Skip to content

Windows: index worker crashes (0xC0000005) in parallel.extract on whole mixed C#/Vue repo; each sub-folder indexes fine alone #2785

Open

Description

@labib03

Summary

On Windows, index_repository on a mixed C# / Vue / JS / Markdown / SQL repo (~990 files after ignores) ends with index worker ended with crash (exit=-1073741819) (0xC0000005, access violation) during parallel.extract. Every top-level sub-folder of the same repo indexes fine on its own; only the whole-repo index crashes. This looks similar in shape to #2349 (each package alone is fine, only the union crashes), but the signature differs: Windows, access violation, extract phase.

Environment

  • codebase-memory-mcp 0.12.1 (install.ps1, --skip-config, checksum verified)
  • Windows 11 Pro 10.0.26300, amd64
  • Repo: .NET Framework 4.8 solution (ASP.NET WebForms + WebApi, several C# projects), Vue 2 single-file components without a bundler, plain JS/CSS, SQL scripts, Markdown docs. Private repo, so I can't attach it.

What I ran

codebase-memory-mcp cli index_repository --repo-path <repo root>, with and without --mode fast.

Result

Attempt Outcome
default mode worker crash, exit=-1073741819
--mode fast same crash
--mode fast + CBM_INDEX_SINGLE_THREAD=1 same crash
--mode fast + CBM_WORKERS=1 + CBM_INDEX_SINGLE_THREAD=1 same crash
each top-level folder indexed separately (5 .NET/web projects, plus docs, Database, specifications, tooling dirs, hidden dirs) all succeed

Notes:

  • The files that were started but not finished when the worker died differ between runs, so I don't think one specific file triggers it.
  • The two env vars come from strings in the binary, not from the docs. The worker log still shows parallel.extract after setting them, so they may be no-ops here.
  • The first lines of stderr are always mem.allocator.preloading_completed still_preloading=false detail=allocator_was_still_preloading,_so_arena_creation_and_purging_were_disabled. Not sure if related.
  • Last log lines before the crash are parallel.extract.file.start ... / ...file.done ... / parallel.extract.progress done=N total=M, with N around 110 to 150 of ~990 to ~1190 depending on the ignore set.

Expected

The whole-repo index completes, as each part does alone.

Questions

  • Is there a supported way to get a crash dump or a more verbose extract-phase trace on Windows (CBM_DIAGNOSTICS?) that I can attach?
  • Is a Windows-specific allocator/arena problem under parallel.extract a known thing?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    stability/performanceServer crashes, OOM, hangs, high CPU/memorywindowsWindows-specific issues

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions