Changelog - Keel AI SRE Platform

May 2026

v1.52026-05-25

Multi-tenancy L2: Full enforcement — TenantDB (RLS), TenantRedis (key prefix), InternalAuth (HMAC S2S), TenantConfig (per-team overrides), per-tenant LLM budgets
Service consolidation: 18→11 services — merged observability-mcp, ticketing-agent, devops-agent, log-analysis-agent into their MCP counterparts
Supervisor decomposition: 1,621→118 line app.py — extracted 16 routers, enrichers, notification modules
Testing: 1,547 unit tests + 104 E2E tests, unified test runner with --unit/--all modes
Real remediation: restart_service + cache_flush wired to devops-mcp (ADR-058)
Prometheus metrics: Supervisor self-observability — enricher errors/duration, notification errors
133 features shipped, 58 ADRs, 20-step demo
Sunset: Grafana OnCall removed (F-102), replaced by Communication MCP

v3.02026-04-24

Hybrid NLP query engine — LLM for complex questions, templates for simple
ML anomaly detection with z-score and time-of-day baselines (F-84)
Multi-strategy alert clustering: temporal + causal + similarity + dependency (F-85)
AI log intelligence with LLM categorization and pattern detection (F-87)
Adaptive learning + post-remediation auto-verify (F-92)
Full RCA explainability API — per-incident /rca/explain (F-76)
AI system observability API — token, cost, quality stats (F-88)
SLO error-budget forecasting with burn-rate + exhaustion ETA (F-83)
NLP-to-PromQL natural language query (F-86)
Adaptive telemetry recommendations for cost/cardinality (F-89)
Synthetic incident simulation suite (F-91)
CLI tool — keel investigate / status / explain / approve / health (F-90)
Decision-grade authorization with approval chains (F-80)
Alert noise reduction engine (F-78)
Remediation execution engine — plan / dry-run / execute / rollback (F-82)
Fixed /logs/analyze 500 error — field name mismatch + response shape
Fixed NLP error rate template — ratio formula replaces raw 5xx filter
Fixed OmniView NLP display — map translation response + show live results
Benchmark v3: 92 criteria across 28 competitors (up from 73 in v2)

v2.02026-03-15

Episodic memory with multi-factor similarity matching (F-67)
Neo4j service topology graph for blast radius analysis (F-64)
Confidence calibration: 40% LLM + 60% signal coverage (F-47)
Multi-backend LLM gateway with circuit breakers (F-12, F-13)
SLO management service with burn-rate alerting (F-21, F-22)
Grafana OnCall integration (F-28) (sunset in v1.5 — replaced by Communication MCP)
OmniView control plane dashboard
Network-adaptive loading for website
E2E test suite: 104 tests across 76 phases

v1.02026-02-01