Skill with disable-model-invocation:true + context:fork gets no invocation args when auto-backgrounded

Description

A plugin skill marked both disable-model-invocation: true and context: fork in its frontmatter reliably fails when the CLI auto-backgrounds it: the forked agent receives only generic session context (memory index, project CLAUDE.md, tool/skill listing) and never the actual $ARGUMENTS/invocation instruction, so it takes no action.

Reproduced 3 times in one session, including once immediately after a plugin update — the update did not change the behavior.

Steps to reproduce

1. Use (or install) a plugin skill whose SKILL.md frontmatter sets both:

``yaml
disable-model-invocation: true
context: fork
`
(Concrete example from the sdlc plugin, dealpath-proddev marketplace, v1.7.73: skills/test-run/SKILL.md`.)

1. Invoke it as a slash command with arguments, e.g. /sdlc:test-run --area=comps or /sdlc:test-run <testrail-run-url>.
2. Let the CLI auto-background the invocation (observed surfacing as an @sdlc-test-run forked agent).
3. Inspect the forked agent's result/transcript.

Expected behavior

The forked agent receives the full invocation — the skill's own instructions plus the arguments/prompt the user supplied — and executes it.

Actual behavior

The forked agent's result indicates it received only generic system context (memory index, CLAUDE.md, tool/skill listing) with no actual instruction payload. It takes no action and effectively returns nothing useful. This happens every time for this skill configuration — not intermittently.

Suspected mechanism

Comparing every skill in the same plugin that sets context: fork: only skills that also set disable-model-invocation: true (test-run, fsd) exhibit this; skills with context: fork and disable-model-invocation: false (deliver, gap, ui, validate) are not reported to have this problem. This suggests the background-fork path that constructs the forked agent's initial prompt doesn't thread the original invocation prompt/arguments through correctly when the skill is also marked as not directly model-invocable — as if the fork is seeded from a different (argument-less) code path than a normal context: fork invocation.

Environment

• Claude Code version: 2.1.222
• Platform: macOS (Darwin 25.2.0)
• Plugin: sdlc (marketplace dealpath-proddev), v1.7.73
• Affected skill(s): skills/test-run/SKILL.md, likely skills/fsd/SKILL.md (same frontmatter combination, not yet independently confirmed)

Impact / workaround

No known workaround for the backgrounded path — retrying the same invocation fails identically each time. Current mitigation is to avoid relying on auto-backgrounding for this skill and instead run it in the foreground, or perform the work manually.