@lexwdex-org/opencode-dcp 3.4.14 → 4.0.0

package/README.en.md

@@ -4,39 +4,41 @@
 
 **English** | [中文](./README.md)
 
-DCP supplies a semantic-pruning policy and proactive triggers to OpenCode's native compaction lifecycle. It does not own message markers, message IDs, compression blocks, anchors, or placeholders.
+DCP manages OpenCode session context with **Dynamic Tiered Compression (DTC)**: folding happens inside every model request, before message serialization, and **never touches the session state machine** — no `session.summarize` calls, no compaction turns, no session writes. Compression is fully transparent to continuous autonomous work: an agent can run for days while the engine absorbs context pressure per request.
 
 ## How it works
 
-OpenCode owns one rolling checkpoint. After compaction, the model sees only:
+The host fires `experimental.chat.messages.transform` before every model request. DCP folds the **request-scoped message copy** inside that hook (the host rebuilds the array from the database on every loop iteration, so all mutations are inherently request-scoped):
 
 ```text
-latest pruned checkpoint + uncompacted recent tail
+┌──────────────┬──────────────┬──────────────┬────────────────┐
+│ D distant    │ M middle     │ C current    │ T protected    │
+│ heavy fold   │ medium fold  │ light fold   │ last 4 turns   │
+│ digest lines │ first lines  │ truncate big │ never touched  │
+└──────────────┴──────────────┴──────────────┴────────────────┘
+        ◄── escalation deepens from the far end under budget pressure ──►
 ```
 
-During `experimental.session.compacting`, DCP asks the summarizer to organize the checkpoint into a fixed three-tier structure:
+- **T tail zone**: the last `dtc.tailTurns` (default 4) conversation turns are **never folded**.
+- **C current-task zone**: turns since the latest topic boundary (lexical drift detection); only oversized tool outputs are head+tail truncated, task details survive.
+- **M middle zone**: long texts keep their first line; tool outputs get the host's native fold marker (rendered by the host as its own `[Old tool result content cleared]`); reasoning is emptied.
+- **D distant zone**: whole turns collapse into one mechanical digest line (intent / actions / files touched / outcome / error count) and tool inputs are cleared — hard facts live on in the digest.
 
-- `## 系统上下文` (System context) — system-level rules such as AGENTS.md, carried forward verbatim from the previous checkpoint;
-- `## 历史概要` (History digest) — early and middle history heavily compressed: one-line outcomes only, no process;
-- Recent tasks — lightly compressed: completed tasks summarized in one line each; in-progress tasks keep full detail (goal, finished steps, file paths, key decisions, blockers, next actions). Details related to the current task are preserved first so work can continue from the checkpoint alone.
+**Dynamic budget**: below `lowWatermarkRatio` of the context window (default 50%) **nothing is folded at all** — short sessions pay zero cost. Above it, folding escalates D→M→C from the far end until the estimate fits `targetRatio` (default 70%). The window size is learned per session from the `chat.params` hook; until it is known the engine fails open (no folding).
 
-Unrelated chat, other-project context, tool retries, and repeated edits are folded away as before.
+**Three structural laws** (the constructive exclusion of the old versions' "lost markers" failure):
 
-## Trigger compaction
+1. Messages and parts are never added, removed, or reordered; IDs never change — tool-call/tool-result pairing cannot break;
+2. Only string payloads are rewritten (text/output/reasoning), and tool-output folding uses the host's native `time.compacted` marker instead of a custom placeholder protocol;
+3. Every mutation lives in the request-scoped copy only — session history in the database stays byte-identical.
 
-- **OpenCode automatic compaction** uses the policy automatically.
-- **The `dcp_prune` model tool** — DCP registers an LLM-invokable tool whose description carries heuristic guidance: call it immediately when the topic clearly changes, when a task wraps up, or when the context has grown noticeably.
-- **Heuristic auto-prune** (`autoPrune`) — observes incoming user messages and triggers native compaction at turn boundaries (`session.idle`) on:
-    - topic drift: lexical similarity between the newest message and recent history drops sharply (CJK bigrams + Jaccard);
-    - message volume since the last prune crossing a threshold;
-    - resuming after a long idle gap.
-      Auto-triggers respect a cooldown; native `/compact` or host compaction also resets the counters.
-- OpenCode's native `/compact` command uses the same path.
-- `/dcp summarize` calls native `session.summarize()` manually.
+## Trigger surfaces
 
-All entries share one coordinator: concurrent requests for the same session share a single native call. A failed call is fail-open and keeps the original context; retries are cooled down for 30 seconds by default.
-
-Pruned prefixes stop being sent to the model, but the original session history is not physically deleted from OpenCode's database. A later compaction replaces the previous checkpoint with one merged checkpoint instead of nesting summaries or building a compression-block graph.
+- **Automatic**: there is nothing to trigger — every request decides its own fold depth from the budget. No signals, no boundaries, no queue.
+- **The `dcp_prune` model tool**: returns instantly. Marks a topic boundary (the current-task zone restarts at the next turn) and raises this session's minimum fold level to M — old-task content folds from the next request onward. Never interrupts the running turn.
+- **`/dcp fold`**: manual variant; raises the minimum fold level to the deepest tier.
+- **`/dcp status`**: shows turns, token estimate, context window, and the manual fold level for the session.
+- **The host's own compaction** (`/compact`, context-overflow fallback): unaffected by DTC and still fully functional. DCP continues to supply the high-quality semantic checkpoint prompt (rolling merge of the previous checkpoint, tiered compression depth, fixed four-section structure) and raises the host's tail protection defaults to 4 turns / 32000 tokens (`compaction.tail_turns` / `preserve_recent_tokens`; explicit user config always wins). When the compaction input passes through the transform hook, DTC skips it via a one-shot flag so the summarizer sees full-fidelity content.
 
 ## Install
 
@@ -46,7 +48,11 @@ opencode plugin @lexwdex-org/opencode-dcp@latest --global
 
 ## Configuration
 
-DCP loads global, custom-config-directory, then project configuration:
+DCP reads these layers in order; later layers override earlier ones:
+
+1. `~/.config/opencode/dcp.jsonc` or `dcp.json`
+2. `$OPENCODE_CONFIG_DIR/dcp.jsonc` or `dcp.json`
+3. Project `.opencode/dcp.jsonc` or `dcp.json`
 
 ```jsonc
 {
@@ -54,40 +60,59 @@ DCP loads global, custom-config-directory, then project configuration:
     "enabled": true,
     "autoUpdate": true,
     "debug": false,
-    "language": "zh",
-    "commands": { "enabled": true },
-    "experimental": { "customPrompts": false },
-    "summarize": { "failureCooldownMs": 30000 },
-    "tool": { "enabled": true },
-    "autoPrune": {
+    "language": "en",
+    "commands": {
+        "enabled": true,
+    },
+    "experimental": {
+        "customPrompts": false,
+    },
+    "dtc": {
         "enabled": true,
-        "minMessages": 8,
-        "volumeThreshold": 30,
+        "tailTurns": 4,
+        "lowWatermarkRatio": 0.5,
+        "targetRatio": 0.7,
         "driftThreshold": 0.18,
-        "idleGapMs": 1800000,
-        "cooldownMs": 300000,
+        "toolOutputKeepChars": 4000,
+    },
+    "tool": {
+        "enabled": true,
     },
 }
 ```
 
-| Key                         | Meaning                                                                                       |
-| --------------------------- | --------------------------------------------------------------------------------------------- |
-| `language`                  | Language of the bundled compaction prompt: `zh` (default) / `en`; custom overrides always win |
-| `tool.enabled`              | Register the model-invokable `dcp_prune` tool with heuristic usage guidance                   |
-| `autoPrune.enabled`         | Enable plugin-side heuristic auto-compaction                                                  |
-| `autoPrune.minMessages`     | Minimum user messages before any auto-prune signal is considered                              |
-| `autoPrune.volumeThreshold` | User messages since the last prune that trigger compaction by volume                          |
-| `autoPrune.driftThreshold`  | Jaccard similarity below which consecutive messages count as a topic change (0–1)             |
-| `autoPrune.idleGapMs`       | Gap between user messages that counts as resuming after a long break                          |
-| `autoPrune.cooldownMs`      | Minimum interval between two automatic prunes of the same session                             |
+| Key                       | Description                                                                                                |
+| ------------------------- | ---------------------------------------------------------------------------------------------------------- |
+| `language`                | Bundled compaction prompt language: `zh` (default) / `en`; custom overrides always win                     |
+| `dtc.enabled`             | Master switch for dynamic tiered request-time compression                                                  |
+| `dtc.tailTurns`           | Turns at the end that are never folded (default 4)                                                         |
+| `dtc.lowWatermarkRatio`   | No folding below window × ratio (default 0.5)                                                              |
+| `dtc.targetRatio`         | Escalate folding until the estimate fits window × ratio (default 0.7)                                      |
+| `dtc.driftThreshold`      | Jaccard similarity below which consecutive user messages start a new current-task zone (0–1, default 0.18) |
+| `dtc.toolOutputKeepChars` | Head+tail characters kept for oversized current-zone tool outputs (default 4000)                           |
+| `tool.enabled`            | Register the model-invokable `dcp_prune` tool (instant boundary mark, never interrupts)                    |
+
+With `experimental.customPrompts` enabled, copy the generated
+`~/.config/opencode/dcp-prompts/defaults/compaction.md` to any override location:
+
+- Project: `.opencode/dcp-prompts/overrides/compaction.md`
+- Custom config dir: `$OPENCODE_CONFIG_DIR/dcp-prompts/overrides/compaction.md`
+- Global: `~/.config/opencode/dcp-prompts/overrides/compaction.md`
 
-When `experimental.customPrompts` is enabled, copy the generated
-`~/.config/opencode/dcp-prompts/defaults/compaction.md` to a project, custom-config, or global `dcp-prompts/overrides/compaction.md` path.
+## Migrating from 3.x
 
-## Migrating from the legacy 3.x pipeline
+The 3.x `summarize` and `autoPrune` config blocks are gone (DCP no longer calls native summarize and has no heuristic triggers); such keys produce a migration warning and are ignored. `autoPrune.driftThreshold` migrates automatically to `dtc.driftThreshold`. Older `compress`, `manualMode`, `strategies`, `turnProtection` keys remain removed.
 
-Legacy `compress`, `manualMode`, `strategies`, `turnProtection`, notification, protected-file, and command-protection settings are removed. DCP reports them as deprecated and ignores them.
+Behavior change: compression no longer produces a visible "checkpoint turn" and needs no at-rest boundaries or continuation machinery — context folds automatically inside each request, invisible to the session and its state machine. Semantic checkpoints are still produced by the host's own compaction (manual `/compact` or the overflow fallback), for which DCP continues to provide the prompt and tail-protection defaults.
 
-The old model tools, range/message compression, `/dcp compress`, decompression/recompression/sweep commands, message markers, and plugin compression persistence are gone. Existing legacy DCP state files are neither read nor modified; remove them manually only after deciding that you will not roll back.
+## Development
+
+```bash
+npm test
+npm run typecheck
+npm run build
+npm run format:check
+npm run check:package
+```
 
 License: AGPL-3.0-or-later.

package/README.md

@@ -4,37 +4,41 @@
 
 [**English**](./README.en.md) | **中文**
 
-DCP 为 OpenCode 原生 compaction 提供语义剪枝策略与主动触发机制。它不维护消息标记、消息 ID、压缩块、锚点或占位符。
+DCP 以**动态分级请求期压缩（DTC）**管理 OpenCode 会话上下文：折叠发生在每次模型请求的消息序列化之前，**永不触碰会话状态机**——不调用 `session.summarize`、不产生压缩回合、不写会话存储。压缩对连续自动工作完全透明：agent 可以不间断运行数日，上下文压力由引擎在请求期自动消化。
 
 ## 工作原理
 
-OpenCode 原生 compaction 负责生成和保存一个滚动检查点，并在后续模型请求中只发送：
+宿主在每次模型请求前触发 `experimental.chat.messages.transform`，DCP 在钩子内对**请求期消息副本**做分级折叠（宿主每轮循环从数据库重建该数组，因此所有修改天然只作用于本次请求）：
 
 ```text
-最新剪枝检查点 + 尚未压缩的近期尾部
+┌────────────┬────────────┬────────────┬──────────────┐
+│ D 远距离区 │ M 中距离区 │ C 当前任务 │ T 尾部保护区 │
+│ 重度折叠   │ 中度折叠   │ 轻度折叠   │ 最后 4 轮    │
+│ 结构化摘要 │ 首行+标记  │ 截长输出   │ 一个字节不动 │
+└────────────┴────────────┴────────────┴──────────────┘
+        ◄── 预算不足时从最远端逐级加深 ──►
 ```
 
-DCP 在 `experimental.session.compacting` 阶段提供专用提示词，把检查点组织为固定三段结构：
+- **T 尾部保护区**：最后 `dtc.tailTurns`（默认 4）轮对话**完全不折叠**。
+- **C 当前任务区**：自最近话题边界（词面漂移检测）以来的轮次，只把超长 tool 输出做首尾截断，任务细节不丢。
+- **M 中距离区**：长文本保留首行；tool 输出打宿主原生折叠标记（渲染为宿主自己的 `[Old tool result content cleared]`）；推理内容清空。
+- **D 远距离区**：整轮塌缩为一行机械摘要（意图 / 动作 / 涉及文件 / 结果 / 错误数），tool 参数一并清空——硬事实保留在摘要里。
 
-- `## 系统上下文` — AGENTS.md、项目规则等系统级内容，从上一份检查点原样保留合并；
-- `## 历史概要` — 早期历史和中部历史高度压缩：每项主题只留一句结论，不含过程；
-- 最近任务 — 轻度压缩：已完成任务一句话概括；进行中任务保留完整详情（目标、已完成步骤、文件路径、关键决策、阻塞、下一步）。与当前任务相关的近期细节优先保留，保证凭检查点即可直接继续工作。
+**动态预算**：估算低于上下文窗口的 `lowWatermarkRatio`（默认 50%）时**完全不折叠**，短会话零开销；超出则按 D→M→C 从远到近逐级加深，直到估算落回 `targetRatio`（默认 70%）以内。窗口大小由 `chat.params` 钩子按会话自动学习，未知时 fail-open 不折叠。
 
-## 触发压缩
+**三条结构铁律**（老版本"丢失标记号"问题的构造性排除）：
 
-- **OpenCode 自动 compaction**：达到宿主阈值时自动运行。
-- **模型调用 `dcp_prune` 工具**：插件注册的 LLM 工具，描述内置启发式规则——话题明显变更、任务收尾、上下文明显变长时立即调用。
-- **插件启发式自动触发**（`autoPrune`）：监听用户消息流，在回合边界（`session.idle`）触发原生压缩：
-    - 话题变更：新消息与近期消息的词面相似度骤降（CJK 二元组 + Jaccard）；
-    - 消息量达到阈值；
-    - 长时间中断后恢复。
-      自动触发带冷却期；OpenCode 原生 `/compact` 或宿主压缩也会重置计数。
-- **OpenCode 原生 `/compact`**：手动触发同一条 compaction 路径。
-- **`/dcp summarize`**：手动调用原生 `session.summarize()`。
+1. 永不增删、重排消息或 part；永不改 ID——tool-call 与 tool-result 的配对不可能断裂；
+2. 只重写字符串载荷（text/output/reasoning），折叠 tool 输出用宿主原生 `time.compacted` 标记，不自造占位符协议；
+3. 一切修改仅存在于请求期副本，数据库中的会话历史逐字节不变。
 
-所有入口都经过同一协调器：同一会话的并发请求合并为一次原生调用；失败不会提交半成品检查点，原始上下文保持可用，默认 30 秒内不重试失败的请求。
+## 触发面
 
-压缩后的旧前缀不会再发送给模型，但 OpenCode 数据库中的原始会话历史不会被物理删除。下一次压缩会把旧检查点和新增尾部合并为一个新检查点，不会形成嵌套摘要。
+- **自动**：无触发概念——每次请求按预算自动决定折叠深度，无需任何信号或边界。
+- **模型调用 `dcp_prune` 工具**：瞬时返回。标记话题边界（当前任务区从下一轮重算）并把本会话折叠下限提到 M 级——旧任务内容从下一次请求起被折叠。绝不打断当前工作。
+- **`/dcp fold`**：手动版本，直接把折叠下限提到最深档。
+- **`/dcp status`**：查看当前会话的轮数、token 估算、窗口与降级档位。
+- **宿主自身的 compaction**（`/compact`、上下文溢出兜底）：不受 DTC 影响照常工作；DCP 仍为其提供高质量语义检查点提示词（滚动合并上一份检查点、三层压缩深度、四段固定结构），并把宿主尾部保护默认提升到 4 轮 / 32000 tokens（`compaction.tail_turns` / `preserve_recent_tokens`，用户显式配置优先）。压缩输入经过 transform 钩子时 DTC 自动跳过，保证摘要器看到全保真内容。
 
 ## 安装
 
@@ -63,33 +67,30 @@ DCP 依次读取以下配置，后面的层覆盖前面的层：
     "experimental": {
         "customPrompts": false,
     },
-    "summarize": {
-        "failureCooldownMs": 30000,
-    },
-    "tool": {
+    "dtc": {
         "enabled": true,
+        "tailTurns": 4,
+        "lowWatermarkRatio": 0.5,
+        "targetRatio": 0.7,
+        "driftThreshold": 0.18,
+        "toolOutputKeepChars": 4000,
     },
-    "autoPrune": {
+    "tool": {
         "enabled": true,
-        "minMessages": 8,
-        "volumeThreshold": 30,
-        "driftThreshold": 0.18,
-        "idleGapMs": 1800000,
-        "cooldownMs": 300000,
     },
 }
 ```
 
-| 键                          | 说明                                                                      |
-| --------------------------- | ------------------------------------------------------------------------- |
-| `language`                  | 内置压缩提示词语言：`zh`（默认）/ `en`；自定义覆盖文件优先于该配置        |
-| `tool.enabled`              | 注册模型可调用的 `dcp_prune` 工具（描述含"话题变更立即使用"的启发式指令） |
-| `autoPrune.enabled`         | 启用插件侧启发式自动压缩                                                  |
-| `autoPrune.minMessages`     | 会话用户消息数达到该值前不做任何自动判定                                  |
-| `autoPrune.volumeThreshold` | 距上次压缩的用户消息量阈值                                                |
-| `autoPrune.driftThreshold`  | Jaccard 相似度低于该值视为话题变更（0–1）                                 |
-| `autoPrune.idleGapMs`       | 用户消息间隔超过该值视为长时间中断后恢复                                  |
-| `autoPrune.cooldownMs`      | 同一会话两次自动压缩的最小间隔                                            |
+| 键                        | 说明                                                                     |
+| ------------------------- | ------------------------------------------------------------------------ |
+| `language`                | 内置压缩提示词语言：`zh`（默认）/ `en`；自定义覆盖文件优先于该配置       |
+| `dtc.enabled`             | 动态分级请求期压缩总开关                                                 |
+| `dtc.tailTurns`           | 尾部保护轮数：最后 N 轮完全不折叠（默认 4）                              |
+| `dtc.lowWatermarkRatio`   | 估算低于窗口 × 该比例时完全不折叠（默认 0.5）                            |
+| `dtc.targetRatio`         | 逐级加深折叠直到估算 ≤ 窗口 × 该比例（默认 0.7）                         |
+| `dtc.driftThreshold`      | Jaccard 相似度低于该值视为话题变更，开启新的当前任务区（0–1，默认 0.18） |
+| `dtc.toolOutputKeepChars` | 当前任务区超长 tool 输出的首尾截断保留字符数（默认 4000）                |
+| `tool.enabled`            | 注册模型可调用的 `dcp_prune` 工具（瞬时标记话题边界，绝不打断工作）      |
 
 启用 `experimental.customPrompts` 后，可将生成的
 `~/.config/opencode/dcp-prompts/defaults/compaction.md` 复制到以下任一覆盖位置：
@@ -98,13 +99,11 @@ DCP 依次读取以下配置，后面的层覆盖前面的层：
 - 自定义配置目录：`$OPENCODE_CONFIG_DIR/dcp-prompts/overrides/compaction.md`
 - 全局：`~/.config/opencode/dcp-prompts/overrides/compaction.md`
 
-## 从 3.x 旧压缩管线迁移
+## 从 3.x 迁移
 
-旧配置 `compress`、`manualMode`、`strategies`、`turnProtection`、
-`pruneNotification`、`protectedFilePatterns` 及旧命令保护字段均已删除。DCP 会显示迁移警告并忽略这些字段，不会静默继续旧行为。
+3.x 的 `summarize`、`autoPrune` 配置块已删除（DCP 不再调用原生 summarize，也不再有启发式触发器）；这些键会显示迁移警告并被忽略。`autoPrune.driftThreshold` 自动迁移为 `dtc.driftThreshold`。更早的 `compress`、`manualMode`、`strategies`、`turnProtection` 等键维持删除状态。
 
-以下功能已删除：`compress` / `compress_range` 模型工具、`/dcp compress`、
-`/dcp decompress`、`/dcp recompress`、`/dcp sweep`、消息标记和插件压缩状态持久化。磁盘上已有的旧 DCP 会话状态文件不会被读取、改写或删除，可在确认不再回退旧版本后自行清理。
+行为变化：压缩不再产生可见的"检查点回合"，也不再需要任何静息边界或续跑机制——上下文折叠在每次请求内自动完成，会话与状态机零感知。语义检查点仍由宿主自身的 compaction（手动 `/compact` 或溢出兜底）产生，DCP 继续为其提供提示词与尾部保护默认值。
 
 ## 开发验证
 

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,qBAAqB,CAAA;AAiBjD,QAAA,MAAM,MAAM,EAAE,MA4DK,CAAA;AAEnB,eAAe,MAAM,CAAA"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,qBAAqB,CAAA;AAgBjD,QAAA,MAAM,MAAM,EAAE,MA0DK,CAAA;AAEnB,eAAe,MAAM,CAAA"}