opencode-acp 1.12.7 → 1.12.8

package/README.md

@@ -422,6 +422,14 @@ For the complete list with root cause analysis, see the [bug tracker](https://gi
 
 ## Changelog
 
+### 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.
+
+**Fix**: Added `checkPhantomBlock()` — a stateless pre-check in `lib/compress/pipeline.ts` that mirrors `applyCompressionState`'s `newlyCompressedMessageIds` computation. For each plan, it builds the effective message set (plan messages + consumed blocks' effective messages) and checks whether ANY message is "new" (i.e., has no active block covering it BEFORE mutation). If no message is new, the plan is a phantom and the entire compress call is rejected with a clear error before any state mutation occurs. Wired into both range-mode (`compress/range.ts`) and message-mode (`compress/message.ts`) after plan preparation, before snapshot. 12 tests cover: empty plans, all-new messages, consumed-block inheritance, GC'd messages (deactivated blocks count as new), and the exact `applyCompressionState` mirroring.
+
+Files: `lib/compress/pipeline.ts`, `lib/compress/range.ts`, `lib/compress/message.ts`. Tests: `tests/phantom-block.test.ts` (NEW, 12 tests). 725 tests pass.
+
 ### v1.12.7 — Smart Recommendation Filter + Dangerous Parameter + Ref-Leak Fix + Phantom Turn Fix (PRs #142, #147, #150)
 
 **Problem**: Four issues. (1) The recommendation filter used a hardcoded 5× growth threshold for single-message ranges (25% of context), leaked context, showed tiny ranges, and contradicted itself by recommending the last segment while blocking it. (2) Nudge text was injected even when the filter suppressed all ranges — wasting context with an empty recommendation list. (3) Compression block metadata leaked message refs (`m01309–m02150`) via `acp_context_recap` tool input, `acp_status` output, and `recap` tool output — the model copied these into compress calls on already-compressed ranges, creating phantom blocks (#93, #135). (4) `sendIgnoredMessage` during the transform hook persisted an ignored user message that resolved async after the model's response — the loop's `lastUser` detection picked it up → phantom LLM call with no new input → confusion → hallucination → feedback loop ("待命" spam).

package/README.zh-CN.md

@@ -395,6 +395,14 @@ 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
 
+### v1.12.8 — 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。
+
+**修复**：新增 `checkPhantomBlock()` —— `lib/compress/pipeline.ts` 中的无状态前置检查，镜像 `applyCompressionState` 的 `newlyCompressedMessageIds` 计算。对每个计划，构建有效消息集（计划消息 + 被消费块的有效消息），检查是否有任何消息是"新的"（即变异前没有活跃块覆盖它）。如果没有新消息，该计划是幽灵的，整个 compress 调用在任何状态变异前以清晰错误被拒绝。接入 range 模式（`compress/range.ts`）和 message 模式（`compress/message.ts`）的计划准备之后、快照之前。12 个测试覆盖：空计划、全新消息、被消费块继承、GC'd 消息（已停用块算新）、以及与 `applyCompressionState` 的精确镜像。
+
+文件：`lib/compress/pipeline.ts`、`lib/compress/range.ts`、`lib/compress/message.ts`。测试：`tests/phantom-block.test.ts`（新增，12 个测试）。725 测试通过。
+
 ### v1.12.7 — 智能推荐过滤 + Dangerous 参数 + Ref 泄漏修复 + Phantom Turn 修复（PR #142, #147, #150）
 
 **问题**：四个问题。（1）推荐过滤器使用硬编码的 5× 增长阈值（上下文的 25%），太大且容易泄露上下文；推荐最后一段的同时又阻止它，自相矛盾。（2）过滤器抑制所有 range 后仍然注入 nudge 文本——用空推荐列表浪费上下文。（3）压缩块元数据通过 `acp_context_recap` 工具输入、`acp_status` 输出和 `recap` 工具输出泄漏消息 ref（`m01309–m02150`）——模型复制这些 ref 到已压缩范围的 compress 调用中，产生幽灵块（#93, #135）。（4）`sendIgnoredMessage` 在 transform hook 中持久化 ignored 用户消息，异步在模型回复后完成——loop 的 `lastUser` 检测拾取它 → 无新输入的 phantom LLM 调用 → 困惑 → 幻觉 → 反馈循环（"待命" 刷屏）。

package/dist/index.js

@@ -5253,6 +5253,30 @@ function getLastVisibleMessageId(rawMessages, state) {
   }
   return null;
 }
+function checkPhantomBlock(state, plans) {
+  for (let i = 0; i < plans.length; i++) {
+    const plan = plans[i];
+    const effective = new Set(plan.messageIds);
+    for (const consumedId of plan.consumedBlockIds) {
+      const block = state.prune.messages.blocksById.get(consumedId);
+      if (block) {
+        for (const mid of block.effectiveMessageIds) {
+          effective.add(mid);
+        }
+      }
+    }
+    const hasNew = [...effective].some((mid) => {
+      const entry = state.prune.messages.byMessageId.get(mid);
+      return !entry || entry.activeBlockIds.length === 0;
+    });
+    if (!hasNew) {
+      return new Error(
+        `Compression range ${i + 1} contains only already-compressed messages (0 new direct messages, 0 tokens saved). Nothing to compress \u2014 pick a range with visible, uncompressed content. Use \`acp_status({scope:"uncompressed"})\` to see which ranges are still compressible.`
+      );
+    }
+  }
+  return null;
+}
 function checkLastSegmentDangerous(ctx, allPlanMessageIds, rawMessages, dangerous) {
   if (ctx.config.compress.lastSegmentSoftBlock === false) return null;
   const lastVisibleId = getLastVisibleMessageId(rawMessages, ctx.state);
@@ -5491,6 +5515,14 @@ function createCompressMessageTool(ctx) {
           summaryWithTools
         });
       }
+      const phantomError = checkPhantomBlock(
+        ctx.state,
+        preparedPlans.map(({ plan }) => ({
+          messageIds: plan.selection.messageIds,
+          consumedBlockIds: []
+        }))
+      );
+      if (phantomError) throw phantomError;
       const snapshot = snapshotCompressionState(ctx.state);
       const runId = allocateRunId(ctx.state);
       try {
@@ -5709,6 +5741,14 @@ function createCompressRangeTool(ctx) {
           consumedBlockIds: mergeConsumedBlockIds
         });
       }
+      const phantomError = checkPhantomBlock(
+        ctx.state,
+        preparedPlans.map((p) => ({
+          messageIds: p.selection.messageIds,
+          consumedBlockIds: p.consumedBlockIds
+        }))
+      );
+      if (phantomError) throw phantomError;
       const snapshot = snapshotCompressionState(ctx.state);
       const runId = allocateRunId(ctx.state);
       try {