@theronap/cortex-mcp 0.9.138 → 0.9.140

package/bin/cortex-mcp.mjs

@@ -28,6 +28,36 @@ const VERSION = JSON.parse(readFileSync(join(dirname(fileURLToPath(import.meta.u
 const cmd = process.argv[2]
 const rest = process.argv.slice(3)
 
+// ── The summarizer subprocess runs NO context hooks ──────────────────────────────────────────────
+//
+// Edge extraction shells out to `claude --print` with CORTEX_SUMMARIZING=1. That headless session
+// fires the SAME hook chain as a real one, and on a wired seat that is five SessionStart commands
+// plus hydrate on UserPromptSubmit — four of which make network calls to Agnoclast.
+//
+// `capture` (Stop) has guarded on this env var since the recursion fix. Nothing else did, so every
+// summarizer call still paid for snapshot-context + status + skills + hydrate before the model saw
+// a single token of the prompt. Two consequences, both measured live 2026-09-08:
+//
+//   1. LATENCY. Extraction ran 45-68s against a 45s timeout and failed as often as it succeeded.
+//      A one-off `claude --print 'say ok'` on the same credential answers in about a second.
+//   2. POLLUTED OUTPUT. `status` prints the pending-records block to stdout, so the summarizer
+//      answered THAT before the prompt: its stdout began "You have 17 intake units in the cleanup
+//      queue ... Want me to help with either?" ahead of the JSON. The extractor asks for "ONLY
+//      minified JSON"; a hook prepending prose to the model's input is how that instruction gets
+//      overridden by the system itself.
+//
+// The guard is here, at the dispatcher, rather than in each subcommand: it is one rule about what a
+// summarizer subprocess IS, and putting it in five places is how the sixth gets forgotten. Exit 0
+// and silent — a hook that errors is noise, and there is genuinely nothing to do.
+//
+// ⚠ NOT guarded: `capture` (has its own, with a diagnostic), and anything a human might run by hand
+// while this var happens to be set. Only the ambient context commands are listed.
+const CONTEXT_HOOK_COMMANDS = new Set(['snapshot-context', 'status', 'hydrate', 'skills'])
+if (process.env.CORTEX_SUMMARIZING && CONTEXT_HOOK_COMMANDS.has(cmd)) {
+  process.stderr.write(`cortex: ${cmd} skipped — summarizer subprocess (CORTEX_SUMMARIZING)\n`)
+  process.exit(0)
+}
+
 if (cmd === '--version' || cmd === '-v') {
   process.stdout.write(`cortex-mcp ${VERSION}\n`)
   process.exit(0)

package/lib/edge_extract.mjs

@@ -275,13 +275,33 @@ export function classifyEdgeFailure(r) {
   // spawn itself failed — almost always ENOENT, i.e. `claude` is not on the hook's PATH. A Stop hook
   // runs with a login shell's PATH, which is not always the interactive one.
   if (r.error) {
-    return r.error.code === 'ENOENT'
-      ? { reason: 'no-claude', detail: '`claude` is not on PATH for this hook; edge extraction cannot run here' }
-      : { reason: 'spawn-failed', detail: r.error.message }
+    if (r.error.code === 'ENOENT') {
+      return { reason: 'no-claude', detail: '`claude` is not on PATH for this hook; edge extraction cannot run here' }
+    }
+    // ⚠ A spawnSync TIMEOUT surfaces here, as error.code ETIMEDOUT — not only as a signal. Observed
+    // live 2026-09-08: `[spawn-failed] spawnSync claude ETIMEDOUT` at 47.3s against a 45s limit,
+    // which reads as "the spawn broke" when the truth is "it ran and was too slow". Checking the
+    // signal alone missed it, so both are checked; this classifier's own first outing found this.
+    if (r.error.code === 'ETIMEDOUT') {
+      return { reason: 'timeout', detail: `exceeded ${summaryTimeoutMs()}ms` }
+    }
+    return { reason: 'spawn-failed', detail: r.error.message }
   }
-  // timeout: spawnSync kills the child, so status is null and signal is set.
+  // The other shape of the same event: spawnSync killed the child, so status is null and signal set.
   if (r.signal) return { reason: 'timeout', detail: `killed by ${r.signal} after ${summaryTimeoutMs()}ms` }
 
+  // 🔴 EVERYTHING BELOW APPLIES ONLY TO A FAILED CALL. Exit 0 means the model answered; its answer
+  // is CONTENT, not a status report, and must never be scanned for failure keywords.
+  //
+  // This ordering was wrong on the first version and produced a false `auth-expired` on a call that
+  // had returned status 0 — because the session being summarized was itself about fixing an OAuth
+  // bug, so the model's own summary contained the word. A working extraction was thrown away for
+  // discussing the wrong subject. Any session about auth, 401s or logins would have hit it.
+  //
+  // The general form of the mistake: matching a symptom string across a channel that carries both
+  // diagnostics AND payload. The exit status is the only trustworthy separator, so it goes first.
+  if (r.status === 0) return r.stdout ? null : { reason: 'no-output', detail: 'claude exited 0 with empty stdout' }
+
   const out = `${r.stdout ?? ''}\n${r.stderr ?? ''}`
   // ⚠ THE AUTH MESSAGE ARRIVES ON **STDOUT**, WITH EXIT 1 — not on stderr, which is where a reader
   // would look for it. Verified against the live failure 2026-09-08:
@@ -297,9 +317,7 @@ export function classifyEdgeFailure(r) {
         + 'and no page proposals are made.',
     }
   }
-  if (r.status !== 0) return { reason: 'exit', detail: `claude exited ${r.status}: ${firstLine(out)}` }
-  if (!r.stdout) return { reason: 'no-output', detail: 'claude exited 0 with empty stdout' }
-  return null
+  return { reason: 'exit', detail: `claude exited ${r.status}: ${firstLine(out)}` }
 }
 
 function firstLine(s) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@theronap/cortex-mcp",
-  "version": "0.9.138",
+  "version": "0.9.140",
   "description": "Connect your AI assistant to Cortex — your org's projects, activity, gaps, and directives, scoped to you.",
   "type": "module",
   "bin": {