Plugin skill via Skill tool still injects no SKILL.md body on 2.1.218 (#25834, #68664 still reproduce)

Status Open
Reported on v2.1.177
Maintainer reply None cached
Activity 0 comments · opened Jul 24, 2026

Summary

Invoking a plugin-provided skill via the Skill tool dispatches the skill but never injects the SKILL.md body into context. The tool result is only:

Launching skill: expandtest:probe

The model is expected to follow instructions it never received. This is the same bug as #25834 (closed 2026-02) and #68664 (closed as duplicate of #25834 in June, reproduced there on 2.1.177). It still reproduces on 2.1.218, so it was either never fixed or regressed; filing fresh since both prior reports are closed.

Environment

  • Claude Code 2.1.218, macOS (Darwin 25.5.0)
  • Reproduced with a minimal local plugin via --plugin-dir (below), and observed in production with a marketplace-installed plugin (three out of three command starts across two instrumented sessions).

Minimal reproduction

  1. Create a minimal plugin:
expandtest/
  .claude-plugin/plugin.json    → {"name": "expandtest", "version": "0.0.1", "description": "probe"}
  skills/probe/SKILL.md

skills/probe/SKILL.md:

---
name: probe
description: probe skill body marker PROBE_SKILL_BODY_LOADED_XYZ
---
PROBE_SKILL_BODY_LOADED_XYZ: if you can read this, say so.
  1. Run headless and capture the full event stream:
claude -p --plugin-dir ./expandtest --permission-mode bypassPermissions \
  --output-format stream-json --verbose \
  "Invoke the Skill tool with skill name 'expandtest:probe', then stop." > stream.jsonl
grep -c "PROBE_SKILL_BODY_LOADED_XYZ" stream.jsonl   # → 0
  1. Observed: the Skill tool_use is answered by a tool_result containing only "Launching skill: expandtest:probe". The marker (i.e. the skill body) appears in zero events in the stream. The same symptom occurs in interactive sessions ("The skill instructions aren't visible in my context"), where the model recovers by find-ing and reading ~/.claude/plugins/cache/.../skills/{name}/SKILL.md by hand.

(Note: the probe skill deliberately declares no allowed-tools, to keep this distinct from #77363.)

Expected

Invoking a plugin skill via the Skill tool makes the SKILL.md body available to the model, as it is for project/user skills.

Related

  • #25834 (canonical, closed), #68664 (duplicate of #25834, closed) — same bug.
  • #15178 (discovery/<available_skills>), #54535 (metadata re-injection), #77363 (allowed-tools headless load failure) — adjacent, distinct.

View original on GitHub ↗