[BUG] claude.ai Chat: org plugin skills shown in autocomplete + Customize but invocation says skill not available — same user works in Cowork

Status Open
Reported on v2.1.98
Maintainer reply None cached
Activity 2 comments · opened Jul 20, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

For some users in our org (HopSkipDrive, Claude for Work), custom org plugin skills are fully visible in the metadata layer but fail at invocation time in Chat (both web browser and desktop app Chat):

  • The plugin and skill appear correctly under Customize > Plugins
  • Typing / shows the skill in autocomplete with its correct description and plugin attribution
  • Selecting it from autocomplete (rendered as a slash pill) produces: "I don't have a skill called it-and-information-security:ai-skill-cybersecurity-review available in my current tool set — it's not among the skills I have access to."
  • Another affected user gets the equivalent for a different plugin's skill: "The plugin skill for the CCDP Discovery Roleplay isn't actually available in this environment — it's listed but the files aren't accessible here, so I can't load its content."

Key observations:

  1. Per-surface, same account: one affected user confirmed the same skill loads and runs fine in Cowork on the Desktop app, while failing in Chat — same account, same group assignments, same synced plugin content.
  2. Not plugin-specific: skills from at least two unrelated plugins in our marketplace (sales-enablement, it-and-information-security) fail identically for affected users.
  3. Per-user: other users in the same groups (including me, an org admin) can invoke the same skills successfully in every surface. Group scoping is confirmed correct — affected users see the plugin listed, which they wouldn't if they weren't in an assigned group.
  4. Fails in brand-new conversations, not just stale sessions.
  5. Persisted across a plugin rename and multiple marketplace.json/plugin.json version bumps (1.0.0 → 1.0.3) on our GitHub-sourced marketplace.

This looks like the claude.ai-surface variant of the resolver mismatch previously reported in #58102 and #46079 (autocomplete/skill registry disagree; #46079 was closed as duplicate of #41842 → #42471 "fixed in Claude Code 2.1.98", but #58102 reported still reproducing on claude.ai/Cowork builds after that fix and was closed as stale, not resolved).

What Should Happen?

If a skill is listed in Customize > Plugins and advertised in / autocomplete with a preview and plugin attribution, invoking it (explicitly or via trigger phrases) should load the skill. The invocation-time skill registry should match what the metadata/autocomplete layer shows.

Error Messages/Logs

I don't have a skill called `it-and-information-security:ai-skill-cybersecurity-review` available in my current tool set — it's not among the skills I have access to.
The plugin skill for the CCDP Discovery Roleplay isn't actually available in this environment — it's listed but the files aren't accessible here, so I can't load its content.

Steps to Reproduce

  1. Configure a private GitHub repo as an org plugin marketplace (standard structure: .claude-plugin/marketplace.json, plugins with .claude-plugin/plugin.json + skills/<name>/SKILL.md), confirmed syncing to claude.ai
  2. Assign plugins to user groups via the admin console; confirm affected users are in the assigned groups
  3. As an affected user, open claude.ai Chat (web or desktop app), start a new conversation
  4. Type / — the skill appears in autocomplete with correct description and plugin attribution
  5. Select it — Claude responds that the skill is not available in its tool set
  6. Same user, same account, opens Cowork in the Desktop app and invokes the same skill — it loads and runs correctly

Not reproducible on demand for all users — some accounts are consistently fine, others consistently affected. Happy to provide org details, affected account emails, and timestamps privately to help correlate telemetry.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

N/A

Claude Code Version

Not applicable — claude.ai Chat (web browser + desktop app, current production builds as of 2026-07-20)

Platform

Other

Operating System

macOS

Terminal/Shell

Other

Additional Information

  • Org: HopSkipDrive (Claude for Work), GitHub-sourced private plugin marketplace confirmed configured and syncing
  • Affected plugins observed so far: sales-enablement (skill ccdp-discovery-roleplay), it-and-information-security (skill ai-skill-cybersecurity-review) — unrelated plugins, identical failure
  • We are separately verifying the org/user "Code execution and file creation" settings, but the same-account Chat-fails/Cowork-works split and the metadata-vs-runtime disagreement match the resolver-bug profile of #58102 / #46079 more than a capability toggle
  • Related prior reports: #58102, #46079, #41842, #42471

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗