fix(sessions): ancestor-chain ownership — Cursor session close / audit / worklog were dead

src/server.ts

@@ -37,6 +37,7 @@ import {
   isPidAlive,
   loadSession,
   closeSession,
+  getOwnAncestorPids,
 } from "./storage/sessions.js";
 import { logEvent } from "./storage/worklog.js";
 import { spawnDetachedAuditWorker } from "./audit-spawner.js";
@@ -78,10 +79,21 @@ const defaultWorkspacePath = isWorkspace ? serverCwd : null;
 // the Claude session_id, which is the key for multi-window isolation.
 //
 // Instead of a single "currentSession", the server owns all AXME sessions
-// created by hooks whose parent process id equals our own parent process
-// id (i.e., the same Claude Code instance that spawned us). At disconnect,
-// we close all of them.
+// created by hooks of the same IDE instance that spawned us. Hooks record
+// ownerPpid via getClaudeCodePid() (their grandparent above the sh wrapper).
+// Under Claude Code that equals our PARENT — claude spawns hooks and the
+// MCP server alike. Cursor adds a layer: hooks hang off the cursor-server
+// main process while we are a child of the extension host, so the recorded
+// owner is our GRANDparent and strict ppid equality never matched — every
+// Cursor extension install had session close / audit / worklog silently
+// dead ("No active AXME session found", QA 2026-06-11). Match against our
+// ancestor chain instead; chain[0] is process.ppid, so the Claude Code
+// behavior is unchanged.
 const OWN_PPID = process.ppid;
+const OWN_ANCESTOR_PIDS = new Set<number>(getOwnAncestorPids(4));
+function isOwnedMapping(ownerPpid: number | undefined): boolean {
+  return ownerPpid != null && OWN_ANCESTOR_PIDS.has(ownerPpid);
+}
 
 // Track which context paths have been delivered in this MCP session.
 // Used by axme_oracle/decisions/memories to avoid duplicating workspace
@@ -121,7 +133,7 @@ void sendStartupEvents();
  */
 function getOwnedSessionIdForLogging(): string | undefined {
   const all = listClaudeSessionMappings(defaultProjectPath);
-  const owned = all.filter(m => m.ownerPpid === OWN_PPID);
+  const owned = all.filter(m => isOwnedMapping(m.ownerPpid));
   if (owned.length > 0) return owned[0].axmeSessionId;
 
   // No mapping for our PID — check for stale mappings from dead Claude Code
@@ -173,7 +185,7 @@ async function cleanupAndExit(reason: string): Promise<void> {
   // detached audit worker for each, clear the mapping, and exit. No awaiting.
   try {
     const mappings = listClaudeSessionMappings(defaultProjectPath);
-    const owned = mappings.filter(m => m.ownerPpid === OWN_PPID);
+    const owned = mappings.filter(m => isOwnedMapping(m.ownerPpid));
 
     // Deduplicate: group AXME sessions by Claude session ID.
     // Multiple AXME sessions can share the same Claude session (race condition
@@ -1197,7 +1209,7 @@ async function auditOrphansInBackground(): Promise<void> {
     // currently owned by this MCP server via an active mapping file.
     const ownedAxmeIds = new Set(
       listClaudeSessionMappings(defaultProjectPath)
-        .filter(m => m.ownerPpid === OWN_PPID)
+        .filter(m => isOwnedMapping(m.ownerPpid))
         .map(m => m.axmeSessionId),
     );
 

src/storage/sessions.ts

@@ -18,6 +18,7 @@
 
 import { join, resolve } from "node:path";
 import { readdirSync, readFileSync, rmSync, openSync, closeSync, unlinkSync, statSync } from "node:fs";
+import { execSync } from "node:child_process";
 import { randomUUID } from "node:crypto";
 import { ensureDir, writeJson, readJson, pathExists, atomicWrite, removeFile, readSafe } from "./engine.js";
 import { logSessionStart } from "./worklog.js";
@@ -104,20 +105,96 @@ export function isRetryableError(errMsg: string): boolean {
  * via ensureAxmeSessionForClaude).
  */
 export function getClaudeCodePid(): number {
+  const parent = readParentPidLinux(process.ppid);
+  if (parent != null) return parent;
+  // /proc missing (macOS, Windows), or parent died before we could read
+  // its stat file.
+  return process.ppid;
+}
+
+/** Parse the parent PID of `pid` from /proc (Linux only). Null elsewhere/on failure. */
+function readParentPidLinux(pid: number): number | null {
   try {
-    const stat = readFileSync(`/proc/${process.ppid}/stat`, "utf-8");
+    const stat = readFileSync(`/proc/${pid}/stat`, "utf-8");
     const closeParen = stat.lastIndexOf(")");
     if (closeParen > 0) {
       // Fields after "(comm) " are space-separated: state ppid pgrp ...
       const parts = stat.slice(closeParen + 2).split(" ");
-      const grandparent = parseInt(parts[1], 10);
-      if (Number.isFinite(grandparent) && grandparent > 1) return grandparent;
+      const parent = parseInt(parts[1], 10);
+      if (Number.isFinite(parent) && parent > 1) return parent;
     }
-  } catch {
-    // /proc missing (macOS, Windows), or parent died before we could read
-    // its stat file. Fall through to fallback.
+  } catch { /* /proc missing or pid gone */ }
+  return null;
+}
+
+/** Parent PID of `pid` via `ps` (macOS and other POSIX without /proc). */
+function readParentPidPosix(pid: number): number | null {
+  try {
+    const out = execSync(`ps -o ppid= -p ${pid}`, { encoding: "utf-8", timeout: 2_000, stdio: ["ignore", "pipe", "ignore"] }).trim();
+    const parent = parseInt(out, 10);
+    if (Number.isFinite(parent) && parent > 1) return parent;
+  } catch { /* ps unavailable or pid gone */ }
+  return null;
+}
+
+/**
+ * Walk this process's ancestor chain — parent, grandparent, … — up to
+ * `maxDepth` levels. The first element is always `process.ppid`.
+ *
+ * Why this exists: hooks record `ownerPpid` via getClaudeCodePid() (their
+ * grandparent — one step above the sh wrapper). Under Claude Code that PID
+ * equals the MCP server's PARENT, because the same claude process spawns
+ * both — so a strict `ownerPpid === process.ppid` equality worked. Cursor
+ * adds a layer: hooks are spawned by the cursor-server main process while
+ * the MCP server is a child of the EXTENSION HOST, so the hook-recorded
+ * owner is the server's GRANDparent and strict equality never matches.
+ * (QA 2026-06-11: `axme_begin_close` returned "No active AXME session
+ * found" on every Cursor extension install — session close, audit spawn
+ * and worklog were all dead.) Ownership checks must match against the
+ * ancestor chain, not a single ppid.
+ *
+ * Platform strategy: Linux walks /proc (microseconds); macOS walks `ps`
+ * (one small exec per level); Windows resolves the whole chain in a single
+ * PowerShell invocation (spawning powershell per level would cost seconds).
+ * Any failure stops the walk — the chain always contains at least
+ * `process.ppid`, so behavior degrades to the old strict equality.
+ */
+export function getOwnAncestorPids(maxDepth = 4): number[] {
+  const chain: number[] = [];
+  const seen = new Set<number>();
+  let current = process.ppid;
+  if (process.platform === "win32") {
+    return getOwnAncestorPidsWindows(maxDepth);
   }
-  return process.ppid;
+  for (let depth = 0; depth < maxDepth; depth++) {
+    if (!Number.isFinite(current) || current <= 1 || seen.has(current)) break;
+    chain.push(current);
+    seen.add(current);
+    const parent = process.platform === "linux"
+      ? readParentPidLinux(current)
+      : readParentPidPosix(current);
+    if (parent == null) break;
+    current = parent;
+  }
+  return chain.length > 0 ? chain : [process.ppid];
+}
+
+/** Windows ancestor chain in ONE PowerShell call (CIM walk). */
+function getOwnAncestorPidsWindows(maxDepth: number): number[] {
+  try {
+    const script =
+      `$p=${process.ppid};$out=@();for($i=0;$i -lt ${maxDepth} -and $p -gt 1;$i++){` +
+      `$out+=$p;$p=(Get-CimInstance Win32_Process -Filter \\"ProcessId=$p\\" -ErrorAction SilentlyContinue).ParentProcessId};` +
+      `$out -join ','`;
+    const out = execSync(`powershell -NoProfile -NonInteractive -Command "${script}"`, {
+      encoding: "utf-8",
+      timeout: 10_000,
+      stdio: ["ignore", "pipe", "ignore"],
+    }).trim();
+    const chain = out.split(",").map(s => parseInt(s.trim(), 10)).filter(n => Number.isFinite(n) && n > 1);
+    if (chain.length > 0) return chain;
+  } catch { /* powershell unavailable — fall back to parent only */ }
+  return [process.ppid];
 }
 
 function sessionsRoot(projectPath: string): string {

test/session-ownership.test.ts

@@ -0,0 +1,58 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import { getOwnAncestorPids, getClaudeCodePid } from "../src/storage/sessions.js";
+
+/**
+ * Ancestor-chain ownership matching (Cursor extension fix, 2026-06-11).
+ *
+ * Hooks record ownerPpid = their grandparent (getClaudeCodePid). Under
+ * Claude Code that equals the MCP server's parent; under Cursor it is the
+ * server's GRANDparent (cursor-server vs extension host). Ownership checks
+ * therefore match against getOwnAncestorPids(), whose first element must be
+ * process.ppid so the historical strict-equality behavior is a subset.
+ */
+describe("getOwnAncestorPids", () => {
+  it("starts with process.ppid", () => {
+    const chain = getOwnAncestorPids();
+    assert.ok(chain.length >= 1, "chain must never be empty");
+    assert.equal(chain[0], process.ppid, "chain[0] must be the direct parent");
+  });
+
+  it("walks beyond the direct parent where the platform allows", () => {
+    const chain = getOwnAncestorPids(4);
+    // Test runners are always at least two levels deep (init → shell/runner
+    // → node). /proc (Linux) and ps (macOS) both support the walk; a
+    // single-element chain there would mean the walk silently broke.
+    if (process.platform === "linux" || process.platform === "darwin") {
+      assert.ok(chain.length >= 2, `expected >=2 ancestors, got: ${chain.join(",")}`);
+    }
+  });
+
+  it("returns finite pids > 1 with no duplicates", () => {
+    const chain = getOwnAncestorPids(4);
+    for (const pid of chain) {
+      assert.ok(Number.isFinite(pid) && pid > 1, `bad pid in chain: ${pid}`);
+    }
+    assert.equal(new Set(chain).size, chain.length, "chain must not contain duplicates");
+  });
+
+  it("respects maxDepth", () => {
+    assert.ok(getOwnAncestorPids(1).length <= 1);
+    assert.ok(getOwnAncestorPids(2).length <= 2);
+  });
+
+  it("contains the hook-side owner pid for same-parent layouts (Claude Code invariant)", () => {
+    // In a Claude Code layout hooks and the server share one parent, and
+    // getClaudeCodePid() from THIS process resolves our grandparent — which
+    // must be inside our own ancestor chain. This is exactly the membership
+    // check isOwnedMapping() performs in server.ts.
+    const chain = new Set(getOwnAncestorPids(4));
+    const hookStyleOwner = getClaudeCodePid();
+    if (process.platform === "linux") {
+      assert.ok(
+        chain.has(hookStyleOwner),
+        `grandparent ${hookStyleOwner} not in ancestor chain ${[...chain].join(",")}`,
+      );
+    }
+  });
+});