用 GPT-6.1 Sol 搭 Codex 多代理树，Astra 只做终审

Vox @Voxyz_ai原推发布于 2026.09.30

在 X 上打开原推

OpenAI just launched GPT-6.1 Sol, and for coding it can 𝗿𝗲𝗽𝗹𝗮𝗰𝗲 𝗔𝘀𝘁𝗿𝗮 𝗼𝘂𝘁𝗿𝗶𝗴𝗵𝘁 at 1/5 the price. OpenAI also announced today that 20x Pro is being cut to 10x. I honestly don't dare use Astra anymore. Switch to 6.1 Sol before Astra eats your whole week. 6.1 Sol is already in Codex. Set it up like the tree: 6.1 Sol runs the root orchestrator and all three subagents, and Astra only does an independent review before a big change ships. Reasoning effort comes from OpenAI's DeepSWE chart: 6.1 Sol scores highest on high, and 𝗴𝗼𝗶𝗻𝗴 𝘂𝗽 𝘁𝗼 𝘅𝗵𝗶𝗴𝗵 𝗼𝗿 𝗺𝗮𝘅 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘀𝗰𝗼𝗿𝗲𝘀 𝗹𝗼𝘄𝗲𝗿, so the orchestrator runs on high. Medium scores only about 2 points lower than high and costs over 30% less per task, so the subagents run on medium. Send the tree and this prompt to Codex 👇 "Set up my Codex to match this tree: → Main session: default to `gpt-6.1-sol` on high reasoning effort. It breaks down tasks, delegates, integrates, and verifies. Handle simple tasks directly; don't delegate them. → Three project-level custom agents, each with model `gpt-6.1-sol` and model_reasoning_effort `medium`: explorer reads code, worker edits code and runs tests, researcher looks up docs. Reuse fitting ones if they exist, and create only the missing ones. → reviewer: also a project-level custom agent, with model `gpt-6-astra` and model_reasoning_effort `xhigh`. Call it only before a big change ships. It reports problems and doesn't edit code. → Put the delegation rule in AGENTS.md: when spawning a subagent, set `fork_turns` to `"none"` or pass only the last few turns. If `fork_turns` is omitted or set to `"all"`, the subagent inherits the main session's model and effort and can't override them, and the tree stops working. Before changing anything, check: → Roles other than the reviewer that still use `gpt-6-astra` or an older Sol (like `gpt-6-sol` or `gpt-5.6-sol`). List them. → Roles moving to 6.1 Sol with effort set to `none` or `minimal`: change them to `low` (6.1 Sol supports neither), then compare the results. → Model assignments written only in AGENTS.md: check recent subagent session logs to confirm which model each role actually ran. If you can't select a model in the tree, stop and ask me. Show me the changes first, and wait for my OK before writing anything."

这条是引用着另一条写的：x.com/openaidevs/status/2104993035507712318