Codex 多模型分工蓝图：Jev 处理确定性分支
把 Sol 做主循环、Astra 做架构审查，Jev 负责文件查找与工具派发等无需推理的分支，附配置提示词。
Codex golden rule: the moment GPT-6.1 Sol becomes your primary driver, stop burning Astra on routine turns keep Astra on standby as an advisory architect let Sol handle the active coding loop Astra only wakes up at three critical decision gates: → prior to planning: does this architecture hold water? → on repeating failures: are we troubleshooting the wrong culprit? → right before ship: where are the blind spots? Sol thinks. Luna builds. Astra audits. Jev handles the plumbing beneath the surface: deterministic branches that require zero reasoning (target file lookup, tool dispatch, loop termination) settle in <16ms via code, while frontier models only step in when paths diverge - the architecture map > primary loop: GPT-6.1 Sol (reasoning: high) > explorer: codebase walkthroughs via Luna > researcher: API specs & documentation via Luna > worker: test execution & code edits via Sol > subagent tier pinned to medium (same output, 1/3 the bill) > architect: Astra on standby for validation > auto_review: gatekeeper for staged changes drop this blueprint into your terminal to configure Codex ↓ "Reconfigure my local Codex environment to match this layout: 1. Audit ~/.codex/agents and .codex/agents for definitions matching explorer, researcher, and worker. > Create TOML manifests strictly for any missing agent roles > Route explorer and researcher to gpt-6-luna, worker to gpt-6.1-sol, setting model_reasoning_effort = "medium" > Provision an architect subagent backed by gpt-6-astra with model_reasoning_effort = "high", constrained entirely to evaluating blueprints, recurring errors, and final diffs > Flag and bypass configurations locking alternate models without touching them 2. Update ~/.codex/config.toml: set model = "gpt-6.1-sol", model_reasoning_effort = "high", and approvals_reviewer = "auto_review" 3. Detect conflicting overrides across active workspace profiles, shell alias exports, and agents.default_subagent_model. Print a summary; apply zero edits 4. Insert a policy into AGENTS.md: invoke the architect subagent before architecture signoff, whenever a test error recurs, and prior to marking complex tasks complete Present all proposed modifications as a unified git diff first. Wait for my explicit confirmation before applying changes." ↳ http://developers.openai.com/codex/subagents
这条是引用着另一条写的：x.com/mirku21/status/2098836008468992471