[Bug] Plugin agents and commands incorrectly included in <available_skills> block
Bug Description
Bug: Plugin commands incorrectly listed as skills in <available_skills>
Expected behavior: Only actual skills (defined by SKILL.md files, invokable via /skill-name) should appear in <available_skills>.
Actual behavior: The list includes ALL plugin components - agents, workflow commands, and other non-skill items - inflating the count from the actual skill count to 81 items.
Proof from /context output:
The context output clearly separates these as distinct categories:
Custom agents · /agents
└ compound-engineering:research:framework-docs-researcher
└ compound-engineering:design:figma-design-sync
...
Slash commands · 60 of 81 commands
Yet in <available_skills>, these same agents appear:
compound-engineering:resolve_parallel
compound-engineering:workflows:work
compound-engineering:workflows:plan
Impact:
- Token waste - 5.6k tokens bloating every request
- Misleading skill counts (81 "skills" when far fewer are actual skills)
- Potential for invoking agents incorrectly via Skill tool instead of Task tool
- Makes the mandatory skill check scan match incorrectly on agent/command keywords
Reproduction:
- Have plugins with multiple component types (skills, agents, commands)
- Check <available_skills> in system prompt
- Note that agents and other non-skill components are listed as skills
Environment Info
- Platform: darwin
- Terminal: ghostty
- Version: 2.0.73
- Feedback ID: 786a3871-b893-4775-8bb1-196221cc2e6a
Errors
[{"error":"Error: Plugin hook \"node \"/Users/kyle/.claude/plugins/cache/thedotmack/claude-mem/7.3.3/scripts/user-message-hook.js\"\" failed with exit code 3\nstderr: 📝 Claude-Mem Context Loaded\n ℹ️ Note: This appears as stderr but is informational only\n\n\n\^[[1m\^[[36m[kylesnowschwartz.github.io] recent context\^[[0m\n\^[[90m────────────────────────────────────────────────────────────\^[[0m\n\n\^[[2mLegend: 🎯 session-request | 🔴 bugfix | 🟣 feature | 🔄 refactor | ✅ change | 🔵 discovery | ⚖️ decision\^[[0m\n\n\^[[1m💡 Column Key\^[[0m\n\^[[2m Read: Tokens to read this observation (cost to learn it now)\^[[0m\n\^[[2m Work: Tokens spent on work that produced this record (🔍 research, 🛠️ building, ⚖️ deciding)\^[[0m\n\n\^[[2m💡 Context Index: This semantic index (titles, types, files, tokens) is usually sufficient to understand past work.\^[[0m\n\n\^[[2mWhen you need implementation details, rationale, or debugging context:\^[[0m\n\^[[2m - Use the mem-search skill to fetch full observations on-demand\^[[0m\n\^[[2m - Critical types (🔴 bugfix, ⚖️ decision) often need detailed fetching\^[[0m\n\^[[2m - Trust this index over re-reading code for past decisions and learnings\^[[0m\n\n\^[[1m\^[[36m📊 Context Economics\^[[0m\n\^[[2m Loading: 10 observations (3,019 tokens to read)\^[[0m\n\^[[2m Work investment: 10,283 tokens spent on research, building, and decisions\^[[0m\n\^[[32m Your savings: 7,264 tokens (71% reduction from reuse)\^[[0m\n\n\^[[1m\^[[36mDec 19, 2025\^[[0m\n\n🎯 \^[[33m#S196\^[[0m User inquired about available skills in the Claude Agent SDK system (Dec 19 at 6:57 PM)\n\n🎯 \^[[33m#S195\^[[0m User requested to activate frontend skills but no frontend-specific skills are currently installed (Dec 19 at 6:57 PM)\n\n\^[[2mGeneral\^[[0m\n \^[[2m#742\^[[0m \^[[2m6:57 PM\^[[0m 🔵 Skills System Query \^[[2m(~243t)\^[[0m \^[[2m(🔍 1,608t)\^[[0m\n \^[[2m#743\^[[0m \^[[2m6:58 PM\^[[0m 🔵 Skills System Query \^[[2m(~200t)\^[[0m \^[[2m(🔍 293t)\^[[0m\n\n🎯 \^[[33m#S199\^[[0m User attempted to activate the frontend-design skill but discovered it is not installed in the environment (Dec 19 at 6:58 PM)\n\n🎯 \^[[33m#S200\^[[0m Locate and read the frontend-design skill definition file to understand its guidelines and capabilities (Dec 19 at 7:22 PM)\n\n\^[[2mGeneral\^[[0m\n \^[[2m#750\^[[0m \^[[2m7:23 PM\^[[0m 🔵 Frontend Design Skill File Not Found \^[[2m(~269t)\^[[0m \^[[2m(🔍 1,720t)\^[[0m\n \^[[2m#751\^[[0m 🔵 Frontend Design Skill File Not Found in Expected Plugin Location \^[[2m(~272t)\^[[0m \^[[2m(🔍 1,703t)\^[[0m\n \^[[2m#752\^[[0
Note: Error logs were truncated.
This issue has 5 comments on GitHub. Read the full discussion on GitHub ↗