@pi-claudian/auto-save-to-markdown 0.1.1 → 0.1.2

package/README.md

@@ -105,6 +105,16 @@ kept in a collapsible `<details>` block) and summarizes each tool call and
 result in one line, so the file stays readable while still showing what the
 agent did.
 
+### Fragmented thinking repair
+
+Some upstream reasoning streams (observed with z-ai/GLM via OpenRouter) store
+thinking with every word — or every CJK character — on its own line: the
+original spaces collapse into leading spaces of one-word fragments joined by
+runs of newlines. The extension detects this corruption (lines starting with a
+single leading space, or a majority of 1–2-character fragment lines) and
+re-joins the fragments into flowing text, so saved thinking reads normally
+instead of one word per line. Clean thinking blocks are written untouched.
+
 `cost` and the token fields cover the whole saved branch and include cached
 tokens (priced at the provider's cache rates), so the totals are comparable
 with provider-side accounting (e.g. OpenRouter activity). Requests that never

package/README.zh.md

@@ -94,6 +94,14 @@ auth 重构之后登录页一直重定向死循环……
 `<details>` 块中），每个工具调用和结果各压缩成一行摘要，既可读又能看出
 agent 做了什么。
 
+### 碎片化 thinking 修复
+
+部分上游推理流（在 z-ai/GLM 经 OpenRouter 的场景中观察到）会把 thinking
+存成一词一行、甚至一字一行：原始空格塌缩成碎片行开头的单个空格，碎片之间
+被成串的换行拼接。扩展会检测这种损坏（依据带单个前导空格的行、或大量
+1–2 字符碎片行），把碎片重新接回通顺的文本，保存的 thinking 不再一行
+一词。正常的 thinking 块原样保存，不做任何改动。
+
 `cost` 和 token 字段统计整条已保存分支，且包含缓存 token（按供应商缓存
 价格计费），因此总计可与供应商侧账单（如 OpenRouter Activity）对照。未
 进入会话树的请求（失败重试、共用同一 API key 的其他会话）不在其中。

package/index.ts

@@ -32,6 +32,9 @@
  * - Compaction: files archive the ORIGINAL messages (getBranch() returns the
  *   raw tree path, not the compaction-aware context), so a compacted session
  *   still exports its complete history.
+ * - Thinking repair: reasoning blocks stored with the upstream
+ *   newline-fragmentation corruption (one word per line) are detected and
+ *   re-joined into flowing text before saving; clean thinking is untouched.
  *
  * Manual command: `/save-conversation` saves the current branch immediately
  * and reports the file path.
@@ -122,6 +125,59 @@ interface BranchMeta {
   projectRoot: string;
 }
 
+// ---------- thinking fragmentation repair ----------
+
+/**
+ * Some upstream reasoning streams (observed with z-ai/GLM via OpenRouter) store
+ * thinking as one word — or one CJK character — per line: the stream splits
+ * tokens into fragments joined by runs of newlines, and the original spaces
+ * survive only as leading spaces of the fragments. The saved markdown then has
+ * every token on its own line, which is miserable to read and bloats storage.
+ *
+ * Detection uses two signatures validated against ~520 real thinking blocks:
+ * lines starting with exactly one space (a survived word separator; blank-ish
+ * " " lines included), and an excess of 1–2-char non-list-marker lines (CJK
+ * fragments carry no leading space). Clean thinking never matches either.
+ *
+ * Repair strips all newlines — run length carries no recoverable meaning (the
+ * same paragraph boundary appears as 1, 2 or 3 newlines, while 4–7 can sit
+ * mid-sentence) — and collapses the doubled spaces left by lone-space
+ * fragments. Clean blocks pass through untouched.
+ */
+
+/** Line whose single leading space is a survived word separator. */
+function isThinkingSigLine(line: string): boolean {
+  return line === " " || /^ [^ *+\-\d]/.test(line);
+}
+
+/** Non-blank line of 1–2 chars that is not a standalone list marker. */
+function isThinkingShortFragment(line: string): boolean {
+  const s = line.trim();
+  if (s.length === 0 || s.length > 2) return false;
+  return !/^([-*+]|\d+[.)])$/.test(s);
+}
+
+/** Whether a thinking block shows the newline-fragmentation corruption. */
+function isFragmentedThinking(s: string): boolean {
+  const lines = s.split("\n");
+  const nonBlank = lines.filter((l) => l.trim().length > 0);
+  if (nonBlank.length === 0) return false;
+  const sig = lines.filter(isThinkingSigLine).length;
+  if (nonBlank.length < 8) return nonBlank.length >= 3 && sig >= 3;
+  if (sig / lines.length >= 0.12) return true;
+  return nonBlank.filter(isThinkingShortFragment).length / nonBlank.length >= 0.4;
+}
+
+/** Repair newline-fragmented thinking; clean thinking is returned unchanged. */
+function repairThinking(s: string): string {
+  if (!isFragmentedThinking(s)) return s;
+  debug("repairing fragmented thinking block:", s.length, "chars");
+  return s
+    .replace(/\n+/g, "")
+    .replace(/[ \t]{2,}/g, " ")
+    .trim();
+}
+
 export default function (pi: ExtensionAPI) {
   /**
    * Resolve the target directory. The env var may hold a relative folder name
@@ -242,7 +298,7 @@ export default function (pi: ExtensionAPI) {
         flushThinking();
         parts.push(b.text);
       } else if (b.type === "thinking") {
-        thinkings.push(b.thinking);
+        thinkings.push(repairThinking(b.thinking));
       } else if (b.type === "toolCall") {
         flushThinking();
         calls.push(`- \`${b.name}\` — ${previewArgs(b.arguments)}`);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-claudian/auto-save-to-markdown",
-  "version": "0.1.1",
+  "version": "0.1.2",
   "description": "Pi extension that automatically saves each completed conversation turn as a markdown file with YAML frontmatter, one file per session-tree branch.",
   "type": "module",
   "license": "MIT",