opencode-acp 1.11.0 → 1.11.2

package/README.md

@@ -422,6 +422,28 @@ For the complete list with root cause analysis, see the [bug tracker](https://gi
 
 ## Changelog
 
+### v1.11.2 — CI Enforcement & Auto-Publish (PR #104)
+
+Added GitHub Actions CI to automate AGENTS.md compliance enforcement:
+
+- **PR validation** (`pr-checks.yml`): every PR to master is checked for branch name convention (`YYYY-MM-DD_short-title`), devlog existence (`devlog/{branch}/REQ.md` + `WORKLOG.md`), and changelog updates on version bumps.
+- **Auto-publish** (`release.yml`): pushing a `v*` tag triggers `npm ci` → `npm run check:package` → `npm test` → `npm publish` → GitHub Release, fully automated.
+- Script: `scripts/ci/check-pr.sh` — reusable PR validation logic.
+
+Requires `NPM_TOKEN` secret in GitHub repo settings.
+
+---
+
+### v1.11.1 — Compress Baseline Fix (PR #99)
+
+**Problem**: When the model called `compress`, both `lastPerMessageNudgeTokens` and `lastToolOutputNudgeTokens` were set to `currentTokens` — the token count from the compress-calling assistant message, which reflects **pre-compression** context. After compressing 100K→50K, the baseline was stuck at 100K, so `growth = 50K - 100K = -50K` and nudges never fired again.
+
+**Fix**: Both baselines are now set to `undefined` on compress detection. The next message-transform run re-establishes the baseline from the real post-compression API token count, without triggering a nudge (`computeShouldNudge` returns `shouldNudge: false` when baseline is `undefined`).
+
+Files: `lib/messages/inject/inject.ts` (line 98-99). Tests: `tests/inject.test.ts` — 3 updated + 2 new (621 total, 0 fail).
+
+---
+
 ### 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.

package/README.zh-CN.md

@@ -395,6 +395,28 @@ 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
 
+### v1.11.2 — CI 自动校验 & 自动发布（PR #104）
+
+新增 GitHub Actions CI 自动执行 AGENTS.md 规范：
+
+- **PR 校验**（`pr-checks.yml`）：每个到 master 的 PR 自动检查分支名规范（`YYYY-MM-DD_short-title`）、devlog 是否存在（`devlog/{分支}/REQ.md` + `WORKLOG.md`）、版本号变更时 changelog 是否更新。
+- **自动发布**（`release.yml`）：push `v*` tag 后自动执行 `npm ci` → `npm run check:package` → `npm test` → `npm publish` → GitHub Release，全自动。
+- 脚本：`scripts/ci/check-pr.sh` — 可复用的 PR 校验逻辑。
+
+需要在 GitHub Secrets 中配置 `NPM_TOKEN`。
+
+---
+
+### v1.11.1 — 压缩基线修复（PR #99）
+
+**问题**：当模型调用 `compress` 时，`lastPerMessageNudgeTokens` 和 `lastToolOutputNudgeTokens` 都被设为 `currentTokens` —— 这是调用 compress 的 assistant 消息的 token 计数，反映的是**压缩前**的上下文。压缩 100K→50K 后，基线卡在 100K，导致 `growth = 50K - 100K = -50K`，nudge 永远不再触发。
+
+**修复**：压缩时将两个基线都设为 `undefined`。下一次 message-transform 运行时从真实的压缩后 API token 计数重建基线，不会误触发 nudge（`computeShouldNudge` 在基线为 `undefined` 时返回 `shouldNudge: false`）。
+
+文件：`lib/messages/inject/inject.ts`（第 98-99 行）。测试：`tests/inject.test.ts` — 3 个更新 + 2 个新增（共 621 个测试，0 失败）。
+
+---
+
 ### 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。

package/dist/index.js

@@ -2161,6 +2161,34 @@ function formatMessageIdTag(ref, attributes) {
   return `
 <${MESSAGE_ID_TAG_NAME}${serializedAttributes}>${ref}</${MESSAGE_ID_TAG_NAME}>`;
 }
+function formatTokenSize(tokens) {
+  if (tokens < 1e3) return String(tokens);
+  if (tokens < 1e4) return `${(tokens / 1e3).toFixed(1)}K`;
+  return `${Math.round(tokens / 1e3)}K`;
+}
+function classifyMessageType(parts) {
+  let hasTool = false;
+  let hasText = false;
+  let hasReasoning = false;
+  const toolNames = [];
+  for (const part of parts) {
+    if (part.type === "tool") {
+      hasTool = true;
+      if (typeof part.tool === "string" && !toolNames.includes(part.tool)) {
+        toolNames.push(part.tool);
+      }
+    } else if (part.type === "text") {
+      hasText = true;
+    } else if (part.type === "reasoning") {
+      hasReasoning = true;
+    }
+  }
+  if (hasTool) {
+    return toolNames.length > 0 ? `tool:${toolNames.join(",")}` : "tool";
+  }
+  if (hasReasoning && !hasText) return "reasoning";
+  return "text";
+}
 function assignMessageRefs(state, messages) {
   let assigned = 0;
   let skippedSubAgentPrompt = false;
@@ -6571,7 +6599,8 @@ var injectCompressNudges = (state, config, logger, messages, prompts, compressio
     state.nudges.contextLimitAnchors.clear();
     state.nudges.turnNudgeAnchors.clear();
     state.nudges.iterationNudgeAnchors.clear();
-    state.nudges.lastPerMessageNudgeTokens = currentTokens;
+    state.nudges.lastPerMessageNudgeTokens = void 0;
+    state.nudges.lastToolOutputNudgeTokens = void 0;
     saveSessionState(state, logger).catch(() => {
     });
     return;
@@ -6860,9 +6889,15 @@ var injectMessageIds = (state, config, messages, compressionPriorities) => {
     }
     const isBlockedMessage = isProtectedUserMessage(config, message);
     const priority = config.compress.mode === "message" && !isBlockedMessage ? compressionPriorities?.get(message.info.id)?.priority : void 0;
+    const msgType = classifyMessageType(message.parts);
+    const msgTokens = Math.round(countMessageCharacters(message) / 4);
     const tag = formatMessageIdTag(
       isBlockedMessage ? "BLOCKED" : messageRef,
-      priority ? { priority } : void 0
+      {
+        priority: priority ?? void 0,
+        type: msgType,
+        tokens: formatTokenSize(msgTokens)
+      }
     );
     if (message.info.role === "user") {
       let injected = false;
@@ -7885,7 +7920,7 @@ You operate in a context-constrained environment. All compression serves the pri
 
 ACP TAGS
 
-\`<acp-context>\` tags wrap ACP (Agent Context Pruning) system metadata \u2014 context management information injected each turn. This is system data, not user input. You may also see \`<dcp-message-id>\` and \`<dcp-system-reminder>\` tags \u2014 these are equivalent (DCP was the previous name for ACP). Treat them as boundary metadata only, not as tool-result content.
+Each message in the conversation is annotated with a <dcp-message-id> tag showing its reference ID, approximate token size, and content type. For example: <dcp-message-id tokens="2.1K" type="tool:bash">m00175</dcp-message-id>. Use these annotations to assess which messages are consuming the most context and prioritize compression accordingly. The token size is approximate \u2014 treat it as a relative guide, not an exact count. You may also see <dcp-system-reminder> tags \u2014 these are system directives. Treat all tags as boundary metadata, not as tool-result content.
 
 COMPRESSION SUMMARIES IN CONTEXT