feat(config): add reserved_context_size setting for configurable compaction trigger

CHANGELOG.md

@@ -11,6 +11,8 @@ Only write entries that are worth mentioning to users.
 
 ## Unreleased
 
+- Config: Add `reserved_context_size` setting to customize auto-compaction trigger threshold (default: 50000 tokens)
+
 ## 0.81 (2026-01-21)
 
 - Skills: Add flow skill type with embedded Agent Flow (Mermaid/D2) in SKILL.md, invoked via `/flow:<skill-name>` commands

docs/en/configuration/config-files.md

@@ -52,6 +52,7 @@ max_context_size = 262144
 max_steps_per_turn = 100
 max_retries_per_step = 3
 max_ralph_iterations = 0
+reserved_context_size = 50000
 
 [services.moonshot_search]
 base_url = "https://api.kimi.com/coding/v1/search"
@@ -117,6 +118,7 @@ capabilities = ["thinking", "image_in"]
 | `max_steps_per_turn` | `integer` | `100` | Maximum steps per turn (alias: `max_steps_per_run`) |
 | `max_retries_per_step` | `integer` | `3` | Maximum retries per step |
 | `max_ralph_iterations` | `integer` | `0` | Extra iterations after each user message; `0` disables; `-1` is unlimited |
+| `reserved_context_size` | `integer` | `50000` | Reserved token count for LLM response generation; auto-compaction triggers when `context_tokens + reserved_context_size >= max_context_size` |
 
 ### `services`
 

docs/en/release-notes/changelog.md

@@ -4,6 +4,8 @@ This page documents the changes in each Kimi CLI release.
 
 ## Unreleased
 
+- Config: Add `reserved_context_size` setting to customize auto-compaction trigger threshold (default: 50000 tokens)
+
 ## 0.81 (2026-01-21)
 
 - Skills: Add flow skill type with embedded Agent Flow (Mermaid/D2) in SKILL.md, invoked via `/flow:<skill-name>` commands

docs/zh/configuration/config-files.md

@@ -52,6 +52,7 @@ max_context_size = 262144
 max_steps_per_turn = 100
 max_retries_per_step = 3
 max_ralph_iterations = 0
+reserved_context_size = 50000
 
 [services.moonshot_search]
 base_url = "https://api.kimi.com/coding/v1/search"
@@ -117,6 +118,7 @@ capabilities = ["thinking", "image_in"]
 | `max_steps_per_turn` | `integer` | `100` | 单轮最大步数（别名：`max_steps_per_run`） |
 | `max_retries_per_step` | `integer` | `3` | 单步最大重试次数 |
 | `max_ralph_iterations` | `integer` | `0` | 每个 User 消息后额外自动迭代次数；`0` 表示关闭；`-1` 表示无限 |
+| `reserved_context_size` | `integer` | `50000` | 预留给 LLM 响应生成的 token 数量；当 `context_tokens + reserved_context_size >= max_context_size` 时自动触发压缩 |
 
 ### `services`
 

docs/zh/release-notes/changelog.md

@@ -4,6 +4,8 @@
 
 ## 未发布
 
+- Config：添加 `reserved_context_size` 配置项，自定义自动压缩触发阈值（默认 50000 tokens）
+
 ## 0.81 (2026-01-21)
 
 - Skills：添加 Flow Skill 类型，在 SKILL.md 中内嵌 Agent Flow（Mermaid/D2），通过 `/flow:<skill-name>` 命令调用

src/kimi_cli/config.py

@@ -55,6 +55,9 @@ class LoopControl(BaseModel):
     """Maximum number of retries in one step"""
     max_ralph_iterations: int = Field(default=0, ge=-1)
     """Extra iterations after the first turn in Ralph mode. Use -1 for unlimited."""
+    reserved_context_size: int = Field(default=50_000, ge=1000)
+    """Reserved token count for LLM response generation. Auto-compaction triggers when
+    context_tokens + reserved_context_size >= max_context_size. Default is 50000."""
 
 
 class MoonshotSearchConfig(BaseModel):

src/kimi_cli/soul/kimisoul.py

@@ -61,8 +61,6 @@ def type_check(soul: KimiSoul):
         _: Soul = soul
 
 
-RESERVED_TOKENS = 50_000
-
 SKILL_COMMAND_PREFIX = "skill:"
 FLOW_COMMAND_PREFIX = "flow:"
 DEFAULT_MAX_FLOW_MOVES = 1000
@@ -110,9 +108,6 @@ def __init__(
         self._context = context
         self._loop_control = agent.runtime.config.loop_control
         self._compaction = SimpleCompaction()  # TODO: maybe configurable and composable
-        self._reserved_tokens = RESERVED_TOKENS
-        if self._runtime.llm is not None:
-            assert self._reserved_tokens <= self._runtime.llm.max_context_size
 
         for tool in agent.toolset.tools:
             if tool.name == SendDMail_NAME:
@@ -348,10 +343,8 @@ async def _pipe_approval_to_wire():
             step_outcome: StepOutcome | None = None
             try:
                 # compact the context if needed
-                if (
-                    self._context.token_count + self._reserved_tokens
-                    >= self._runtime.llm.max_context_size
-                ):
+                reserved = self._loop_control.reserved_context_size
+                if self._context.token_count + reserved >= self._runtime.llm.max_context_size:
                     logger.info("Context too long, compacting...")
                     await self.compact_context()
 

tests/test_config.py

@@ -37,6 +37,7 @@ def test_default_config_dump():
                 "max_steps_per_turn": 100,
                 "max_retries_per_step": 3,
                 "max_ralph_iterations": 0,
+                "reserved_context_size": 50000,
             },
             "services": {"moonshot_search": None, "moonshot_fetch": None},
             "mcp": {"client": {"tool_call_timeout_ms": 60000}},
@@ -62,3 +63,13 @@ def test_load_config_text_invalid():
 def test_load_config_invalid_ralph_iterations():
     with pytest.raises(ConfigError, match="max_ralph_iterations"):
         load_config_from_string('{"loop_control": {"max_ralph_iterations": -2}}')
+
+
+def test_load_config_reserved_context_size():
+    config = load_config_from_string('{"loop_control": {"reserved_context_size": 30000}}')
+    assert config.loop_control.reserved_context_size == 30000
+
+
+def test_load_config_reserved_context_size_too_low():
+    with pytest.raises(ConfigError, match="reserved_context_size"):
+        load_config_from_string('{"loop_control": {"reserved_context_size": 500}}')