billion-context-dsh 0.1.4 → 0.1.6

package/README.en.md

@@ -3,7 +3,7 @@
 [English](./README.en.md) | [中文](./README.md)
 
 > **⚠️ Beta notice — not for production use**
-> This project (**v0.1.4**) is a work-in-progress beta. The [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) itself is also in **public beta**. **Do not use either in engineering / production environments** — expect breaking changes and rough edges.
+> This project (**v0.1.6**) is a work-in-progress beta. The [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) itself is also in **public beta**. **Do not use either in engineering / production environments** — expect breaking changes and rough edges.
 
 <p align="center">
 <strong>Built with gratitude on top of these projects</strong> — please give them a ⭐:
@@ -73,7 +73,7 @@ That's it. Then add a composition row where a compaction backend is expected —
     - id: compaction-acp
       name: 'billion-context-dsh'
       config:
-        modelContextLimit: 128000   # default; the pressure window
+        modelContextLimit: 128000   # optional; omit to auto-detect the model's real window (fallback 128000)
 ```
 
 **Per-mode — an agent preset's `compaction` realm.** First *disable (or delete) the realm's existing `dsh-compaction-basic` row*, then mount this engine — two backends cannot coexist in the same realm:
@@ -87,7 +87,7 @@ That's it. Then add a composition row where a compaction backend is expected —
 - id: compaction-acp
   name: 'billion-context-dsh'
   config:
-    modelContextLimit: 128000   # default; the pressure window
+    modelContextLimit: 128000   # optional; omit to auto-detect the model's real window (fallback 128000)
 ```
 
 > **One context manager per agent.** Two backends providing `ctx.compaction` collide — never run both in the same realm. Full install & verification guide: [docs/INSTALL.md](docs/INSTALL.md).
@@ -105,6 +105,7 @@ DSH derives every model request from its append-only session log (the *surface*)
 | `search_context` | scores block summaries + originals rebuilt from the log |
 | `acp_status` | block ledger + context pressure |
 | block state | in-memory kernel state + **log-rebuilt ledger** (no sidecar files) |
+| tiered distillation (T2/T3) | re-compressing a block's summary node distills that block (tier 2); distilling a tier-2 block yields tier 3. Tier + kernel block ids are persisted to the log, so kernel state rehydrates from the log after a restart and stays distillable |
 
 The load-bearing compression guidance (tools, philosophy, summary rules) is registered as a one-time system-prompt section, so nudges stay short. There is deliberately **no automatic summarization**: automatic policy only nudges the model (`compactIfNeeded` returns null).
 
@@ -118,7 +119,7 @@ A walkthrough of the ACP philosophy this project inherits — how active context
 
 | Tool | What it does |
 | --- | --- |
-| `compress` | Replace a seq range with a dense summary you write (edges auto-balanced to tool-pair boundaries) |
+| `compress` | Replace a seq range with a dense summary you write (edges auto-balanced to tool-pair boundaries); re-compressing a block's summary node distills it (tier 2/3) |
 | `decompress` | Restore a previously compressed block's original content (read-only) |
 | `search_context` | Search compressed block summaries and originals by keyword |
 | `acp_status` | Context usage, compressed blocks, compressible ranges |
@@ -141,10 +142,11 @@ This project reuses `acp-kernel`'s compression core and `billion-context-pi`'s d
 
 | Key | Default | Meaning |
 |---|---|---|
-| `modelContextLimit` | `128000` | Context window used for the kernel's pressure decisions |
-| `nudgeMinContextLimitPct` | kernel default `0.45` | Nudge window lower bound (usage fraction) — same default as billion-context-pi |
-| `nudgeMaxContextLimitPct` | kernel default `0.75` | Over-limit line: above this the nudge fires regardless of growth |
-| `nudgeEmergencyThresholdPct` | kernel default `0.95` | Emergency nudge (bypasses the per-turn dedup) |
+| `modelContextLimit` | auto-detected (fallback `128000`) | Context window used for the kernel's pressure decisions; an explicit value wins and skips the probe |
+| `autoModelContextLimit` | `true` | Probe the model's real window from the model API (`agent.ctx.llm.resolveModelInfo`); fall back to the default on failure, `acp_status` shows the window source |
+| `nudgeMinContextLimitPct` | kernel default `0.45` | Nudge window lower bound (usage fraction) — validation only; the growth-driven trigger has no percentage floor — same default as billion-context-pi |
+| `nudgeMaxContextLimitPct` | engine default `0.70` (kernel/pi default `0.75`) | Over-limit line: above this the nudge fires regardless of growth — deliberately below the host compaction-basic 80% auto-compaction line so the forced nudge fires first; an explicit value wins |
+| `nudgeEmergencyThresholdPct` | engine default `0.85` (kernel/pi default `0.95`) | Emergency nudge (bypasses the per-turn dedup) — lowered from `0.95`: at 95% the model has no room to act and the 80% auto-compaction line shadows it; an explicit value wins |
 | `coreOverrides` | — | Any other acp-kernel `Config` override (billion-context-pi's `coreOverrides` escape hatch) |
 | `autoTools` | `true` | Register the four model tools on `ctx.tools` |
 | `autoCommand` | `true` | Register the `/acp` command on `ctx.commands` |

package/README.md

@@ -3,7 +3,7 @@
 [中文](./README.md) | [English](./README.en.md)
 
 > **⚠️ 测试版声明——请勿用于生产环境**
-> 本项目（**v0.1.4**）仍处于开发中的测试版。[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 本身也处于**公开测试版**阶段。**请勿将两者用于工程化 / 生产环境**——预期会有破坏性变更与粗糙之处。
+> 本项目（**v0.1.6**）仍处于开发中的测试版。[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 本身也处于**公开测试版**阶段。**请勿将两者用于工程化 / 生产环境**——预期会有破坏性变更与粗糙之处。
 
 <p align="center">
 <strong>衷心感谢以下项目——请给它们一个 ⭐：</strong>
@@ -75,7 +75,7 @@ npm install billion-context-dsh
     - id: compaction-acp
       name: 'billion-context-dsh'
       config:
-        modelContextLimit: 128000   # 默认；压力窗口
+        modelContextLimit: 128000   # 可选；省略时自动探测模型真实窗口（回退 128000）
 ```
 
 **单模式生效（agent preset 的 `compaction` realm）**。先在该 realm 内*禁用（或删除）原有的 `dsh-compaction-basic` 行*，再插入本引擎——同一 realm 内两个后端不能并存：
@@ -89,7 +89,7 @@ npm install billion-context-dsh
 - id: compaction-acp
   name: 'billion-context-dsh'
   config:
-    modelContextLimit: 128000   # 默认；压力窗口
+    modelContextLimit: 128000   # 可选；省略时自动探测模型真实窗口（回退 128000）
 ```
 
 > **每个 agent 只留一个上下文管理器。** 两个后端同时 provide `ctx.compaction` 会冲突——同一 realm 内切勿并存。完整安装与验证指南见 [docs/INSTALL.md](docs/INSTALL.md)。
@@ -107,6 +107,7 @@ DSH 的每个模型请求都派生自其 append-only 会话日志（*surface*）
 | `search_context` | 对从日志重建的块摘要与原文打分 |
 | `acp_status` | 块账本与上下文压力 |
 | 块状态 | 内存内核状态 + **日志重建账本**（无旁车文件） |
+| 分层蒸馏（T2/T3） | 再次压缩某块的摘要节点 = 蒸馏该块（tier 2），蒸馏 tier-2 块得 tier 3；tier 与内核块 id 持久化进日志，重启后内核状态从日志再水合、可继续蒸馏 |
 
 承载性的压缩指引（工具、哲学、摘要规则）注册为一次性系统提示段，因此 nudge 保持简短。刻意**不做自动摘要**：自动策略只 nudge 模型（`compactIfNeeded` 返回 null）。
 
@@ -120,7 +121,7 @@ DSH 的每个模型请求都派生自其 append-only 会话日志（*surface*）
 
 | 工具 | 作用 |
 | --- | --- |
-| `compress` | 用你书写的紧凑摘要替换 seq 范围（边界自动平衡到 tool-call/result 配对点） |
+| `compress` | 用你书写的紧凑摘要替换 seq 范围（边界自动平衡到 tool-call/result 配对点）；对某块的摘要节点再次压缩 = 分层蒸馏（tier 2/3） |
 | `decompress` | 恢复已压缩块的原始内容（只读） |
 | `search_context` | 按关键词搜索压缩块摘要与原文 |
 | `acp_status` | 上下文占用、压缩块、可压缩范围 |
@@ -143,10 +144,11 @@ DSH 的每个模型请求都派生自其 append-only 会话日志（*surface*）
 
 | 键 | 默认值 | 含义 |
 |---|---|---|
-| `modelContextLimit` | `128000` | 用于内核压力决策的上下文窗口 |
-| `nudgeMinContextLimitPct` | 内核默认 `0.45` | Nudge 窗口下界（用量占比）——与 billion-context-pi 相同的默认值 |
-| `nudgeMaxContextLimitPct` | 内核默认 `0.75` | 过限线：超过此值则无论增长与否都触发 nudge |
-| `nudgeEmergencyThresholdPct` | 内核默认 `0.95` | 紧急 nudge（绕过每轮去重） |
+| `modelContextLimit` | 自动探测（回退 `128000`） | 用于内核压力决策的上下文窗口；显式配置时优先且跳过探测 |
+| `autoModelContextLimit` | `true` | 从模型 API 自动探测真实窗口（`agent.ctx.llm.resolveModelInfo`）；探测失败回退默认值，`acp_status` 展示窗口来源 |
+| `nudgeMinContextLimitPct` | 内核默认 `0.45` | Nudge 窗口下界（用量占比）——仅作配置校验，增长路径的触发没有百分比下限——与 billion-context-pi 相同的默认值 |
+| `nudgeMaxContextLimitPct` | engine 默认 `0.70`（内核/pi 默认 `0.75`） | 过限线：超过此值则无论增长与否都触发 nudge——刻意低于宿主 compaction-basic 的 80% 自动压缩线，保证强制 nudge 先触发；显式配置优先 |
+| `nudgeEmergencyThresholdPct` | engine 默认 `0.85`（内核/pi 默认 `0.95`） | 紧急 nudge（绕过每轮去重）——从 `0.95` 下调：95% 时模型已无操作空间且会被 80% 自动压缩线遮蔽；显式配置优先 |
 | `coreOverrides` | — | 任何其他 acp-kernel `Config` 覆盖（billion-context-pi 的 `coreOverrides` 逃生口） |
 | `autoTools` | `true` | 在 `ctx.tools` 注册四个模型工具 |
 | `autoCommand` | `true` | 在 `ctx.commands` 注册 `/acp` 命令 |

package/dist/config.d.ts

@@ -7,15 +7,19 @@
  * growth ratio 5%, protected last messages 5). Every knob is optional — an
  * omitted value keeps the kernel default, so the behavior matches the Pi
  * adapter exactly unless a deployment opts out.
+ *
+ * NOTE: `AcpCompactionEngine` (src/index.ts) ships its own engine-level
+ * defaults 0.70/0.85 for the two nudge thresholds on top of this layer, so an
+ * engine with no explicit config lands on 0.70/0.85, not 0.75/0.95.
  * @module billion-context-dsh/config
  */
 import { type Config } from 'acp-kernel';
 /** The kernel-facing knobs shared by the nudge path and the compress tool. */
 export interface KernelConfigInput {
     readonly modelContextLimit: number;
-    /** Nudge window lower bound (usage fraction). Kernel default: 0.45. */
+    /** Nudge window lower bound (usage fraction; validation only — the growth-driven trigger has no percentage floor). Kernel default: 0.45. */
     readonly nudgeMinContextLimitPct?: number;
-    /** Nudge window upper bound — over-limit. Kernel default: 0.75. */
+    /** Nudge window upper bound — over-limit guarantee line. Kernel default: 0.75. */
     readonly nudgeMaxContextLimitPct?: number;
     /** Emergency nudge threshold (bypasses per-turn dedup). Kernel default: 0.95. */
     readonly nudgeEmergencyThresholdPct?: number;

package/dist/index.d.ts

@@ -28,23 +28,44 @@
 import type { Context } from '@deepseek-ai/cordis';
 import { CompactionEngine, type CompactionAgentContext, type CompactionResult, type CompactionTrigger, type ManualCompactAgentContext } from '@deepseek-ai/dsh-compaction';
 import { type CompressionCore } from 'acp-kernel';
+import type { Agent } from '@deepseek-ai/dsh-agent';
 import { AcpStateStore } from './state.ts';
+import { type AcpWindow } from './window.ts';
 export { AcpStateStore } from './state.ts';
 export { kernelConfigFor, type KernelConfigInput } from './config.ts';
 export { ACP_SYSTEM_PROMPT, ACP_SYSTEM_PROMPT_ORDER } from './system-prompt.ts';
 export { makeTools, type ToolEnvironment } from './tools.ts';
 export { acpCommand } from './commands.ts';
 export { buildNudge, type NudgeEnvironment, type NudgeOutcome } from './nudge.ts';
-export { rebuildBlockLedger, resolveSurfaceRange, runCompactionTransaction, shadowedSeqsOf, findOpenTurn, assertNoActiveCompaction, type AcpBlockLedgerEntry, type CompactionTransactionInput, } from './region.ts';
+export { DEFAULT_CONTEXT_WINDOW, detectContextWindow, windowSourceLabel, type AcpWindow, } from './window.ts';
+export { rebuildBlockLedger, resolveSurfaceRange, runCompactionTransaction, shadowedSeqsOf, findOpenTurn, assertNoActiveCompaction, blockRegistry, blockRefForSummarySeq, compactionIdsOfKernelBlocks, summarySeqOfKernelBlock, expandShadowedSeqs, type AcpBlockLedgerEntry, type CompactionTransactionInput, } from './region.ts';
 export { eventsToCoreMessages, projectEvent, surfaceEventsOf, extractEventText } from './messages.ts';
 export interface AcpConfig {
-    /** The context window used for pressure decisions. Default 128000. */
-    readonly modelContextLimit: number;
-    /** Nudge window lower bound (usage fraction). Kernel default 0.45 — same as billion-context-pi. */
+    /**
+     * The context window used for pressure decisions, in tokens. When omitted,
+     * `autoModelContextLimit` (default true) probes the model's real window via
+     * `agent.ctx.llm.resolveModelInfo(provider, model)`; an explicit value
+     * always wins and disables the probe.
+     */
+    readonly modelContextLimit?: number;
+    /** Probe the model's real context window from the LLM runtime. Default true. */
+    readonly autoModelContextLimit: boolean;
+    /** Nudge window lower bound (usage fraction; validation only — the growth-driven trigger has no percentage floor). Kernel default 0.45 — same as billion-context-pi. */
     readonly nudgeMinContextLimitPct?: number;
-    /** Nudge window upper bound (over-limit). Kernel default 0.75 — same as billion-context-pi. */
+    /**
+     * Nudge window upper bound — over-limit guarantee line: above this the
+     * kernel injects a nudge regardless of growth or cadence. Engine default
+     * 0.70 (deliberately BELOW the kernel/billion-context-pi default 0.75 and
+     * the host compaction-basic auto-compaction line 0.80, so the forced nudge
+     * always fires first); an explicit value wins.
+     */
     readonly nudgeMaxContextLimitPct?: number;
-    /** Emergency nudge threshold (bypasses per-turn dedup). Kernel default 0.95 — same as billion-context-pi. */
+    /**
+     * Emergency nudge threshold (bypasses the per-turn dedup). Engine default
+     * 0.85 (down from the kernel/billion-context-pi default 0.95: 95% leaves
+     * the model no room to act before the API rejects, and the host's 80%
+     * compaction-basic line shadows it in standard/code/cordis modes).
+     */
     readonly nudgeEmergencyThresholdPct?: number;
     /** Any other acp-kernel Config override (billion-context-pi's `coreOverrides` escape hatch). */
     readonly coreOverrides?: Partial<import('acp-kernel').Config>;
@@ -69,7 +90,17 @@ export declare class AcpCompactionEngine extends CompactionEngine {
     /** Resolved engine configuration. */
     readonly config: AcpConfig;
     private readonly lastNudgeTurn;
+    /** Per provider/model route the resolved window (probe failures cached too). */
+    private readonly windowCache;
     constructor(ctx: Context, config?: Partial<AcpConfig>);
+    /**
+     * Resolve the effective context window for an agent. An explicitly
+     * configured `modelContextLimit` always wins (no probe). Otherwise probe the
+     * model's real window via `agent.ctx.llm.resolveModelInfo` (cached per
+     * provider/model route, probe failures cached too) and fall back to
+     * DEFAULT_CONTEXT_WINDOW when auto-detection is disabled or unavailable.
+     */
+    windowFor(agent: Agent): Promise<AcpWindow>;
     /** ACP is model-driven: automatic pressure policy never summarizes by itself. */
     compactIfNeeded(_agent: CompactionAgentContext, _trigger: CompactionTrigger, signal: AbortSignal): Promise<CompactionResult | null>;
     /** Explicit idle-session compaction: ACP leaves the decision to the model. */