[asonawalla]

The thread panel's Changes row showed +2,381 −1,165 for #14, whose diff is +58 −63. The base name (main) was only ever looked up on the primary remote or as a local branch. So a branch cut from fork/main was diffed against origin/main and counted the fork's 13 commits that upstream doesn't have. Even gh-merge-base=fork/main resolved to origin/main.

Summary

resolveBaseBranchForNoUpstream (apps/server/src/vcs/GitVcsDriverCore.ts) feeds the Changes row, the diff panel's Changes view, and the ahead count:

 name = gh-merge-base ?? primary remote's default ?? main ?? master
-return origin/<name> ?? local <name>
+if origin lacks <name> and local <name> exists
+  return local <name>
+copies = [origin/<name>, <remote local <name> tracks>/<name>]
+return the copy that leaves the fewest of the branch's commits outside it

Setup	Base before	Base after
One remote	origin/main	origin/main, plus one cheap for-each-ref
main tracks fork/main, branch cut from fork/main	origin/main ✗	fork/main
Same repo, branch cut from origin/main	origin/main	origin/main
Same repo, on main itself	origin/main ✗	fork/main
Checkout of a PR opened from its author's main	origin/main	origin/main: the author's remote is never compared

Only the remote the local base branch tracks counts. The push remote is excluded: in the common contributor setup (main tracks origin, pushes to a fork), counting it would leave Changes on main with no committed work after a push.

Evidence

• Before: T3's real code on the PR fix(server): one runaway agent no longer freezes every agent #14 history, with origin/main and fork/main as they were in the screenshot:

  remove-agent-memory-limits {"baseRef":"origin/main","insertions":2381,"deletions":1165}
  

  After:

  remove-agent-memory-limits {"baseRef":"fork/main","insertions":58,"deletions":63}
  

• Before: counts Changes from the copy of main a branch was cut from fails on the old code:

  × expected { baseRef: 'fork/main', insertions: 1 }
    actual   { baseRef: 'origin/main', insertions: 4 }   # 1 feature line + 3 fork-only lines
  

  An earlier version of this fix compared every remote. never counts Changes from a pull request author's copy of main fails on that version with { baseRef: 'author/main', insertions: 0 }.

  After: both pass. GitVcsDriverCore.test.ts passes 156/156 and VcsStatusBroadcaster.test.ts 27/27. The server typecheck, lint and format are clean.

• GitManager.test.ts: passed 143/143 in most runs. Two tests with tight timeouts ("stops probing head selectors…" at 12 s, "does not reuse a cross-repo PR…" at 20 s) failed once while the Proxmox host's load average was about 300. In an A/B run under that load, the unchanged code also took 22.3 s on the first one, and the fix took 1.3–2.6 s.

Merge Danger

Door: two-way

Read-only base selection. It changes no stored state, so reverting restores the old numbers on the next status refresh.

Blast Radius: multi-remote

Repos with one remote get the same base as before. Repos whose local base branch tracks a second remote can get a different base in three places:

• the Changes row
• the diff panel's Changes view
• aheadCount / aheadOfDefaultCount, which gate the Create PR button

Known limit: a fork main that already merged the branch through a merge commit has none of the branch's commits outside it, so it wins and Changes shows only uncommitted work. The old code does the same on origin. This fork squash-merges.

Not changed here: new worktrees start from the primary remote's default branch, and PR creation picks its base in GitManager.resolveBaseBranch.

Model and harness: Claude Opus 5.5 in Claude Code via T3 Code.

🤖 Generated with Claude Code