Model routing: route Designer Drift Audit, Daily Documentation Diagram and Chaos PR Bundle Fuzzer on OpenAI instead of pinning GPT-6 Astra, GPT-5.6 Sol and GPT-5.6 Terra

已关闭 适合新手
  • 哪里不对？
#74,866 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

@SivaKesava1 已经在做这个了。

开始于 2026年11月6日。

  • #75461 来自 @copilot-swe-agent —— 已合并

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
70/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
github-actions, go, yaml
领域
ci-cd, devtools

调研方向

阅读 .github/workflows/designer-drift-audit.md、daily-documentation-diagram.md 和 chaos-pr-bundle-fuzzer.md；删除每个文件顶层的 model: 固定配置，并添加指定的 model-routing 块，同时保留 model-provider: openai。运行 gh aw compile designer-drift-audit daily-documentation-diagram chaos-pr-bundle-fuzzer，并确认编译后的工作流使用 OpenAI 路由且没有新增警告。三个工作流都按描述编译完成后，任务即告完成；该 issue 还要求维护者审查首次路由运行，并审阅图表 PR。

由索引模型根据 Issue 内容生成。

描述

ai-inspected community model-routing

Estimated saving: about 750-850 AIC/week. These three Codex workflows run on the OpenAI key (model-provider: openai) and pin a model. AWF also routes natively on OpenAI (smoke-codex-routed-openai), so route them with economy and keep the OpenAI provider and key.

Workflow Output Now Mode Agent runs/wk AIC/run now Est. saving/wk
designer-drift-audit issue (drift between docs and the designer skill) openai/gpt-6-astra economy 5 ~122 440-490
daily-documentation-diagram PR (one docs diagram) openai/gpt-5.6-sol economy 7 ~66 300-340
chaos-pr-bundle-fuzzer draft test PRs (expire in 4h) openai/gpt-5.6-terra economy ~1 ~32 20-35

AIC/run includes threat detection.

Why

  • OpenAI routing is ready. smoke-codex-routed-openai has passed every run since 25 October. It limits candidates to gpt-5.4, but in its runs:

    • AWF's OpenAI catalogue lists all ten OpenAI models the router ranks (GPT-5.4, 5.4-mini and 5.5; GPT-5.6 Luna, Sol and Terra; GPT-6 Astra, Luna and Sol; and GPT-6.1 Sol).
    • Classification wasn't degraded.
    • The detection job resolved detection to gpt-5-mini.

    These three are the first real workflows to route on OpenAI. None has sub-agents, and each runs at most once a day, so a failed routed run costs one report, not a burst of runs.

  • Routed runs cost much less. On Copilot, routing has cut cost per run for the same pins. GPT-6 Astra to routed: Auto-Triage Issues 79 to 20 AIC. GPT-5.6 Sol: Terminal Stylist 106 to 35. Terra: Daily Go Test Parallelizer 55 to 11-33. The router uses the same per-model costs on OpenAI.

  • economy for all three. Designer Drift Audit files an issue, Daily Documentation Diagram opens a PR with one diagram that a maintainer reviews, and Chaos PR Bundle Fuzzer opens draft test PRs confined to its chaos folders. For this work, economy picks GPT-5.6 Luna (xhigh or max). A hard cross-system label picks gpt-5.6-sol:max, which still costs less than GPT-6 Astra.

  • Detection gets cheaper. It inherits today's pin: about 29 AIC per run for Designer Drift Audit, 11 for Daily Documentation Diagram and 6 for Chaos PR Bundle Fuzzer. Routed, it uses gpt-5-mini, about 0.5.

  • Not tied to one model. Copilot's 16 November retirement of GPT-5.6 Terra doesn't cover openai/ pins, but routing removes the dependency on a single model name anyway. No model restrictions.

Change

In .github/workflows/designer-drift-audit.md, daily-documentation-diagram.md and chaos-pr-bundle-fuzzer.md, delete the top-level model: line and add model-routing to the existing engine block, keeping model-provider: openai:

engine:
  id: codex
  model-provider: openai
  model-routing:
    goal: cost
    mode: economy

No permission changes: inference stays on the OpenAI key. Keep the prompts, steps and safe outputs. Don't add model restrictions.

Run gh aw compile designer-drift-audit daily-documentation-diagram chaos-pr-bundle-fuzzer. On main (fdfc5d5a0d) these compile with no new warnings: routing is enabled with "provider":"openai" and gpt-* candidates, and detection uses the detection model.

Check

  • Audit the first routed run of each with gh aw audit <run>. The selection in usage/aw_session.jsonl should be an OpenAI Luna model. If routing fails on OpenAI (the run fails at AWF startup, or the selection is rejected), restore the three pins and file the failure in gh-aw-firewall.
  • Compare AIC per run with the table. If Daily Documentation Diagram keeps getting gpt-5.6-sol:max and costs more than about 66 AIC per run, restore its pin.
  • Review Daily Documentation Diagram's next PRs.

Estimates use gh-aw usage from 19 October to 3 November and AWF v0.29.6's routing tables; the routed cost per run is estimated from routed Copilot workflows on the same pins.

主要语言
Go
星标
5.6k
派生
615
平均合并
5 小时 44 分钟
30 天内合并 PR
808

环境准备

在 Codespaces 中打开

在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。

  • 提供 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

github/gh-aw 的其他 Issue

查看 github/gh-aw 的全部 Issue

相似的 Issue

更多 Go Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。