infra: agent backend cannot scale beyond a few concurrent users

Start with apps/agent/main.py, apps/agent/src/bounded_memory_saver.py, render.yaml, and apps/app/src/app/api/copilotkit/route.ts; reproduce the concurrency failure described in #63 before choosing a phase. The work is done only when the selected scaling, persistence, backpressure, or load-testing goals are implemented and verified at the stated target concurrency.

Problem

The agent backend runs on a single uvicorn worker process with an in-memory checkpointer on a 512MB Render starter instance. This is a global bottleneck — not per-user. All concurrent users share the same event loop, the same memory pool, and the same 200-thread checkpoint limit.

Currently the app breaks at ~3 concurrent connections (#63). At production scale (100-1000 users), it would be effectively unusable.

Architecture bottlenecks

1. Single worker process

• uvicorn runs with 1 worker (default) — all requests share one Python event loop
• Each GPT-5.4 visualization call takes 10-30s
• LangGraph has synchronous sections that block the event loop
• Throughput: ~2-6 visualization requests/minute

2. In-memory checkpointer (BoundedMemorySaver)

• All conversation state stored in RAM — shared global pool of 200 threads
• FIFO eviction: after 200 conversations across ALL users, oldest threads are silently deleted
• Users lose conversation context mid-session with no error
• Not thread-safe — designed for single-process async only
• On 512MB starter plan, memory pressure builds well before 200 threads

3. No backpressure or error surfacing

• When the backend is saturated, requests hang silently — no timeout, no error, no retry
• Frontend shows no indication that the agent is overloaded
• Health check at /health returns 200 even when the event loop is blocked

Scale projections

Proposed solution

Phase 1 — Quick wins (config changes only)

• Add --workers 4 to uvicorn startCommand in render.yaml — multiplies throughput ~4x
• Upgrade agent service from starter (512MB) to standard (1GB+) in render.yaml
• Enable rate limiting (RATE_LIMIT_ENABLED=true) with reasonable limits (e.g. 20 req/min per IP)

Phase 2 — Persistent checkpointer

• Replace BoundedMemorySaver with PostgreSQL or SQLite async checkpointer
• Conversation state survives restarts and doesn't consume RAM
• No more silent thread eviction — threads persist until explicitly cleaned up
• Render already supports managed Postgres — can add as a service in render.yaml

Phase 3 — Error handling and backpressure

• Add frontend timeout — show error after ~30s of no response instead of hanging forever
• Add backend concurrency limit — return 503 "busy" when at capacity rather than queuing indefinitely
• Add connection health monitoring — detect dropped SSE connections and surface to user
• Reuse thread IDs per browser tab (sessionStorage) to avoid creating unnecessary threads

Phase 4 — Horizontal scaling

• Use Gunicorn with uvicorn workers for proper process management
• Verify Render auto-scaling (1-3 instances) works correctly with persistent checkpointer
• Add Redis or Postgres for shared state across instances
• Load test at target concurrency (100+ users) to validate

Related issues

• #63 — Agent stops responding after multiple concurrent tabs (symptom of this)
• #58 — Quality regression (long-running visualization calls exacerbate the single-worker bottleneck)
• #62 — Planning step before visualization (adds an extra round-trip, making concurrency even more critical)

Key files

• apps/agent/main.py — uvicorn config, BoundedMemorySaver(max_threads=200)
• apps/agent/src/bounded_memory_saver.py — FIFO eviction logic
• render.yaml — Render service config (starter plan, no worker config)
• apps/app/src/app/api/copilotkit/route.ts — Frontend → agent connection