cc @JesperSchulz
Noticed while walking the routing flow for a community-layer PR: the two contracts that specify layer precedence disagree on the middle tier.
READ (skills/read.md, knowledge files)
/custom/ wins over /microsoft/ and /community/.
/microsoft/ wins over /community/.
Ordering: custom > microsoft > community.
Entry (skills/entry.md, action skills)
Skill layer precedence is /custom/ over /community/ over /microsoft/ — the same ordering READ defines for knowledge files.
Ordering: custom > community > microsoft. The inline claim that this matches READ is incorrect.
Why it matters
- Both orderings are individually defensible. Knowledge prefers authority (
/microsoft/ wins). Skills plausibly prefer local customization (/community/ overrides Microsoft defaults). But the two contracts shouldn't silently disagree, and Entry's comment shouldn't assert an alignment that doesn't hold.
- This becomes a real bug when a community contributor ships a skill with the same
id as a Microsoft skill: under Entry's current rule the community skill wins, which is a surprising outcome given READ's framing.
Options
- Align on READ's ordering (
custom > microsoft > community for both) — skills layer behaves the same as knowledge; Microsoft-endorsed skills always win unless a consumer fork overrides in /custom/.
- Keep Entry's ordering and fix the text — keep skills as
custom > community > microsoft with an explicit rationale, and correct READ's claim of equivalence.
- Document both and explain the split — if the asymmetry is intentional, call it out in both files.
No preference from me — just flagging before the first community skill lands, since a community al-performance-review under today's rule would supersede the Microsoft one.
cc @JesperSchulz
Noticed while walking the routing flow for a community-layer PR: the two contracts that specify layer precedence disagree on the middle tier.
READ (
skills/read.md, knowledge files)Ordering: custom > microsoft > community.
Entry (
skills/entry.md, action skills)Ordering: custom > community > microsoft. The inline claim that this matches READ is incorrect.
Why it matters
/microsoft/wins). Skills plausibly prefer local customization (/community/overrides Microsoft defaults). But the two contracts shouldn't silently disagree, and Entry's comment shouldn't assert an alignment that doesn't hold.idas a Microsoft skill: under Entry's current rule the community skill wins, which is a surprising outcome given READ's framing.Options
custom > microsoft > communityfor both) — skills layer behaves the same as knowledge; Microsoft-endorsed skills always win unless a consumer fork overrides in/custom/.custom > community > microsoftwith an explicit rationale, and correct READ's claim of equivalence.No preference from me — just flagging before the first community skill lands, since a community
al-performance-reviewunder today's rule would supersede the Microsoft one.