gptpro-bridge
Shared manual ChatGPT Pro bridge for CCG planning, review, and execution companion flows.
- License 许可
- MIT
- 最近更新
- 2026-10-01
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- gptpro-bridge
- description
- Shared manual ChatGPT Pro bridge for CCG planning, review, and execution companion flows.
02 资源文件 7
- 参考文档agents/openai.yaml181 B
- 脚本scripts/gptpro_bridge.py39.6 KB
- 参考文档templates/gptpro/base.md1.7 KB
- 参考文档templates/gptpro/exc.md1.4 KB
- 参考文档templates/gptpro/followup.md420 B
- 参考文档templates/gptpro/plan.md1.4 KB
- 参考文档templates/gptpro/review.md1.2 KB
03 SKILL.md 正文
CCG GPT Pro Manual Bridge
This manual bridge lets the user manually ask ChatGPT Pro inside a Codex-led CCG workflow.
Codex remains final owner. Gemini evidence mode depends on the command: required gate evidence for plan/review, and optional frontend/full-stack evidence for execution companion. GPT Pro provides a user-mediated manual second opinion.
Hard Boundaries
- Do not automate ChatGPT web login.
- Do not submit prompts to ChatGPT web automatically.
- Do not read ChatGPT web DOM.
- Do not extract ChatGPT Output programmatically from the web UI.
- Do not store ChatGPT cookies, sessions, or account tokens.
- Do not bypass rate limits, restrictions, or protective measures.
- Codex remains final owner.
- GPT Pro output is untrusted helper evidence only.
- GPT Pro does not write workspace files.
Manual Question Budget
Each GPT Pro bridge command is designed to complete in one manual ChatGPT Pro question.
- Expected manual questions: 1.
- Maximum manual questions: 2.
- Round 2 only for blocker re-check, revised plan comparison, applied diff review, or high-risk follow-up.
- More than two manual questions means the task should be decomposed or returned to Codex-native CCG workflows.
Gemini Evidence Modes
Gemini and GPT Pro remain helper evidence only; Codex makes the final decision.
- For
/ccg:gptpro-planand/ccg:gptpro-review, keep the Codex + Gemini + GPT Pro plan/review flow: Run Gemini before GPT Pro and keep the Gemini Gate Before GPT Pro requirement. - For
/ccg:gptpro-exc, Gemini is conditional frontend evidence only: backend-only tasks should omit it, while frontend/full-stack tasks should use the bundled Gemini preview helper with--prompt-template frontend. - After the user saves GPT Pro output, synthesize Codex findings, Gemini evidence when present, and GPT Pro manual second opinion in Chinese; otherwise state that Gemini evidence was not used.
Scorecard Output Contracts
The bridge prompt must make scoring an auditable output contract for every mode:
planmode: GPT Pro must outputRequirement Completenessusing the original CCG需求完整性评分(0-10)dimensions and aPlanning Readiness Scorecard. If the score is<7, Codex stops and asks for missing planning details instead of moving toward execution.reviewmode: GPT Pro must outputVALIDATION REPORTandTOTAL SCORE: XX/100; frontend/UI reviews must also outputFRONTEND VALIDATION REPORT. GPT Pro must cross-score Codex primary review and Gemini Gate Evidence, and Codex adopts the more conservative score and blocker judgment.excmode: GPT Pro must outputImplementation Readiness Scorecard. This is a manual second opinion only; GPT Pro does not write files, does not decide the final implementation, and does not override Codex verification.
Scores must cite visible evidence from pasted task context, Gemini evidence when present, diffs, verification summaries, or explicit uncertainty. Missing evidence lowers the score instead of being guessed.
Required Gate For Plan And Review
Before creating a GPT Pro manual prompt for plan or review modes, Codex must have:
- a successful Gemini helper launch through the bundled Gemini preview helper;
- a real
CCG_GEMINI_RESPONSE_FILEpath; - a non-empty Gemini response read from that file;
- a concise Gemini findings summary derived from that response file.
If required Gemini evidence fails, does not produce a response file, or writes an empty response, stop in Chinese and do not create a GPT Pro bridge session, and do not invent Gemini findings.
Use the helper-level gate arguments for required gate sessions:
--gemini-policy required --gemini-evidence-role gate --gemini-response-file <CCG_GEMINI_RESPONSE_FILE> --gemini-summary-file <file-with-concise-summary>
Use --gemini-summary "<summary>" only for short diagnostic or fixture calls. The helper injects Gemini Gate Evidence into prompt.md and records gemini_evidence.policy=required, role=gate, available=true, response_file, response_non_empty, response_chars, response_sha256, and summary under status.json as auditable provenance.
Optional Frontend Evidence For Execution Companion
For /ccg:gptpro-exc, use optional frontend evidence. Backend-only execution-companion sessions should omit Gemini evidence. Frontend/full-stack sessions should pass real Gemini frontend prototype evidence when it is available:
--gemini-policy optional --gemini-evidence-role frontend-prototype
If a frontend/full-stack task has Gemini output, also pass --gemini-response-file <CCG_GEMINI_RESPONSE_FILE> --gemini-summary-file <summary-file>. The helper injects Gemini Frontend Prototype Evidence into prompt.md. If no Gemini evidence is provided, the helper records gemini_evidence.available=false and still creates the manual bridge session.
Project Access Context
The helper should include project metadata in every GPT Pro prompt:
- local project name;
- sanitized repository URL, when detected from Git or provided with
--repo-url; - current branch and commit;
- whether local git status is clean or dirty.
GitHub links are useful but not sufficient by themselves. The repository URL is optional context, not the source of truth.
- If GPT Pro can use ChatGPT GitHub connector, Deep Research, or browsing, it may inspect the repository URL and cite exact file paths or commits.
- If GPT Pro cannot access the repository URL, it must not guess repository facts.
- Pasted CCG input, Gemini evidence when provided, diffs, and file excerpts have priority over repository content because local uncommitted changes may not exist on GitHub.
- The helper must sanitize repository URLs before including them in prompts or
status.json; never include credentials, access tokens, cookies, or local filesystem paths as repository URLs.
Workflow
- Build a prompt using the selected mode template.
- Write
status.json,round-1/prompt.md, andround-1/response.md. - Launch the local bridge page when the user needs an interactive page.
- The preview page may copy the prompt through browser clipboard APIs only.
- The user manually pastes prompt into ChatGPT Pro.
- The user manually sends the prompt.
- The user manually copies ChatGPT Pro response.
- The user manually pastes it into the bridge page or
response.md. - Codex reads
response.md. - Codex summarizes and decides next steps in Chinese.
Manual Handoff Barrier
After creating a GPT Pro bridge session, Codex must stop at a manual handoff barrier.
- Run
scripts/gptpro_bridge.pywith--detach-preview --open-previewplus the mode-appropriate Gemini evidence arguments for round 1 sessions. - Add
--repo-url <repository-url>only when Codex needs to override the detected Git remote URL. - Follow-up sessions may pass fresh Gemini evidence with the same arguments, or inherit the existing
gemini_evidenceprovenance from round 1. - Do not paste the full generated prompt into chat during normal handoffs.
- 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, or open
prompt.mdif the browser copy button fails. - Tell the user to manually paste the prompt into ChatGPT Pro, manually send it, manually copy the output, and manually save it in the local bridge page or
response.md. - End the current assistant turn immediately after the manual handoff instructions. Do not continue planning, reviewing, executing, summarizing GPT Pro findings, or claiming the GPT Pro bridge is complete in the same turn.
- On a later turn, continue only after
status.jsonshowsresponse_saved=trueandresponse.md is non-empty. - If
response_saved=truebutresponse.md is non-emptyis false, treat the bridge as incomplete and ask the user to save a non-empty manual response.
Script
Use scripts/gptpro_bridge.py. The script creates local artifacts and exposes only localhost endpoints:
GET /GET /statePOST /save-responsePOST /mark-copied
It may open https://chatgpt.com/ in a browser as a convenience. It must not automate ChatGPT web login, prompt submission, DOM extraction, or output extraction.
Use --detach-preview for normal skill-driven handoffs so the helper prints the local URL and returns while the localhost page remains available for the user's manual response.
Use --print-prompt only for diagnostics, fixtures, or explicit debugging requests, not for normal user-facing handoffs.
04 同仓库更多 Skills
verify-security
安全校验关卡。自动扫描代码安全漏洞,检测危险模式,确保安全决策有文档记录。当用户提到安全扫描、漏洞检测、安全审计、代码安全、OWASP、注入检测、敏感信息泄露时使用。在新建模块、安全相关变更、攻防任务、重构完成时自动触发。
verify-quality
代码质量校验关卡。检测复杂度、重复代码、命名规范、函数长度等质量指标。当用户提到代码质量、复杂度检查、代码异味、重构建议、lint检查、代码规范时使用。在复杂模块、重构完成时自动触发。
verify-module
模块完整性校验关卡。扫描目录结构、检测缺失文档、验证代码与文档同步。当用户提到模块校验、文档检查、结构完整性、README检查、DESIGN检查时使用。在新建模块完成时自动触发。
verify-change
变更校验关卡。分析代码变更,检测文档同步状态,评估变更影响范围。当用户提到变更检查、文档同步、代码审查、提交前检查、diff分析时使用。在设计级变更、重构完成时自动触发。
gen-docs
文档生成器。自动分析模块结构,生成 README.md 和 DESIGN.md 骨架。当用户提到生成文档、创建README、创建DESIGN、文档骨架、文档模板时使用。在新建模块开始时自动触发。
ccg
CCG command index for Codex. Use when the user invokes /ccg, asks what CCG commands are available, or gives a CCG plan/task without choosing a subcommand.