[BUG] Auto-compaction still drops the entire Available skills listing (2.1.281, Windows) — #74990 was bot-closed, not fixed
What's Wrong?
Still reproducible on 2.1.281 (Windows 11, Claude desktop app → Code tab): after auto-compaction the "Available skills" system-reminder is not re-injected, so the model has no skills listing at all for the rest of the session.
This is the same bug as #74990 (reproduced by a maintainer on 2.1.233, assessed as "intended today, but the expectation is reasonable and we should change it"). #74990 was then auto-closed as inactive by the bot (see #87647), not fixed. Related: #76161.
Real-world consequence in our setup: we have 10 skills installed (7 user-level in ~/.claude/skills/, 3 project-level). One of them is a review/dispatch skill whose purpose is to tell the model how to request and evaluate code reviews. After auto-compaction the model saw zero skills, never loaded it, and for the rest of the session wrote review requests that violated that skill's guidance. Nothing told it a skill existed. The listing only reappeared, for that single skill, after the model happened to edit that skill's SKILL.md. The other 9 never came back.
What Should Happen?
After auto and manual compaction, re-announce the full Available skills listing, the same as at session start. At minimum, document in the compaction table (https://code.claude.com/docs/en/context-window) that the listing is dropped.
Steps to Reproduce
- Install a few user-level skills.
- Work in a long session until auto-compaction happens.
- Ask the model, with no tools: "List the skills shown in the Available skills system-reminder in your current context."
- It reports there is no such listing.
Workaround
The SessionStart hook with matcher: "compact" returning {"hookSpecificOutput": {"hookEventName": "SessionStart", "reloadSkills": true}}, from #74990, or /reload-skills after each compaction. Users shouldn't need to discover this to keep their skills working.
Claude Code Version
2.1.281
Platform
Windows 11 (desktop app, Code tab)
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗