Sprint 1 Complete: 46 Issues, 42K Lines

Sprint 1 is done. 46 issues closed, 235 files changed, approximately 42,700 lines of code added across the entire platform. Here’s what shipped.

Gateway: 34 New Modules

The AI gateway grew from a core proxy into a full governance platform:

• Security scanner — prompt injection, jailbreak, and data exfiltration detection with multi-signal risk scoring
• DAG orchestrator — conditional branching workflows with per-stage cost budgets
• Trace context — W3C Trace Context propagation across fleet agents
• Fleet cost tracker — cross-agent cost attribution for multi-runner fleets
• Unified cache — Redis cache layer with provider-aware TTL
• Connection pool — managed httpx connection pools for upstream providers
• SSE heartbeat — keepalive for long-running streaming responses
• Error classifier — normalized error codes and automatic retry classification
• Error replay — replay failed requests with modified parameters
• Model recommender — suggest cheaper models based on usage patterns
• Cost attribution — project-based cost tagging with custom labels
• Ephemeral tokens — short-lived tokens for browser-based gateway access
• Policy YAML — define governance policies as YAML files
• Dashboard subscriptions — real-time WebSocket updates for dashboard widgets
• CloudEvents — standardized event format for webhooks

Plus 19 more modules covering warmup, route registry, API versioning, webhook event storage, and more.

27 New Gateway Routes

Every new module is exposed via gateway routes, bringing the total to 130 route files. Key additions:

• /gateway/security-scanner/* — scan status, configuration, test endpoints
• /gateway/dag/* — create, execute, and monitor DAG workflows
• /gateway/traces/* — distributed tracing, span ingestion, HITL checkpoints
• /gateway/fleet-costs/* — fleet-wide cost aggregation and breakdown
• /gateway/cost-attribution/* — query costs by model, provider, tag, key
• /gateway/error-replay/* — replay and debug failed requests
• /gateway/model-recommendations/* — cost optimization suggestions

Observer SDK v2

A new distributed tracing SDK for Python and TypeScript:

• Python: CurateMeTracer with span context managers, @trace decorator, per-span cost tracking, and @hitl_checkpoint for human-in-the-loop gating
• TypeScript: Observer class with WebSocket streaming, 15+ tracking methods, privacy-aware redaction, and automatic reconnection
• W3C Trace Context: Full traceparent header propagation across services

CLI Commands

New gateway-focused CLI commands:

curate gateway test --full     # Health check + smoke test
curate gateway status          # Latency and pool status
curate costs today             # Today's spend
curate costs by-model          # Cost breakdown by model
curate runners health          # Runner fleet health

MCP Server: 12 New Tools

The in-container MCP server gained tools for desktop automation, BYOVM management, warm pool control, and compliance:

• curate_desktop_click, curate_desktop_type, curate_desktop_screenshot_region
• curate_byovm_status, curate_byovm_capabilities
• curate_warm_pool_status, curate_warm_pool_request
• curate_compliance_check, curate_compliance_report, curate_audit_log

Total MCP tools: 43 across 7 tiers.

Developer Documentation

• Quickstart guides — gateway, Python, TypeScript, cURL
• Governance reference — full 13-step chain documentation
• Cost tracking guide — attribution, budgets, alerts
• 5 operational runbooks — budget exceeded, high latency, PII blocked, rate limits, stuck runners
• Framework examples — LangChain, CrewAI, OpenAI Agents, Vercel AI, Claude SDK

Dashboard: 10 New Pages

• Cost attribution dashboard with tag-based drill-down
• DAG orchestration builder, list, and detail views
• Onboarding verification test page
• Latency benchmarks visualization
• Runner inspector with live resource monitoring
• Pricing calculator

Tests: 380+ New

45 new test files covering all Sprint 1 modules. Total test count across the platform: 20,800+ backend, 2,118 frontend, 87+ E2E.

What’s Next

Sprint 1 built the foundation. The platform now has:

• 85 gateway modules with 130 routes
• 51 LLM providers
• 121 runner control plane modules
• 229 dashboard pages
• Full SDK coverage (Python, TypeScript, CLI, MCP, Observer, Embed)

Next up: design partner outreach, production hardening, and the features that design partners request. The platform is ready.