Šķîþ ţö ḿàîñ çöñţéñţ

Šķîļļ éṽàļ

Àñ àĝéñţ îš ĥàñđéđ à þŕöḿþţ àñđ à ŵöŕķšþàçé. Éṽéŕý ŕöŵ éẋþàñđš îñţö ţĥé éẋàçţ þŕöḿþţ, ţĥé ƒîļéš ţĥé àĝéñţ šàŵ, ţĥé ţööļš àñđ çöḿḿàñđš îţ ŕéàçĥéđ ƒöŕ, àñđ à đ öƒ ŵĥàţ îţ çĥàñĝéđ.

21 þàšš
1 ƒļàķý
0 ƒàîļ
1 ƒàļšé ţŕîĝĝéŕš
24đ šîñçé ţĥîš ŕüñ
ŕüñ 2026-08-27T22:34:43Z · 3 þàššéš þéŕ šçéñàŕîö àĝéñţ 2.1.228 (Claude Code) · the claude CLI's default, not pinned šķîļļ 4c682ac83 · ļàšţ éđîţéđ 2026-08-15 kapi bin/kapi · kapi v1.2.0-rc29-9-g4b8194cf7 (beta, commit: 4b8194cf7, built: 2026-08-27T22:02:21Z) ĥöšţ Darwin arm64 šéţţîñĝš claude -p, bypassPermissions, 3 repeat(s), trigger cap 4 turns (MCP scenarios use their own, since picking a tool can take a step or two). Sampling follows the CLI's defaults and is not pinned.

The agent has to notice kapi is relevant. The only lever is the skill's description, and a false trigger costs more than a miss: missing a positive wastes one kapi user's prompt, firing on a code task wastes everyone's.

Ŵĥàţ kapi àđđéđ

Éàçĥ šçéñàŕîö ŕàñ à šéçöñđ ţîḿé ŵîţĥ ñö šķîļļ, ñö ḾÇÞ šéŕṽéŕ, àñđ ñö kapi öñ ÞÀŢĤ. Ţĥé çöḿþàŕîšöñ îš đéļîƃéŕàţéļý çöñšéŕṽàţîṽé: kapi ƒàîļîñĝ îš ñéṽéŕ çöüñţéđ àš à ŵîñ, ţĥé ḿéššàĝé çöüñţš àŕé ḿéđîàñš, àñđ à šçéñàŕîö ŵîţĥ ñö ĝàţé îš üñķñöŵñ ŕàţĥéŕ ţĥàñ àššüḿéđ.

Ŵĥàţ îţ çöšţ. Àçŕöšš 22 šçéñàŕîöš ţĥé àĝéñţ šéñţ 197 ḿéššàĝéš ŵîţĥ kapi àñđ 164 ŵîţĥöüţ, àñđ ţĥé üñàîđéđ àŕḿ ŵàš šĥöŕţéŕ öñ 15 öƒ ţĥéḿ. Ţĥàţ îš ţĥé ĥöñéšţ çöüñţéŕŵéîĝĥţ ţö ţĥé çöüñţš ƃéļöŵ: öñ ţĥîš šüîţé kapi ŕéàçĥéš àñšŵéŕš ţĥé üñàîđéđ àĝéñţ çàññöţ, àñđ îţ îš ñöţ ţĥé çĥéàþéŕ ŕöüţé ţö ţĥé öñéš îţ çàñ.

0
kapi enabled it
the unaided agent could not finish and the one with kapi did
0
kapi eased it
both finished, and kapi took materially fewer messages
0
kapi hindered it
the unaided agent finished and the one with kapi did not
17
no difference
both finished, and kapi saved nothing measurable
5
not comparable
no gate, so there is no outcome to compare

Ḿüšţ ƒîŕé (17)

Ḿüšţ šţàý ǫüîéţ (5)