Skip to content

feat(cmind): add Codex support and multi-agent backend switchingFeat/codermind codex support - #96

Open
HYS (HuYaSen) wants to merge 10 commits into
mainfrom
feat/codermind-codex-support
Open

feat(cmind): add Codex support and multi-agent backend switchingFeat/codermind codex support#96
HYS (HuYaSen) wants to merge 10 commits into
mainfrom
feat/codermind-codex-support

Conversation

@HuYaSen

Copy link
Copy Markdown
Collaborator

Summary

Add Codex CLI as a supported CoderMind runtime while preserving the existing Claude Code and GitHub Copilot integrations.

CoderMind now provisions Claude, Copilot, and Codex together while allowing the active encoder/decoder backend to be switched independently through .cmind/config.toml.

Key changes

  • Add a Codex runtime adapter based on codex exec.
  • Generate project-scoped Codex skills under .agents/skills/.
  • Register rpg-tools safely in .codex/config.toml.
  • Add backend switching for copilot, claude, and codex.
  • Provision all supported agent integrations regardless of the active backend.
  • Isolate per-agent MCP configuration failures so one client cannot block the others.
  • Add the standard MCP resource rpg://tree while retaining all existing MCP tools.
  • Reload MCP graph state when rpg.json changes.

Reliability improvements

  • Make forced planning bypass stale checkpoints consistently.
  • Require complete data-flow type coverage during base-class generation.
  • Resolve qualified interface edges during call-graph validation.
  • Persist RPG Edit code and review results for reliable resume behavior.
  • Harden code-generation retries and branch naming.
  • Recover corrupted RPG data through the central loader.
  • Preserve dependency-graph file stems.
  • Keep Codex sandbox bypass as an explicit opt-in for trusted automation.

Compatibility

Shared planning, encoding, code generation, MCP, and RPG Edit paths remain backend-neutral.

  • GitHub Copilot continues to use the existing MCP tools.
  • Claude Code retains its command and MCP configuration.
  • Codex uses project skills and the MCP resource compatibility path.
  • User-owned MCP configuration is preserved.
  • A configuration conflict for one agent no longer prevents other agents from being configured.

Validation

  • CoderMind suite: 1526 passed
  • cmbench suite: 243 passed
  • Codex Encoder E2E: PASS
    • full encode
    • MCP access
    • CRUD hooks
    • RPG Edit and final synchronization
  • Codex Decoder E2E: PASS
    • 48/48 code-generation tasks
    • 178 final tests passed
    • repository and RPG heads synchronized
  • Live GitHub Copilot CLI MCP smoke: PASS
    • invoked list_rpg_tree
    • recorded client_context=copilot-agent
  • git diff --check: PASS

- add explicit Codex sandbox bypass support for trusted automation
- avoid codegen batch branch collisions and preserve failed retry branches
- recover corrupted RPG data through the central service loader
- make multilingual and planning regressions deterministic
- rebuild interface subtrees instead of restoring stale output under --force
- forward force semantics through the consolidated plan workflow
- record MCP client context for real agent-call telemetry
- add regression coverage for force and MCP evidence
Keep MCP provisioning isolated across supported agents, refresh long-lived MCP graph state, enforce planning coverage, resolve qualified interface edges, and persist RPG Edit resume artifacts.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant