feat(claude): discover skills for the $ picker via SDK reloadSkills

[colonelpanic8]

What Changed

• The Claude provider's capability probe now issues the Claude Agent SDK's reloadSkills control request after initializationResult() and maps the returned entries into ServerProviderSkills, populating skills on the provider snapshot.
• New pure mapper parseClaudeReloadedSkills (apps/server/src/provider/Layers/ClaudeSkillDiscovery.ts):
  • strips the trailing scope suffix reloadSkills appends to descriptions (e.g. (user)) and maps it to a snapshot scope (user; project/local → project; builtin/bundled/system → system);
  • resolves each skill's path best-effort against the workspace .claude/skills/<name>, then the effective Claude config dir's skills/<name>, falling back to a constructed config-dir path (the contract requires a non-empty path); plugin-qualified names skip the filesystem probe;
  • dedupes by name and skips empty names, mirroring the existing slash-command handling.
• The control request is bounded by its own 5s timeout and any failure falls back to an empty list, so older CLIs that don't answer reload_skills never fail the capability probe.
• Slash commands are left untouched — skills legitimately also appear as / commands for Claude, matching upstream CLI behavior.

Why

Only the Codex provider populated skills, so the composer's $ skill picker was always empty on Claude threads even though the Claude CLI discovers the same ~/.claude/skills / .claude/skills entries (refs #1480, #2736). With this change Claude skills get the same $ autocomplete, inline pill rendering, and scope labels (Personal/Project/System) that Codex skills already have.

Checklist

• Focused tests: vp test run ClaudeSkillDiscovery.test.ts ClaudeCapabilitiesProbe.test.ts ProviderRegistry.test.ts → 3 files, 53 tests passed (plus ClaudeAdapter.test.ts regression guard → 60 passed)
• Typecheck: tsgo --noEmit in apps/server → 0 errors
• Lint/format: vp lint on changed files, vp fmt → clean

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches provider health probing and filesystem path resolution for user config; skill reload is isolated with timeouts so auth/command discovery should remain stable, but probe timing and path edge cases could affect the picker on some setups.

Overview
Claude skills now populate the provider snapshot so the composer $ picker can list them like Codex already does.

After a successful SDK initialization, probeClaudeCapabilities issues a best-effort reloadSkills call with its own timeout; failures yield an empty skill list but no longer fail the whole probe. Account info and slash commands are unchanged.

New parseClaudeReloadedSkills maps SDK entries to ServerProviderSkill: strips description scope suffixes (e.g. (user)), resolves paths under workspace .claude/skills or the effective config dir (with CLAUDE_CONFIG_DIR / HOME / USERPROFILE handling), and dedupes by name. checkClaudeProviderStatus threads skills into every built snapshot.

^{Reviewed by Cursor Bugbot for commit aac809e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Discover Claude skills for the $ picker via SDK reloadSkills

• After a successful capabilities probe, probeClaudeCapabilities issues a reloadSkills request to the Claude SDK subprocess and parses the response into ServerProviderSkill objects via the new parseClaudeReloadedSkills utility in ClaudeSkillDiscovery.ts.
• Skill paths are resolved by checking <cwd>/.claude/skills/<name> then <configDir>/skills/<name>; scope is normalized from description suffix tokens or inferred from the resolved path.
• The Claude config directory is derived from CLAUDE_CONFIG_DIR (with tilde expansion) or defaults to <home>/.claude, using a new resolveClaudeEnvironmentHomePath helper that prefers HOME over USERPROFILE.
• Separate timeouts are now used for initialization and skills reload; a skills reload failure is non-fatal and only the initialization timeout causes the probe to return undefined.
• The skills array is threaded through checkClaudeProviderStatus and included in the provider snapshot (empty when capabilities are absent).

^{Macroscope summarized aac809e.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: eacd9fbe-7634-4689-be12-8058c40b45c7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces a new feature for skill discovery via the Claude SDK's reloadSkills API, adding new runtime behavior including SDK calls, filesystem checks, and new data flowing through provider capabilities. As a new capability from a first-time contributor, it warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 138250a. Configure here.}

[colonelpanic8]

Superseded by the maintainer implementation in #4414, which is now merged. Closing this independent implementation in favor of the upstream-selected approach.