Architectural update from the user. ## New worker: engineer A new worker is online: **engineer**. Confirm by listing your tools — `delegate_engineer(prompt, wait_seconds=...)` should be available alongside the others. Engineer's CLAUDE.md is at `/workspaces/engineer/CLAUDE.md` if you want to read it. **What engineer is for:** tier-2/3 escalations from helpdesk. Root-cause analysis, multi-system correlations, advanced remediation, scripted fixes, network diagnosis, infrastructure-as-code reasoning. Engineer has `mcp-datto-rmm`, `mcp-itglue`, `mcp-autotask`, `mcp-unifi`. **Why this worker exists:** the user wants helpdesk to stay LEAN (lower context, fewer tokens) for tier-1 routine work. Engineer absorbs the deep technical load. Helpdesk should escalate to engineer (via your dispatch) rather than try to solve everything itself. ## Queue-learning split (corrected scope) **Helpdesk worker** learns from: - General Support queue - Earney IT Support Queue (the legacy 110k-ticket monthly-paced track) **Engineer worker** learns from: - Advanced Support queue - Critical Support queue Drop "Advanced" and "Critical" from the helpdesk-learning directive's Phase 1 scope. They belong to engineer's KB now. The original helpdesk-learning directive's Phase 1 (scope discovery), Phase 2 (per-client learning), Phase 3 (cross-client aggregation), and Phase 4 (graph refresh) all stand — just with the queue list pruned to General Support only (plus the EarneyIT track per the addendum). ## New parallel directive: engineer-learning **After** the helpdesk Phase 4 lands (so the box isn't running both heavy tracks at once), kick off a parallel directive for engineer: > "Read /workspace/CLAUDE.md, /workspace/knowledge/, /workspace/skills/. Build out your knowledge base from Autotask tickets in queues [Advanced Support queueID, Critical Support queueID] for ACTIVE clients only (`Companies.isActive=true`), last 24 months. For each ticket: read title, description, resolution, last note. Extract per-client engineering notes into /workspace/knowledge/clients/<slug>.md (sections: `## Recurring escalation patterns`, `## Root causes seen`, `## Resolution playbooks (engineering tier)`, `## Open issues`). Then aggregate cross-client into /workspace/knowledge/issues/ and /workspace/knowledge/resolutions/ same as the helpdesk learning. Then run /graphify . to build the engineering graph. Reply with: clients, issues, resolutions counts, top 3 root causes by frequency. Under 10 lines." Use the same phased approach as helpdesk: - Phase 1: scope discovery — queue IDs + active-client ticket counts - Phase 2: per-client learning, sequential, batched per Phase 1 scale - Phase 3: cross-client aggregation - Phase 4: refresh engineer's graph Same constraints: sonnet (orchestrator default), 60-min hard cap, sequential not parallel, ACTIVE clients only, durable .md files. ## Routing rule going forward When a tier-2/3 ticket comes in (anything that's not routine helpdesk-tier work), prefer `delegate_engineer` over `delegate_helpdesk`. Helpdesk's KB will tell you what's tier-1 vs not — if helpdesk's playbook says "escalate to engineer," respect that. If a worker reports back "this isn't really my domain," redirect rather than retry on the same worker. ## Surfacing rules (unchanged) - Phase 1 of each track surfaces the scope estimate to the user - Phase 4 of each track surfaces the final tally - Otherwise silently archive each phase's inbox completion — Signed: deployment automation, on Tristen's behalf