diff --git a/PROGRESS.md b/PROGRESS.md
index b7510fd..51c5718 100644
--- a/PROGRESS.md
+++ b/PROGRESS.md
@@ -2,7 +2,7 @@
 
 > 配合 `DESIGN.md` 阅读。本文件记录已完成的事、关键决策、与原设计的偏差。
 
-最后更新: 2026-05-06 (PPT skill v3:红色硬约束 + ⛔ blocking + canvas 合并 + apply_brand 品牌条 + 强制尾页 + Iconify 图标库)
+最后更新: 2026-05-06 (Phase 4 + Phase 6 task/state.json 落地;移除 legacy session 兼容;Phase 5 evals 决定不做)
 
 ---
 
@@ -13,9 +13,9 @@
 | 1 | 最小可用骨架 | ✅ 完成 | 全部验收点过 |
 | 2 | Skill 系统 + 三个 skill | ✅ 完成 | Anthropic 格式;coding/ppt/proposal |
 | 3 | Hybrid 范式 (run_python) | ✅ 完成 | subprocess + 敏感 env 过滤 |
-| 4 | 演化性能力 | 🟡 部分 | Model Profile 已就位;capability probing 未做;版本化 prompts 未做 |
-| 5 | Eval Suite | ❌ 未开始 | |
-| 6 | 长任务工程化 | 🟡 部分 | session 中断恢复已完成;context 压缩、双层记忆未做 |
+| 4 | 演化性能力 | 🟡 部分 | Model Profile 已就位;capability probing ✅;版本化 prompts 未做 |
+| 5 | Eval Suite | ⏸ 不做 | 个人工具,造作 case 没区分度;真换模型用 dogfooding 判断 |
+| 6 | 长任务工程化 | 🟡 部分 | task + state.json ✅;session 中断恢复 ✅;context 压缩、双层记忆未做 |
 | 7 | 打磨 | ❌ 未开始 | Docker 沙盒 / 更多 skill / Web UI |
 
 ---
@@ -86,6 +86,29 @@
 - `prompts/system/general_v1.md`(无版本化软链接,直接引用 v1)
 - 启动时拼接顺序: 通用指引 → discovery 块(skill 列表) → 当前工作目录
 
+### 8. Capability Probing(Phase 4)
+- `core/probe.py`:四项探测 + ProbeReport
+  - `probe_basic_chat` —— 连通性,失败则跳过其余
+  - `probe_parallel_tools` —— 给两个独立工具,看 single response 里 tool_calls 数 ≥2
+  - `probe_thinking_mode` —— 对 declared=True 的模型传 reasoning_effort,看 API 接受 + 是否产出 reasoning_content/thinking
+  - `probe_long_context` —— needle-in-haystack 简化版,默认 reliable_context * 4 / 8 字符,上限 200K(opt-in,需 `--long-context`)
+- `cli.py probe [--model X] [--long-context]`:rich Table 输出;退出码 0=全 ok / 2=有 mismatch / 3=有 error
+- 不修改 yaml,只报告差异——是否调档案由用户决定
+- 不进启动路径(每次启动跑会烧 API),用户显式触发
+
+### 9. Task 概念 + state.json(Phase 6)
+- `core/task.py`:`TaskState` dataclass(对齐 DESIGN.md §7.1)
+  - 字段:task_id / mode / description / status (active/completed/abandoned) / model / model_profile / reasoning_effort / cwd / created_at / updated_at / tokens_prompt / tokens_completion / cost_usd
+  - `save(task_dir)` / `load(task_dir)`,加载时容忍未知字段(schema 演化)
+- 存储布局:`workspace/tasks/<task_id>/{state.json, messages.json}`
+- `main.build_agent` 返回 5 元组 `(agent, session, sid, task_state, task_dir)`
+- `main.sync_task_tokens(state, dir, llm)`:每轮 agent.run 后调,把 LLM 累计 tokens 写回 state.json
+- CLI:
+  - `chat --mode coding --desc "修一处 bug"`(可选元数据)
+  - REPL:`/status` 看 state、`/done` 标完成、`/abandon` 标弃、`/desc <文本>` 改描述
+  - `cli.py tasks [--status active|completed|abandoned]`:列任务,显示 mode/model/msgs/tokens/desc
+- 已知小坑:`Session.save()` 不原子,write_text 抛错(如 stdin 注入了 UTF-8 surrogate 半字符)会留下 0 字节文件——交互使用没问题,后续可加 tmp+rename
+
 ---
 
 ## 关键决策与偏差
@@ -93,8 +116,9 @@
 | 项 | 决策 | 与设计差异 |
 |---|------|-----------|
 | 工具基目录 | 用户当前 cwd,不是 workspace/ | 设计未明说;选 cwd 是因为 agent 该操作用户的项目 |
-| Workspace 用途 | 只存 sessions/(暂时) | 设计含 `tasks/ memory/ logs/`,后续 Phase 6 再加 |
-| Session 粒度 | 一个文件一个 session,无 task 概念 | 设计有 task_id / state.json,Phase 6 再加 |
+| Workspace 用途 | `tasks/<id>/{state.json, messages.json}`;memory/ 待 Phase 6 双层记忆 | 设计含 `tasks/ memory/ logs/`,部分落地 |
+| Session 粒度 | 一个目录一个 task,含 state.json | Phase 6 落地;DESIGN.md §7.1 字段对齐 |
+| Eval Suite | 决定不做 | DESIGN.md §6.3 设计为团队/产品场景;个人工具用 dogfooding 替代,probe 覆盖健康检查 |
 | 版本化 prompt | 直接 general_v1.md,无 active.md 软链接 | Windows 软链接麻烦;后续要切版本时再做 |
 | run_python 沙盒 | subprocess + env 过滤 | 设计阶段 1 就是这套,未升级 Docker |
 | 工具数 | 8 个 (read/write/edit/glob/grep/shell/run_python/load_skill) | 设计上限 ≤10 同时可见,目前刚好 |
@@ -106,10 +130,14 @@
 - 全项目 `ast.parse` 语法 OK
 - yaml 配置可解析
 - 所有 import 链路在 venv 中跑通
-- `cli.py --help` / `cli.py chat --help` / `cli.py sessions --help` 正常
+- `cli.py --help` / `cli.py chat --help` / `cli.py tasks --help` / `cli.py probe --help` 正常
 - `SkillRegistry` 识别出 3 个 skill,discovery 块拼装正确
 - 缺 `DEEPSEEK_API_KEY` 时报清晰错误
-- 实测 DeepSeek API 接通(`deepseek-v4-flash` 模型 ID 被认),仅因账户余额不足而返回 InsufficientBalance —— **接入路径已通**
+- 实测 DeepSeek API 接通,`flash` 和 `pro` 两档都能调通
+- **真实 probe 结果**(2026-05-06):
+  - `deepseek_v4.flash`:basic_chat ok / **parallel_tools mismatch**(yaml=false, observed=true,2 个 tool_calls)/ thinking_mode skip(declared false)
+  - `deepseek_v4.pro`:basic_chat ok / parallel_tools ok / thinking_mode ok(reasoning_content 返回)
+  - flash 的 mismatch 暂不自动改 yaml —— `parallel_tools=true` 会影响所有实际调用,需更多场景观察后再决定
 
 ---
 
@@ -117,13 +145,11 @@
 
 按性价比排序:
 
-1. **Phase 4 capability probing**(~半天)—— 启动时跑 needle-in-haystack / 并行 tool 探测,把 yaml 声称的能力对账
-2. **Phase 5 Eval Suite**(~2 天)—— 模型升级决策的依据。每类任务 3-5 个 case,客观 + LLM judge 双评分
-3. **Phase 6 task 概念 + state.json**(~1 天)—— 让 session 升级为任务,workspace 加 `tasks/<task_id>/`
-4. **Phase 6 context 三层压缩**(~1 天)—— 兜底用,V4 长上下文一般用不到
-5. **Phase 6 双层记忆**(~半天)—— `workspace/memory/core.md` 注 prompt + `extended/` 按需读
-6. **Phase 7 Docker 沙盒**(~1 天)—— 替换 subprocess,run_python 安全升级
-7. **Phase 7 更多 skill / 模型档案**(持续)
+1. **Phase 6 双层记忆**(~半天)—— `workspace/memory/core.md` 注 prompt + `extended/` 按需读
+2. **Phase 6 context 三层压缩**(~1 天)—— 兜底用,V4 长上下文一般用不到
+3. **小修打磨**(~半小时)—— `Session.save()` 改原子写(tmp + rename),防 surrogate 等异常 truncate
+4. **Phase 7 Docker 沙盒**(~1 天)—— 替换 subprocess,run_python 安全升级
+5. **Phase 7 更多 skill / 模型档案**(持续)
 
 ---
 
@@ -133,17 +159,19 @@
 core/capabilities.py        71 行
 core/llm.py                 89 行
 core/loop.py                99 行
+core/probe.py              243 行   ← Phase 4
 core/session.py             77 行
 core/skills.py              81 行
+core/task.py                63 行   ← Phase 6
 tools/base.py               34 行
 tools/fs.py                182 行
 tools/shell.py              63 行
 tools/run_python.py         84 行
 tools/skill_tool.py         45 行
-main.py                    120 行
-cli.py                     138 行
+main.py                    175 行   ← +tasks 布局 / TaskState 装配
+cli.py                     265 行   ← +probe / +tasks 子命令
 ─────────────────────────────────
-合计 Python               1083 行
+合计 Python               ~1571 行
 
 prompts/system/general_v1.md
 skills/coding/SKILL.md
@@ -154,4 +182,4 @@ config/models/deepseek_v4.yaml
 requirements.txt
 ```
 
-设计预估 Phase 1-3 大约 800-1000 行,实际 1083 行,略多但仍在可读范围。
+设计预估 Phase 1-3 大约 800-1000 行,实际 1083 行,加上 Phase 4 + Phase 6 约 1571 行 Python。
diff --git a/cli.py b/cli.py
index de5b994..16e2817 100644
--- a/cli.py
+++ b/cli.py
@@ -1,23 +1,33 @@
 """CLI 入口: 简单 REPL。
 
 用法:
-    python cli.py chat                       # 新建一个 session
-    python cli.py chat --resume last         # 恢复最近一个
-    python cli.py chat --resume 20260506_141523
+    python cli.py chat                                    # 新建一个 task
+    python cli.py chat --mode coding --desc "修一处 bug"  # 带元数据建任务
+    python cli.py chat --resume last                      # 恢复最近一个 task
+    python cli.py chat --resume 20260506_141523           # 显式 task_id
     python cli.py chat --model deepseek_v4.pro
-    python cli.py sessions                   # 列出历史 session
+    python cli.py tasks                                   # 列出 task
+    python cli.py probe                                   # 实测对账 yaml 声称的能力
 """
 from __future__ import annotations
 
 import json
 import sys
-from pathlib import Path
 
 import click
 from rich.console import Console
 from rich.prompt import Prompt
+from rich.table import Table
 
-from main import build_agent, load_config, resolve_workspace, sessions_dir
+from core.task import TaskState
+from main import (
+    ROOT,
+    build_agent,
+    load_config,
+    resolve_workspace,
+    sync_task_tokens,
+    tasks_dir,
+)
 
 
 @click.group()
@@ -27,18 +37,22 @@ def cli() -> None:
 
 @cli.command()
 @click.option("--model", default=None, help="模型档案,如 deepseek_v4.flash 或 deepseek_v4.pro")
-@click.option("--workspace", default=None, help="工作目录(存 sessions/)")
-@click.option("--resume", default=None, help="恢复某个 session: 'last' 或 session_id")
-def chat(model: str, workspace: str, resume: str) -> None:
-    """启动交互式 REPL。每次启动默认开新 session,用 --resume 接老的。"""
+@click.option("--workspace", default=None, help="工作目录(存 tasks/ 和 sessions/)")
+@click.option("--resume", default=None, help="恢复 task: 'last' 或 task_id")
+@click.option("--mode", default="", help="任务模式标签(coding/ppt/proposal/...自由形式)")
+@click.option("--desc", default="", help="一句话任务描述,便于 tasks 列表识别")
+def chat(model: str, workspace: str, resume: str, mode: str, desc: str) -> None:
+    """启动交互式 REPL。每次启动默认开新 task,用 --resume 接老的。"""
     console = Console()
     try:
-        agent, session, sid = build_agent(
+        agent, session, sid, task_state, task_dir = build_agent(
             model_name=model,
             workspace=workspace,
             console=console,
             session_id=resume,
             resume=bool(resume),
+            mode=mode,
+            description=desc,
         )
     except Exception as e:
         console.print(f"[red]启动失败:[/red] {type(e).__name__}: {e}")
@@ -46,15 +60,21 @@ def chat(model: str, workspace: str, resume: str) -> None:
 
     if resume:
         console.print(
-            f"[green]恢复 session[/green] [bold]{sid}[/bold] ({len(session.messages)} 条消息) "
+            f"[green]恢复 task[/green] [bold]{sid}[/bold] ({len(session.messages)} 条消息) "
             f"model: [bold]{agent.caps.model_id}[/bold]"
         )
     else:
+        meta_tail = ""
+        if task_state.mode or task_state.description:
+            meta_tail = f"  mode={task_state.mode!r}  desc={task_state.description!r}"
         console.print(
-            f"[green]新 session[/green] [bold]{sid}[/bold] "
-            f"model: [bold]{agent.caps.model_id}[/bold]"
+            f"[green]新 task[/green] [bold]{sid}[/bold] "
+            f"model: [bold]{agent.caps.model_id}[/bold]{meta_tail}"
         )
-    console.print("[dim]/exit 退出  /reset 清空当前对话  /new 开一个新 session  /id 显示 session id[/dim]\n")
+    console.print(
+        "[dim]/exit 退出  /reset 清空对话(保留 task)  /new 开新 task  /id /status 查看  "
+        "/done /abandon 改状态  /desc <文本> 设描述[/dim]\n"
+    )
 
     while True:
         try:
@@ -68,22 +88,48 @@ def chat(model: str, workspace: str, resume: str) -> None:
             break
         if cmd == "/reset":
             session.reset(keep_system=True)
-            console.print("[dim]当前 session 已重置(保留 system)[/dim]")
+            console.print("[dim]当前 task 对话已重置(保留 system 和 state)[/dim]")
             continue
         if cmd == "/new":
             try:
-                agent, session, sid = build_agent(
-                    model_name=model, workspace=workspace, console=console
+                agent, session, sid, task_state, task_dir = build_agent(
+                    model_name=model, workspace=workspace, console=console,
+                    mode=mode, description=desc,
                 )
             except Exception as e:
                 console.print(f"[red]新建失败:[/red] {type(e).__name__}: {e}")
                 continue
-            console.print(f"[green]新 session[/green] [bold]{sid}[/bold]")
+            console.print(f"[green]新 task[/green] [bold]{sid}[/bold]")
             continue
         if cmd == "/id":
             cwd_disp = session.meta.get("cwd", "?")
             model_disp = session.meta.get("model", agent.caps.model_id)
-            console.print(f"[dim]session: {sid}  model: {model_disp}  cwd: {cwd_disp}[/dim]")
+            console.print(f"[dim]task: {sid}  model: {model_disp}  cwd: {cwd_disp}[/dim]")
+            continue
+        if cmd == "/status":
+            console.print(
+                f"[dim]task {task_state.task_id}  status={task_state.status}  "
+                f"mode={task_state.mode!r}  desc={task_state.description!r}\n"
+                f"  model={task_state.model}  tokens={task_state.tokens_total} "
+                f"(p={task_state.tokens_prompt}/c={task_state.tokens_completion})  "
+                f"created={task_state.created_at}  updated={task_state.updated_at}[/dim]"
+            )
+            continue
+        if cmd == "/done":
+            task_state.status = "completed"
+            task_state.save(task_dir)
+            console.print(f"[green]task {sid} marked completed[/green]")
+            break
+        if cmd == "/abandon":
+            task_state.status = "abandoned"
+            task_state.save(task_dir)
+            console.print(f"[yellow]task {sid} marked abandoned[/yellow]")
+            break
+        if cmd.startswith("/desc"):
+            new_desc = cmd[len("/desc"):].strip()
+            task_state.description = new_desc
+            task_state.save(task_dir)
+            console.print(f"[dim]description set: {new_desc!r}[/dim]")
             continue
         if not cmd:
             continue
@@ -91,47 +137,127 @@ def chat(model: str, workspace: str, resume: str) -> None:
         try:
             agent.run(user_input)
         except KeyboardInterrupt:
-            console.print("\n[yellow]已中断本轮。下一条输入会继续这个 session。[/yellow]")
+            console.print("\n[yellow]已中断本轮。下一条输入会继续这个 task。[/yellow]")
         except Exception as e:
             console.print(f"[red]运行错误:[/red] {type(e).__name__}: {e}")
+        finally:
+            sync_task_tokens(task_state, task_dir, agent.llm)
 
 
 @cli.command()
 @click.option("--workspace", default=None, help="工作目录")
 @click.option("--limit", default=20, help="显示最近 N 个")
-def sessions(workspace: str, limit: int) -> None:
-    """列出已有 session。"""
+@click.option("--status", default=None, help="只看某状态: active / completed / abandoned")
+def tasks(workspace: str, limit: int, status: str) -> None:
+    """列出已有 task(新格式,workspace/tasks/<id>/state.json)。"""
     cfg = load_config()
     ws = resolve_workspace(workspace, cfg)
-    sdir = sessions_dir(ws)
+    tdir = tasks_dir(ws)
 
-    items = sorted(sdir.glob("*.json"), reverse=True)[:limit]
-    if not items:
-        click.echo(f"(no sessions in {sdir})")
-        return
-
-    click.echo(f"{'session id':<18}  {'msgs':>4}  {'cwd':<32}  preview")
-    click.echo("-" * 100)
-    for p in items:
+    rows = []  # (mtime, task_id, status, mode, model, tokens, n_msgs, desc)
+    for d in tdir.iterdir():
+        if not d.is_dir():
+            continue
+        msg_path = d / "messages.json"
+        if not msg_path.exists():
+            continue
+        st = TaskState.load(d)
+        if st is None:
+            continue
+        if status and st.status != status:
+            continue
         try:
-            data = json.loads(p.read_text(encoding="utf-8"))
-            if isinstance(data, list):
-                messages, meta = data, {}
-            else:
-                messages = data.get("messages", []) or []
-                meta = data.get("meta", {}) or {}
-            n = len(messages)
-            preview = ""
-            for m in messages:
-                if isinstance(m, dict) and m.get("role") == "user":
-                    preview = (m.get("content") or "")[:50].replace("\n", " ")
-                    break
-            cwd = meta.get("cwd") or "?"
-            if len(cwd) > 32:
-                cwd = "..." + cwd[-29:]
-        except Exception as e:
-            n, preview, cwd = -1, f"[parse error: {e}]", "?"
-        click.echo(f"{p.stem:<18}  {n:>4}  {cwd:<32}  {preview}")
+            data = json.loads(msg_path.read_text(encoding="utf-8"))
+            n = len(data.get("messages", []))
+        except Exception:
+            n = -1
+        rows.append((
+            msg_path.stat().st_mtime, st.task_id, st.status, st.mode,
+            st.model_profile or st.model, st.tokens_total, n, st.description,
+        ))
+    rows.sort(reverse=True)
+    rows = rows[:limit]
+
+    if not rows:
+        click.echo(f"(no tasks in {tdir})")
+        return
+    tbl = Table(show_lines=False)
+    tbl.add_column("task id", style="bold")
+    tbl.add_column("status")
+    tbl.add_column("mode")
+    tbl.add_column("model")
+    tbl.add_column("msgs", justify="right")
+    tbl.add_column("tokens", justify="right")
+    tbl.add_column("desc")
+    sc = {"active": "cyan", "completed": "green", "abandoned": "dim"}
+    for _, tid, st, mode, model, tok, n, desc in rows:
+        c = sc.get(st, "white")
+        d_show = desc if len(desc) <= 50 else desc[:47] + "..."
+        tbl.add_row(tid, f"[{c}]{st}[/{c}]", mode, model, str(n), str(tok), d_show)
+    Console().print(tbl)
+
+
+@cli.command()
+@click.option("--model", default=None, help="模型档案,如 deepseek_v4.flash 或 deepseek_v4.pro")
+@click.option("--long-context", is_flag=True, help="加跑 needle-in-haystack(费 token,默认关)")
+def probe(model: str, long_context: bool) -> None:
+    """实测对账模型 yaml 声称的能力。会调用 LLM,有 API 开销。"""
+    from core.capabilities import ModelCapabilities
+    from core.llm import LLM
+    from core.probe import probe_capabilities
+
+    cfg = load_config()
+    name = model or cfg["default_model"]
+
+    console = Console()
+    try:
+        caps = ModelCapabilities.load(name, ROOT / cfg["models_dir"])
+    except Exception as e:
+        console.print(f"[red]档案加载失败:[/red] {type(e).__name__}: {e}")
+        sys.exit(1)
+
+    console.print(
+        f"[bold]probing[/bold] [cyan]{caps.model_id}[/cyan] (profile: {name}) "
+        f"[dim]long-context={long_context}[/dim]\n"
+    )
+
+    try:
+        llm = LLM(caps)
+    except Exception as e:
+        console.print(f"[red]LLM 构造失败:[/red] {type(e).__name__}: {e}")
+        sys.exit(1)
+
+    with console.status("[dim]running probes...[/dim]", spinner="dots"):
+        report = probe_capabilities(caps, llm, include_long_context=long_context)
+
+    tbl = Table(show_lines=False)
+    tbl.add_column("capability", style="bold")
+    tbl.add_column("declared")
+    tbl.add_column("observed")
+    tbl.add_column("status")
+    tbl.add_column("detail")
+    color = {"ok": "green", "mismatch": "yellow", "error": "red", "skip": "dim"}
+    for r in report.results:
+        c = color.get(r.status, "white")
+        tbl.add_row(
+            r.name,
+            str(r.declared),
+            str(r.observed),
+            f"[{c}]{r.status}[/{c}]",
+            r.detail,
+        )
+    console.print(tbl)
+
+    if report.has_mismatch:
+        console.print(
+            "\n[yellow]存在能力对账差异 —— 看 detail,必要时改 "
+            f"config/models/{caps.family}.yaml[/yellow]"
+        )
+        sys.exit(2)
+    if any(r.status == "error" for r in report.results):
+        console.print("\n[red]部分探测出错(见 detail)[/red]")
+        sys.exit(3)
+    console.print("\n[green]全部能力声明与实测一致。[/green]")
 
 
 if __name__ == "__main__":
diff --git a/core/probe.py b/core/probe.py
new file mode 100644
index 0000000..554b274
--- /dev/null
+++ b/core/probe.py
@@ -0,0 +1,243 @@
+"""能力探测: 用真实 LLM 调用对账 yaml 中声称的能力。
+
+只在用户显式触发(`cli.py probe`)时跑——会花 API 额度,不进启动路径。
+不修改 yaml,只输出对比报告;让用户自己判断要不要改档案。
+
+四项探测:
+- basic_chat:连通性。失败则跳过其余。
+- parallel_tools:给两个独立工具,看 single response 里 tool_calls 数量。
+- thinking_mode:对 declared=True 的模型传 reasoning_effort,看 API 是否接受 + 是否产出 thinking。
+- long_context(opt-in):needle-in-haystack 简化版,默认探 reliable_context 的 1/8。
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, List, Optional
+
+from .capabilities import ModelCapabilities
+from .llm import LLM
+
+
+@dataclass
+class ProbeResult:
+    name: str
+    declared: Any
+    observed: Any
+    status: str  # "ok" / "mismatch" / "skip" / "error"
+    detail: str = ""
+
+
+@dataclass
+class ProbeReport:
+    model: str
+    results: List[ProbeResult] = field(default_factory=list)
+
+    def add(self, r: ProbeResult) -> None:
+        self.results.append(r)
+
+    @property
+    def has_mismatch(self) -> bool:
+        return any(r.status == "mismatch" for r in self.results)
+
+
+def _msg_dict(msg: Any) -> dict:
+    if hasattr(msg, "model_dump"):
+        return msg.model_dump()
+    if hasattr(msg, "dict"):
+        return msg.dict()
+    return {}
+
+
+# ----- 单项 probe -----
+
+def probe_basic_chat(llm: LLM) -> ProbeResult:
+    try:
+        resp = llm.chat(
+            messages=[{"role": "user", "content": "Reply with exactly the word: pong"}],
+        )
+        text = (resp.choices[0].message.content or "").strip()
+        ok = "pong" in text.lower()
+        return ProbeResult(
+            name="basic_chat",
+            declared="reachable",
+            observed=text[:40] or "<empty>",
+            status="ok" if ok else "mismatch",
+            detail="" if ok else f"expected 'pong', got: {text[:80]!r}",
+        )
+    except Exception as e:
+        return ProbeResult(
+            name="basic_chat",
+            declared="reachable",
+            observed=None,
+            status="error",
+            detail=f"{type(e).__name__}: {e}",
+        )
+
+
+def probe_parallel_tools(llm: LLM, caps: ModelCapabilities) -> ProbeResult:
+    declared = caps.parallel_tools
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "get_weather",
+                "description": "Get current weather for a city.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"city": {"type": "string"}},
+                    "required": ["city"],
+                },
+            },
+        },
+        {
+            "type": "function",
+            "function": {
+                "name": "get_time",
+                "description": "Get current time in a timezone.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"tz": {"type": "string"}},
+                    "required": ["tz"],
+                },
+            },
+        },
+    ]
+    user_msg = (
+        "I need two independent pieces of information at the same time: the weather "
+        "in Beijing AND the current time in Tokyo. Please call BOTH tools in this "
+        "single turn (in parallel)."
+    )
+    try:
+        resp = llm.chat(
+            messages=[{"role": "user", "content": user_msg}],
+            tools=tools,
+            parallel_tool_calls=True,
+        )
+        tool_calls = getattr(resp.choices[0].message, "tool_calls", None) or []
+        n = len(tool_calls)
+        observed = n >= 2
+        return ProbeResult(
+            name="parallel_tools",
+            declared=declared,
+            observed=observed,
+            status="ok" if observed == bool(declared) else "mismatch",
+            detail=f"{n} tool_calls in single response",
+        )
+    except Exception as e:
+        return ProbeResult(
+            name="parallel_tools",
+            declared=declared,
+            observed=None,
+            status="error",
+            detail=f"{type(e).__name__}: {e}",
+        )
+
+
+def probe_thinking_mode(llm: LLM, caps: ModelCapabilities) -> ProbeResult:
+    declared = caps.thinking_mode
+    if not declared:
+        return ProbeResult(
+            name="thinking_mode",
+            declared=False,
+            observed=None,
+            status="skip",
+            detail="declared false; skipping (cap-side flag controls API forwarding)",
+        )
+    effort = (
+        caps.default_reasoning_effort
+        or (caps.reasoning_effort_levels[0] if caps.reasoning_effort_levels else "medium")
+    )
+    try:
+        resp = llm.chat(
+            messages=[{"role": "user", "content": "Briefly: what is 17 * 23?"}],
+            reasoning_effort=effort,
+        )
+        msg = resp.choices[0].message
+        d = _msg_dict(msg)
+        rc = (
+            getattr(msg, "reasoning_content", None)
+            or getattr(msg, "thinking", None)
+            or d.get("reasoning_content")
+            or d.get("thinking")
+        )
+        observed = bool(rc)
+        return ProbeResult(
+            name="thinking_mode",
+            declared=True,
+            observed=observed,
+            status="ok" if observed else "mismatch",
+            detail=(
+                f"reasoning_effort={effort} accepted; "
+                + ("thinking content returned" if observed else "no thinking content in response")
+            ),
+        )
+    except Exception as e:
+        return ProbeResult(
+            name="thinking_mode",
+            declared=True,
+            observed=False,
+            status="mismatch",
+            detail=f"reasoning_effort rejected: {type(e).__name__}: {e}",
+        )
+
+
+def probe_long_context(
+    llm: LLM, caps: ModelCapabilities, target_chars: Optional[int] = None
+) -> ProbeResult:
+    """needle-in-haystack 简化版。默认探 reliable_context * 4 / 8 字符,上限 200K。"""
+    if target_chars is None:
+        target_chars = caps.reliable_context * 4 // 8
+    target_chars = max(2_000, min(target_chars, 200_000))
+    SECRET = "K7-ZULU-9213"
+    pad = "The quick brown fox jumps over the lazy dog. " * 200
+    n_blocks = max(1, target_chars // len(pad))
+    middle = n_blocks // 2
+    parts: List[str] = []
+    for i in range(n_blocks):
+        if i == middle:
+            parts.append(f"\n>>> SECRET TOKEN: {SECRET} <<<\n")
+        parts.append(pad)
+    haystack = "".join(parts)
+    prompt = (
+        "Below is a long block of text. Somewhere in it a SECRET TOKEN is recorded "
+        "after the marker '>>> SECRET TOKEN:'. Reply with ONLY the token value, "
+        "nothing else.\n\n" + haystack
+    )
+    try:
+        resp = llm.chat(messages=[{"role": "user", "content": prompt}])
+        text = (resp.choices[0].message.content or "").strip()
+        ok = SECRET in text
+        return ProbeResult(
+            name="long_context",
+            declared=f"reliable_context={caps.reliable_context}",
+            observed=f"{len(haystack)} chars sent; secret {'recovered' if ok else 'missed'}",
+            status="ok" if ok else "mismatch",
+            detail=f"reply head: {text[:80]!r}",
+        )
+    except Exception as e:
+        return ProbeResult(
+            name="long_context",
+            declared=f"reliable_context={caps.reliable_context}",
+            observed=None,
+            status="error",
+            detail=f"{type(e).__name__}: {e}",
+        )
+
+
+# ----- 顶层入口 -----
+
+def probe_capabilities(
+    caps: ModelCapabilities,
+    llm: LLM,
+    *,
+    include_long_context: bool = False,
+) -> ProbeReport:
+    report = ProbeReport(model=caps.model_id)
+    report.add(probe_basic_chat(llm))
+    if report.results[0].status == "error":
+        return report
+    report.add(probe_parallel_tools(llm, caps))
+    report.add(probe_thinking_mode(llm, caps))
+    if include_long_context:
+        report.add(probe_long_context(llm, caps))
+    return report
diff --git a/core/task.py b/core/task.py
new file mode 100644
index 0000000..aed061e
--- /dev/null
+++ b/core/task.py
@@ -0,0 +1,63 @@
+"""任务状态: DESIGN.md §7.1 规约,落 `<task_dir>/state.json`。
+
+Task 是 Session 的上层概念 —— Session 只管对话消息,Task 还管 mode/description/
+status/tokens/cost/timestamps,这些是跨轮次共享、和文件系统状态对齐的元数据。
+
+文件路径约定(workspace/ 下):
+    tasks/<task_id>/state.json     ← 此模块负责
+    tasks/<task_id>/messages.json  ← Session 落
+"""
+from __future__ import annotations
+
+import json
+from dataclasses import asdict, dataclass, fields
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+
+@dataclass
+class TaskState:
+    task_id: str
+    mode: str = ""               # 自由形式: coding / ppt / proposal / general / 自定
+    description: str = ""        # 一句话描述,便于列表识别
+    status: str = "active"       # active / completed / abandoned
+    model: str = ""              # caps.model_id
+    model_profile: str = ""      # 档案名,如 deepseek_v4.flash
+    reasoning_effort: str = ""
+    cwd: str = ""                # 任务的工作基目录
+    created_at: str = ""         # ISO 时间戳
+    updated_at: str = ""
+    tokens_prompt: int = 0
+    tokens_completion: int = 0
+    cost_usd: float = 0.0        # 暂不算,留位
+
+    @property
+    def tokens_total(self) -> int:
+        return self.tokens_prompt + self.tokens_completion
+
+    def save(self, task_dir: Path) -> None:
+        task_dir.mkdir(parents=True, exist_ok=True)
+        self.updated_at = datetime.now().isoformat(timespec="seconds")
+        (task_dir / "state.json").write_text(
+            json.dumps(asdict(self), ensure_ascii=False, indent=2),
+            encoding="utf-8",
+        )
+
+    @classmethod
+    def load(cls, task_dir: Path) -> Optional["TaskState"]:
+        p = task_dir / "state.json"
+        if not p.exists():
+            return None
+        try:
+            data = json.loads(p.read_text(encoding="utf-8"))
+        except Exception:
+            return None
+        if not isinstance(data, dict):
+            return None
+        # 容忍 schema 演化:只取已知字段,缺失字段用 dataclass 默认
+        known = {f.name for f in fields(cls)}
+        kwargs = {k: v for k, v in data.items() if k in known}
+        if "task_id" not in kwargs:
+            kwargs["task_id"] = task_dir.name
+        return cls(**kwargs)
diff --git a/main.py b/main.py
index fc892f8..cbda6e1 100644
--- a/main.py
+++ b/main.py
@@ -1,4 +1,9 @@
-"""装配入口: 读 config → 加载 capabilities/skills → 构造 LLM/tools/session/loop。"""
+"""装配入口: 读 config → 加载 capabilities/skills → 构造 LLM/tools/session/loop。
+
+存储布局:
+    workspace/tasks/<task_id>/state.json     ← TaskState
+    workspace/tasks/<task_id>/messages.json  ← Session 消息
+"""
 from __future__ import annotations
 
 from datetime import datetime
@@ -13,6 +18,7 @@ from core.llm import LLM
 from core.loop import AgentLoop
 from core.session import Session
 from core.skills import SkillRegistry
+from core.task import TaskState
 from tools.fs import EditTool, GlobTool, GrepTool, ReadTool, WriteTool
 from tools.run_python import RunPythonTool
 from tools.shell import ShellTool
@@ -32,28 +38,38 @@ def resolve_workspace(workspace: Optional[str], cfg: Optional[dict] = None) -> P
     return p
 
 
-def sessions_dir(workspace_dir: Path) -> Path:
-    d = workspace_dir / "sessions"
+def tasks_dir(workspace_dir: Path) -> Path:
+    d = workspace_dir / "tasks"
     d.mkdir(parents=True, exist_ok=True)
     return d
 
 
-def resolve_session_path(workspace_dir: Path, session_id: Optional[str], resume: bool) -> Tuple[Path, str]:
-    """返回 (path, session_id)。resume=True 时找现有文件,否则新建一个时间戳 id。"""
-    sdir = sessions_dir(workspace_dir)
+def resolve_task_messages_path(
+    workspace_dir: Path, task_id: Optional[str], resume: bool
+) -> Tuple[Path, str]:
+    """返回 (messages_file_path, task_id)。
+    新建:tasks/<id>/messages.json;Resume:tasks/<id>/messages.json,'last' 取最新。
+    """
+    tdir = tasks_dir(workspace_dir)
     if resume:
-        if session_id in (None, "", "last"):
-            existing = sorted(sdir.glob("*.json"))
-            if not existing:
-                raise FileNotFoundError(f"{sdir} 下没有任何 session 可恢复")
-            path = existing[-1]
-            return path, path.stem
-        path = sdir / f"{session_id}.json"
-        if not path.exists():
-            raise FileNotFoundError(f"session 不存在: {path}")
-        return path, session_id  # type: ignore[return-value]
-    sid = session_id or datetime.now().strftime("%Y%m%d_%H%M%S")
-    return sdir / f"{sid}.json", sid
+        if task_id in (None, "", "last"):
+            candidates = []
+            for d in tdir.iterdir():
+                mf = d / "messages.json"
+                if mf.is_file():
+                    candidates.append((mf.stat().st_mtime, mf, d.name))
+            if not candidates:
+                raise FileNotFoundError(f"无可恢复的 task: {tdir} 下无 task")
+            candidates.sort(key=lambda x: x[0], reverse=True)
+            _, path, sid = candidates[0]
+            return path, sid
+        task_msg = tdir / task_id / "messages.json"
+        if not task_msg.exists():
+            raise FileNotFoundError(f"task 不存在: {task_msg}")
+        return task_msg, task_id
+
+    sid = task_id or datetime.now().strftime("%Y%m%d_%H%M%S")
+    return tdir / sid / "messages.json", sid
 
 
 def build_agent(
@@ -62,7 +78,11 @@ def build_agent(
     console: Optional[Console] = None,
     session_id: Optional[str] = None,
     resume: bool = False,
-) -> Tuple[AgentLoop, Session, str]:
+    tool_base: Optional[Path] = None,
+    mode: str = "",
+    description: str = "",
+) -> Tuple[AgentLoop, Session, str, TaskState, Path]:
+    """返回 (agent, session, task_id, task_state, task_dir)。"""
     cfg = load_config()
     model = model_name or cfg["default_model"]
 
@@ -70,38 +90,65 @@ def build_agent(
     llm = LLM(caps)
 
     workspace_dir = resolve_workspace(workspace, cfg)
-    session_path, sid = resolve_session_path(workspace_dir, session_id, resume)
+    session_path, sid = resolve_task_messages_path(workspace_dir, session_id, resume)
 
-    # 工具基目录: 用户当前 cwd —— agent 操作的是用户项目,不是 zcbot 仓库本身
-    tool_base = Path.cwd()
+    tool_base = Path(tool_base) if tool_base else Path.cwd()
 
     skills = SkillRegistry(ROOT / cfg.get("skills_dir", "skills"))
 
+    task_dir = session_path.parent
+
     if resume:
-        # 恢复: 直接加载老 session,不再注入新的 system prompt
         session = Session.load(session_path)
         saved_cwd = session.meta.get("cwd")
         if saved_cwd and console is not None and saved_cwd != str(tool_base):
             console.print(
-                f"[yellow]提示:[/yellow] 当前 cwd 与 session 记录不同 —— "
+                f"[yellow]提示:[/yellow] 当前 cwd 与 task 记录不同 —— "
                 f"工具基于 current cwd,不会自动切回。\n"
-                f"  session cwd: [dim]{saved_cwd}[/dim]\n"
+                f"  task cwd: [dim]{saved_cwd}[/dim]\n"
                 f"  current cwd: [dim]{tool_base}[/dim]"
             )
+        task_state = TaskState.load(task_dir)
+        if task_state is None:
+            # messages.json 存在但 state.json 缺失:用 session.meta 兜底重建
+            task_state = TaskState(
+                task_id=sid,
+                mode=mode,
+                description=description,
+                status="active",
+                model=session.meta.get("model", caps.model_id),
+                model_profile=session.meta.get("model_profile", model),
+                cwd=session.meta.get("cwd", str(tool_base)),
+                created_at=session.meta.get("created_at", datetime.now().isoformat(timespec="seconds")),
+            )
+            task_state.save(task_dir)
     else:
         system_prompt = (ROOT / cfg["system_prompt"]).read_text(encoding="utf-8")
         if skills.skills:
             system_prompt += f"\n\n## 可用 skill (用 load_skill 加载完整指引)\n{skills.discovery_block()}"
         system_prompt += f"\n\n## 当前工作目录\n{tool_base}"
+        now_iso = datetime.now().isoformat(timespec="seconds")
         meta = {
             "id": sid,
-            "created_at": datetime.now().isoformat(timespec="seconds"),
+            "created_at": now_iso,
             "cwd": str(tool_base),
             "model": caps.model_id,
             "model_profile": model,
         }
         session = Session(system_prompt=system_prompt, path=session_path, meta=meta)
-        session.save()  # 立刻落盘,占住文件名
+        session.save()  # 占住文件名
+        task_state = TaskState(
+            task_id=sid,
+            mode=mode,
+            description=description,
+            status="active",
+            model=caps.model_id,
+            model_profile=model,
+            reasoning_effort=caps.default_reasoning_effort or "",
+            cwd=str(tool_base),
+            created_at=now_iso,
+        )
+        task_state.save(task_dir)
 
     tools = {}
     for cls in (ReadTool, WriteTool, EditTool, GlobTool, GrepTool, ShellTool):
@@ -117,4 +164,12 @@ def build_agent(
         tools[rp.name] = rp
 
     agent = AgentLoop(llm, tools, session, caps, console=console)
-    return agent, session, sid
+    return agent, session, sid, task_state, task_dir
+
+
+def sync_task_tokens(task_state: TaskState, task_dir: Path, llm: LLM) -> None:
+    """每轮 agent.run 后调,把 LLM 累计 tokens 写回 state.json。"""
+    tc = llm.token_counter
+    task_state.tokens_prompt = tc.prompt_tokens
+    task_state.tokens_completion = tc.completion_tokens
+    task_state.save(task_dir)