ranxianglei · ranxianglei · Jun 29, 2026 · Jun 28, 2026 · Jun 29, 2026 · Jun 29, 2026
diff --git a/README.md b/README.md
@@ -82,18 +82,20 @@ Or add to your opencode config:
 
 ACP hands the context-compression tool directly to the model. The model is
 **100% responsible** for context compression. The model's available tools are
-mainly: **compress**, **decompress**, and **delete** (`mark_block` / `unmark_block`).
+mainly: **compress** and **decompress**. A hardcoded 100% GC fallback acts as
+a safety net when the context window is completely full.
 
 ### Lifecycle
 
-Three operations: **compress**, **decompress**, and **delete**. Content loops
-between raw and compressed, and eventually terminates in deletion:
+Two operations: **compress** and **decompress**. Content loops between raw and
+compressed. When context hits 100%, old-gen block summaries are truncated as
+a last resort:
 
 ```mermaid
 stateDiagram-v2
     Raw --> Compressed : compress
     Compressed --> Raw : decompress
-    Compressed --> Deleted : delete
+    Compressed --> Truncated : GC at 100%
 ```
 
 ### Compression strategy
@@ -305,7 +307,7 @@ Each level overrides the previous, so project settings take priority over global
             "protectedTools": [],
         },
     },
-    // Garbage collection and batch cleanup
+    // Garbage collection — hardcoded 100% fallback only
     "gc": {
         "algorithm": "truncate",
         // young → old generation promotion after this many survivals
@@ -314,18 +316,8 @@ Each level overrides the previous, so project settings take priority over global
         "maxBlockAge": 15,
         // truncate old-gen summaries exceeding this length (chars)
         "maxOldGenSummaryLength": 3000,
-        // run major GC when context usage exceeds this
+        // run major GC when context usage exceeds this (hardcoded, not configurable)
         "majorGcThresholdPercent": "100%",
-        // Three-tier batch merge-cleanup for blocks flagged via mark_block.
-        // Accepts a number or "X%" of the model context window.
-        "batchCleanup": {
-            // At/above this usage, remind the model about marked blocks
-            "lowThreshold": "60%",
-            // At/above this usage, auto merge-compress all marked blocks into one
-            "highThreshold": "75%",
-            // At/above this usage, force-merge all old-gen blocks (before GC)
-            "forceThreshold": "90%",
-        },
     },
 }
 ```
@@ -354,7 +346,7 @@ To reset an override, delete the matching file from your overrides directory.
 ### Protected Tools
 
 By default, these tools are always protected from pruning:
-`task`, `skill`, `todowrite`, `todoread`, `compress`, `decompress`, `mark_block`, `unmark_block`, `batch`, `plan_enter`, `plan_exit`, `write`, `edit`
+`task`, `skill`, `todowrite`, `todoread`, `compress`, `decompress`, `batch`, `plan_enter`, `plan_exit`, `write`, `edit`
 
 The `protectedTools` arrays in `commands` and `strategies` add to this default list.
 

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -73,17 +73,17 @@ opencode plugin opencode-acp@latest --global
 
 ## 工作原理
 
-ACP 把上下文压缩工具直接交给模型。模型对上下文压缩**负全责**。模型可用的工具主要是：**compress**、**decompress** 和 **delete**（`mark_block` / `unmark_block`）。
+ACP 把上下文压缩工具直接交给模型。模型对上下文压缩**负全责**。模型可用的工具主要是：**compress** 和 **decompress**。当上下文达到 100% 时，系统自动触发 GC 截断作为兜底。
 
 ### 生命周期
 
-三个操作：**压缩**、**解压缩**、**删除**。内容在原始与压缩之间循环，最终以删除终结：
+两个操作：**压缩**、**解压缩**。内容在原始与压缩之间循环。当上下文达到 100% 时，GC 自动截断老年代 block 作为兜底：
 
 ```mermaid
 stateDiagram-v2
     Raw --> Compressed : compress
     Compressed --> Raw : decompress
-    Compressed --> Deleted : delete
+    Compressed --> GC_Truncated : GC (100%)
 ```
 
 ### 压缩策略
@@ -104,9 +104,9 @@ stateDiagram-v2
 
 由模型决定何时解压。当上下文大到足以干扰模型的 self-attention 时，简短的 block 会让模型先压缩一部分内容，处理完紧急事务，再在后续工作中按需解压。
 
-### 删除策略
+### GC 兜底
 
-为了应对大量小块历史内容的堆积，新版本增加了删除策略。由模型决定是否删除。**一旦删除，内容不可恢复。** 这取代了原先的强制 GC，使得强制垃圾回收不再删除模型认为重要的内容。
+当上下文达到 100% 时，系统自动截断老年代 block 摘要，防止上下文溢出。这是最后的兜底机制，不影响模型的正常压缩/解压操作。
 
 ---
 
@@ -289,18 +289,8 @@ ACP 使用自己的配置文件，按以下顺序搜索：
         "maxBlockAge": 15,
         // 截断超过此长度（字符）的老年代摘要
         "maxOldGenSummaryLength": 3000,
-        // 上下文使用率超过此值时执行主 GC
+        // 上下文使用率超过此值时执行主 GC（兜底，硬编码为 100%）
         "majorGcThresholdPercent": "100%",
-        // 通过 mark_block 标记的块的三级批量合并清理阈值。
-        // 接受数字或 "X%"（模型上下文窗口的百分比）。
-        "batchCleanup": {
-            // 达到此使用率时，提醒模型已标记的块
-            "lowThreshold": "60%",
-            // 达到此使用率时，自动将所有已标记块合并压缩为一个
-            "highThreshold": "75%",
-            // 达到此使用率时，强制合并所有老年代块（GC 之前）
-            "forceThreshold": "90%",
-        },
     },
 }
 ```
@@ -329,7 +319,7 @@ ACP 暴露六个可编辑的 prompt：
 ### 受保护工具
 
 默认情况下，以下工具始终受保护不被剪枝：
-`task`、`skill`、`todowrite`、`todoread`、`compress`、`decompress`、`mark_block`、`unmark_block`、`batch`、`plan_enter`、`plan_exit`、`write`、`edit`
+`task`、`skill`、`todowrite`、`todoread`、`compress`、`decompress`、`batch`、`plan_enter`、`plan_exit`、`write`、`edit`
 
 `commands` 和 `strategies` 中的 `protectedTools` 数组会添加到此默认列表。
 

diff --git a/devlog/2026-06-29_context-optimization/REQ.md b/devlog/2026-06-29_context-optimization/REQ.md
@@ -0,0 +1,31 @@
+# Context Optimization — Reduce Token Waste
+
+## Problem
+
+Session ses_102504697ffeYg89Sn0k8aknYg grew to 47% context usage. Root cause analysis revealed systematic token waste:
+
+1. **Compress summaries too verbose**: avg 579 chars (~145 tokens), some up to 2011 chars. Include unnecessary metrics, reviewer quotes, experimental parameters.
+2. **Compress tool calls are pure overhead**: 344 calls × 813 chars avg = 280K chars. Each stores full summary in input — duplicated with block summary.
+3. **Step markers waste space**: 4698 step-start/step-finish parts × ~88 chars avg = 413K chars (~103K tokens). Only mark boundaries, no useful content.
+4. **Large tool outputs not compressed**: Model keeps 20-50K char outputs "just in case".
+5. **No minimum compress range**: Model compresses tiny ranges (<2K chars) where overhead exceeds savings.
+6. **ACP guidance too verbose**: Multi-paragraph nudge text wastes ~200 tokens/turn.
+
+## Requirements
+
+1. **R1**: Limit compress summary length to configurable max (default 100 chars). Reject if exceeded.
+2. **R2**: ~~Truncate compress tool input after execution~~ — NOT FEASIBLE (no API to modify stored parts).
+3. **R3**: Strengthen nudge to target large tool outputs (>5K chars) explicitly.
+4. **R5**: Truncate step markers in context construction (skip step-start, truncate step-finish to 50 chars).
+5. **R6**: Shorten ACP guidance text (pressure levels + per-message guidance).
+6. **R7**: Enforce minimum compress range (default 2000 chars). Reject if below.
+
+## Cache Safety
+
+All fixes are either cache-neutral (only affect future operations) or one-time breaks that stabilize after deployment. No recurring cache breaks.
+
+## Non-Goals
+
+- Excluding old reasoning from context (causes recurring cache breaks — cancelled).
+- Modifying block ID list (accuracy risk — kept as-is).
+- compress tool input cleanup (not feasible with current API).
diff --git a/devlog/2026-06-29_context-optimization/WORKLOG.md b/devlog/2026-06-29_context-optimization/WORKLOG.md
@@ -0,0 +1,47 @@
+# Worklog — Context Optimization
+
+## Changes (8 files, +186/-8 lines)
+
+### Fix 1: Summary length limit (R1)
+- **config.ts**: Added `maxSummaryLength` (default 100) to CompressConfig
+- **config-validation.ts**: Type + key validation
+- **compress/message.ts, compress/range.ts**: Check `summary.length > maxSummaryLength` → throw error before creating block
+
+### Fix 2: Compress tool cleanup (R2) — NOT FEASIBLE
+- ToolContext API only allows modifying output/title/metadata, NOT input args
+- Added TODO comments in both handlers noting `experimental.chat.messages.transform` as alternative
+- Documented for future investigation
+
+### Fix 3: Nudge strengthening (R3)
+- **inject/utils.ts**: Guidance text now explicitly mentions ">5000 characters" tool outputs
+- Changed from generic "compress tool outputs" to targeted "if any tool output >5000 chars and you've finished reading, compress it into a summary NOW"
+
+### Fix 5: Step marker truncation (R5)
+- **prune.ts**: New `stripStepMarkers()` function
+  - Skips `step-start` parts entirely (zero-value boundary markers)
+  - Truncates `step-finish` reason to 50 chars (was avg 155 chars)
+  - Called from `prune()` before context injection
+- Estimated savings: ~90K tokens per session with heavy reasoning
+
+### Fix 6: ACP simplification (R6)
+- **system.ts**: Pressure level descriptions shortened to 1 sentence each
+  - Normal: "Be frugal — compress tool outputs you've finished using into summaries."
+  - Elevated: "Context is growing — compress larger ranges you no longer need."
+  - Critical: "Compress aggressively now — target the largest visible ranges first."
+- **inject/utils.ts**: Per-message guidance reduced from 5+ to 3 sentences
+- Block ID list: UNCHANGED (accuracy requirement)
+
+### Fix 7: Minimum compress range (R7)
+- **config.ts**: Added `minCompressRange` (default 2000) to CompressConfig
+- **config-validation.ts**: Type + key validation
+- **compress/message.ts, compress/range.ts**: Calculate total message chars via `countMessageCharacters()` → throw error if < minCompressRange
+- **token-utils.ts**: New `countMessageCharacters()` helper
+
+## Verification
+- `npm run typecheck`: clean ✅
+- `npm run test`: 487 pass, 0 fail ✅
+- Block ID list: verified unchanged (empty git diff on nudge.ts)
+
+## Not Implemented
+- **Fix 4 (exclude old reasoning)**: Cancelled — causes recurring cache breaks as reasoning crosses age threshold every turn.
+- **Fix 2 (compress input cleanup)**: Not feasible with current OpenCode plugin API. Needs `experimental.chat.messages.transform` hook investigation.
diff --git a/index.ts b/index.ts
@@ -4,8 +4,6 @@ import {
     createCompressMessageTool,
     createCompressRangeTool,
     createDecompressTool,
-    createMarkBlockTool,
-    createUnmarkBlockTool,
 } from "./lib/compress"
 import {
     compressDisabledByOpencode,
@@ -91,8 +89,6 @@ const server: Plugin = (async (ctx) => {
                         ? createCompressMessageTool(compressToolContext)
                         : createCompressRangeTool(compressToolContext),
                 decompress: createDecompressTool(compressToolContext),
-                mark_block: createMarkBlockTool(compressToolContext),
-                unmark_block: createUnmarkBlockTool(compressToolContext),
             }),
         },
         config: async (opencodeConfig) => {
@@ -113,7 +109,7 @@ const server: Plugin = (async (ctx) => {
 
             const toolsToAdd: string[] = []
             if (config.compress.permission !== "deny" && !config.experimental.allowSubAgents) {
-                toolsToAdd.push("compress", "decompress", "mark_block", "unmark_block")
+                toolsToAdd.push("compress", "decompress")
             }
 
             if (toolsToAdd.length > 0) {

diff --git a/lib/compress/index.ts b/lib/compress/index.ts
@@ -2,4 +2,3 @@ export { ToolContext } from "./types"
 export { createCompressMessageTool } from "./message"
 export { createCompressRangeTool } from "./range"
 export { createDecompressTool } from "./decompress"
-export { createMarkBlockTool, createUnmarkBlockTool } from "./mark-block"
diff --git a/lib/compress/mark-block.ts b/lib/compress/mark-block.ts