压缩

上下文压缩

运行时上下文压缩：对话在数据库中全量保留；仅裁切发给 LLM 的四段视图。 关联：sleep.md、memory.md。

设计原则

原则	说明
不删消息	历史保留；压缩仅改变运行时视图与 conversation meta
四段	LLM 上下文 = system + summary + slim + raw
按需触发	用量接近窗口上限时压缩；工具循环内阈值更高
与记忆独立	压缩不触发语义提取；浅睡 cron 独立运行

运行时四段

① system  — system prompt (self layer + resident memory + project context)
② summary — text summary of compressed portion (synthetic; not stored in messages)
③ slim    — middle segment: trimmed user/assistant messages
④ raw     — recent segment: full messages including tool calls

非压缩路径：新消息仅追加到 raw；边界不变。

压缩路径：用量超过阈值时，较早消息移入 summary/slim；摘要文本异步更新。

配置

Adjust in Habitat runtime (habitat_runtime_config; Shell Settings → Habitat 服务 → 服务配置):

# habitat_runtime_config fragment (not config.yaml)
models:
  deepseek-v4-flash:
    context_window: 1000000

compression:
  enabled: true
  reserved_tokens: 8192
  trigger_low: 0.60 # outside tool loop: compress at 60% usage
  trigger_high: 0.80 # inside tool loop: compress at 80% usage
  emergency_ratio: 0.92
  raw_min_messages: 5
  slim_min_messages: 50
  summary_max_tokens: 4000

设置	默认值	说明
trigger_low	0.60	普通对话中的阈值
trigger_high	0.80	工具循环内的阈值
raw_min_messages	5	raw 段最少消息数
slim_min_messages	50	裁剪后 slim 段最少消息数

Token counting

Compression and embedding batch packing use heuristic token estimates (tokenx in-process, id __estimate__:tokenx), not full HuggingFace vocabularies. Counts are approximate (good enough for budget triggers and chunk sizing); they are not provider-billing exact.

上下文窗口解析（token 模式）

估算压缩预算时的优先级：

1. models.<model>.context_window in Habitat runtime config
2. compression.default_context_window（适用于所有模型）
3. Provider /models 目录中的 contextWindow（Habitat 已注册 lookup 时）
4. 以上皆无时的消息数回退（max_rounds 阈值）

Per-model config always wins over dynamically fetched catalog values; catalog is read-only and never written back to runtime config.

无任何上下文窗口来源时，压缩回退到按消息数触发。

在对话中强制压缩：/compress（--force 忽略迟滞）。

Manual summarize (Cursor-style): /summarize collapses history into the runtime summary without waiting for automatic thresholds. When the turn is idle (last message is a completed assistant reply), boundaries become l2 = l3 = l4 and slim/raw are empty. During an in-progress turn, only content through the last completed assistant is summarized; the unfinished tail stays in raw. Summary text is merged incrementally (same as automatic compression), and the command waits for the summary LLM to finish.

Mid-turn protection vs empty raw

Automatic boundary derivation keeps a non-empty raw segment (with a leading user message) so turns can continue while usage-based compression runs. That rule applies when some raw tail must remain. It is not a ban on empty raw: when the user explicitly /summarize on an idle conversation, full collapse (l2 = l3 = l4) is allowed. Mid-turn protection is implemented by not folding incomplete turns into the summary, not by permanently requiring a user in raw.

与记忆管道的关系

压缩与浅睡/深睡独立运行：压缩管理当前 conversation 的 LLM 窗口；记忆提取每晚从完整对话存档运行。