Context
Codex Desktop support was bootstrapped locally during session onboarding, but the repository does not yet contain the reusable Codex harness bridge files. The current local checkout contains the support surface needed for future Codex sessions to start from the same MCP and shared skill substrate instead of re-deriving it manually.
Observed local change surface:
.codex/CODEX.md documents the Codex identity (neo-gpt) and the GitHub auth caveat specific to Codex sandboxing..codex/config.template.toml provides the committed, reusable Codex MCP server template for the Neo GitHub Workflow, Knowledge Base, Memory Core, and Neural Link servers..codex/config.toml is the per-user customized local config and must stay ignored..agents/skills/* symlinks expose shared Progressive Disclosure skills for harnesses that look under .agents/skills..claude/skills/debugging-antigravity should be removed because that skill is Antigravity IDE specific, not a generic Claude or Codex shortcut.
The Problem
Without these files in the repo, every Codex Desktop session has to reconstruct the same local wiring: minimal Codex fallback identity notes, MCP server definitions, npm run --silent handling to avoid stdout pollution, and shared skill discovery through .agents/skills. That creates onboarding friction, increases the chance of divergent local configs, and weakens the A2A continuity path for the new @neo-gpt swarm member.
This is distinct from #10478. That ticket seeds the Memory Core graph identity for Codex. This ticket codifies the repo-local harness support that lets Codex sessions boot with the correct identity note, MCP tools, and shared skills.
The ticket also corrects one harness-boundary leak discovered during intake: debugging-antigravity is useful for Gemini running inside Google Antigravity, but it should not be exposed as a default Claude or Codex skill shortcut.
The Architectural Reality
The existing repo already treats harness bootstrap as first-class infrastructure:
.claude/CLAUDE.md delegates to the shared AGENTS.md operational mandate..claude/skills/* exposes many Progressive Disclosure skills as symlinks to .agent/skills/*..gemini/settings.template.json provides a committed harness template while .gemini/settings.json remains local.AGENTS_STARTUP.md documents harness memory-file wiring and worktree/bootstrap behavior.
Codex needs the same general pattern, but with Codex-specific names and config shape:
.codex/CODEX.md is intentionally tiny. It should only carry Codex-specific identity and sandbox caveat details because the full operational rules are already loaded through turn-level memory..codex/config.template.toml is the committed template. It must be visible to Git and tracked in this PR..codex/config.toml is the customized per-user config copied from the template. It must be ignored so repo users can customize it locally..agents/skills/* is a compatibility bridge for Codex-style skill discovery while preserving .agent/skills/* as the canonical source of truth.- Harness-specific tactical skills should not be mirrored blindly.
debugging-antigravity remains canonical under .agent/skills/debugging-antigravity, but the shortcut should not appear under Claude or Codex harness folders.
The Fix
Commit the existing local Codex support surface and the harness-boundary cleanup:
- Add
.codex/CODEX.md.
- Add
.codex/config.template.toml.
- Add
.codex/config.toml to .gitignore.
- Add
.agents/skills/* symlinks to shared canonical .agent/skills/* directories, excluding debugging-antigravity.
- Remove
.claude/skills/debugging-antigravity.
No framework runtime logic changes are required.
Acceptance Criteria
Out of Scope
- No changes to Memory Core identity roots; #10478 covered the
@neo-gpt graph identity.
- No implementation of trio A2A wakeups; that remains the next coordination phase and should use its own ticket/intake path.
- No rename of
.agent or migration away from the existing .claude / .gemini conventions.
- No deletion of the canonical
.agent/skills/debugging-antigravity skill.
- No duplication of
AGENTS.md or AGENTS_STARTUP.md content inside .codex/CODEX.md.
- No committing
.codex/config.toml; it is the customized local config, not the template.
Avoided Traps / Gold Standards Rejected
- Do not ignore the template. The template is the shareable artifact and must be committed.
- Do ignore the customized config. Repo users need a local
.codex/config.toml they can customize without dirtying the worktree.
- Do not duplicate the skill contents under
.agents/skills. Symlinks preserve .agent/skills as the single source of truth and avoid skill drift.
- Do not blindly mirror harness-specific skills across every harness. Antigravity debugging belongs to Antigravity context; exposing it as a Claude or Codex shortcut increases false triggers.
- Do not turn
.codex/CODEX.md into another rules document. The turn-loaded memory already supplies the canonical rules.
- Do not treat this as a graph-identity ticket. Repo harness bootstrap and Memory Core AgentIdentity are separate substrates.
Related
Origin Session ID: 049d9856-daa8-4151-b089-8ce8c7ed56f4
Handoff Retrieval Hints:
query_raw_memories("Codex onboarding .codex config template .agents skill symlinks neo-gpt")query_raw_memories("debugging-antigravity Claude Codex harness-specific skill exposure")query_raw_memories("NL is the embodiment layer A2A is the continuity layer Codex")
Context
Codex Desktop support was bootstrapped locally during session onboarding, but the repository does not yet contain the reusable Codex harness bridge files. The current local checkout contains the support surface needed for future Codex sessions to start from the same MCP and shared skill substrate instead of re-deriving it manually.
Observed local change surface:
.codex/CODEX.mddocuments the Codex identity (neo-gpt) and the GitHub auth caveat specific to Codex sandboxing..codex/config.template.tomlprovides the committed, reusable Codex MCP server template for the Neo GitHub Workflow, Knowledge Base, Memory Core, and Neural Link servers..codex/config.tomlis the per-user customized local config and must stay ignored..agents/skills/*symlinks expose shared Progressive Disclosure skills for harnesses that look under.agents/skills..claude/skills/debugging-antigravityshould be removed because that skill is Antigravity IDE specific, not a generic Claude or Codex shortcut.The Problem
Without these files in the repo, every Codex Desktop session has to reconstruct the same local wiring: minimal Codex fallback identity notes, MCP server definitions,
npm run --silenthandling to avoid stdout pollution, and shared skill discovery through.agents/skills. That creates onboarding friction, increases the chance of divergent local configs, and weakens the A2A continuity path for the new@neo-gptswarm member.This is distinct from #10478. That ticket seeds the Memory Core graph identity for Codex. This ticket codifies the repo-local harness support that lets Codex sessions boot with the correct identity note, MCP tools, and shared skills.
The ticket also corrects one harness-boundary leak discovered during intake:
debugging-antigravityis useful for Gemini running inside Google Antigravity, but it should not be exposed as a default Claude or Codex skill shortcut.The Architectural Reality
The existing repo already treats harness bootstrap as first-class infrastructure:
.claude/CLAUDE.mddelegates to the sharedAGENTS.mdoperational mandate..claude/skills/*exposes many Progressive Disclosure skills as symlinks to.agent/skills/*..gemini/settings.template.jsonprovides a committed harness template while.gemini/settings.jsonremains local.AGENTS_STARTUP.mddocuments harness memory-file wiring and worktree/bootstrap behavior.Codex needs the same general pattern, but with Codex-specific names and config shape:
.codex/CODEX.mdis intentionally tiny. It should only carry Codex-specific identity and sandbox caveat details because the full operational rules are already loaded through turn-level memory..codex/config.template.tomlis the committed template. It must be visible to Git and tracked in this PR..codex/config.tomlis the customized per-user config copied from the template. It must be ignored so repo users can customize it locally..agents/skills/*is a compatibility bridge for Codex-style skill discovery while preserving.agent/skills/*as the canonical source of truth.debugging-antigravityremains canonical under.agent/skills/debugging-antigravity, but the shortcut should not appear under Claude or Codex harness folders.The Fix
Commit the existing local Codex support surface and the harness-boundary cleanup:
.codex/CODEX.md..codex/config.template.toml..codex/config.tomlto.gitignore..agents/skills/*symlinks to shared canonical.agent/skills/*directories, excludingdebugging-antigravity..claude/skills/debugging-antigravity.No framework runtime logic changes are required.
Acceptance Criteria
.codex/CODEX.mdfallback note that identifies theneo-gptGitHub identity and auth verification path without duplicating the canonical agent instructions..codex/config.template.tomlcontaining the four Neo MCP server definitions usingnpm run --silent..codex/config.template.tomlis tracked and not ignored..codex/config.tomlis ignored as the per-user customized config..agents/skills/exposes shared Progressive Disclosure skills through symlinks back to.agent/skills/..agents/skills/debugging-antigravityis absent..claude/skills/debugging-antigravityis removed.devand uses a commit subject ending in this ticket ID.Out of Scope
@neo-gptgraph identity..agentor migration away from the existing.claude/.geminiconventions..agent/skills/debugging-antigravityskill.AGENTS.mdorAGENTS_STARTUP.mdcontent inside.codex/CODEX.md..codex/config.toml; it is the customized local config, not the template.Avoided Traps / Gold Standards Rejected
.codex/config.tomlthey can customize without dirtying the worktree..agents/skills. Symlinks preserve.agent/skillsas the single source of truth and avoid skill drift..codex/CODEX.mdinto another rules document. The turn-loaded memory already supplies the canonical rules.Related
Origin Session ID: 049d9856-daa8-4151-b089-8ce8c7ed56f4
Handoff Retrieval Hints:
query_raw_memories("Codex onboarding .codex config template .agents skill symlinks neo-gpt")query_raw_memories("debugging-antigravity Claude Codex harness-specific skill exposure")query_raw_memories("NL is the embodiment layer A2A is the continuity layer Codex")