Codex CLI fails on deployed app: "Error loading configuration: No such file or directory"

[dgokeeffe]

Symptom

Running codex inside the deployed CODA app's terminal returns:

Error loading configuration: No such file or directory (os error 2)

Diagnosis

Confirmed empirically via the deployed app's terminal:

$ ls -la ~/.codex/
-rw------- 1 app app  114 May  6 08:42 .env
-rw-r--r-- 1 app app    3 May  6 05:56 .personality_migration
-rw-r--r-- 1 app app 3678 May  6 05:53 AGENTS.md
-rw-r--r-- 1 app app  535 May  6 05:53 config.toml
drwxr-xr-x 2 app app 4096 May  6 05:56 memories
drwxr-xr-x 3 app app 4096 May  6 05:56 tmp
$ [ -f ~/.codex/databricks-models.json ] && echo OK || echo MISSING
MISSING

config.toml references model_catalog_json = "databricks-models.json". Codex resolves that relative to CODEX_HOME (~/.codex), tries to open ~/.codex/databricks-models.json, fails — hence the error.

The deploy doesn't have it because databricks sync skipped it. The .codex/ directory doesn't appear in the synced workspace at all:

$ databricks workspace list /Workspace/Users/<owner>/apps/coding-agents/.codex/ -p daveok
Error: Path doesn't exist.

Root cause

.gitignore has:

.codex/*
!.codex/databricks-models.json

Git honors the negation — .codex/databricks-models.json is tracked. databricks sync (which respects .gitignore to decide what to upload) appears to honor the blanket-ignore but not the negation pattern, so it skips the catalog along with everything else. Net effect: file is in your local repo, in the GitHub repo, and in git ls-tree, but never makes it to the deployed app's filesystem.

setup_codex.py:100's guard then short-circuits silently:

if catalog_src.exists() and catalog_src.resolve() != catalog_dst.resolve():
    shutil.copyfile(catalog_src, catalog_dst)

Since catalog_src doesn't exist on the synced machine, the copy is skipped, no warning, no error — ~/.codex/ ends up without the catalog.

Fix (PR coming)

Replace the .codex/* + !X negation pattern in .gitignore with explicit per-file/dir ignores:

.codex/config.toml
.codex/.env
.codex/AGENTS.md
.codex/.personality_migration
.codex/memories/
.codex/tmp/
.codex/sessions/
.codex/cron/
.codex/skills/

Same runtime artifacts ignored. databricks-models.json is no longer touched by any pattern → both git and databricks sync track it.

This is a general antipattern lesson: gitignore negations work in git but several gitignore-honoring tools skip them. Better to enumerate ignores explicitly than rely on * + !.

Out of scope

• setup_codex.py could also defensively log a warning when catalog_src doesn't exist (it currently passes silently). Worth a follow-up but not required for the fix to land.
• Same antipattern probably exists in other .gitignore-honoring tools we use (Docker, etc.). Not auditing here.