codex - 💡(How to fix) Fix Codex App truncates injected Available skills metadata and deterministically drops the same repo-local skills [1 comments, 2 participants]

What version of the Codex App are you using?

Codex App 26.417.41555 (CFBundleVersion 1858)

What subscription do you have?

Pro

What platform is your computer?

macOS / Darwin arm64

What issue are you seeing?

Codex App is not failing to discover my repo-local skills on disk. Instead, it appears to truncate the injected ### Available skills block before the model sees it, which causes the same tail subset of skills to be omitted from every desktop-app session.

This is deterministic in my workspace:

• Repo-local skills on disk: 60
• System skills also available: 5
• In the desktop app session prompt, only 55 unique skill names are injected
• The same 9 repo-local skills are always missing from the injected list
• In CLI, these skills do not go missing in the same way

Missing repo-local skills in the desktop app session:

thesis-tracker
trello
vault-architecture
web-image-search
x-rebound-framework
xlsx
yimeng-tst-decoder
zbgd-video-tts
zbgd-writer

This breaks implicit/adaptive skill routing in the app, because the model never sees those skills as available even though they are present on disk.

Why I think this is injection truncation rather than discovery failure

I checked the local rollout file for the affected desktop-app thread and found that the developer-injected prompt block includes:

"truncation_policy": { "mode": "tokens", "limit": 10000 }

In the same rollout, the injected ### Available skills section ends after theme-factory / tracking-report-writing, and the 9 skills above are absent.

That strongly suggests the desktop app is truncating the injected skills metadata block rather than failing to discover those skill folders.

I also see desktop logs like:

Skills/list request cwdsCount=1 forceReload=false
Skills/list missing short_description count ... missingShortDescriptionCount=63
Skills/list ok entryCount=1

So the app is clearly going through skills/list, but the final session-visible skill list is still incomplete.

Reproduction

In a repo/workspace with many repo-local skills:

1. Put many valid skills under the repo-local skill directory (in my case the source of truth is .shared/skills/ and the project mapping points to the same content).
2. Open the workspace in Codex App.
3. Start a fresh conversation and ask the agent to list the explicitly available skills.
4. Compare that list with the actual folders under the repo-local skill directory.
5. Repeat in a fresh desktop-app thread.

Actual behavior

The app injects a truncated Available skills list, and the same trailing subset of repo-local skills is silently omitted from the model prompt on every fresh app thread.

Expected behavior

The desktop app should not silently drop repo-local skills from the model-visible Available skills list just because the metadata block gets large.

If there is a prompt budget, I would expect one of these instead:

• do not truncate the skills list at all
• compress/summarize metadata without dropping skills
• inject a manifest that preserves all skill names/paths and lazy-loads longer descriptions separately
• fall back to discovery-at-use time rather than deterministically dropping tail entries

A fixed truncation cap here is especially painful because skill bodies are already loaded lazily; the app is only truncating the metadata that tells the model what exists.

Additional context

This is easy to misdiagnose as:

• broken symlink handling
• malformed SKILL.md
• missing frontmatter
• missing UI metadata like short_description

But in this case the missing set is deterministic and aligned with prompt truncation, not random discovery failure.

If useful, I can provide the exact injected skill list vs. on-disk skill list from the affected workspace.