Skip to content

Commit c60ccae

Browse files
committed
docs(config): skills-index measurement extended to frontier open models
kimi-k2.6/k3 and glm-5.2 join the 2026-07-10 finding: the passive index never fires organically (0/9 baseline). An assertive header and tool fronting, alone and combined, were measured too: kimi 1/12 total, glm-5.2 4/9 only with both levers (0/9 for each alone). Nothing reliable, so nothing shipped; always / --skill / /name stay the delivery paths when a skill must apply.
1 parent ec7688e commit c60ccae

1 file changed

Lines changed: 8 additions & 0 deletions

File tree

‎docs/config.md‎

Lines changed: 8 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -314,6 +314,14 @@ compliance (byte-verified delivery, MOOSE positive control). `always`, `/name`,
314314
and `--skill` are the reliable delivery paths for such models; an irrelevant
315315
index measurably distracted mistral-small. Prefer a small index on weak models.
316316

317+
Re-measured on frontier open models (2026-07-18, task tailor-made for one
318+
installed skill, wire delivery verified): kimi-k2.6, kimi-k3, and glm-5.2 all
319+
0/3 organic invocations under the shipped index. An assertive header
320+
("your FIRST action is use_skill"), fronting the tool in the list, and both
321+
combined moved kimi to 1/12 total; glm-5.2 reached 4/9 only with both levers
322+
(neither alone: 0/9). No lever made invocation reliable, so none shipped; the
323+
delivery paths above remain the answer when a skill must apply.
324+
317325
### Config profiles
318326

319327
A profile presets many settings at once so a task picks a strategy with one knob.

0 commit comments

Comments
 (0)