Skip to content

feat: v0.8 roadmap — Codex MCP/plugin integration and explicit session wake #46

Closed

Description

@abruption

Prerequisites: #44 (v0.6 Codex adapter) and #45 (v0.7 diagnostics and reply/status semantics).

Motivation

An agent using session-peer through shell commands has to assemble executable paths, environment overrides, session IDs, and SSH arguments. It can also encounter sandbox restrictions even when the destination exists. A structured tool interface would make this integration easier to configure and inspect.

Target v0.8: a Codex-facing MCP integration and plugin packaging, plus a separately opt-in --wake capability for queued Codex sessions.

Does the official path already cover it?

Codex provides native queueing, MCP integration, and session-resume facilities. Reuse those supported mechanisms rather than inventing a model runner or terminal-input transport. The new value is exposing session-peer's cross-agent/local/SSH capabilities as narrow tools and offering explicit activation when a user chooses it.

An MCP process having different access from a Codex shell sandbox is a hypothesis to verify, not a reason to describe MCP as an approval bypass.

Proposed behaviour

  • Expose structured session listing and message sending through MCP; expose status/wait only where the underlying implementation can provide it. Reuse the CLI's target resolution, result semantics, and errors.
  • Package the integration for Codex with setup and usage instructions. Keep core CLI operation available without MCP/plugin dependencies; isolate optional integration requirements.
  • Configure allowed destinations and capabilities rather than exposing arbitrary shell execution through a tool. Preserve the host application's approval model and existing SSH authentication.
  • Add explicit session-peer send --to codex:<thread-uuid> --wake "message" support using supported activation/resume mechanisms. Plain send must remain queue-only.
  • Explain that wake can incur model usage and modify session history. Resolve the selected home and original working directory before activation; do not silently bypass project trust, enable unsafe approval flags, fork the thread, or create concurrent writers.
  • Define already-active, inactive, missing/uninitialised, archived, and locked-session behaviour from native capabilities and tests. Fail clearly when safe activation cannot be established. A queue-success/wake-failure result must preserve the queue ID so callers do not resend.

Evidence and limits

Earlier manual testing queued messages for a remote Codex session and processed them through explicit resume in the original project directory. A resume attempt outside that directory failed its project-trust check. A Codex-to-Claude UDS send was denied in one sandbox configuration and succeeded through approved execution in another.

These are feasibility observations, not a guarantee that one resume consumes an entire queue, that lock-file presence proves activity, or that MCP removes access restrictions. Establish those boundaries before shipping automatic activation.

Acceptance criteria

  • A Codex MCP client lists and messages Claude/Codex targets locally and over a configured SSH route using the same core behaviour as the CLI.
  • Installation, optional dependencies, destination restrictions, and approval behaviour are documented and exercised; actual UDS access from the MCP process is measured.
  • Plain send never activates a session; --wake is explicit and tested for inactive, active, missing, and locked targets without concurrent resume or duplicate submission.
  • Wake uses the selected Codex home and correct project directory; trust/approval failures are surfaced without automatic bypass.
  • Queue success and activation failure remain distinguishable in human/JSON/tool results; no automatic resend on timeout or ambiguous outcome.
  • Optional plugin/MCP installation does not change the standalone CLI's dependency requirements; integration tests and PR verification describe supported versions/platforms.

Implementation follows linked issue → branch → PR changes; delivery is a roadmap target, not an automatic configuration change on users' machines.

Activity

  1. added 2 commits that reference this issue on Sep 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions