Repository navigation
[v0.9.0][Windows] persistence=true crashes the indexing worker for repositories under CJK paths #1171
Description
Version
codebase-memory-mcp 0.9.0
Platform
Windows (x64)
Install channel
GitHub release archive / install.sh / install.ps1
Binary variant
standard
What happened, and what did you expect?
Summary
On Windows 11 x64 with the official v0.9.0 Windows amd64 standard binary, a clean full index of the public DeusData/codebase-memory-mcp repository succeeds from a CJK path when persistence=false, but the supervised worker exits nonzero when the same repository is indexed from the same CJK path with persistence=true.
An ASCII NTFS junction pointing to that exact CJK directory succeeds with persistence=true and writes .codebase-memory/graph.db.zst. The three-way result reproduced twice from clean project names, isolating the failure to persistence export for the non-ASCII path.
The failed run leaves a readable SQLite database (PRAGMA integrity_check = ok) and index_status reports ready, but the explicit index call returns exit_nonzero and no artifact is created.
Expected behavior
- A clean persistent index should work for a repository under a CJK path and write
.codebase-memory/graph.db.zst. - The canonical CJK path and an ASCII junction to the same directory should behave identically.
- If export fails, the response and worker log should expose the actual Windows path/file error.
- A failed explicit index should not be reported as
readywithout making the failed persistence state clear.
Query-phase CJK path behavior is tracked separately by #903 and is out of scope here.
Related: #400, #434, #903, and PR #891.
Reproduction
Environment
- OS: Windows 11 x64, build
10.0.26200 - PowerShell:
5.1.26100.8875; active code page:936 - Binary: official v0.9.0 Windows amd64 standard release
- SHA-256:
9A205FA5AE759FBC866BFE1554F0C05A303BE9AE6E0A00F94D875DC0C25E0680 - Public repository:
https://github.com/DeusData/codebase-memory-mcp.git - Tag / commit:
v0.9.0/b637e3330c96cfe452da623db068c241aaa3ec01
Setup
$cbm = "$env:USERPROFILE\.local\bin\codebase-memory-mcp.exe"
$cjkRepo = Join-Path $env:LOCALAPPDATA 'Temp\cbm-持久化复现'
$asciiRepo = Join-Path $env:LOCALAPPDATA 'Temp\cbm-cjk-repro'
git clone --depth 1 --branch v0.9.0 --single-branch `
https://github.com/DeusData/codebase-memory-mcp.git $cjkRepo
git -C $cjkRepo rev-parse HEAD
# b637e3330c96cfe452da623db068c241aaa3ec01
New-Item -ItemType Junction -Path $asciiRepo -Target $cjkRepoEvery scenario started with delete_project, so every index was a clean full run rather than an incremental run.
1. CJK path, persistence=false
$project = 'cbm-v090-public-cjk-pfalse-r2'
& $cbm cli delete_project --project $project
& $cbm cli index_repository --repo-path $cjkRepo --mode moderate --name $project --persistence falseResult: status=indexed, nodes=13007, edges=73886, skipped_count=0, artifact_present=false; SQLite integrity_check=ok; File nodes: 556.
2. Same CJK path, persistence=true
$project = 'cbm-v090-public-cjk-ptrue-r2'
& $cbm cli delete_project --project $project
& $cbm cli index_repository --repo-path $cjkRepo --mode moderate --name $project --persistence true{"status":"error","outcome":"exit_nonzero","hint":"Indexing worker crashed on a file. The crash was contained (the server survived). Re-run to retry; a future release isolates the culprit file.","repo_path":"%LOCALAPPDATA%\\Temp\\cbm-持久化复现"}Result: worker exit 1; empty worker log; database integrity_check=ok; nodes=13007, edges=73864, File nodes 556; skipped_count not reported; artifact absent. index_status immediately returned ready.
3. ASCII junction to the same CJK directory, persistence=true
$project = 'cbm-v090-public-ascii-ptrue-r2'
& $cbm cli delete_project --project $project
& $cbm cli index_repository --repo-path $asciiRepo --mode moderate --name $project --persistence trueResult: status=indexed, nodes=13007, edges=73754, skipped_count=0, artifact_present=true; SQLite integrity_check=ok; File nodes 556; artifact 9,162,274 bytes, compression level 9, commit b637e3330c96cfe452da623db068c241aaa3ec01.
Repeated control matrix
| Scenario | Run 1 | Run 2 |
|---|---|---|
| CJK, persistence=false | indexed; 13,007 nodes / 74,585 edges; skipped 0; no artifact; SQLite ok | indexed; 13,007 / 73,886; skipped 0; no artifact; SQLite ok |
| CJK, persistence=true | worker exit 1; 13,007 / 73,895; skipped not reported; no artifact; SQLite ok | worker exit 1; 13,007 / 73,864; skipped not reported; no artifact; SQLite ok |
| ASCII junction, persistence=true | indexed; 13,007 / 73,941; skipped 0; artifact 9,156,778 bytes; SQLite ok | indexed; 13,007 / 73,754; skipped 0; artifact 9,162,274 bytes; SQLite ok |
The edge-count variation occurs across all three scenarios and is not used as the pass/fail criterion. The invariant signals are the worker exit status and artifact creation.
Workaround
Create an ASCII NTFS junction to the CJK repository directory, delete the old project, and run the full persistent index through the junction.
Logs
level=info msg=index.supervisor.reap outcome=exit_nonzero exit_code=1 signal=0
level=warn msg=index.supervisor.worker_failed outcome=exit_nonzero exit_code=1 log=%USERPROFILE%/.cache/codebase-memory-mcp/logs/.worker-42772.log
{"status":"error","outcome":"exit_nonzero","hint":"Indexing worker crashed on a file. The crash was contained (the server survived). Re-run to retry; a future release isolates the culprit file.","repo_path":"%LOCALAPPDATA%\\Temp\\cbm-持久化复现"}
Referenced worker log:
%USERPROFILE%\.cache\codebase-memory-mcp\logs\.worker-42772.log
Length: 0 bytes
Failed database:
integrity=ok
nodes=13007
edges=73864
files=556
skipped_count=not reported in error JSON
artifact_exists=false
ASCII junction control:
level=info msg=index.supervisor.reap outcome=clean exit_code=0 signal=0
{"project":"cbm-v090-public-ascii-ptrue-r2","skipped_count":0,"nodes":13007,"edges":73754,"artifact_present":true,"status":"indexed"}
Diagnostics trajectory (memory / performance / leak issues)
Not applicable. This is a deterministic path- and persistence-dependent worker exit; no memory growth, hang, or leak was observed.
Project scale (if relevant)
Public repository at b637e33:1,818 tracked files; 556 indexed File nodes; 13,007 graph nodes; 73,754–74,585 graph edges across clean runs; 8 excluded directories; 0 skipped files on successful runs.
Confirmations
- I searched existing issues and this is not a duplicate.
- My reproduction uses shareable code (a dummy snippet or a public OSS repository), not proprietary code.
Activity
[-][v0.9.0][Windows] persistence export crashes the indexing worker for repositories under CJK paths[/-][+][v0.9.0][Windows] persistence=true crashes the indexing worker for repositories under CJK paths[/+]DeusData commentedon Jul 19, 2026
Thanks for the controlled CJK-path versus ASCII-junction matrix. This is now high priority in
0.9.1-rcunder the Windows platform epic #394. The evidence isolates the persistent-artifact export path; the empty worker log andreadystatus after a failed persistence request are part of the bug, not separate cosmetic concerns.Akii-kawaii commentedon Jul 20, 2026
DeusData commentedon Oct 1, 2026
Thank you again, @Akii-kawaii, for the CJK-path versus ASCII-junction matrix. Isolating the failure to the persistence export made the cause easy to find, and I'm sorry this took so long.
The artifact code still used ANSI-code-page file APIs on Windows, so any non-ASCII path failed:
fopen) and the rename into place (MoveFileExA);rename) and the existence check (stat);.gitattributesfile created next tograph.db.zst(open).So with
persistence=trueunder a CJK path, the export could not create its temp file, and nograph.db.zstwas written. The ASCII junction worked because the same bytes then stayed inside the code page.The fix in #2461 routes all of these through the project's UTF-8 path layer, which uses the wide-character Windows APIs (
_wfopen,MoveFileExW,_wstat64). We verified it on a real Windows machine whose ANSI code page cannot represent the path (cp1252):open_temp errno=2, 3 runs out of 3;.gitattributesgets itsmerge=oursline, 3 runs out of 3.If you still have your matrix set up, a retest with the next release (or a build of #2461) would be very welcome.
Merge pull request #2461 from DeusData/fix/issue-1171
DeusData commentedon Oct 7, 2026
The fix from #2461 has now landed on main in d19cab8 and will ship in the next release. @Akii-kawaii, thank you again for your patience and for the offer to retest: once the release is out, a run of your CJK-path versus ASCII-junction matrix with
persistence=truewould be the best confirmation we could get. If anything still fails, just comment here and we will reopen it.fix(artifact): write and detect artifacts through wide paths on Windo…
fix(artifact): create .gitattributes through the UTF-8 path layer