Skip to content

Codex

meshpoint connect codex

writes three things:

  • ~/.codex/hooks.json: the hook entries, with additionalContextLimit raised from Codex’s 2,500-character default so the recall inject fits.
  • ~/.codex/rules/meshpoint.rules: prefix rules allowing meshpoint items put and meshpoint items retract. Codex has no per-skill allowed tools, and “allow” also runs the command outside the sandbox so it can reach the platform.
  • ~/.agents/skills/meshpoint/SKILL.md: the checkpoint skill.

With --project <repo> the first two land in <repo>/.codex/ and the skill in <repo>/.agents/skills/.

Codex speaks the same hook protocol as Claude Code: event names, payload fields, hookSpecificOutput.additionalContext, and Stop decision: block. meshpoint tap codex is the same tap with a reader for Codex rollouts (response_item messages, apply_patch paths). Session start injects recall and presence; prompts and replies are spooled; a due checkpoint blocks the stop once with the nudge; compaction and session end flush.

Project-level .codex/ files load only once the repository is trusted in ~/.codex/config.toml:

[projects."/abs/path/to/repo"]
trust_level = "trusted"

A -c projects.<dir>.trust_level=trusted command-line override does not load them. User-level hooks in ~/.codex/hooks.json run regardless.

Codex asks you to review new hooks with /hooks before it runs them. codex exec runs hooks too; --dangerously-bypass-hook-trust skips the review for automation.

  • If every turn fails with model_not_found, the model slug pinned in config.toml is missing from the account’s model list (~/.codex/models_cache.json). Pick one that exists with -m.
  • Redirect stdin from /dev/null in scripts; codex exec otherwise waits on it.