@theronap/cortex-mcp 0.9.12 → 0.9.14

package/lib/capture.mjs

@@ -4,6 +4,20 @@ import { resolve } from 'path'
 import { createHash } from 'crypto'
 import { fetchCortex, classify, resolveBase } from './diagnose.mjs'
 import { extractSession } from './edge_extract.mjs'
+import { extractTyped } from './extract_typed.mjs'
+
+// Fetch the node-type registry (global + this org's custom types) — the catalog the typed extractor
+// produces against. Best-effort: null on any failure (typed extraction is then skipped, never blocks capture).
+async function fetchRegistry(base, token) {
+  try {
+    const res = await fetchCortex(`${base}/api/node-types`, { headers: { Authorization: `Bearer ${token}` } })
+    if (!res.ok) return null
+    const j = await res.json().catch(() => ({}))
+    return Array.isArray(j.types) ? j.types : null
+  } catch {
+    return null
+  }
+}
 
 // Project name from the hook cwd — cross-platform. Windows hooks send backslash paths,
 // and splitting on '/' alone turned the WHOLE path into one garbage project slug
@@ -112,6 +126,19 @@ export async function runCapture() {
     ? { ...common, summary: extracted.summary, people: extracted.people, entities: extracted.namedEntities }
     : { ...common, transcript }
 
+  // Parallel typed extraction (OPT-IN via CORTEX_TYPED): registry-driven typed notes, sent ALONGSIDE the
+  // people/entities above. The server's typedNotes receiver persists them additively. Off by default so it
+  // never adds a 2nd `claude` call / latency until verified; flip to default once typed ≥ the blob path.
+  if (process.env.CORTEX_TYPED && transcript) {
+    try {
+      const registry = await fetchRegistry(base, token)
+      if (registry?.length) {
+        const typed = extractTyped(transcript, registry)
+        if (typed?.notes?.length) ingestBody.typedNotes = typed.notes
+      }
+    } catch { /* best-effort — never block capture */ }
+  }
+
   let res
   try {
     res = await fetchCortex(`${base}/api/ingest`, {

package/lib/edge_extract.mjs

@@ -6,7 +6,7 @@ import { spawnSync } from 'child_process'
 // both session people AND the entity catalogue down; extracting here keeps the engine on `claude -p`
 // (the standing preference) and removes that dependency. Node-native (no bun) to match capture.mjs.
 //
-// edgeSafeEnv() strips EVERY ANTHROPIC_* var + CLAUDE_CODE_OAUTH_TOKEN before spawning, so the call
+// edgeSafeEnv() strips EVERY ANTHROPIC_* var (but KEEPS CLAUDE_CODE_OAUTH_TOKEN — the headless
 // runs on the file-based subscription login and a stray ANTHROPIC_BASE_URL (e.g. a keytunnel proxy)
 // can't reroute "free, local" inference through a billed proxy or ship raw text off-machine. The
 // spawn also sets CORTEX_SUMMARIZING=1 so the headless session's own Stop hook no-ops (runCapture
@@ -21,7 +21,12 @@ export function edgeSafeEnv(base = process.env, extra = {}) {
   const env = {}
   for (const [k, v] of Object.entries(base)) {
     if (v === undefined) continue
-    if (k.startsWith('ANTHROPIC_') || k === 'CLAUDE_CODE_OAUTH_TOKEN') continue
+    // Strip billed/proxy routing only (ANTHROPIC_API_KEY metered, ANTHROPIC_BASE_URL keytunnel,
+    // ANTHROPIC_AUTH_TOKEN/CUSTOM_HEADERS proxy) so edge inference runs on the SUBSCRIPTION, never a
+    // metered/proxy path. KEEP CLAUDE_CODE_OAUTH_TOKEN — it IS the headless subscription credential
+    // (from `claude setup-token`); stripping it leaves `claude --print` with no usable auth → 401.
+    // (The earlier "subscription is file-based" assumption was wrong on macOS — headless needs this token.)
+    if (k.startsWith('ANTHROPIC_')) continue
     env[k] = v
   }
   return { ...env, ...extra }

package/lib/extract_typed.mjs

@@ -0,0 +1,68 @@
+import { spawnSync } from 'child_process'
+import { edgeSafeEnv } from './edge_extract.mjs'
+
+// Typed extraction (synthesis slice 2) — registry-driven. Given a transcript + the node_types registry,
+// the LLM emits TYPED NOTES (one per real thing that happened / was mentioned) conforming to the type
+// definitions, instead of the old loose {summary, people, namedEntities} blob. Runs locally on the
+// subscription (edgeSafeEnv strips ANTHROPIC_*). PARALLEL to edge_extract.mjs — does NOT replace it yet;
+// built to be tested in isolation, then wired into capture once proven (slice 2b).
+//
+// Per the ontology (cortex-node-ontology-spec): programmatic fields are filled by the caller from source
+// metadata BEFORE this runs; here the LLM fills only what the text supports, self-flags llm_filled fields,
+// and uses "unknown" for required fields it can't determine (→ red-link downstream). It may also COIN a
+// new type (open registry) when nothing fits.
+
+function typeCard(t) {
+  const fields = (arr) => (Array.isArray(arr) ? arr : []).map((f) => `${f.field}${f.note ? ` (${f.note})` : ''}`).join(', ')
+  return `• ${t.name} [${t.layer}]${t.applies_to?.length ? ` applies_to:${t.applies_to.join('/')}` : ''}\n` +
+    `    required: ${fields(t.required_fields) || '—'}\n` +
+    `    optional: ${fields(t.optional_fields) || '—'}\n` +
+    `    how: ${t.extraction_instructions}`
+}
+
+export function extractTyped(transcript, registry, opts = {}) {
+  const text = (transcript ?? '').trim()
+  if (!text || !Array.isArray(registry) || !registry.length) return null
+  const catalog = registry.map(typeCard).join('\n')
+  const prompt =
+    'You are building an organization knowledge graph from a work session. Using ONLY the node/note types ' +
+    'in the CATALOG below, emit the typed notes this session supports — one per real thing that happened ' +
+    '(actions, decisions, communications) or entity referenced (people, projects, etc.).\n\n' +
+    'RULES:\n' +
+    '- Fill a field ONLY if the text supports it. If a REQUIRED field is unknowable, set it to "unknown".\n' +
+    '- List the fields YOU inferred (vs. ones obviously given) in "llm_filled".\n' +
+    '- Reference other things by name (e.g. actors:["Jane Doe"]) — do not invent ids.\n' +
+    '- NEVER invent numbers (amounts, values, metrics) — "unknown" if not explicitly stated.\n' +
+    '- Be conservative: only emit a note for something genuinely present. Empty array is fine.\n' +
+    '- If something important fits NO type, you may coin one: add it to "new_types" with ' +
+    '{name, layer, required_fields:[{field}], why}.\n\n' +
+    'Return ONLY minified JSON: {"notes":[{"type","fields":{...},"llm_filled":[...]}],"new_types":[...]}.\n\n' +
+    '--- CATALOG ---\n' + catalog + '\n\n--- SESSION ---\n' + text.slice(0, 12000) + '\n--- END ---'
+  try {
+    const r = spawnSync(
+      'claude',
+      ['--print', '--model', opts.model ?? process.env.CORTEX_SUMMARY_MODEL ?? 'claude-haiku-4-5', prompt],
+      { env: edgeSafeEnv(process.env, { CORTEX_SUMMARIZING: '1' }), encoding: 'utf8', timeout: 120_000, maxBuffer: 8 * 1024 * 1024 },
+    )
+    if (r.status !== 0 || !r.stdout) return null
+    return parseTyped(r.stdout)
+  } catch {
+    return null
+  }
+}
+
+export function parseTyped(out) {
+  if (!out) return null
+  const start = out.indexOf('{')
+  const end = out.lastIndexOf('}')
+  if (start < 0 || end <= start) return null
+  try {
+    const o = JSON.parse(out.slice(start, end + 1))
+    return {
+      notes: Array.isArray(o.notes) ? o.notes : [],
+      new_types: Array.isArray(o.new_types) ? o.new_types : [],
+    }
+  } catch {
+    return null
+  }
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@theronap/cortex-mcp",
-  "version": "0.9.12",
+  "version": "0.9.14",
   "description": "Connect your AI assistant to Cortex — your org's projects, activity, gaps, and directives, scoped to you.",
   "type": "module",
   "bin": {