Codex 多智能体配置：Sol 主跑、Luna 群跑、Astra 当对手

mirku @mirku21原推发布于 2026.10.12

在 X 上打开原推

holy f*ck. Official OpenAI tip for Codex: with GPT Astra 6.1 dropping soon, stop leaving GPT-6 Astra sitting idle and burning GPT-6.1 Sol context on tasks GPT-6 Luna can swarm put the adversary on call with custom agents codex --model gpt-6.1-sol --agent adversary --subagents gpt-6-luna GPT-6.1 Sol drives the primary session on high effort, writing diffs and synthesizing PR merges ($2 / $10 per 1M tokens) GPT-6 Luna subagents swarm the repo in parallel at 240+ tok/s, handling file discovery and test fixtures GPT-6 Astra stays on call in ~/.codex/agents/ as the adversary, only stepping in at three critical moments: → before a plan locks: does the implementation plan miss auth invariants or schema contracts? → when a test breaks twice: are we fixing the root cause or falling into a recursive rabbit hole? → before calling done: did the full diff introduce hidden regressions or break pre-flight rules? GPT-6 Astra attacks. GPT-6.1 Sol builds. GPT-6 Luna swarms Codex subagent architecture runs the identical pattern through isolated contexts: high-volume operations (reading 30 files, grepping AST, running test fixtures) execute in separate git worktrees, so 400k+ tokens stay in Luna workers while the primary session keeps a pristine 14k context Plan on high. Delegate on medium. Keep Astra on call. - the full subagent swarm setup > GPT-6.1 Sol on high runs the main session ($2 / $10 per 1M tokens) > explorer reads the code in AST graph on GPT-6 Luna > worker edits and runs test suites in isolated worktrees on GPT-6 Luna > researcher pulls OpenAI specs and docs on GPT-6 Luna > all three on GPT-6 Luna at medium effort (240+ tok/s) > GPT-6 Astra on call as the adversary (upgradeable to GPT Astra 6.1) paste the setup and this prompt into Codex CLI below: "Configure my Codex workspace for hierarchical model selection and subagents: 1. Audit ~/.codex/agents and .codex/agents for subagent profiles fitting lead, workers, and adversary: > Give explorer, worker, and researcher model = 'gpt-6-luna', model_reasoning_effort = 'medium' > Assign worker an isolated git worktree branch to eliminate concurrent file conflicts > Create .codex/agents/adversary.toml with model = 'gpt-6-astra' (upgradeable to 'gpt-6.1-astra'), model_reasoning_effort = 'high', sandbox_mode = 'read-only' 2. Configure model selection in ~/.codex/config.toml: [session] model = 'gpt-6.1-sol' model_reasoning_effort = 'high' [agents] default_subagent_model = 'gpt-6-luna' default_subagent_reasoning_effort = 'medium' max_concurrent_threads_per_session = 8 3. Verify environment variables and routing: check for flags overriding reasoning effort or disabling multi-agent tools. Report them, change nothing 4. Add automatic delegation rules to AGENTS.md: > Summon @ adversary before finalizing multi-file architecture plans > Automatically trigger @ adversary when a test or compiler error repeats twice > Enforce @ adversary contract audit before declaring tasks complete or staging git commits Show every configuration diff first. Do not apply edits until confirmed" ↳ http://developers.openai.com/api/docs/guides/model-selection

这条是引用着另一条写的：x.com/mirku21/status/2098836008468992471