verification pass), not one holistic reviewer. Convergence is the resolution status of an enumerated finding set, never a cross-round count comparison. (Distinct from /audit: temper drives delve-engine's instance-bug fan-out — one-reproduction defects; /audit runs systemic lenses — different machines.)

Renamed from /code-review (2026-05-17) to avoid collision with Claude Code's built-in /review command. Same iteration behavior; the argument shape and platform-agnostic PR support are new.

Non-Goals