refactor(skill): steer find-then-act instead of open-ended discovery

The bundled skill's Discovery section told the agent to run the bare
dashboard and reach for --help proactively, and advertised overlapping
find-paths — inducing exploratory commands that made the gitea-axi arm the
most expensive of the benchmark's four. Replace it with a "find the target,
then act" section, name the non-obvious mutation flags so common edits do
not need --help, and drop the setup line and the over-tea/raw/git bullets
that only duplicated the description. A same-time A/B cut cost-equivalent
tokens ~10% and collapsed bare-dashboard use from 60% to 7% with no loss of
success.
This commit is contained in:
2026-07-19 09:41:13 -04:00
parent dd58cd9dad
commit d56b1d0707
2 changed files with 32 additions and 37 deletions

View File

@@ -36,19 +36,23 @@ describe("bundled Agent Skill markdown", () => {
it("references each command group as a one-liner", () => {
const body = skill.toLowerCase();
for (const group of ["issue", "pr", "label", "search", "setup"]) {
for (const group of ["issue", "pr", "label", "search"]) {
expect(body, `expected the skill to mention the ${group} command group`).toContain(
group,
);
}
});
it("points at the bare dashboard and per-command help for discovery", () => {
it("steers find-then-act and does not push exploratory discovery", () => {
const body = skill.toLowerCase();
// Bare dashboard: running the binary with no arguments.
expect(body).toContain("no argument");
expect(body).toContain("dashboard");
// Per-command help.
expect(body).toContain("--help");
// The intended steering: find the target, then act on it.
expect(body).toContain("find the target");
expect(body).toContain("act on it");
// The find/act discipline: not a full survey, and not both find commands at once.
expect(body).toContain("not a survey");
expect(body).toContain("not both");
// The removed exploratory anti-pattern must be gone.
expect(body).not.toContain("dashboard");
expect(body).not.toContain("no argument");
});
});