fix: scope CodebaseGraphManager to current repo for concurrent sessions (alternative approach)

[lchachurski]

Problem

When two Shotgun sessions run simultaneously on different repositories in the same parent directory, a startup health check scans every .kuzu file in ~/.shotgun-sh/codebases/. Kuzu uses exclusive file locks, so any database held open by a peer session is reported as "locked" — even though the current repo's database is perfectly fine. This shows a blocking dialog and prevents the second session from starting.

Reported in #534.

Root cause

Three methods in CodebaseGraphManager glob the entire storage directory:

• detect_database_issues() — the startup health check
• list_graphs() — used by shotgun codebase list and graph service
• cleanup_corrupted_databases() — called during startup and recovery

This fix (alternative to #535)

Add a single process-wide class variable _scope_graph_ids to CodebaseGraphManager, set once at process startup from the current working directory. A new _iter_databases() method replaces all three .glob("*.kuzu") calls. Any future method that enumerates databases uses _iter_databases() and automatically inherits the scope — no parameter threading required.

main.py (2 lines)
  └─ CodebaseGraphManager._scope_graph_ids = [generate_graph_id(cwd)]

manager.py (~15 lines)
  └─ _scope_graph_ids: ClassVar[list[str] | None] = None
  └─ _iter_databases() → replaces 3 glob sites

cli/codebase/commands.py (1 line)
  └─ list_codebases(): _scope_graph_ids = None  # needs all DBs

Total: 3 files changed, 43 insertions(+), 9 deletions(-)

Comparison with PR #535

	PR #535	This PR
Approach	graph_ids parameter on each method + .meta.json sidecars	Process-wide ClassVar, set once in main.py
Files changed	8	3
New future-proofing	Must thread parameter / update sidecar for every new method	_iter_databases() handles it automatically
shotgun codebase list	Shows all codebases (uses full glob)	Explicitly unscopes for this one command

How it works

Every Shotgun process starts from a CWD. main.py calls generate_graph_id(os.getcwd()) and stores the result in the class variable. Because _scope_graph_ids is a ClassVar, every instance of CodebaseGraphManager created anywhere in the process automatically respects the scope — TUI, shotgun run, shotgun codebase index, benchmarks, all covered by the two lines in main.py.

The only exception is shotgun codebase list, which legitimately needs to show all indexed repositories. It clears the scope before querying.

Tests

Standalone test script using real_ladybug==0.15.1 and multiprocessing to hold real Kuzu file locks. 3 simulated repos (project-alpha, project-beta, project-gamma). All 8 tests pass:

#	Scenario	Result
T1	Bug reproduced: OLD globs all, 2 peers locked, blocks 3rd	PASS
T2	NEW scope set → alpha/beta never opened, gamma starts clean	PASS
T3	list_graphs scoped: opens 1/3 not 3/3	PASS
T4	Each CWD's scope targets only its own graph	PASS
T5	Unrelated CWD returns empty	PASS
T6	Scoped cleanup avoids locked peers	PASS
T7	Backward compat: scope=None still full-scans	PASS
T8	graph_id uniqueness across 3 repos	PASS

Closes #534

Made with Cursor

[lchachurski]

Hey team! I hd a bit more time to think about the problem today and above is a cleaner fix approach.