Inconsistencies/bugs in the vscode codex extension

[Enferlain]

Version

codebase-memory-mcp 0.8.2

Platform

Windows (x64) in wsl

Install channel

GitHub release archive / install.sh / install.ps1

Binary variant

ui

What happened, and what did you expect?

GPT appears inconsistent when using codebase-memory-mcp tools. Sometimes it forgets that the tools exist or avoids using them even when project instructions say to prefer them. In this case, it did call query_graph, but the result looked incorrectly serialized: the Cypher query requested multiple projected columns with aliases, but the returned response contained only one column, labels(n), losing the other requested fields.

Expected behavior: the model should reliably use available codebase-memory tools when the context has the usage steps described, and query_graph should return all requested projected columns/aliases from the Cypher query.

Not exactly sure how I'm supposed to troubleshoot these incosistencies/issues. I will update the thread with news when I encounter them

Reproduction

Project:

/home/imi/Projects/sd-scripts
Indexed project name: home-imi-Projects-sd-scripts

Tool call:

{
  "project": "home-imi-Projects-sd-scripts",
  "max_rows": 80,
  "query": "MATCH (n) WHERE (n:Class OR n:Function OR n:Method OR n:Module) AND n.file_path STARTS WITH 'library/metadata' RETURN labels(n)[0] AS label, n.name AS name, n.qualified_name AS qn, n.file_path AS file, n.start_line AS line ORDER BY n.file_path, n.start_line LIMIT 80"
}

Actual output:

{
  "columns": ["labels(n)"],
  "rows": [
    ["[\"Class\"]"],
    ["[\"Class\"]"],
    ["[\"Class\"]"],
    ["[\"Class\"]"]
  ],
  "total": 4
}

Expected output: rows with all requested columns, something like:

{
  "columns": ["label", "name", "qn", "file", "line"],
  "rows": [
    ["Class", "...", "...", "library/metadata/...", 123]
  ]
}

A similar query scoped to library/logging/resource_monitor also collapsed to only labels(n).

The graph index itself seemed healthy:

{
  "project": "home-imi-Projects-sd-scripts",
  "nodes": 11074,
  "edges": 50079,
  "status": "ready"
}

Other tools worked correctly afterward, including search_code, get_code_snippet, trace_path, and index_status.

Logs

No stderr/logs captured from the MCP binary during this run.

Project scale (if relevant)

Nodes: 11074, Edges: 50079, Files: graph schema reported 748 File nodes

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.

[DeusData]

Thanks. I’ve triaged this under #395 and #594. The concrete, testable bug is the query_graph projection/alias collapse; the agent-usage variability is related integration behavior but the query serialization issue should get its own focused fixture.