@x12i/memorix-skill-process

Select Memorix records by OT + structured CT filter, run each through @x12i/memorix-skills, convert markdown skill output to JSON, and write one output content type at the same OT + recordId.

Requires @x12i/memorix-client@^3.8.0, @x12i/ai-dispatcher@^2.0.0, @x12i/memorix-scoper@^2.0.0. See docs/prerequisites.md.

Scoper 2.0 fetches stored records by default. Set options.synthesize: true and objective only when the skill must summarize records it cannot name. The host wrapper does that alignment on scoper__scope / scoper__subject_data.

Flow

1. Select — explore.queryContent on query.contentType + query.filter → recordId[]
2. Fetch — getContent for each inputContentTypes entry
3. Run skill — skills.run (dispatcher has Scoper MCP attached unless scoperMcp: false)
4. Convert — markdown → JSON (nx-json-parser)
5. Push — writeContent to outputContentType only (OT + recordId known)

Usage

import { createMemorixClient } from "@x12i/memorix-client";
import { createMemorixSkillProcess } from "@x12i/memorix-skill-process";

const memorix = createMemorixClient({ /* baseUrl, orgId, agentIds, caller */ });

// Omit scoperMcp to auto-attach from memorix config; or pass createUniqueTraceScoperMcp(...).
// Pass dispatcher options (not a live dispatcher) so Scoper can be merged.
const process = await createMemorixSkillProcess({
  memorix,
  skills: {
    dispatcher: {
      openrouter: { apiKey: process.env.OPENROUTER_API_KEY },
    },
  },
});

const result = await process.run({
  skill: "echo-records",
  objectType: "employees",
  inputContentTypes: ["raw"],
  query: {
    contentType: "raw",
    filter: { "data.status": "open" },
  },
  outputContentType: "skill-summary",
  request: { provider: "openrouter", model: "openai/gpt-5.2" },
});

Scoper MCP attach

• Omit scoperMcp: auto-build from Memorix config when this factory creates the dispatcher.
• Pass servers: merge into dispatcher options.
• scoperMcp: false: skip Scoper (unit tests / preload-only).
• Do not pass a live AiDispatcher / MemorixSkills together with scoperMcp — that throws SCOPER_MCP_ATTACH_SKIPPED. Pre-wire with createDispatcherWithScoperMcp and omit top-level scoperMcp instead.

Cloud models only see tool schemas; callTool runs on the host (VPN reachability to Memorix is required on the host, not on OpenRouter).

Notes

• Output catalog: skill-api startup and each analysis sequence load the full agent pack, add every pipeline output content type that is missing ({ id } only; other content types stay), and call ensureMetadata once (confirm: true). The same checksum is a noop. ready: false stops the process before it writes. A write refused with ROUTE_UNKNOWN_CONTENT_TYPE declares that one content type and retries once.
• Per-record isolation: one skill run per recordId. Errors on one record do not stop the loop.
• concurrency (default 1): bounded parallel fetch → skill → write; result order matches recordIds.
• skipIfOutputExists: skip when the output CT already exists (safe restarts).
• Empty skill text or non-object markdown→JSON fails that record (no raw markdown push).
• Write-back: skill JSON is written as record.data only.
  • If the output CT already exists → op: "patch" with [{ op: "set", path: "data", value }].
  • If absent → op: "add" with { recordId, data } plus concept / dataCategory copied from an input CT for the same subject (never invented).
  • Memorix owns format, revision, _system, and keeps concept on updates (buildMutatedRecord).
• Filter accepts equality maps or SafeFilter (@x12i/memorix-client ContentQueryFilter).

Analysis sequence

# Smoke (2 records per step)
node --env-file=.env packages/skill-process/scripts/run-analysis-sequence.mjs

# Full estate, concurrency 8 — effort: low by default; high for
# attack surface, asset triage, and deepen on exploitable+reachable assets
CONCURRENCY=8 node --env-file=.env packages/skill-process/scripts/run-analysis-sequence.mjs --full

Generic single step:

node --env-file=.env packages/skill-process/scripts/run-skill-batch.mjs \
  --skill=analyze-endpoint-asset --objectType=endpoint-asset \
  --inputContentTypes=inventory --queryContentType=inventory \
  --outputContentType=analysis --discriminatorValue=asset-exposure \
  --reasoningEffort=high --maxRecords=1