Athena Dev

/athena-dev — Athena 主入口 (Codex, v9.8.0)

触发

用户进任意项目, 说 "开始", "做个 X", "帮我 Y" 等. 主 agent 进入路由分诊.

路由分诊 (铁律[分诊])

def route(user_input, ai_state):
    # 0. 检查是否在 Athena 项目中
    if not has_ai_state_dir():
        suggest("先跑 /athena-init 初始化项目")
        return None

    # 1. 显式信号优先 (用户直接说)
    if explicit_kws(["想法不清楚", "先 brainstorm", "讨论", "聊聊"]):
        return start_stage("brainstorm")
    if explicit_kws(["路线图", "拆分", "分步推进"]):
        return start_stage("roadmap")
    if explicit_kws(["生产事故", "线上故障", "hotfix"]):
        return start_stage("plan", path="Hotfix")
    if explicit_kws(["bug", "缺陷", "修复"]):
        return start_stage("plan", path="Bugfix")
    if explicit_kws(["重构", "refactor"]):
        return start_stage("plan", path="Refactor")
    if explicit_kws(["系统级", "跨模块", "架构"]):
        return start_stage("plan", path="System")

    # 2. 隐式判断 (铁律[分诊])
    if len(user_input.split()) < 8 and not has_concrete_verb(user_input):
        # 单词级模糊 → brainstorm
        return start_stage("brainstorm")

    if mentions_modules(user_input) >= 3:
        # ≥3 模块 → roadmap
        return start_stage("roadmap")

    # 3. 默认: 按改动量分类
    estimated = estimate_scope(user_input)
    if estimated == "tiny":
        return start_stage("plan", path="Quick")  # ≤3 文件
    if estimated == "single_module":
        return start_stage("plan", path="Feature")
    if estimated == "multi_module":
        return start_stage("plan", path="System")

    return start_stage("plan", path="Feature")  # fallback

路由决定后的动作

→ brainstorm

# 创建 sprint 目录
slug="$(date +%Y-%m-%d)-$(slugify '$user_topic')"
mkdir -p ".ai_state/sprints/${slug}"
cp ~/.codex/skills/pace/templates/sprints/brainstorm.md ".ai_state/sprints/${slug}/"

# 更新 _index.md
update_field stage "brainstorm"
update_field current_sprint_slug "${slug}"

# 进 brainstorm skill
read ~/.codex/skills/brainstorm/SKILL.md
# 多轮对话

→ roadmap

slug="$(slugify '$user_topic')"
mkdir -p ".ai_state/roadmap/${slug}/drafts"
cp ~/.codex/skills/pace/templates/roadmap/{roadmap.md,items.yaml} ".ai_state/roadmap/${slug}/"

update_field stage "roadmap"
update_field current_roadmap_slug "${slug}"

read ~/.codex/skills/roadmap/SKILL.md

→ plan / design (需求清晰)

slug="$(date +%Y-%m-%d)-$(slugify '$task_name')"
mkdir -p ".ai_state/sprints/${slug}/reviews"
cp ~/.codex/skills/pace/templates/sprints/{design.md,checklist.yaml} ".ai_state/sprints/${slug}/"

update_field path "${path_type}"
update_field stage "plan"
update_field current_sprint_slug "${slug}"

# Codex plan_mode_reasoning_effort = "xhigh" 已生效 (config.toml)
# 进 pace skill

next_action 处理

主 agent 进 athena-dev 时, 先读 _index.next_action:

推理力度提示

plan/design stage 由 config.toml plan_mode_reasoning_effort = "xhigh" 保证最大推理预算. SessionStart hook (session-start.py) 在 stage_hints 中自动提示.

与其他 skill 联动

例外

• 用户直接说 "直接做:" 或 --skip-brainstorm → 跳过 brainstorm
• 用户显式说 "我自己拆 roadmap" → 跳过 roadmap, 直接 plan
• Hotfix 路径: 跳过所有分诊, 直接进 plan (生产事故无时间分诊)

主 agent 行为约束

• ✅ 必须先读 _index.md 确定当前状态
• ✅ 路由判断必须基于实际输入 + ai_state, 不能"我觉得"
• ✅ 模糊时优先 brainstorm, 不要直接猜想用户意图
• ❌ 不允许跳过分诊直接进 plan (铁律[分诊])
• ❌ 写入不按红黄绿区路由 (铁律[零写入]: 绿区主 thread 直做, 黄/红区 spawn_agent)