@pi-claudian/auto-save-to-markdown 0.1.0 → 0.1.2

package/README.md

@@ -67,10 +67,12 @@ cost: 0.023401
 tokens: 18745
 tokens_input: 15230
 tokens_output: 3515
+tokens_cache_read: 0
+tokens_cache_write: 0
 messages: 8
 created: "2026-08-29T05:05:12.000Z"
 updated: "2026-08-29T05:42:10.000Z"
-cwd: "/Users/me/project"
+project_root: "/Users/me/project"
 session_file: "~/.pi/agent/sessions/--Users-me-project-20260829-050500_ab12.jsonl"
 ---
 
@@ -103,6 +105,22 @@ 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
+landed in the session tree (failed retries, other sessions sharing the same
+API key) are necessarily excluded.
+
 ## Branch behavior
 
 Pi sessions are trees: `/tree` navigates to an earlier point and a new prompt

package/README.zh.md

@@ -57,10 +57,12 @@ cost: 0.023401
 tokens: 18745
 tokens_input: 15230
 tokens_output: 3515
+tokens_cache_read: 0
+tokens_cache_write: 0
 messages: 8
 created: "2026-08-29T05:05:12.000Z"
 updated: "2026-08-29T05:42:10.000Z"
-cwd: "/Users/me/project"
+project_root: "/Users/me/project"
 session_file: "~/.pi/agent/sessions/--Users-me-project-20260829-050500_ab12.jsonl"
 ---
 
@@ -92,6 +94,18 @@ 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。
+
 ## 分支行为
 
 Pi 会话是树：`/tree` 导航到更早的位置后再提问就分出新的分支。每个

package/index.ts

@@ -18,8 +18,8 @@
  *   of the deepest message entry at file creation, and <time> is the local
  *   file-creation timestamp (YYYYMMDD-HHmmss).
  * - Frontmatter: title, session id, tree (branch key), model, provider,
- *   cumulative cost and tokens, message count, created/updated timestamps,
- *   cwd and session file.
+ *   cumulative cost and tokens (input, output, cache read/write), message
+ *   count, created/updated timestamps, project root and session file.
  * - Branching: each file records exactly ONE branch (the root→leaf path
  *   returned by sessionManager.getBranch()). State is persisted via
  *   `pi.appendEntry()` custom entries, which are part of the session tree
@@ -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.
@@ -114,10 +117,65 @@ interface BranchMeta {
   cost: number;
   tokensInput: number;
   tokensOutput: number;
+  tokensCacheRead: number;
+  tokensCacheWrite: number;
   messages: number;
   created: string;
   updated: string;
-  cwd: string;
+  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) {
@@ -221,23 +279,32 @@ export default function (pi: ExtensionAPI) {
   // ---------- markdown rendering ----------
 
   function renderAssistant(m: AssistantMessage, t: string): string {
-    const texts: string[] = [];
+    // Render blocks in their original chronological order: thinking always
+    // precedes the text it produced, instead of being grouped after the fact.
+    const header = `## Assistant · ${t}${m.model ? ` · ${m.model}` : ""}`;
+    const parts: string[] = [];
     const thinkings: string[] = [];
+    const flushThinking = () => {
+      if (thinkings.length) {
+        parts.push(
+          `<details>\n<summary>Thinking</summary>\n\n${thinkings.join("\n\n")}\n\n</details>`,
+        );
+        thinkings.length = 0;
+      }
+    };
     const calls: string[] = [];
     for (const b of m.content) {
-      if (b.type === "text") texts.push(b.text);
-      else if (b.type === "thinking") thinkings.push(b.thinking);
-      else if (b.type === "toolCall") calls.push(`- \`${b.name}\` — ${previewArgs(b.arguments)}`);
-    }
-
-    const header = `## Assistant · ${t}${m.model ? ` · ${m.model}` : ""}`;
-    const parts: string[] = [];
-    if (texts.length) parts.push(texts.join("\n\n"));
-    if (thinkings.length) {
-      parts.push(
-        `<details>\n<summary>Thinking</summary>\n\n${thinkings.join("\n\n")}\n\n</details>`,
-      );
+      if (b.type === "text") {
+        flushThinking();
+        parts.push(b.text);
+      } else if (b.type === "thinking") {
+        thinkings.push(repairThinking(b.thinking));
+      } else if (b.type === "toolCall") {
+        flushThinking();
+        calls.push(`- \`${b.name}\` — ${previewArgs(b.arguments)}`);
+      }
     }
+    flushThinking();
     if (calls.length) parts.push(`**Tool calls**\n\n${calls.join("\n")}`);
     if (m.errorMessage) parts.push(`> Error: ${m.errorMessage.replace(/\s+/g, " ").trim()}`);
     if (parts.length === 0) parts.push("_(empty response)_");
@@ -292,13 +359,19 @@ export default function (pi: ExtensionAPI) {
     if (meta.model) lines.push(`model: ${yamlQuote(meta.model)}`);
     if (meta.provider) lines.push(`provider: ${yamlQuote(meta.provider)}`);
     lines.push(`cost: ${meta.cost.toFixed(6)}`);
-    lines.push(`tokens: ${meta.tokensInput + meta.tokensOutput}`);
+    // `tokens` counts everything billed, including cached tokens, so it is
+    // comparable with provider-side token totals (e.g. OpenRouter activity).
+    lines.push(
+      `tokens: ${meta.tokensInput + meta.tokensOutput + meta.tokensCacheRead + meta.tokensCacheWrite}`,
+    );
     lines.push(`tokens_input: ${meta.tokensInput}`);
     lines.push(`tokens_output: ${meta.tokensOutput}`);
+    lines.push(`tokens_cache_read: ${meta.tokensCacheRead}`);
+    lines.push(`tokens_cache_write: ${meta.tokensCacheWrite}`);
     lines.push(`messages: ${meta.messages}`);
     lines.push(`created: ${yamlQuote(meta.created)}`);
     lines.push(`updated: ${yamlQuote(meta.updated)}`);
-    lines.push(`cwd: ${yamlQuote(meta.cwd)}`);
+    lines.push(`project_root: ${yamlQuote(meta.projectRoot)}`);
     if (meta.sessionFile) lines.push(`session_file: ${yamlQuote(meta.sessionFile)}`);
     lines.push("---");
     return lines.join("\n");
@@ -323,18 +396,21 @@ export default function (pi: ExtensionAPI) {
     let cost = 0;
     let tokensInput = 0;
     let tokensOutput = 0;
+    let tokensCacheRead = 0;
+    let tokensCacheWrite = 0;
     for (const e of pathMessages) {
       const m = e.message;
+      const usage = m.role === "assistant" ? m.usage : m.role === "toolResult" ? m.usage : null;
       if (m.role === "assistant") {
         model = m.model;
         provider = m.provider;
-        cost += m.usage?.cost?.total ?? 0;
-        tokensInput += m.usage?.input ?? 0;
-        tokensOutput += m.usage?.output ?? 0;
-      } else if (m.role === "toolResult" && m.usage) {
-        cost += m.usage.cost?.total ?? 0;
-        tokensInput += m.usage.input ?? 0;
-        tokensOutput += m.usage.output ?? 0;
+      }
+      if (usage) {
+        cost += usage.cost?.total ?? 0;
+        tokensInput += usage.input ?? 0;
+        tokensOutput += usage.output ?? 0;
+        tokensCacheRead += usage.cacheRead ?? 0;
+        tokensCacheWrite += usage.cacheWrite ?? 0;
       }
     }
     const now = new Date().toISOString();
@@ -348,10 +424,12 @@ export default function (pi: ExtensionAPI) {
       cost,
       tokensInput,
       tokensOutput,
+      tokensCacheRead,
+      tokensCacheWrite,
       messages: pathMessages.length,
       created: created ?? now,
       updated: now,
-      cwd: ctx.cwd,
+      projectRoot: ctx.cwd,
     };
   }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pi-claudian/auto-save-to-markdown",
-  "version": "0.1.0",
+  "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",