Codex 里用 GPT-6.1 Sol 替代 Astra 的分工方案

damkin @damkina7原推发布于 2026.10.01

在 X 上打开原推

OpenAI just dropped GPT-6.1 Sol in Codex, and for repository coding it replaces Astra outright at 1/5 the token cost With OpenAI cutting Pro rate limits down from 20x to 10x, using Astra for routine implementation burns your entire weekly quota in 48 hours. Switch your default runtime to 6.1 Sol before your limits evaporate. The architecture secret is in OpenAI's DeepSWE benchmark curve: GPT-6.1 Sol peaks on high reasoning effort. Pushing it to xhigh or max actually drops the score by 6.6 points: recursive reasoning overthinks simple diffs and pollutes the context window. Medium effort scores within 1.8 points of high while cutting token cost by 32% and halving latency. The winning move is an asymmetric tree: 6.1 Sol on high runs the root orchestrator. 6.1 Sol on medium swarms the routine subagent work. Astra on xhigh stays locked in reserve, acting strictly as an independent reviewer before code ships to main. Plan on high. Swarm on medium. Verify in worktrees. Keep Astra for the final gate. - the full codex tree > root orchestrator: gpt-6.1-sol (effort: high) -> scopes tasks, coordinates subagents, merges diffs > explorer: gpt-6.1-sol (effort: medium) -> reads AST maps, finds references, never edits > worker: gpt-6.1-sol (effort: medium) -> writes minimal diffs and runs unit tests > researcher: gpt-6.1-sol (effort: medium) -> fetches docs and API contracts > verifier: gpt-6.1-sol (effort: low) -> validates tests in an ephemeral git worktree > reviewer: gpt-6-astra (effort: xhigh) -> independent read-only gate before final commit The quota killer to avoid: When spawning subagents in Codex, always set `fork_turns: "none"`. If omitted or set to "all", every subagent inherits the parent session's full KV-cache, forces the parent model parameters, and burns 100k tokens per call. paste the tree and this prompt into Codex CLI ↓ "Rebuild my Codex setup around this asymmetric tree: 1. Check ~/.codex/agents and .codex/agents for subagents that already fit explorer, worker, researcher, and verifier. > Draft new ones only for missing roles > For explorer, worker, and researcher: set model = 'gpt-6.1-sol', model_reasoning_effort = 'medium' > For verifier: set model = 'gpt-6.1-sol', model_reasoning_effort = 'low', sandbox_mode = 'git-worktree' > Create .codex/agents/reviewer.toml with model = 'gpt-6-astra', model_reasoning_effort = 'xhigh', sandbox_mode = 'read-only' 2. In ~/.codex/config.toml, set: [agents] default_subagent_model = 'gpt-6.1-sol' default_subagent_reasoning_effort = 'medium' fork_turns = 'none' max_concurrent_threads_per_session = 8 3. Set main session model to 'gpt-6.1-sol' with model_reasoning_effort = 'high'. Check for environment variables that disable multi-agent tools (agents.enabled) or override reasoning effort. Report them, change nothing. 4. Add rules to AGENTS.md: > Always dispatch subagents with fork_turns: 'none' to eliminate KV-cache bleed > Run full test suites strictly through the verifier agent in isolated worktrees > Consult the reviewer agent (gpt-6-astra, effort: xhigh) before final merge to main Show me every change as a diff first. No edits until I say go." ↳ https://developers.openai.com/codex/multi-agent

这条是引用着另一条写的：x.com/damkina7/status/2102860956724335039