Claude Skills Manager

A VS Code extension that bundles a personal Claude Code skill library and helps you install the skills relevant to whatever project you have open.

This is the VS Code front-end for the claude-skills-deployer repo. The skill content lives in ../skills_library/; this extension bundles a synced copy (skills_library/, generated by npm run sync-skills, gitignored).

Features

• Activity bar view ("Claude Skills") — lists every skill in the bundled library with its status for the current workspace:

  • 💡 relevant to this workspace but not yet installed
  • ✅ installed in <workspace>/.claude/skills/
  • ◦ available in ~/.claude/skills/ but not relevant/installed here
  • ⊘ not yet installed in ~/.claude/skills/ at all

• Project-local skills — if <workspace>/.claude/skills/ already contains skills that aren't part of the bundled library (e.g. this extension was installed into a project that already has its own custom skills, or skills installed from elsewhere), they're listed too, marked project-only. Their description is read from the skill's own SKILL.md frontmatter when available. This means installing the extension into an existing project never hides skills that are already there.

• On/off checkboxes — each skill in the tree has a checkbox: check it to install ("enable") the skill into <workspace>/.claude/skills/, uncheck it to remove ("disable") it from there. This is the simplest way to turn individual skills on or off per project.

  Note: <workspace>/.claude/skills/ is normally git-tracked and shared with your team — unchecking a skill here deletes its files, and if you commit that, it removes the skill for everyone after merge. To turn a skill off only for yourself, without touching the shared files, use the eye icon described below instead.

• Multi-agent clones (auto-sync by default) — one skills_library/ fans out to Claude, Cursor, Kiro (SKILL.md), and Copilot (.github/instructions/*.instructions.md). When claudeSkills.features.multiAgent is on and claudeSkills.agents.syncWorkspaceToAll is true (default), your effective skill set (enabled for you, respecting skillOverrides) is mirrored to all enabled agent paths: .cursor/skills/, .kiro/skills/, .github/instructions/ (Copilot). Sync runs on checkbox/eye toggle, branch profile apply, extension startup, and .claude/skills / settings.local.json changes. Manual: Sync Workspace Skills to All Agents. .claude/skills/ stays the git-tracked file source of truth.

• Per-branch skill profiles — each git branch can have its own skill layout stored in ~/.claude/learning/branch-profiles.json (global, personal, not committed). A Branch profiles section at the top of the Skills tree shows the current branch and saved profiles; toolbar icons provide show / save / apply. The extension auto-saves when you install/remove skills or change local overrides; on branch switch it restores the saved profile for the new branch (adds missing skills by default, mirrored to other agents when sync is on). Committed .claude/skills/ on the branch remains the team-shared source of truth after you git commit.

• Team branch profiles in git — commit a shared layout per branch in .claude/skills-profile.json via Export Team Branch Profile to Git. Shown in a Team profiles tree section. On branch switch: team profile first, then personal profile from ~/.claude/learning/branch-profiles.json.

• Multi-root workspaces — uses the active editor's workspace folder; Pick Active Workspace Folder (or status bar) when multiple roots are open.

• SKILL.md lint — source + Cursor/Kiro/Copilot mirror checks on sync (claudeSkills.lint.*); blocks multi-agent sync only when blockSyncOnError is on.

• Attribution v2 hooks — Enable Attribution Hooks (v2) logs each Skill tool invoke to runs.jsonl for per-skill cost attribution without equal-split transcript guessing. Tokens are backfilled from session transcripts via tool_use_id when the hook row has zero tokens. Auto-installed when claudeSkills.agents.syncHooksOnSkillChange is on.

• Copilot bootstrap — multi-agent sync writes .github/copilot-instructions.md (always-on index) plus per-skill .github/instructions/*.instructions.md.

• Local skill set vs. git branch — the checkbox means enabled for you, not "change what the branch commits":

  • Uncheck a branch-committed skill → writes "off" to personal .claude/settings.local.json (skillOverrides). Files stay in .claude/skills/ for teammates; your git status stays clean.
  • Check a skill not on the branch → copies into .claude/skills/ and adds it to .git/info/exclude so it stays personal-only on your machine.
  • Uncheck a personal-only skill → removes the directory from your workspace. Branch profiles in ~/.claude/learning/branch-profiles.json store this per-branch effective set (including overrides), restored on branch switch.

• Eye icon — same as uncheck/check for branch-committed skills (local skillOverrides only). Disabled skills show "disabled locally" and a closed-eye icon; the checkbox is unchecked while files remain on the branch.

• Status bar items:

  • How many relevant skills are pending install; click it to install them.
  • A skill usage summary (active / needs review / unused), based on .claude/learning/runs.jsonl written by the self-learning skill; click it to open the full usage/KPI report.
  • Today's estimated Claude spend (tokens + USD, with % of daily budget when configured); click for the full usage report.
  • Budget mode (Economy / Normal / Unlimited); click to cycle modes.

• Claude Credits Usage report — the usage report also includes a "Claude Credits Usage" section: total tokens and an estimated cost for the last 14 days across all your projects, broken down by day and by model, read directly from Claude Code's session transcripts (~/.claude/projects/**/*.jsonl). The per-skill table also has a Tokens column (populated once the self-learning skill records the optional tokens field on a run).

• Click any skill to open its SKILL.md (workspace copy if installed, else global copy, else the bundled copy).

How to use this extension

1. Install the .vsix (Extensions view -> ... menu -> "Install from VSIX...") or get it from the Marketplace, then open a workspace folder.
2. Open the Claude Skills view in the activity bar. It lists the full bundled skill library plus any skills already in <workspace>/.claude/skills/, each annotated with its status for this workspace (relevant / installed / available / project-only).
3. First time only: run "Install Skill Library to ~/.claude/skills" to seed your personal global library (subsequent steps copy from there).
4. Run "Install Relevant Skills for Workspace" to auto-install the skills that match files in your project (or "Preview Skill Detection (Dry Run)" first to see what would happen without writing anything).
5. Fine-tune per skill: tick/untick a skill's checkbox in the tree to enable/disable it for this workspace, or right-click -> "Install to Workspace .claude/skills" for a single skill.
6. Click "Show Skill Usage Report" (graph icon, also in the status bar) to see which installed skills are active, underused, or worth removing, plus your overall Claude credits/token usage for the last 14 days.
7. Optionally run "Enable Cost Control Hooks" to install session-size and daily-budget UserPromptSubmit hooks (warn at 80% of budget, auto-disable high-tier skills when the cap is hit).
8. Set budget caps under Settings → Extensions → Claude Skills Manager → Budget, or click the Economy / Normal / Unlimited status bar item to cycle modes.

Commands (Command Palette / view toolbar)

Weekly AI usage report (informative email)

When claudeSkills.weeklyReport.enabled is on (default), the extension checks every 15 minutes while VS Code/Cursor is open. On the configured day (default Monday 9:00 local), it sends a plain informative email about AI agent usage and cost — no GitHub/GitLab issues.

One-time setup

1. Open a workspace with a GitHub or GitLab origin remote.
2. Run Configure Weekly Report Email (sidebar key icon or Command Palette).
3. Paste a personal access token (stored in VS Code Secret Storage — not in settings.json), or reuse gh auth login / GITLAB_TOKEN.
4. Complete SMTP (Gmail app password, Microsoft 365, or custom). Also stored in secrets.

Full token types, scopes, and settings table: see Weekly AI usage report in the repo README.

GitHub PAT (minimum scopes): read:user, user:email (classic) or fine-grained with Email addresses: Read + Profile: Read. repo is not required.

GitLab PAT (minimum scope): read_user.

SMTP is required — the git token finds your address; SMTP sends the mail.

Prerequisites

1. One-time Configure Weekly Report Email (PAT + SMTP).
2. VS Code/Cursor open at the scheduled time, or run Send Weekly AI Usage Report manually.

Advanced: claudeSkills.weeklyReport.emailTo or CLAUDE_SKILLS_SMTP_* env vars instead of the wizard.

Settings highlights

Find all options under Settings → Extensions → Claude Skills Manager (or search @ext:serhiivoinolovych.claude-skill-deployer). Sections: Budget, Agents, Features, Lint, Optimizer, Weekly Report, and more.

Checkbox toggles your effective skill set: enable installs or clears skillOverrides; disable on a branch-committed skill sets skillOverrides: off locally (files stay on the branch). Personal-only skills are removed from disk and added to .git/info/exclude. Changes propagate to other enabled agents automatically.

All results are logged to the "Claude Skills" output channel.

Onboarding & recovery (v1.0.x)

First activation prompts Get Started if ~/.claude/skills/ is empty.

Cost attribution notes (v1.0.1+)

The background collector attributes session tokens only to skills with evidence of actual use (SKILL.md reads, Skill tool calls, etc.) — not the full skill_listing catalog. Unattributed sessions are labeled unattributed in reports. Record runs with metadata.invoked: true via the self-learning skill for best accuracy. Keep claudeSkills.optimizer.autoApply off until attribution looks correct.

What this tool does NOT do

• SKILL.md lint is advisory by default (claudeSkills.lint.blockSyncOnError to hard-block) — covers the .claude/skills source of truth plus Cursor/Kiro SKILL.md mirrors and Copilot .instructions.md mirror presence.
• Community benchmark upload requires you to configure endpoints (no default public server).
• PR comments require GitHub CLI and explicit feature enable.
• Copilot clones are instruction files, not native Copilot skills.

Performance & compatibility

See root README.md and CHANGELOG.md.

Development

npm install
npm run sync-skills   # copy ../skills_library -> ./skills_library
npm run compile       # or: npm run watch

Press F5 in VS Code (with this folder open) to launch an Extension Development Host.

Packaging

npm run package   # runs sync-skills, then `vsce package` (requires @vscode/vsce)

Produces a .vsix you can install via "Extensions: Install from VSIX...".