[Bug]: Project skills that a new worktree receives after creation never appear in the $ picker

[caitlon]

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/server

Steps to reproduce

1. Use the Claude provider in the desktop app with a repository whose .claude/skills/ is gitignored. Ours is gitignored on purpose; a new worktree gets it from a Claude Code SessionStart hook that symlinks .claude/skills from the main checkout.
2. Start a new thread in New worktree mode and send the first message. T3 runs git worktree add, starts the Claude session, and the hook creates <worktree>/.claude/skills within the same second.
3. In that thread, type $ in the composer.

Expected behavior

The $ picker lists the project skills from <worktree>/.claude/skills, the same ones Claude Code itself loads in that session.

Actual behavior

The picker shows only the user-scope skills from ~/.claude/skills, and it stays that way for the life of the server process. Claude Code in the same thread does see the project skills: they are in the session's skill listing, and typing $name by hand still invokes them. A thread whose worktree already existed when T3 started shows all project skills, which is how we noticed the difference (two threads created after the last app launch are affected, one created before it is not).

Mechanism, read from ProviderRegistry.refreshWorkspaceSnapshot and ClaudeDriver.snapshotForCwd: the workspace snapshot for a cwd is scanned once, and every later call returns early because a snapshot for that cwd already exists. If the first scan runs before .claude/skills is present in the new worktree, the empty project list is cached for that cwd. In server.trace.ndjson every refreshWorkspaceSnapshot span for the affected thread lasts 0.005 to 0.017 ms and has no discoverClaudeSkills child, so the folder is never read again.

#14542 (merged 2026-10-01) adds Restart agent session with a fresh rescan, which should work around this by hand. Nothing rescans automatically, so every new worktree thread starts with a picker that is missing the project skills until the user knows to run it. This is the skills counterpart of #11575, where the same early return freezes the / menu.

A possible fix: rescan the workspace snapshot when a thread's provider session starts (after its hooks have run), or treat a snapshot whose project root did not exist or was empty as stale.

Impact

Minor bug or occasional failure

Version or commit

Desktop 0.0.44; the early return is still on main @ 921cb3c without the fresh flag.

Environment

macOS 27, T3 Code desktop 0.0.44, Claude Code 2.1.286, Claude provider (Fable 5.1 and Opus 5.5).

Workaround

Type the skill name in full ($name), or restart the T3 app. On builds with #14542, Restart agent session from cmd+k should rescan.

[juliusmarminge]

Note

Grok responding on behalf of Julius.

Triage

Thanks for tracking this down so precisely, @caitlon. Your reading is right, and it reproduces from source on current main (54084ae1).

Here's the order of events. A new-worktree send publishes worktreePath as soon as git worktree add finishes, before the turn starts (apps/server/src/ws.ts). The composer picks that up as its cwd and right away asks for a workspace snapshot without fresh. After that it stops asking (ChatComposer.tsx on web, use-composer-command-menu.ts on mobile). refreshWorkspaceSnapshot then returns early for any cwd it has already scanned unless fresh is set (ProviderRegistry.ts). That fits the very short spans you saw with no discoverClaudeSkills child.

The refresh at session start doesn't fix it either. ensureSessionForThread calls the same method without fresh, and startSession returns once the Claude query is spawned, without waiting for SessionStart hooks. So by the time your hook adds the .claude/skills symlink, the user-only list is already cached. Typing $name still works because sendTurn discovers skills separately when it decides which tokens to forward, but it never updates the picker's snapshot.

On current main, only Restart agent session (merged PR #14542) passes fresh: true. Nothing on the automatic path sets it.

This is related to a few other reports, but it isn't a duplicate of any of them:

• Open issue [Bug]: Workspace snapshot taken before the first Claude probe finishes leaves the / menu without provider commands for the whole session #11575 has the same never-rescanned rule, but there the / menu freezes on the placeholder Claude probe at launch.
• Closed issue [Bug]: Provider refresh does not update cached skills after installing skills or plugins #11497 is the manual rescan after adding skills, which merged PR feat(web): restart the agent session from cmd+k to load new skills and plugins #14542 addressed. That's the workaround here, not the fix.
• Open issue [Bug]: Plugin-provided Claude skills (e.g. mattpocock-skills) never appear in the $ picker #5622 is about plugin skills that filesystem discovery never finds.
• Open PRs fix(server): Claude picks up project slash commands #13077 and fix(server): refresh held workspace skill lists on provider refresh #13251 would rebuild held snapshots on a later refresh, but neither rescans once the session's hooks have run.

Workaround: on 0.0.44, relaunch T3 or type the full $name. On a build that includes #14542, run Restart agent session from cmd+k. It rescans the worktree and keeps the conversation.

Likely fix: rescan the cwd once the provider session is actually up, with hooks finished, and don't treat a scan that ran before the project skills folder existed as final. The fresh flag is already the right hook for that.