fix(mcp): reopen the database when it's replaced on disk instead of serving a deleted inode (#925) (#940)

A long-lived `serve --mcp` process opens `.codegraph/codegraph.db` and holds
the fd for its whole life. If `.codegraph/` is removed and recreated AT THE
SAME PATH while it runs — `git worktree remove <p>` + re-add, or `rm -rf
.codegraph` + `codegraph init` — the held fd points at the now-unlinked inode
and can never see the new index. The server serves the pre-removal snapshot
(renamed/removed symbols still "live", new ones missing); `codegraph sync`
can't refresh it and the CLI (a fresh process) diverges. Only a restart fixed
it — and because the daemon registry is keyed by path, a same-path recreate
routes new clients straight back to the same stale daemon, so the fix has to
self-heal inside the running process.

- DatabaseConnection records the DB file's (dev, ino) at open and exposes
  isReplacedOnDisk() — a different inode now at the same path. POSIX-gated:
  Windows can't unlink an open file and its st_ino is unreliable, so it never
  fires there.
- CodeGraph.reopenIfReplaced() opens the live file first, then swaps the
  connection + query layers IN PLACE (via the new wireLayers() helper), so
  every holder of the instance (the daemon's default project, cached
  projectPath connections) heals without a restart. Closing the dead handle
  also frees the leaked db/-wal/-shm fds pinning the unlinked inode.
- ToolHandler.getCodeGraph calls it (freshen) before serving — one stat() per
  call, a no-op unless the inode actually changed, never throws into a tool.

Tests cover isReplacedOnDisk (unchanged / replaced / absent / Windows-gated)
and an end-to-end reopen that heals a held instance after a same-path recreate
(asserts the pre-heal staleness too). Validated on macOS with a dist probe of
the raw instance and the MCP serving path; full suite green.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Colby Mchenry
2026-06-21 14:03:07 -05:00
committed by GitHub
co-authored by Claude Opus 4.8
parent 62d5cdde2c
commit e43ac82cdf
5 changed files with 252 additions and 21 deletions
+62 -18
View File
@@ -132,11 +132,14 @@ export class CodeGraph {
private db: DatabaseConnection;
private queries: QueryBuilder;
private projectRoot: string;
private orchestrator: ExtractionOrchestrator;
private resolver: ReferenceResolver;
private graphManager: GraphQueryManager;
private traverser: GraphTraverser;
private contextBuilder: ContextBuilder;
// Assigned via wireLayers() from the constructor (and again on reopen) — the
// `!` tells TS these are definitely set even though the assignment is one
// method call away from the constructor body.
private orchestrator!: ExtractionOrchestrator;
private resolver!: ReferenceResolver;
private graphManager!: GraphQueryManager;
private traverser!: GraphTraverser;
private contextBuilder!: ContextBuilder;
// Mutex for preventing concurrent indexing operations (in-process)
private indexMutex = new Mutex();
@@ -155,27 +158,68 @@ export class CodeGraph {
this.db = db;
this.queries = queries;
this.projectRoot = projectRoot;
// Down-weight the project name as a query term in search ranking — it names
// the whole repo, not a symbol, so it has no discriminative value (#720).
try {
this.queries.setProjectNameTokens(deriveProjectNameTokens(projectRoot));
} catch {
// Best-effort: ranking still works without it.
}
this.fileLock = new FileLock(
path.join(getCodeGraphDir(projectRoot), 'codegraph.lock')
);
this.orchestrator = new ExtractionOrchestrator(projectRoot, queries);
this.resolver = createResolver(projectRoot, queries);
this.graphManager = new GraphQueryManager(queries);
this.traverser = new GraphTraverser(queries);
this.wireLayers();
}
/**
* (Re)build the query/extraction/graph layers over the current `this.queries`
* (which wraps `this.db`). Factored out of the constructor so `reopenIfReplaced`
* can rebuild them against a fresh connection without duplicating the wiring.
* The path-based `fileLock` is independent of the DB handle, so it stays put.
*/
private wireLayers(): void {
// Down-weight the project name as a query term in search ranking — it names
// the whole repo, not a symbol, so it has no discriminative value (#720).
try {
this.queries.setProjectNameTokens(deriveProjectNameTokens(this.projectRoot));
} catch {
// Best-effort: ranking still works without it.
}
this.orchestrator = new ExtractionOrchestrator(this.projectRoot, this.queries);
this.resolver = createResolver(this.projectRoot, this.queries);
this.graphManager = new GraphQueryManager(this.queries);
this.traverser = new GraphTraverser(this.queries);
this.contextBuilder = createContextBuilder(
projectRoot,
queries,
this.projectRoot,
this.queries,
this.traverser
);
}
/**
* Heal a stale database handle in place. If `.codegraph/` was removed and
* recreated at the SAME path while this instance held the DB open — a git
* worktree removed and re-added, or `rm -rf .codegraph` + `codegraph init` —
* our open fd points at the now-unlinked inode and can never see the new
* index, so every query returns the pre-removal snapshot until the process
* restarts (#925). When that's detected, open the live file at the same path,
* rebuild the query layers, and swap them IN PLACE, so every holder of this
* instance (the MCP daemon's default project, cached projectPath connections)
* heals without a restart. Returns true iff it reopened.
*
* POSIX-only in practice: `isReplacedOnDisk` never fires on Windows (an open
* file can't be unlinked there, and st_ino is unreliable).
*/
reopenIfReplaced(): boolean {
if (!this.db.isReplacedOnDisk()) return false;
const dbPath = this.db.getPath();
// Open the live file FIRST — if that throws (e.g. mid-recreate), the old
// handle stays in place and the caller retries on the next query, rather
// than leaving this instance with no connection at all.
const fresh = DatabaseConnection.open(dbPath);
const stale = this.db;
this.db = fresh;
this.queries = new QueryBuilder(fresh.getDb());
this.wireLayers();
// Releasing the dead handle also frees the leaked db/-wal/-shm fds that were
// pinning the unlinked inode (#925).
try { stale.close(); } catch { /* the old inode is gone; closing just frees fds */ }
return true;
}
// ===========================================================================
// Lifecycle Methods
// ===========================================================================