codex-orchestrate

Codex Orchestrate

One skill for reaching Codex on the MCP rail. Read the mode from the user's intent, then dispatch it as a framing inside a single mcp__codex__codex call.

Rebased 2026-08-09. This skill used to map each mode to a distinct codex_* MCP tool (codex_brainstorm, codex_discuss, codex_coordinate, codex_review, codex_curate, codex_consult). Those tools were exposed by tools/mcp/codex-orchestrator/, which was deleted with the rest of the ChatGPT-web lane; they were never mounted on any registered server even before that. The mode is now a prompt framing, not a tool name.

The rail

• mcp__codex__codex - start a Codex thread.
• mcp__codex__codex-reply - continue by threadId. The name is hyphenated; the underscore form does not resolve.

Each invocation is single-threaded and fresh-context. State that must persist goes in the prompt or in a file - there is no server-side session store any more.

Modes - what to put in the prompt

Intent / trigger	Mode	How to dispatch
Generate ideas, options, or approaches	brainstorm	ask for N distinct options with tradeoffs, or fan out 2 parallel calls with different framings and synthesize
Debate tradeoffs, pros and cons, X vs Y	discuss	2 parallel calls with opposed framings, then synthesize yourself
Split a task across roles and consolidate	coordinate	1 call per role in parallel, then consolidate yourself
Review or critique a plan / design / diff	review	1 call: hand over the artifact, ask for a critic pass with severities
Audit for risk, gaps, blind spots	audit	same as review, framed as "what is missing / what could fail"
Pick the single best option from candidates	curate	1 call: hand over the option set, ask for a selection plus justification
A governance decision weighing options vs risks	governance	same as curate, framed against the stated constraints
Fast classification of next steps	triage	1 call: ask for an ordered shortlist, nothing else

(audit==review, governance==curate, and a fix path is just coordinate; the labels exist so the router matches the user's wording.)

Behavior

• Keep the task scoped and decision-oriented; pass one clear prompt.
• Pick the mode from the user's verb; default to review for a generic "have Codex look at this".
• You are the synthesizer. The retired server did multi-peer debate and consolidation server-side; on this rail you fan out the calls and combine the results.
• Say which rail you used and whether the final synthesis was yours or Codex's.

Not this skill

• Per-call scoping - restricting a run's MCP servers, sealing its base, allowlisting its skills, or taking structured --output-schema output - use codex-invoke (exec rail). The MCP rail inherits the full base and cannot scope.
• Codex environment / session troubleshooting - codex-runtime-recovery, codex-session-reunion, codex-vscode-host-routing.
• A browser-mediated ChatGPT web lane - retired 2026-08-09, no replacement in this repo.

Replaces

Supersedes the per-mode wrapper skills codex-audit, codex-brainstorm, codex-coordinate, codex-curate, codex-discuss, codex-fix, codex-governance, codex-review, codex-triage, and - since the 2026-08-09 retirement - the PromptBridge-lane skill codex-consult, which is now a redirect tombstone pointing here.