Changelog

Product and SDK release notes for Maev.

2026-04-15

Terminal confirmation on connect: maev.run() now prints a confirmation line to stderr immediately after connecting — ✓ Maev connected agent=<name> dashboard=<url> — so you know Maev is active without opening the dashboard.
Session link after each run: After your agent finishes, Maev prints the session URL directly to your terminal — ✓ Maev session recorded status=completed calls=N cost=$0.0042 view=<url> — for instant access.

Email alerts now fire in production: A missing environment variable was silently disabling all email alerts in production. Fixed.
Slack alerts no longer throw errors for unconnected orgs: Previously, Maev attempted to send a Slack alert for every org, including those that hadn't connected Slack — resulting in logged exceptions on every alert attempt. Maev now checks for an active Slack connection first and skips silently if none is found.
Slack message format corrected: Alert messages now use Slack's native link format (<url|text>) instead of Markdown, so links render correctly in Slack.
Session links in alert emails now point to home.maev.dev instead of maev.dev.

Hybrid SDK Architecture: maev.run(agent, gateway=True) now supports dual-layer proxying. You get absolute execution control (circuit breakers) via the local SDK, while LLM traffic is automatically proxied through gateway.maev.dev.
Multi-provider Gateway Routing: Gateway now fully natively routes across Anthropic, OpenAI, and Gemini with baked-in fallback mappings.
Backend Self-Reflection Loop: Failure self-reflection now runs as an asynchronous offline optimization cycle on our backend, automatically generating and testing improved prompt candidates.
Automated Prompt Rollouts: Newly synthesized prompts automatically shift into shadow mode, scaling to live traffic dynamically via Canary rollouts.

Inspector completely deprecated: The old post-run inspector endpoints have been replaced with the live transparent Gateway architecture.

maev.run() is now the single public API. Observability and Autopilot activate together in one call. There is no separate maev.init() step.
Plan tiers renamed: developer is now dev, team is now teams

Autopilot learning loop: after 20+ runs, Maev analyzes failure patterns and applies winning strategies automatically
Strategy explanations: plain English descriptions of each learned strategy, visible in the Autopilot tab
Global strategy pool: Teams plan agents share effective strategies across the organization
Plan-based limits enforced at the API layer: agent caps, data retention windows, autopilot access
Data retention enforced per plan: Free 5 days, Dev 30 days, Teams 90 days

Gateway callback: fixed record_call argument count mismatch
Webhook signature verification replaced with a more reliable implementation (was broken)
Autopilot stats: interventions were being queried by agent_id instead of run_id

Import-time patch matching now supports nested module targets like google.generativeai
Supported Libraries docs now reflect direct injection coverage for the new providers

OpenAI prompt injection now covers both chat completions and responses APIs
Fix injection is idempotent, preventing duplicate correction blocks when multiple hook layers are active
Import-time patching is more reliable when libraries are imported after setup
Docs now distinguish telemetry support from direct Maev Fix injection support

Active fixes no longer depend on process restart to begin applying in common SDK flows
Fix generation now validates the requested session against the selected agent on the backend
Autofix behavior docs now reflect background refresh instead of implying fixes only apply on the next init