Optimize broker concurrency and add rate limiting middleware

[artugro]

Summary

This PR improves performance and reliability of the agent broker by batching independent async lookups, adds Redis-backed rate limiting middleware, enhances the concurrency limiter with pub/sub notifications, and adds smoke tests for CI.

Key Changes

Broker Performance Optimization (src/services/broker.py)

• Batch async lookups: Refactored invoke_agent() to use asyncio.gather() for independent database/repository calls (config, agent, conversation) instead of sequential awaits, reducing latency
• Conditional gathering: Only fetches conversation when conv_id is provided, avoiding unnecessary queries
• Removed duplicate agent fetch: Agent is now fetched once via the batched gather instead of separately later in the function
• Code formatting improvements for readability (line wrapping for long expressions)

Rate Limiting Middleware (src/core/rate_limit.py - new)

• Added RateLimitMiddleware using Redis fixed-window counters per client (by user ID or IP)
• Graceful degradation: allows all requests when Redis is unavailable
• Returns 429 with Retry-After and rate limit headers when limit exceeded
• Skips health check endpoints to avoid false positives
• Configurable via RATE_LIMIT_ENABLED and RATE_LIMIT_REQUESTS_PER_MINUTE settings

Concurrency Limiter Enhancement (src/workflow/utils/concurrency.py)

• Pub/sub notifications: RedisSemaphore now uses Redis pub/sub to notify waiters when slots are released instead of polling
• Fallback to polling: If pub/sub subscription fails, automatically falls back to polling with configurable interval
• Improved acquire logic: Tries immediate acquisition first, then waits for notifications with timeout
• Better docstrings explaining the pub/sub + polling hybrid approach

API Metadata Endpoints (src/main.py)

• Added /.well-known/agent.json endpoint for A2A (agent-to-agent) discovery with AgentCard format
• Reformatted existing /.well-known/mcp/server-card.json for readability
• Added RateLimitMiddleware to middleware stack

Settings (src/core/settings.py)

• Added RATE_LIMIT_ENABLED (default: True) and RATE_LIMIT_REQUESTS_PER_MINUTE configuration options

Testing (tests/test_smoke.py - new, tests/conftest.py - new)

• Added smoke tests that verify FastAPI app imports and basic endpoints respond without a live backend
• Tests cover: health endpoint, OpenAPI schema, A2A agent card, MCP server card
• Added pytest configuration to auto-mark integration tests so CI can skip them with -m "not integration"

CI (github/workflows/ci.yml)

• Re-enabled test job (was temporarily skipped)

Implementation Details

• The broker optimization reduces sequential I/O by parallelizing independent lookups, improving P99 latency for agent invocations
• Rate limiting uses a simple fixed-window counter (not sliding window) for Redis efficiency; window resets every 60 seconds
• Concurrency limiter's pub/sub approach reduces polling overhead in high-concurrency scenarios while maintaining correctness via race condition handling
• All new features include graceful degradation when Redis is unavailable

https://claude.ai/code/session_015zTWwVFtUWeN5ameM6teEa

[artugro]

@claude