ask → work → result → wait

work → check → continue or complete

CREATE TABLE thread_goals (
    thread_id TEXT PRIMARY KEY NOT NULL REFERENCES threads(id) ON DELETE CASCADE,
    goal_id TEXT NOT NULL,
    objective TEXT NOT NULL,
    status TEXT NOT NULL CHECK(status IN ('active','paused','budget_limited','complete')),
    token_budget INTEGER,
    tokens_used INTEGER NOT NULL DEFAULT 0,
    time_used_seconds INTEGER NOT NULL DEFAULT 0,
    created_at_ms INTEGER NOT NULL,
    updated_at_ms INTEGER NOT NULL
);

{
  "remainingTokens": 6750,
  "completionBudgetReport": "Goal achieved. Report final budget usage to the user: tokens used: 3250 of 10000; time used: 75 seconds."
}

Continue working toward your goal: {objective}

Current status: {status}
Tokens used: {tokensUsed} of {tokenBudget}
Time used: {timeUsedSeconds}s

Use create_goal, get_goal, or update_goal(complete) to manage this goal.

[system message — 由 Responses API 服务端控制]
  ↓
[tools — 由客户端提供的工具定义]
  ↓
[instructions — developer/system 指令]
  ↓
[input — 列表形式]

input 的结构（依次）：
  1. [developer] sandbox 描述（内嵌模板：workspace_write.md, on_request.md）
  2. [developer] config.toml 中的 developer_instructions（可选）
  3. [user] user instructions（AGENTS.md + skills 定义等）
     → 限制 32 KiB，逐文件夹向上搜索
  4. 实际用户消息

You're approaching the token budget for your goal: {objective}
Current usage: {tokensUsed}/{tokenBudget}
Consider completing soon or the goal will be marked budget_limited.

用户输入 /goal ...
    ↓
TUI 解析 slash command
    ↓
JSON-RPC: thread/goal/set(→ 写入 SQLite)
    ↓
通知 thread/goal/updated
    ↓
Runtime: TurnStarted → 捕获 goal 状态 + 基线
    ↓
Responses API 请求（input 中包含延续 prompt + goal objective）
    ↓
模型生成：推断 → 工具调用 → 工具完成 → 记账（自动）
    ↓
TurnFinished → 最终记账
    ↓
MaybeContinueIfIdle（检查：非终态 + 未抑制 + 非 plan 模式）
    ↓
注入 continuity prompt → 新一轮 Responses API 请求
    ↓
[循环直至 complete/budget_limited/抑制]

sequenceDiagram
    actor 用户
    participant TUI as TUI终端
    participant SQL as SQLite
    participant Runtime as 运行时
    participant Model as 模型

    Note over 用户,Model: 阶段A：澄清
    用户->>TUI: /goal 迁移支付到Paddle
    TUI->>Runtime: 普通prompt推理
    Runtime->>Model: POST /v1/responses
    Model-->>用户: 追问六要素（Outcome/Verification/...）
    用户->>Model: 确认选A，API兼容，只改后端

    Note over 用户,Model: 阶段B：创建Goal
    Model->>Runtime: create_goal({objective: "...", token_budget: 100000})
    Runtime->>SQL: INSERT thread_goals (active)
    SQL-->>Runtime: goal_id: a1b2c3d4
    Runtime-->>TUI: thread/goal/updated 通知
    TUI-->>用户: 状态栏显示 🎯 active

    Note over 用户,Model: 阶段C：执行与自动延续
    Model->>Runtime: 第一轮推理（工具调用：读代码、写适配器）
    Runtime->>SQL: ToolCompleted → UPDATE tokens_used
    Runtime->>Runtime: TurnFinished → 检查自动延续条件

    loop 自动延续（检查四个条件）
        Runtime->>SQL: SELECT objective, status, tokens_used
        SQL-->>Runtime: active, 4520/100000
        Runtime->>Runtime: 填充 continuation.md 模板
        Runtime->>Model: POST /v1/responses（含goal prompt）
        Model->>Runtime: 工具调用 → 跑测试
        Runtime->>SQL: ToolCompleted → 记账
        Runtime->>Runtime: TurnFinished → 再检查延续条件
    end

    Note over 用户,Model: 阶段D：完成
    Model->>Runtime: update_goal({status: "complete"})
    Runtime->>SQL: UPDATE status='complete'
    Runtime-->>TUI: goal/cleared 通知
    TUI-->>用户: 状态栏消失，显示完成报告

graph TD
    TF[TurnFinished] --> Q1{goal.status?}

    Q1 -->|complete| STOP_OK[✅ 停]
    Q1 -->|paused| WAIT[⏸ 等用户 resume]
    WAIT --> RESUME[ThreadResumed → active]
    RESUME --> Q2
    Q1 -->|budget_limited| STOP_BUDGET[⛔ 停 预算强制]
    Q1 -->|active| Q2{continuation_suppressed?}

    Q2 -->|true 零工具轮| SKIP1[跳过]
    Q2 -->|false| Q3{Plan 模式?}

    Q3 -->|yes| SKIP2[跳过]
    Q3 -->|no| Q4{有 token_budget?}

    Q4 -->|超了| SET_BL[设 budget_limited → 停]
    Q4 -->|null 或没超| Q5{线程空闲?}

    Q5 -->|不空闲| SKIP3[跳过]
    Q5 -->|空闲| CONTINUE[MaybeContinueIfIdle]

    CONTINUE --> SQL_READ[读 SQLite SELECT goal 状态]
    SQL_READ --> FILL[填充 continuation.md 模板]
    FILL --> APPEND[追加到 input 数组]
    APPEND --> POST[POST /v1/responses]
    POST --> LOOP[新一轮推理循环]
    LOOP --> TF

sequenceDiagram
    participant SQL as SQLite
    participant Runtime as 运行时
    participant Model as 模型

    Note over SQL,Model: 阶段A：用户输入 /goal 之前
    Runtime->>Model: POST input=[sandbox][AGENTS.md][用户消息]
    Note over Model: ❌ 无 goal 内容

    Note over SQL,Model: 阶段B：create_goal 之后的第一次推理
    Model->>Runtime: create_goal(...)
    Runtime->>SQL: INSERT thread_goals
    Runtime->>Model: POST input=[sandbox][AGENTS.md][对话历史(含create_goal工具调用)]
    Note over Model: ⚡ 模型知道goal存在，因为自己调的create_goal
    Note over Model: ❌ 但 continuation.md 尚未注入
    Model->>Runtime: 第一轮工具调用
    Runtime->>SQL: ToolCompleted → 记账
    Runtime->>Runtime: TurnFinished → 检查延续条件 ✓

    Note over SQL,Model: 阶段C：自动延续后（关键变化）
    Runtime->>SQL: SELECT objective, status, tokens_used
    SQL-->>Runtime: "迁移支付...", active, 4520/100000
    Runtime->>Runtime: 填充 continuation.md 模板
    Runtime->>Model: POST input=[...对话历史][🎯continuation.md][续轮消息]
    Note over Model: ✅ goal 文本第一次以结构化方式出现！

TUI 状态栏（thread 底部）:
   ╔═══════════════════════════════════════╗
   ║ 🎯 迁移支付模块到 Paddle  █ active    ║
   ║    37,250 / 100,000 tokens · 847s     ║
   ╚═══════════════════════════════════════╝

Slash 命令:
   /goal              → 查看当前 goal 状态
   /goal pause        → 暂停（系统设 paused）
   /goal resume       → 恢复（系统设 active）
   /goal clear        → 删除 goal 记录

进度输出（模型每轮报告的摘要）:
   第1轮  ✅ 读取 Stripe 配置
   第2轮  ✅ 创建 Paddle 适配器骨架
   第3轮  ❌ 测试失败: 签名不兼容
   第4轮  ✅ 修复签名逻辑
   第5轮  ✅ 32/32 测试全部通过