Refresh system prompt template variables on context compaction

What feature would you like to see?

Summary

Currently, BuiltinSystemPromptArgs is rendered once at session start (agent.py::Runtime.create) and stored as an immutable string on the frozen Agent dataclass. This means edits made during a long session to files that feed the system prompt — most notably AGENTS.md — do not take effect until the user restarts the session.

I'd like to propose refreshing these template variables when context compaction occurs. Compaction is a natural "reset point": the context is rebuilt and the system prompt is rewritten via KimiSoul.compact_context → Context.write_system_prompt anyway, which already invalidates the KV cache, so re-rendering at that moment is almost free conceptually.

Motivation

• AGENTS.md is the file users most actively edit during a session to steer the agent. The current "edit, then restart to see the effect" flow is friction-heavy.
• KIMI_NOW is frozen at session start, so multi-day sessions show a stale date in the system prompt.
• Other fields (KIMI_WORK_DIR_LS, KIMI_ADDITIONAL_DIRS_INFO) also drift over time.
• This would likely also mitigate https://github.com/MoonshotAI/kimi-cli/issues/2168, since that issue is in the same family of "state captured at session start becomes stale".

Open design questions

I'd like to align on the direction before writing a PR. The main forks:

1. Where to store the re-render inputs (system_prompt_path, system_prompt_args)?

   • (a) Add fields to Agent (keeps the spec close to the loaded agent, but Agent is frozen and arguably should stay "loaded snapshot").
   • (b) Add fields to Runtime (mutable already; subagents inherit naturally via copy_for_subagent).
   • My preference: (b), so that subagents stay in sync.

2. Partial refresh vs. full refresh of BuiltinSystemPromptArgs.

   • Refreshing only KIMI_AGENTS_MD + KIMI_NOW is the smallest change but leaves KIMI_WORK_DIR_LS / KIMI_ADDITIONAL_DIRS_INFO / KIMI_SKILLS inconsistent.
   • Full refresh is a few tens of ms (one list_directory + possibly skills rescan).

3. Failure semantics. Compaction is invoked when the context is near the limit; any new I/O on this path must be wrapped so a stale-but-working prompt is preferred over a hard failure.

Rough scope estimate

About two files (soul/agent.py, soul/kimisoul.py), well under 100 LoC if option (b) + partial-refresh-with-AGENTS.md-only is chosen; closer to ~80 LoC with full refresh. Tests would cover: AGENTS.md edit → compact → new prompt; read failure → compact still succeeds with old prompt; subagent inheritance behavior.

Happy to take this on if the direction is approved. Per CONTRIBUTING.md I'm raising an issue first since the change touches Agent / Runtime lifecycle and I'd rather not write a PR that gets reorganized in review.

Related: https://github.com/MoonshotAI/kimi-cli/issues/2168 — same root cause (state captured once at session start).

概述

目前 SystemPrompt 在会话开始时一次性渲染（见 agent.py::Runtime.create），并以字符串形式存放在 frozen 的 Agent 上。长会话中用户对 AGENTS.md 等文件的修改在当前会话内不会生效，必须重启会话。

希望能在 context compaction 触发时刷新这些模板变量。compaction 本身就会重写 system prompt（KimiSoul.compact_context → Context.write_system_prompt）导致 KV cache 失效，刚好可以用来刷新SystemPrompt。

动机

• AGENTS.md 是用户在会话过程中最常编辑、用于引导 agent 行为的文件。当前"改完要重启会话才生效"的流程体验割裂。
• KIMI_NOW 在会话开始时被冻结，跨天的长会话里 system prompt 中的日期就是错的。
• 其他字段（KIMI_WORK_DIR_LS、KIMI_ADDITIONAL_DIRS_INFO）也会随时间漂移。
• 此修改可能也会缓解 https://github.com/MoonshotAI/kimi-cli/issues/2168，因为该 issue 与本提案属于同一类问题——"会话开始时捕获的状态会过期"。

待讨论的设计点

希望先就方向对齐再动手写 PR。主要的分歧点：

1. 重新渲染所需的输入（system_prompt_path、system_prompt_args）放在哪里？

   • (a) 加在 Agent 上： agent自带从哪个模版上渲染的提示词，但 Agent 是 frozen 的，修改会破坏"加载即不变"的语义，而且subagent的system prompt不会同步更新。
   • (b) 加在 Runtime 上（本身就可变；subagent 通过 copy_for_subagent 天然继承），但是会让subagent的kv cache失效。

2. BuiltinSystemPromptArgs 部分刷新还是全量刷新？

   • 只刷新 KIMI_AGENTS_MD + KIMI_NOW 改动最小，但会让 KIMI_WORK_DIR_LS / KIMI_ADDITIONAL_DIRS_INFO / KIMI_SKILLS 处于不一致状态。
   • 全量刷新成本约几十毫秒（一次 list_directory，可能加 skills 重扫）。

3. 失败语义。 compaction 是在上下文濒临爆掉时触发的；该路径上新增的 I/O 必须做兜底，使"用旧 prompt 但 compact 成功"优先于"硬失败"。

改动量预估

约改动 2 个文件（soul/agent.py、soul/kimisoul.py），如果选 (b) + 只刷新 AGENTS.md 则远少于 100 行；全量刷新约 80 行。测试覆盖三个核心场景：AGENTS.md 编辑后 compact 生效、读失败时 compact 仍成功并降级为旧 prompt、subagent 继承行为。

Additional information

No response

Start by reading soul/agent.py, especially Runtime.create and copy_for_subagent, then trace compaction through soul/kimisoul.py and Context.write_system_prompt. The design choices for stored inputs, refresh scope, and failure handling still need maintainer agreement. Done would include agreed behavior and tests for AGENTS.md edits, refresh failures, and subagent inheritance.