Session import skips Claude Code / Codex sessions that ran in a linked git worktree

[ahorovit]

Problem

Session import never finds a session whose working directory is a linked git worktree of the project. Such worktrees are created by claude --worktree, by the Claude desktop app, or by hand with git worktree add. The scanner does not offer the worktree as its own project, and it does not count the session toward the main checkout's project either, so the session cannot be imported at all.

For people who run every task in a worktree, this leaves out most of their history. Claude Code on my machine has 95 open sessions. 43 of them ran in worktrees of repositories I have already added to T3, and 14 of my 15 pinned sessions are among those 43. All 43 worktree directories still exist on disk. The import brought in 6 threads in total, none of them from a worktree.

Why it happens (at 32a05317)

• readGitIdentities in apps/server/src/project/AgentSessionScanner.ts already resolves a linked worktree's commondir back to its main repository. It then tags every worktree of a non-bare repository as { _tag: "Worktree" }, and the candidate loop drops that tag (key = ""). The comment explains that this is deliberate: only worktrees of a bare repository are offered on their own. Keeping worktrees out of the project list makes sense, but nothing then assigns their sessions to the main repository.
• prepareRecentThreads accepts a transcript only when directoryIdentity(cwd) equals the project root's identity. That identity is an inode or a real path, so a worktree directory never matches. The check runs once on the candidate and again on the transcript's own cwd.
• AgentSessionImporter.ts creates every imported thread with worktreePath: null, branch: null and runtimePayload: { cwd: project.workspaceRoot }.

Proposed behavior

When the project root is the main checkout, import sessions whose cwd is a linked worktree of that root's repository (the same common git directory). Each imported thread records the worktree as its worktreePath and the worktree's branch, and resumes with that worktree as its working directory. Claude Code stores transcripts by working directory, so resuming from the project root would likely not find them.

Things that stay the same:

• Worktrees still do not appear as separate projects in the import list.
• T3-managed worktrees stay excluded.
• Sessions whose worktree directory no longer exists stay out of scope. The scanner cannot tie them to a repository without reading a provider's own metadata.
• The 30-day window, the 100-transcript cap and the other import limits do not change.

Questions for maintainers

1. Is attaching a worktree session to its main checkout's project the direction you want, rather than offering each worktree as its own import candidate?
2. Should an imported worktree thread take over that worktree as its worktreePath, so T3's worktree lifecycle and cleanup then apply to a directory another tool created? Or should it record the worktree only as the thread's working directory?

If the direction is approved, I can open one focused PR covering the scanner, the importer and their tests (AgentSessionScanner.test.ts, AgentSessionImporter.test.ts), with evidence from a real import.

Related: #10961 (import is unreachable after onboarding), #27165 (configurable import limits), #6873 (continuing threads from Claude Code).

[Qusic]

my agent helped me to import by temporarily changing worktree session cwd to main repo path and changing them back after import. but yeah good to have native support for worktree sessions