Copy only what a repository asks for when forking a worktree

[ShravanSunder]

What

Rule-based copying for copy-on-write worktree forks, so a warm agentstudio worktree new copies what the repository asks for and nothing it shouldn't:

• GitWorktreeCopyRules on GitForkWorktreeRequest.copyRules: .copyAll (previous behavior) or .copyMatching([GitPathPattern]).
• GitPathPattern: gitignore grammar without negation (brackets, POSIX classes, **, anchoring); a leading ! is rejected. Patterns compile once.
• Copy filter, over the walked plan, before topology capture:
  • tracked paths (HEAD ∪ every index entry) are always copied;
  • submodules are always kept;
  • ignored untracked paths are copied only when they or an ancestor match an include;
  • an independent nested repository follows its location's rule;
  • hard-link groups re-elect a kept primary;
  • core.ignorecase is followed.
• Always, for both modes: nested linked worktrees of the same repository are skipped and reported in nestedWorktreesSkipped.
• Report fields: ignoredIncludedPatterns, ignoredExcludedCount, nestedWorktreesSkipped.
• New rejections: sourceIndexUnreadable, sourceIndexUnsupported (sparse/split index; .copyMatching only).
• .copyAll doesn't read the source index. The app's Fork button keeps working on sparse and split-index checkouts.

Contract: agentstudio Spec LR28 and Program Design "How new copies" (worktree-lifecycle design, r23/r29).

Proof (final head f259a9a)

• CI check: green on f259a9a (Homebrew git 2.55.0), including the AddressSanitizer and ThreadSanitizer lanes and the downstream-consumer checks.

• Local: mise run check and mise run test-tsan each ran 613 tests with 0 failed (89 suite runs).

• Matching follows Git exactly. GitPathPattern is a byte-oriented port of Git's wildmatch plus the match_pathname one-byte prefix context (git 1940a02dc1, shipped in Git 2.52). Patterns compile once; there is no regex.

• Permanent native-Git differentials in the normal gates:

  Suite	Pairs	Mismatches
  check-ignore	6,144	0
  sparse check-rules	2,560	0
  real sparse-checkout reapply	20,100	0

  The suites are version-aware. Below Git 2.52, the one class whose behavior changed in Git is asserted against recorded modern values.

• Lead's independent harness (compiled from the committed sources): about 182k check-ignore decisions, 0 mismatches.

• Real checkout (agent-studio main, 123k files, release build):

  • .copyMatching is within a few percent of .copyAll;
  • Frameworks/GhosttyKit.xcframework is present;
  • all four include patterns are reported;
  • the source is unchanged, and the temp forks are removed.

• .copyAll doesn't read the source index. The original sparse-index topology and clean-adoption proofs are byte-identical to before.

Review

• Independent Review Sidekick (Claude Opus xhigh), six rounds:
  • F1–F7, then R1/R2, then ready, then D1/D2/N1, then E1, then ready.
  • It then retracted D1, which had copied a pre-2.52 Git bug.
  • Final: ready.
• Advisor (Sol xhigh):
  • A1/A2 HOLD (the sparse matcher and grammar parity);
  • then MERGE on every later delta.
  • Final: merge.

🤖 Generated with Claude Code