switching memory off switches skills off, and the chat then denies the feature exists

[santoshkumarradha]

What happened

A person with eighty-one skills on disk asked the chat "can you use skills ?" and
was told:

If you meant skills in the Claude Code SKILL.md sense, codeaf has no such
mechanism — harnesses and the notebook are what it has instead.

The binary was built from the head that carries the skills wave and its manual
pages. The feature was there. The answer denies it exists.

Why

memory.enabled was off.

~/.codeaf/config.json   memory.enabled = off
~/.codeaf/graph.db      59.7 MB, facts table: 0 rows
~/.claude/skills        40 direct-child SKILL.md
~/.agents/skills        41 direct-child SKILL.md

importForeignSkillsBeforeFirstMessage (cmd/codeaf/chatv3.go:1169) returns on
its first line when memory is nil. Nothing is recorded, so
renderSkillCatalog (internal/session/skillcatalog.go:64) reads an empty
shelf and emits zero bytes, and useSkillTool (internal/session/tools_skill.go:71)
withholds the verb. Every one of those is behaving as designed.

The model then has no shelf, no verb, and no sentence about either, so it
reasons from the silence and invents a denial.

The two defects

One. Skills are gated on a setting that names something else, and nothing says
so. A person switches memory off to stop the program remembering things about
them. They also switch off skills, and are not told. That is one decision
producing two outcomes with only one of them disclosed.

Two. "Absent because impossible here" and "absent because it does not exist"
are the same silence. The absent-not-broken law works: it stops a model
planning a reply around a call that can only refuse. It does not stop the model
reasoning from the silence and concluding the capability was never built. The
law needs a companion clause rather than an exception: WHEN A CAPABILITY IS
WITHHELD BY A SETTING, SOMETHING HAS TO SAY SO.

This is not the emptiness law's territory. That law is about a number nobody has
yet: draw nothing rather than $0.00. It was never a licence to hide the
difference between "you have no skills" and "skills are switched off here".

Replication

1. A profile with memory.enabled = off and any number of SKILL.md folders
   under a scanned root.
2. Ask the chat whether it can use skills.
3. It denies the feature exists, rather than saying it is switched off.

Acceptance

End to end, through the real surface, in BOTH states.

• With memory off: the answer says skills are switched off on this machine and
  names the setting. It does not say codeaf has no such mechanism.
• With memory on: the shelf is there and the answer is yes.
• The /skill picker with memory off says the same thing rather than drawing an
  empty list.

Then whatever mechanism carries it. One candidate: the catalog emits a single
line when the store is absent and the roots are not, which costs a person with
no skills nothing, because that person has no roots either.

Not in scope

#1378 is the manual side: no heading answers the capability question in either
state. Fixing the sentence here does not fix retrieval there, and neither is a
substitute for the other.