Skip to content
GRPNR.

Ground Truth — What Exists Today (2026-07-18)

Recon of groupon/groupon-monorepo, groupon/encore-ts-sample, groupon/architecture (cloud github.com, cloned in ../repos/) + full GHE estate survey (github.groupondev.com). Raw agent reports: raw-recon-output.json.

1. The three estates

Estate Where Size State
Legacy (Continuum) GHE, 20,463 repos / 949 orgs ~70M LOC claim plausible; only 581 repos (2.8%) pushed in last 90 days ~93–97% dormant. Real platform = I-Tier (466 repos, Node/JS/Coffee) + J-Tier (372 repos, Java) + Rails remnants. “Continuum” is the C4-model name only — zero GHE hits.
Target (Stack 2.0) Cloud groupon/groupon-monorepo 13,292 TS files ≈ 1.44M lines + 79k lines Go; 124 Encore services (SERVICES.md says 93 — drifted), 2,060 test files Created 2025-02, 101 contributors, ~290 PRs/month, default branch staging, staging→production release-train PRs, no merge queue.
Architecture model groupon/architecture (Structurizr C4) 141 legacy services / 21 domains federated; 206 Encore containers Auto-maintained by a Claude agent workflow (automap-repos.yml, Mon/Thu, opus-pinned) — agentic architecture federation already runs in production.

2. Target architecture (encore-ts-sample = codified rules)

  • App-per-system (~10 systems), service = dir with encore.service.ts; one Gateway + authHandler per app; every app has health service.
  • Per-service package.json = dependency encapsulation boundary, enforced by dependency-cruiser gate (service-deps-must-be-declared), not physics (node-linker=hoisted load-bearing).
  • Clean architecture per service: domain/ (framework-free) ← application/ ← infrastructure/ + composition/container.ts + presentation/. Declarative arch-contract.yaml per service; pnpm arch:check.
  • Services born compliant: pnpm gen:service scaffolder (plop); hand-creating services forbidden by AGENTS.md “Prime Directive”.
  • Data: raw parameterized sqldb default, Drizzle opt-in (ADR-0004) — conflicts with monorepo .rules/encore-ts-standards.md which mandates Drizzle. Must pick a winner.
  • Seams: intra-system ~encore/clients RPC + pub/sub; cross-system generated @acme/client-<system> packages; expand-contract migrations with destructive-DDL footgun lint.
  • Testing: integration-first encore test vs real Dockerised Postgres; pnpm verify = build:libs → typecheck → lint → depcruise → arch:check → footgun.

3. Conformance delta (monorepo vs target rules) — the actual refactor

The monorepo violates every headline rule of its own sample today:

  1. 2 encore.app files, not ~10 — all 124 TS services in ONE app (groupon-encore-83x2) + one Go app. The core refactor is a 1→N app split.
  2. 0 per-service package.json (target: one per service).
  3. No dependency-cruiser — only bespoke check scripts with .exceptions.json escape hatches.
  4. No node-linker=hoisted; .npmrc contains an expired hard-coded GitHub Packages token (housekeeping).
  5. Drizzle-mandated vs raw-SQL-default rules conflict.

At ~290 PRs/month, a long-lived restructure branch has a merge-conflict half-life of hours → incremental service-by-service moves through the existing train only.

4. Migration doctrine

  • Vision: _documentation/Stack 2.0 Groupon vision.md (Zaruba & Sima, 2025–2026, 2-year plan). “Calm revolution”: Encore grows around legacy; GAPI/UMAPI chokepoints die; per-tribe 2-year plans (B2B done, B2C underway, Core Ops preparing Order 2.0).
  • Observed pattern: wrapper-first strangler — _proxy_to_legacy/ with ~24 proxies (dc, dmapi, umapi, m3, gims, mds, gazebo, lazlo, orders, accounting, vss…) + MBUS WebSocket bridge for legacy events.
  • Sanctioned doctrine (GHE zblut/encore_migration_guide, a Claude Code plugin, 2026-03): “Default recommendation is a Clean Break — not a cautious strangler fig”, numeric risk score (0–2 clean break; 5+ justify strangler), Phase-0 “should this service exist at all” check, DELETE-first audit, worked examples (bhuvan, barcode-service).
  • Strangler frontier is SOX-scoped: most wrapped legacy sources live in sox-inscope (simultaneously the most-active GHE org — 66 repos active) → migrations carry compliance overhead. Several proxied sources are already dead upstream (dmapi last push 2016, deal-catalog 2022, GIMS 2018) → clean-break candidates.

5. Agent infrastructure already institutionalized

  • .agents/skills/ — ~50 skills, single source shared by Claude Code, Codex, Cursor (AGENTS.md + CLAUDE.md + .cursor/ + .junie/ + .aiassistant/). Ops fleet (encore-deploy-monitor, log-scanner, trace-monitoring, incident-management, release-train skills), autonomous partner-onboarding pipeline with a hard gate (“any gap stops the run… posts missing fields on the Jira” — an existing question-triage precedent), BYOD program (agents operate APIs, no admin UI).
  • Temporal in production: @temporalio/* 1.13, 13 workflow workers tribe-sharded, Workflow Management service + AdminIQ UI, docs incl. agentic_workflow_patterns.md (LLM agents as durable workflows — action-discriminator + MAX_ITERATIONS loops); live example workflow_oncall_agent. Vision: “Temporal = how AI thinks, retries, corrects itself, and keeps state.”
  • AI gateway: _core_system/ai-gateway → LiteLLM, Langfuse project registry (LANGFUSE_PROJECT_ECHELON et al.). Mastra app hosts oncall-triage agent. Production MCP server groupon-iq-mcp at api.enc.groupon.com. CodeRabbit reviews PRs.
  • CI: 36 workflows; architecture gates (controller→repository layering, cross-service imports, duplicate types), doc-drift-check, ORR metadata check, diff coverage, Happo visual, Playwright cron; hall-of-shame culture bot.

6. Access map (for the agent fleet)

Surface Token State
Cloud groupon/* (monorepo, architecture, sample) GROUPON_GITHUB_CLOUD_READ_TOKEN (ervojacek-png) ✅ read-all, approved
GHE 20k legacy repos GROUPON_GHE_READ_TOKEN (e-rvojacek) ✅ broad read
monorepo-development + monorepo-infra (the refactor repos, per Robert’s screenshot) GROUPON_GITHUB_CLOUD_GRPN2_TOKEN ❌ org-blind as of 2026-07-18 — pending org approval or wrong resource owner; org name unknown. Blocker for direct recon of the actual refactor repos.

monorepo-infra (per screenshot description): “Infrastructure monorepo: Temporal, Grafana, GKE, everything self-hosted” — matches the Stack 2.0 self-hosted direction; unverified beyond that.