fix(web): a new worktree thread opened during a clone bases on the real default branch - #36379

The proof item is right and is fixed. Reproduced on a2cd749 in the real app: when the clone finishes within about 5 s of the draft opening, the picker label reads 'From staging' (it follows live status) but the draft stores no branch, and sending shows 'Select a base branch before sending in New worktree mode.' Holding the clone 8 s made the same build seed 'staging', which is why the earlier runs passed. Cause: the server caches the ref snapshot and repository paths read while the clone was empty, and a client refresh within LIST_REFS_REFRESH_COALESCE_TTL (5 s) of that read is coalesced with it, so the client-side refresh added in a2cd749 got the empty list back. Every other git operation that changes refs invalidates that snapshot in the driver; the clone runs through raw git.execute and did not. Fix (fa6ba12): new GitVcsDriver.invalidateRefs(cwd), called by SourceControlRepositoryService.cloneRepository after a successful clone and before the tracker reports 'done', so the client's refresh reads the real refs. 1ac4526 adds invalidateRefs to the git driver mock in apps/server/src/mcp/toolkits/project/handlers.test.ts, which my change had broken (the Test Server 5 failure and CodeRabbit's first Major finding). c01b609 settles CodeRabbit's second Major finding, which was correct: invalidateListRefsSnapshot looks the repository up through the path cache and returns early on a cached 'not a repository', so invalidateRefs now clears repositoryPathsRefreshCache and repositoryPathsCache first, the order initRepoWithListRefsInvalidation already uses; CodeRabbit marked it addressed. The web changes from a2cd749 are unchanged: GitActionsControl skips its live branch sync while the project clone is not done, the worktree base is seeded only from a listed ref (resolveDefaultWorktreeBaseBranch), and the selector calls refreshLocalStatusForPicker when the clone reaches 'done'. Final state: all 37 checks pass on c01b609; CodeRabbit reports no actionable comments and merge risk Minimal up to c01b609; Macroscope's only comment is the approvability verdict 'Needs Human Approval', which is not a code finding. The real-app run was on fa6ba12; the two later commits (a test mock and the cache-clearing order) were covered by tests, not a second app run. No new regex or parser.

Fixes #25719

Reproduction on main

Commit origin/main 9efd960, web client (headless Chromium via Playwright, since the T3 preview browser is blocked by AppArmor on this VM) against T3CODE_DEV_SEED=0 vp run dev with a fresh database. Built a local repo whose only branch is staging and served it over smart HTTP (git http-backend), with a 60 s delay before the upload-pack response so the clone stays in its unborn-HEAD window; init.defaultBranch is unset, so the unborn HEAD is master. Steps: Add project → Git URL → entered the http URL → destination folder → Create & Clone. The app immediately opened a draft thread for the project; I switched Workspace to 'New worktree' while the clone was still running ('Cloning … Connecting'). The base picker read 'From master'. After the clone finished (git showed refs/heads/staging and origin/HEAD -> origin/staging, no master anywhere), the picker still read 'From master' 15 s later and again after reopening the draft URL. I picked Replay (dev) → Showcase and sent 'go'. The run was stored with workspacePreparation {"type":"worktree","baseRef":"master","startFromOrigin":true}. The first attempt failed on a fetch timeout caused by my artificial delay; after removing the delay, Retry failed with 'Workspace preparation failed during provision worktree: Git command failed in GitVcsDriver.createWorktree.createBranch (...): git branch failed (ref_not_found)'.

Evidence

Real web client, headless Chromium, 1280×800, light theme, fresh databases, Replay (dev). Add project → Git URL → Create & Clone a staging-only repository over gated smart HTTP. While cloning, select New worktree. Release the clone, wait for completion and another 18 seconds, then inspect the base picker.

Proven: the scenario fails on base c83e78fe07 and passes on head c01b609b73.

	Before c83e78fe07	After c01b609b73
Result	Stays on nonexistent master	Automatically settles on staging; worktree creation succeeds
During clone
Completed clone base picker
Assertion failure
Worktree created

Recordings of cloning and settling: Before · After.

Scenario and smart-HTTP fixture. Set PROOF_SCRATCH to a scratch directory before running it.

The head replay completed successfully. Its branch-renaming warning is expected because the scenario disables paid text-generation providers.

Not shown:

• The base scenario stops at the stale-picker assertion, so its subsequent send/retry ref_not_found error and reopening the draft URL were not separately exercised.

Tests

Real app (headless Chromium via Playwright, T3CODE_DEV_SEED=0 vp run dev, staging-only repo over git http-backend with upload-pack held until New worktree was selected, then released). On a2cd749: after 18 s picker 'From staging', draft branch null, send refused with 'Select a base branch before sending in New worktree mode.'; with an 8 s hold the draft stored 'staging'. On fa6ba12, fast clone: picker 'From staging (remote)', draft branch 'staging', Replay Showcase selected, 'go' sent, the turn completed and git worktree list showed a new worktree on branch t3/go at the staging commit. Unit: apps/server/src/sourceControl/SourceControlRepositoryService.test.ts, 32 passed, including two new real-git cases (ref list read before git creates the repository, and while HEAD is unborn); the unborn case fails with the invalidateRefs call removed and the pre-init case fails with the path-cache clearing removed. The 8 server test files importing the clone service: 136 passed. vp run check:changed on the final code: format, lint, knip 1/2 and 2/2, graphql, typecheck (4 packages), test (7 files) and react compiler all pass, run with GIT_CONFIG_GLOBAL=/dev/null because this VM's global url.insteadOf config makes an unrelated GitVcsDriver.test.ts remote-URL case fail. CI on c01b609: all 37 checks passed.

Not in this change

• The server does not invalidate its listRefs snapshot or repository-paths cache when a project clone finishes; the web client works around it with a refresh, but other clients (mobile, MCP) can still get the empty snapshot for up to 2 minutes.
• The picker label still shows the unborn HEAD name ('From master') while the clone is running, because it falls back to live status for display; not changed since it is a visible change beyond the bug.
• Mobile was not changed or tested: its new-task flow seeds a worktree base only from listed refs, so it does not appear to store the unborn name, but whether its ref list refreshes after a clone was not checked.
• Drafts that already stored a phantom base from an older build are not repaired.
• The vcs status contract has no flag for an unborn HEAD; adding one would let every consumer tell a real branch from an unborn name.
• The server does not invalidate its listRefs snapshot or repository-paths cache when a project clone finishes; the web client works around it with a refresh, but other clients can still get the empty snapshot for up to 2 minutes.
• The picker label still shows the unborn HEAD name while the clone is running, because it falls back to live status for display.
• Mobile was not changed or tested.
• Macroscope's approvability verdict asks for human approval of the changed default behaviour and the new lint suppression; that is a reviewer decision, not something to change in code.
• Mobile was not changed or tested; it now benefits from the server invalidation, but whether its ref list re-requests after a clone was not checked.
• The clone does not invalidate GitManager's status cache; the status stream already updated to the real branch in every run, so it was left alone.
• GitVcsDriver.test.ts 'lists the fetch and push URLs of a partial clone remote' fails on this VM because global git config rewrites [email protected]: to https; it passes in CI and with GIT_CONFIG_GLOBAL=/dev/null.

Who did the work

• Written by Claude Fable 5.2, in Claude Code
• Reviewed by GPT-6.1 Sol, in Codex, and by Claude Opus 5.5, each on its own
• Proof captured by GPT-6.1 Sol and checked by Claude Opus 5.5

Run by the t3code factory in T3 Code.