opencode-acp 1.10.1 → 1.11.0

package/README.md

@@ -422,7 +422,45 @@ For the complete list with root cause analysis, see the [bug tracker](https://gi
 
 ## Changelog
 
-### v1.10.1 — Over-Compression Fix & More Compression Candidates
+### v1.11.0 — Tool-Result Recap Injection, Context Breakdown & Fork Rebuild
+
+This release fixes two critical compression-injection bugs (#20 echo, #78 drift), adds a visible context breakdown to `acp_status`, and introduces a fork-rebuild mechanism.
+
+#### Tool-Result Recap Injection — Fixes #20 & #78 (PR #95)
+
+**Problem**: Compression summaries were injected as text-based `role:assistant` or `role:user` messages. Both roles misled the model:
+- `role:assistant` (Bug 37 path) → model treated summaries as its own prior output and echoed them verbatim (#20, GLM-5.2).
+- `role:user` (Bug 36 merge path) → model treated summaries as user instructions and chased old topics (#78, gpt-5.5).
+
+**Fix**: Summaries are now injected as synthetic **tool-call + tool-result** pairs via `acp_context_recap`. At the API level, the model sees `role:"tool"` — a neutral third role meaning "data from a tool", neither instruction nor own voice. This eliminates both echo (#20) and drift (#78) without breaking prefix caching (mid-stream injection, system prompt unchanged) and works across all providers (OpenAI `role:"tool"`, Anthropic `tool_result` content type).
+
+Files: `lib/messages/utils.ts` (`createSyntheticToolRecap`), `lib/messages/prune.ts` (single-path injection), `lib/messages/query.ts` (`isSyntheticMessage` recognizes `msg_acp_recap_` prefix), `lib/prompts/system.ts` (updated tool description). Removed dead code: `prependCompressionSummary`, `MERGED_SUMMARY_HEADER/FOOTER`, `[ACP system message]` notification wrapper.
+
+Tests: 600 total (10 updated for tool-result format + 1 new multi-block test). TypeScript: 0 errors. Dual-agent reviewed.
+
+#### acp_status Visible Context Breakdown (PR #91)
+
+`acp_status` now shows a per-category token breakdown (tool/code/text/summaries) with largest-item identification. Added drilldown params: `scope:"uncompressed"` with optional `tool:"bash"` filter and `sort:"size"`. The nudge injection was simplified — removed mini breakdown and Top blocks, replaced with a cleaner per-tool-type breakdown. Fixed tool-type detection to only count `type:"tool"` parts (skip step-start/step-finish). Added a dynamic drill-down hint that shows the session's actual top tool.
+
+Files: `lib/compress/status.ts`, `lib/messages/inject/inject.ts`, `lib/messages/inject/utils.ts`, `tests/acp-status.test.ts`.
+
+#### Fork Rebuild & Prune Tool (PR #90)
+
+Added `lib/state/rebuild.ts` — rebuilds compression state after a session fork to prevent context overflow. Added `lib/compress/prune-tool.ts` — a standalone `prune` tool that removes old tool outputs by type (`toolType` param), separate from the `compress` tool for safety.
+
+Files: `lib/state/rebuild.ts`, `lib/compress/prune-tool.ts`, `lib/compress/index.ts`, `index.ts`, `tests/rebuild.test.ts`.
+
+#### Remove todowrite/todoread from compress.protectedTools defaults (PR #87)
+
+`todowrite` and `todoread` were in the `compress.protectedTools` default list, which prevented them from being compressed. Removed so that old todowrite states can be pruned normally.
+
+---
+
+### v1.10.2 — Protected Tools Default Update (PR #87)
+
+Removed `todowrite` and `todoread` from the `compress.protectedTools` default configuration. These tools' outputs accumulate over long sessions and should be eligible for compression like other tool outputs. Users who want to keep them protected can set `compress.protectedTools: ["todowrite", "todoread"]` in their config.
+
+---
 
 Fixes the over-compression bug reported in issue #18 and GitHub #85, where the `toolOutputReminder` nudge bypassed the adaptive 5%-of-context growth protection and fired ~10x too often on large-context models.
 

package/README.zh-CN.md

@@ -395,7 +395,37 @@ ACP 在首次启动时自动将配置从 `dcp.jsonc` 迁移到 `acp.jsonc`，将
 
 ## 更新日志
 
-### v1.10.0 — 受保护工具硬排除、压缩提示词重写、后缀与部署修复
+### v1.11.0 — 工具结果注入、上下文分解 & Fork 重建
+
+本次发布修复了两个关键压缩注入 bug（#20 复读、#78 漂移），为 `acp_status` 添加了可见上下文分解，并引入了 fork 重建机制。
+
+#### 工具结果注入 — 修复 #20 & #78（PR #95）
+
+**问题**：压缩摘要以文本形式的 `role:assistant` 或 `role:user` 消息注入。两种角色都会误导模型：
+- `role:assistant`（Bug 37 路径）→ 模型将摘要当作自己的前文，逐字复读（#20，GLM-5.2）。
+- `role:user`（Bug 36 合并路径）→ 模型将摘要当作用户指令，去执行旧话题（#78，gpt-5.5）。
+
+**修复**：摘要现在以合成的 **tool-call + tool-result** 对注入（`acp_context_recap`）。在 API 层面，模型看到的是 `role:"tool"` —— 一个中立角色，表示「工具返回的数据」，既不是指令也不是自己的前文。这同时消除了复读（#20）和漂移（#78），不破坏前缀缓存（mid-stream 注入，system prompt 不变），且跨所有 provider 兼容。
+
+#### acp_status 可见上下文分解（PR #91）
+
+`acp_status` 现在显示按类别（tool/code/text/summaries）的 token 分解，并标识最大项。新增下钻参数：`scope:"uncompressed"` + 可选 `tool:"bash"` 过滤和 `sort:"size"`。简化了 nudge 注入 —— 移除了 mini breakdown 和 Top blocks，替换为更清晰的按工具类型分解。
+
+#### Fork 重建 & Prune 工具（PR #90）
+
+新增 `lib/state/rebuild.ts` —— 在 session fork 后重建压缩状态以防止上下文溢出。新增 `lib/compress/prune-tool.ts` —— 独立的 `prune` 工具，按类型（`toolType` 参数）移除旧工具输出，与 `compress` 工具分离以提高安全性。
+
+#### 移除 todowrite/todoread 的默认保护（PR #87）
+
+从 `compress.protectedTools` 默认配置中移除了 `todowrite` 和 `todoread`，使旧 todowrite 状态可以正常被压缩。
+
+---
+
+### v1.10.2 — 受保护工具默认配置更新（PR #87）
+
+从 `compress.protectedTools` 默认配置中移除了 `todowrite` 和 `todoread`。这些工具的输出在长会话中累积，应像其他工具输出一样可被压缩。如需保持保护，可在配置中设置 `compress.protectedTools: ["todowrite", "todoread"]`。
+
+---
 
 本次发布合并了 7 个 PR。核心变更是**受保护工具消息硬排除**；其余为同期合入的修复和提示词重写。
 

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;AA2BjD,QAAA,MAAM,MAAM,EAAE,MAqHK,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;AA4BjD,QAAA,MAAM,MAAM,EAAE,MAsHK,CAAA;AAEnB,eAAe,MAAM,CAAA"}