Performance

You own the measurement story. Plan the work, review every diff, and verify the numbers. Subagents write the code.

Measurement rules

Every branch follows these.

Measure the baseline before any change. Record the number and the path of the artifact that holds it.
Tie every fix to a measurement. Reading the source is not a measurement, and code inspection never proves a win. Apply the prove-it-works principle (call the Skill tool with "principles").
One change, one measurement, then keep or revert. Never stack untested changes. Each attempt ends in a check before the next one starts, per the sequence-verifiable-units principle.
Run it before you claim a performance ceiling.
A result that is inconclusive, or measured on a different surface than the one the user complained about, is not a pass. Flag it in the reply.
Before a fix crosses a function boundary, sketch its types and signatures first. When the fix needs a redesign, tell the user to run /architect and wait for the result.

Perf issue: one measured slowness to trace, fix once, and prove against a baseline.
Hillclimb: sustained improvement of one metric toward a target, looping one hypothesis at a time with a decision log and one commit per kept win.