You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 0ba94f5
Browse filesBrowse the repository at this point in the historyBrowse files
fix: disable Codex's Visualize skill, whose output ACP clients show as raw text (spacering-net#667)
Codex's bundled Visualize skill (`visualize@openai-bundled`, installed
by the ChatGPT and Codex desktop apps) ends a reply with a content
reference that only the OpenAI apps render. ACP clients showed it as raw
`visualize{"path":...}` text. Upstream: openai/codex#40525.
- Every new, resumed, loaded, and forked thread gets the config rule
`skills.config = [{name: "visualize:visualize", enabled: false}]`. Codex
applies session rules after the user's config, and a later rule for the
same selector replaces an earlier one, so this rule wins.
- The skill commands and `/skills` leave the skill out, because
`skills/list` still reports it as enabled.
- Matching by name needs no `skills/list` request at session start, and
still applies if the plugin is installed mid-session. Limitation: a
skill of another plugin with the same qualified name is disabled too.
Documented in `readme-dev.md`.
- Tests for v1 and v2. Verified with real Codex 0.164.0: without the
rule, the model lists `visualize:visualize` among its skills; with it,
the skill is gone and the other skills stay.
0 commit comments