opencode-acp 1.12.8 → 1.12.10-dev.1

package/README.md

@@ -422,6 +422,22 @@ For the complete list with root cause analysis, see the [bug tracker](https://gi
 
 ## Changelog
 
+### v1.12.10-dev.1 — Decompress Range Mode + Token Classification + Protected Label Accuracy + Nudge Suppression + Discrete Intervals (PRs #73, #155, #157, #158, #159)
+
+**Problem**: Five issues across decompress ergonomics, token accounting, and nudge quality. (1) `decompress` required a per-block `acp_status` → decompress-per-block loop to restore multiple compressed blocks. (2) Since v1.12.9 (compress-as-anchor), compress tool `summary` content was misclassified as `toolTokens` instead of `summaryTokens`, inflating tool% and deflating summary% in the context breakdown — misleading the model's compression decisions. (3) The `[PROTECTED: ...]` label listed every tool in a protected message (e.g. `[PROTECTED: grep, skill]`), misleading the model into thinking non-protected tools like `grep` were the reason for protection. (4) When all visible content was protected, the nudge still fired with an empty recommendation list — wasting context and confusing the model. (5) When a nudge was suppressed (filter removed all recommendations), the next-turn check re-evaluated every turn because the baseline never advanced, causing repeated wasted computation.
+
+**Fix**: (1) **PR #73** — Added optional `startId`/`endId` to `decompress` schema; range mode batch-restores every active block whose `effectiveMessageIds` overlaps the resolved range in one call. New pure helper `findActiveBlocksOverlappingMessages` in `decompress-logic.ts`; backward compatible (`blockId` path unchanged). (2) **PR #155** — In `estimateContextComposition` (`lib/messages/inject/utils.ts`), when `toolName === "compress"`, extract `summary` text from `part.state.input.content[].summary` and classify it as `summaryTokens`. Structural overhead remains `toolTokens`. (3) **PR #157** — `buildCompressibleRanges` now only adds tools that actually trigger protection via `isToolNameProtected` or `isFilePathProtected`, mirroring `messageContainsProtectedTool` exactly — the label now shows `[PROTECTED: skill]` only. (4) **PR #158** — Added `allProtected = compressible.length === 0 && protected.length > 0` check; `nothingToCompress = filterSuppressed || allProtected` gates nudge injection so soft nudges are suppressed when there's genuinely nothing to compress (the `protected.length > 0` discriminator preserves the "no refs assigned yet" edge case). (5) **PR #159** — When a nudge is suppressed, advance `lastPerMessageNudgeTokens` to `currentTokens` and clear `lastNudgeShownTokens`, creating discrete 5% check intervals instead of every-turn re-evaluation.
+
+Files: `lib/compress/decompress.ts`, `lib/compress/decompress-logic.ts`, `lib/messages/inject/utils.ts`, `lib/messages/inject/inject.ts`. Tests: 757 pass (11 new for decompress-logic, 5 new for token classification, 4 "all protected" tests fixed to actually exercise the `allProtected` branch, +1 new Scenario A test for compress-after-suppression).
+
+### v1.12.9 — Compress-as-Anchor (PR #153)
+
+**Problem**: Since v1.12.1, compression summaries were injected as synthetic tool-result messages via a registered `acp_context_recap` tool, while `stripStaleCompressCalls` removed past `compress` tool calls from the API context to avoid duplication. This doubled summary overhead: each block's summary existed both as a synthetic recap message AND in the original (now-stripped) compress call's `summary` parameter. For sessions with many compressions, this "recap overhead" consumed 10–20% of context with no additional information value. The `acp_context_recap` tool description also claimed it "automatically injects" summaries, which was misleading.
+
+**Fix**: Removed the synthetic recap injection entirely. Compression summaries now live **inside the model's own past `compress` tool calls** — the `summary` parameter of each historical `compress({ summary: "..." })` call serves as the anchor, visible to the model like any other tool call. Deleted `createSyntheticToolRecap` (prune.ts), `stripStaleCompressCalls` (prune.ts), and the automatic recap injection path. The `acp_context_recap` tool is now manual-only (model-callable for re-fetching summaries that scrolled out of context). Updated `system.ts` prompt to describe compress-as-anchor behavior and warn against reusing historical `startId`/`endId` without `acp_status` verification. Updated `RECAP_TOOL_DESCRIPTION` to reflect manual-only usage. Net effect: ~50% reduction in summary overhead for compression-heavy sessions.
+
+Files: `lib/messages/prune.ts`, `lib/messages/utils.ts`, `lib/compress/recap.ts`, `lib/prompts/system.ts`. Tests: updated for compress-anchor behavior; 725 pass.
+
 ### v1.12.8 — Phantom Block Rejection (PR #148)
 
 **Problem**: When the model called `compress` on a range that was already covered by an active compression block, `applyCompressionState` still created a new block with `directMessageIds: []`, `compressedTokens: 0`, and `effectiveMessageIds` inherited from the consumed block. The model saw "0 tokens removed" in the notification, retried the same range, and entered a death loop: each phantom block added ~1K of summary overhead while compressing nothing, causing context to *grow* with every compression call (issues #93, #135). User sessions showed 9 consecutive phantom compressions (b12–b20) on the same range before the user manually intervened.

package/README.zh-CN.md

@@ -395,6 +395,22 @@ 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
 
+### v1.12.10-dev.1 — Decompress 范围模式 + Token 分类 + Protected 标签精确性 + Nudge 抑制 + 离散间隔（PR #73, #155, #157, #158, #159）
+
+**问题**：五个问题，涉及 decompress 易用性、token 统计和 nudge 质量。（1）`decompress` 需要先 `acp_status` 再逐块 decompress 的循环才能恢复多个压缩块。（2）自 v1.12.9（compress-as-anchor）起，compress 工具的 `summary` 内容被错误分类为 `toolTokens` 而非 `summaryTokens`，导致上下文分布中 tool% 虚高、summary% 虚低 —— 误导模型的压缩决策。（3）`[PROTECTED: ...]` 标签列出受保护消息中的所有工具（例如 `[PROTECTED: grep, skill]`），让模型误以为 `grep` 等非保护工具也是保护原因。（4）当所有可见内容都是受保护内容时，nudge 仍然以空推荐列表注入 —— 浪费上下文且让模型困惑。（5）当 nudge 被抑制（过滤器移除所有推荐）时，下一轮检查每轮都重新评估，因为 baseline 从不前进，导致重复浪费计算。
+
+**修复**：（1）**PR #73** —— 为 `decompress` schema 新增可选 `startId`/`endId`；范围模式批量恢复所有 `effectiveMessageIds` 与解析范围重叠的活跃块，一次调用完成。新纯函数 `findActiveBlocksOverlappingMessages` 在 `decompress-logic.ts`；向后兼容（`blockId` 路径不变）。（2）**PR #155** —— 在 `estimateContextComposition`（`lib/messages/inject/utils.ts`）中，当 `toolName === "compress"` 时，从 `part.state.input.content[].summary` 提取 `summary` 文本并分类为 `summaryTokens`。结构开销仍计为 `toolTokens`。（3）**PR #157** —— `buildCompressibleRanges` 现在只添加通过 `isToolNameProtected` 或 `isFilePathProtected` 实际触发保护的工具，与 `messageContainsProtectedTool` 逻辑完全一致 —— 标签现在只显示 `[PROTECTED: skill]`。（4）**PR #158** —— 新增 `allProtected = compressible.length === 0 && protected.length > 0` 检查；`nothingToCompress = filterSuppressed || allProtected` 门控 nudge 注入，当确实没有可压缩内容时抑制软 nudge（`protected.length > 0` 判别式保留了"尚未分配 ref"的边缘情况）。（5）**PR #159** —— 当 nudge 被抑制时，将 `lastPerMessageNudgeTokens` 前进到 `currentTokens` 并清除 `lastNudgeShownTokens`，创建离散 5% 检查间隔而非每轮重新评估。
+
+文件：`lib/compress/decompress.ts`、`lib/compress/decompress-logic.ts`、`lib/messages/inject/utils.ts`、`lib/messages/inject/inject.ts`。测试：757 通过（decompress-logic 新增 11 个，token 分类新增 5 个，4 个 "all protected" 测试修复为真正触发 `allProtected` 分支，+1 个 Scenario A 测试覆盖抑制后压缩）。
+
+### v1.12.9 — Compress-as-Anchor（PR #153）
+
+**问题**：自 v1.12.1 起，压缩摘要通过注册的 `acp_context_recap` 工具以 synthetic tool-result 消息注入，同时 `stripStaleCompressCalls` 从 API 上下文中移除历史 `compress` 工具调用以避免重复。这导致摘要开销翻倍：每个块的摘要同时存在于 synthetic recap 消息和原始（已移除的）compress 调用的 `summary` 参数中。对于压缩频繁的会话，这个 "recap 开销" 占用 10–20% 上下文却没有额外信息价值。`acp_context_recap` 工具描述也声称它"自动注入"摘要，具有误导性。
+
+**修复**：完全移除 synthetic recap 注入。压缩摘要现在保留在**模型自己的历史 `compress` 工具调用中** —— 每个历史 `compress({ summary: "..." })` 调用的 `summary` 参数作为锚点，像其他工具调用一样对模型可见。删除了 `createSyntheticToolRecap`（prune.ts）、`stripStaleCompressCalls`（prune.ts）和自动 recap 注入路径。`acp_context_recap` 工具改为手动调用（模型可调用以重新获取滚动出上下文的摘要）。更新 `system.ts` 提示描述 compress-as-anchor 行为，警告不要不经 `acp_status` 验证就重用历史 `startId`/`endId`。更新 `RECAP_TOOL_DESCRIPTION` 反映手动调用语义。净效果：压缩频繁会话的摘要开销减少约 50%。
+
+文件：`lib/messages/prune.ts`、`lib/messages/utils.ts`、`lib/compress/recap.ts`、`lib/prompts/system.ts`。测试：更新为 compress-anchor 行为；725 通过。
+
 ### v1.12.8 — 幽灵块拒绝（PR #148）
 
 **问题**：当模型对已被活跃压缩块覆盖的范围调用 `compress` 时，`applyCompressionState` 仍然创建新块，`directMessageIds: []`、`compressedTokens: 0`、`effectiveMessageIds` 从被消费的块继承。模型在通知中看到"移除 0 tokens"，重试同一范围，进入死亡循环：每个幽灵块增加约 1K 摘要开销却什么都不压缩，导致上下文随每次压缩调用*增长*（issues #93, #135）。用户会话显示同一范围连续 9 次幽灵压缩（b12–b20），直到用户手动干预。