jed-zed/ccg-codex-workflow/ccg-gptpro-exc

CCG GPT Pro Execution Companion

Load and follow skills/ccg-gptpro-bridge/SKILL.md.

This is a Codex-led execution-companion workflow: Codex controls implementation, Gemini only participates for frontend/full-stack frontend prototype or review evidence, GPT Pro provides one manual second opinion, and Codex makes the final implementation, verification, and delivery decision.

Behavior

• Treat input as an implementation companion request.
• Codex owns the implementation context, file edits, verification, final landing, and final decision.
• For backend-only tasks, do not run Gemini by default; Codex may create the GPT Pro bridge without a Gemini response file.
• For frontend or full-stack tasks, Gemini is read-only frontend evidence only: run Gemini through the bundled Gemini preview helper with --prompt-template frontend before GPT Pro, and summarize its UI/UX prototype findings.
• Include Codex's implementation context, target files, constraints, existing patterns, and any available Gemini Frontend Prototype Evidence in the GPT Pro prompt.
• If Gemini frontend evidence is provided, it must come from a real, non-empty response file with a concise summary; do not invent Gemini findings.
• Gemini is not a gate for /ccg:gptpro-exc, is not a general execution participant, and must not apply workspace changes.
• GPT Pro is a manual second opinion only; it does not write workspace files.
• Expected manual questions: 1.
• Maximum manual questions: 2.
• Round 2 should be converted into /ccg:gptpro-review whenever possible; use Gemini --prompt-template review and --gemini-evidence-role frontend-review for frontend review evidence over the applied diff.
• Use scripts/gptpro_bridge.py --mode exc --detach-preview --open-preview --gemini-policy optional --gemini-evidence-role frontend-prototype.
• When frontend/full-stack Gemini output is available, add --gemini-response-file <CCG_GEMINI_RESPONSE_FILE> --gemini-summary-file <summary-file>.
• GPT Pro output is a sketch, pseudo patch, test idea list, or edge-case review.
• GPT Pro output must include an Implementation Readiness Scorecard for plan fit, implementation completeness, verification readiness, risk, and adoption recommendation.
• The Implementation Readiness Scorecard is a manual second-opinion signal only; it does not make GPT Pro an executor and does not decide the final implementation.
• Report in Chinese and synthesize Codex, Gemini frontend evidence, and GPT Pro manual second opinion when Gemini evidence exists; otherwise synthesize Codex and GPT Pro findings and state that Gemini frontend evidence was not used.
• Codex remains final owner.
• Do not automate ChatGPT web login.
• Do not read ChatGPT web DOM.
• Do not extract ChatGPT Output programmatically.

Manual Handoff Barrier

• After creating the bridge artifacts, show only handoff metadata.
• Do not paste the full generated prompt into chat.
• Show the preview URL, session directory, prompt file path, response file path, and status file path.
• Tell the user to open the preview page and use the preview page Copy Prompt button, then manually submit the prompt to ChatGPT Pro and manually save the response.
• End the current assistant turn after the handoff. Do not continue the execution-companion analysis in the same turn.
• Continue only after status.json shows response_saved=true and response.md is non-empty.