Day 60: Project: Migration Strangler Assistant

Lab / Project: Migration Strangler Assistant

Project day. Ship a two-agent strangler assistant: a Cut Planner that proposes a service boundary, and an Implementer that emits intents + contract tests — so a monolith endpoint moves behind a façade without a big-bang rewrite.

⚡ TL;DR: Planner outputs StranglerPlan (old route, new service, façade mapping, test list). Implementer applies intents only. Contract tests must pass before merge. Human approves the cut. Emphasize evidence bundles.

Architecture

Ticket → CutPlanner → strangler-plan.json
                   → Implementer → intents + contract tests
                   → Critic/CI → human merge gate

# project/strangler_plan.py
from pydantic import BaseModel

class StranglerPlan(BaseModel):
    monolith_route: str
    new_service: str
    facade_path: str
    data_owner: str
    contract_tests: list[str]
    rollback: str

Step 1 — Planner

python agents/cut_planner.py \
  --ticket MIG-220 \
  --repo . \
  --out artifacts/strangler-plan.json

Require the plan to cite call-graph edges (who imports the handler) — no vibes-only cuts.

Step 2 — Implementer intents

[
  {"kind": "add_facade", "path": "apps/api/facades/invoices.ts", "params": {"target": "http://invoices:8080"}},
  {"kind": "codemod", "path": "apps/api", "params": {"name": "redirect-invoice-reads"}},
  {"kind": "add_contract_test", "path": "tests/contract/test_invoice_cut.py", "params": {}}
]

Step 3 — Contract tests (non-negotiable)

# tests/contract/test_invoice_cut.py
def test_facade_matches_monolith_golden(client, golden):
    new = client.get("/facade/invoices/1").json()
    assert new == golden["invoices/1"]  # ✅ behavior lock

Step 4 — Evidence bundle

On exit, write artifacts/evidence.json: plan hash, intents, test JUnit, critic verdict, git SHAs. Reviewers open one folder.

Acceptance criteria

• [ ] Plan schema validates
• [ ] No direct prod IAM tools available to agents
• [ ] Contract tests pass in CI
• [ ] Human approved merge
• [ ] Rollback steps documented in plan

Closing checklist

• [ ] Two agents + contract tests
• [ ] Façade before deleting monolith code
• [ ] Evidence bundle per run
• [ ] Human gate on cut
• [ ] Keep rollback boring and tested

Series navigation

Day 59: Multi-Agent Eval: Task Suites, Not Vibes · Day 61: Prompt Injection in Jira, Email, and RAG

Last updated September 11, 2026