Subagent `skills:` frontmatter preload does not inject skill content when the skill has `disable-model-invocation: true`

Summary

When a subagent's frontmatter declares skills: [<name>], the documented behavior
is that the full SKILL.md content is injected into the subagent's context at
startup (see sub-agents.md → Preload skills into subagents):

"The full content of each skill is injected into the subagent's context, not just made available for invocation."

In practice, if the target skill has disable-model-invocation: true in its
SKILL.md frontmatter, this explicit preload is silently skipped and the
skill is also unusable via the Skill tool from inside the subagent.

This conflates two orthogonal concerns:

• disable-model-invocation is documented as preventing the **model from

auto-discovering/auto-invoking** a skill based on its description.

• Subagent skills: is an explicit, developer-authored preload — the

decision has already been made at configuration time; no model auto-invocation
is involved.

Blocking the second on the basis of the first breaks the documented subagent
preload contract.

Related but distinct: #43875 covers the top-level session case (slash command
blocked, skill hidden from session listing). This issue is specifically about
the subagent skills: preload path, which the docs describe as content
injection rather than invocation.

Reproduction

1. Create ~/.claude/skills/demo-skill/SKILL.md:

```yaml
---
name: demo-skill
description: Demo skill for repro.
disable-model-invocation: true
---

## Core Concept
CANARY_STRING_12345 — this line should appear in the subagent's context
if preload works.
```

1. Create a project-level subagent at

.claude/agents/demo-agent.md (or the equivalent in your project config):

```yaml
---
name: demo-agent
description: Demo subagent that should have demo-skill preloaded.
skills:

• demo-skill

---

You are a demo agent.
```

1. Invoke the subagent via the Agent tool with

subagent_type: "demo-agent". Ask it to:

• (a) quote any line containing CANARY_STRING_12345 from its system context;
• (b) try to call the Skill tool with skill: "demo-skill".

Observed

• (a) The canary string is not present in the subagent's context. The

SKILL.md body is not injected.

• (b) The Skill tool call is rejected with:

``
Skill demo-skill cannot be used with Skill tool due to disable-model-invocation
``

• MCP tools that the skill normally coordinates with (e.g.

mcp__chrome-devtools__* for a chrome-devtools skill) remain available via
ToolSearch in the subagent, but they are granted independently and carry
none of the SKILL.md guidance.

Flipping the SKILL.md to disable-model-invocation: false makes the SKILL.md
body appear in the subagent's context as documented.

Expected

Either:

1. The subagent skills: preload should honor the explicit declaration and

inject SKILL.md content regardless of disable-model-invocation, since
preload is not model auto-invocation; or

1. Document this interaction explicitly in both

skills.md
and sub-agents.md,
and recommend a workaround for developers who want a skill that is both
(a) not auto-invoked in the parent conversation and (b) preloaded into a
specific subagent.

Impact

The common use case this breaks: a skill that is expensive or side-effectful
at the top level (so you want disable-model-invocation: true), but which is
exactly the right playbook for a dedicated subagent that the parent delegates
to. Currently the skills: frontmatter silently no-ops in this configuration,
which is especially confusing because the docs explicitly promise content
injection.

Environment

• Claude Code CLI 2.1.114, macOS (Darwin 25.3.0)
• Verified locally with both project-level and user-level combinations

(project subagent + project skill, user subagent + user skill). The
remaining cross-scope combinations (user/project, project/user) are likely
affected as well but were not exhaustively tested.