Session resume overrides newly generated system prompt, preventing skill/config updates from taking effect / 恢复会话时旧系统提示覆盖新生成的提示，导致技能与配置更新无法生效

Problem Description

When a user resumes an old session, Kimi CLI reads the stale _system_prompt from context.jsonl and unconditionally overrides the freshly generated system prompt from load_agent(). This causes:

1. New skills never appear: Adding a skill to ~/.kimi/skills/ and restarting/resuming the session still shows the old skill list in the system prompt.
2. Config changes are silently ignored: Modifying settings that affect the system prompt (e.g. merge_all_available_skills) has no effect after resume.
3. /fork cannot fix it: /fork copies the old context.jsonl verbatim, including the stale system prompt.

Reproduction Steps

1. Start a session (system prompt loads N skills).
2. Add a new skill to ~/.kimi/skills/.
3. Exit and re-enter Kimi CLI (auto-resumes last session).
4. Observe ## Available skills in the system prompt — still only N skills, the new one is missing.
5. Check logs: Discovered {N+1} skill(s) is logged, but the system prompt does not reflect it.

Root Cause

The root cause is in src/kimi_cli/app.py:

context = Context(session.context_file)
await context.restore()

if context.system_prompt is not None:
    agent = dataclasses.replace(agent, system_prompt=context.system_prompt)  # stale overrides fresh
else:
    await context.write_system_prompt(agent.system_prompt)

Runtime.create() and load_agent() do re-discover skills from disk and render a fresh system prompt. However, context.restore() reads the persisted _system_prompt and overwrites the fresh one unconditionally.

Impact

• Users cannot update skills via normal restart.
• Users cannot update skills via /fork.
• Workaround requires manually deleting context.jsonl or using kimi --new.

Suggested Fixes

Option A (recommended): When resuming, compare the skills section of the old vs. new system prompt. If they differ (or if KIMI_AGENTS_MD / other external-state sections differ), use the newly generated prompt and persist it:

context = Context(session.context_file)
await context.restore()

if context.system_prompt is not None:
    if _skills_section_changed(context.system_prompt, agent.system_prompt):
        await context.write_system_prompt(agent.system_prompt)
    else:
        agent = dataclasses.replace(agent, system_prompt=context.system_prompt)
else:
    await context.write_system_prompt(agent.system_prompt)

Option B: Add a config flag always_regenerate_system_prompt = true for power users.

Option C: During /fork, do not copy _system_prompt from context.jsonl; let the forked session regenerate it.

Environment

• kimi-cli version: 1.46.0
• OS: Windows 11

---

問題描述

當用戶恢復舊 session 時，Kimi CLI 會從 context.jsonl 中讀取舊的 _system_prompt，並用它無條件覆蓋 load_agent() 剛剛新生成的系統提示。這導致：

1. 新增的技能永遠不會出現：用戶在 ~/.kimi/skills/ 新增了 skill，重啟/恢復 session 後系統提示仍然顯示舊的 skill 列表。
2. 配置變更被靜默忽略：修改 merge_all_available_skills 等影響系統提示的配置後，重啟仍然使用舊配置生成的系統提示。
3. /fork 也無法解決：/fork 直接複製舊的 context.jsonl，包括舊的系統提示。

復現步驟

1. 創建一個 session（此時系統提示載入 N 個 skills）。
2. 在 ~/.kimi/skills/ 新增一個 skill。
3. 退出 Kimi CLI，重新進入（會自動恢復上個 session）。
4. 觀察系統提示的 ## Available skills 區塊——仍然只有 N 個 skills，新增的 skill 沒有出現。
5. 檢查日誌可見 Discovered {N+1} skill(s)，但系統提示中並未反映。

根因分析

問題出在 src/kimi_cli/app.py：

context = Context(session.context_file)
await context.restore()

if context.system_prompt is not None:
    agent = dataclasses.replace(agent, system_prompt=context.system_prompt)  # 用舊的覆蓋新的
else:
    await context.write_system_prompt(agent.system_prompt)

Runtime.create() 和 load_agent() 確實會根據當前磁盤狀態重新發現 skills 並生成新的系統提示，但 context.restore() 讀取的舊 _system_prompt 會無條件覆蓋它。

影響範圍

• 用戶無法通過正常重啟更新 skills。
• 用戶無法通過 /fork 更新 skills。
• 必須手動刪除 context.jsonl 或使用 kimi --new 才能看到新技能。

建議修復方案

方案 A（推薦）：恢復 session 時比較新舊系統提示，若 KIMI_SKILLS 或 KIMI_AGENTS_MD 等依賴外部狀態的部分發生變化，則使用新生成的系統提示：

context = Context(session.context_file)
await context.restore()

if context.system_prompt is not None:
    if _skills_section_changed(context.system_prompt, agent.system_prompt):
        await context.write_system_prompt(agent.system_prompt)
    else:
        agent = dataclasses.replace(agent, system_prompt=context.system_prompt)
else:
    await context.write_system_prompt(agent.system_prompt)

方案 B：提供配置選項 always_regenerate_system_prompt = true，讓進階用戶選擇始終使用最新生成的系統提示。

方案 C：/fork 時不複製 context.jsonl 中的 _system_prompt，而是讓新會話重新生成。

環境

• kimi-cli version: 1.46.0
• OS: Windows 11

Start in src/kimi_cli/app.py at the session restore path, then trace Context.restore(), Runtime.create(), load_agent(), and the /fork handling described in the issue. Verify resumed and forked sessions use the freshly generated prompt when skills or prompt-related configuration has changed, while preserving the stored prompt when it has not.