Roadmap
From foundation through memory, multi-provider, and production hardening — phases, milestones, and what ships next.
Foundation
Phase 0: Foundation — complete.
Core Platform
Phase 1: Core Platform — complete as of 2026-06-05.
- 1.0.1 GO Integration
- 1.1.1 Postgres Migrations
- 1.1.2 Auth Proto
- 1.1.3 Auth Token
- 1.1.4 Token Creation
- 1.1.5 Permission Bitmap
- 1.1.6 Argon2id Parameters
- 1.1.7 Users Agents
- 1.2.1 Proxy Auth
- 1.2.2 Proxy Request
- 1.2.3 Proxy Input
- 1.2.4 Proxy Rate
- 1.2.5 Proxy Agent
- 1.2.6 Request ID
- 1.2.7 Graceful Shutdown
- 1.3.1 Otel Tracer
- 1.3.2 Prometheus Metric
- 1.3.3 Shared Logger
- 1.4.1 Developer Experience
- 1.4.2 Shared Config
- 1.4.3 Health Check
- 1.5.1 Security Integration
Public Web Product
Phase 1.5: Public Web Product — complete.
- d1.1 Pnpm Workspace
- d1.2 Cursor Rules
- d2.1 Bootstrap Fumadocs
- d2.2 Matte Graphite
- d2.3 MDX Components
- d2.4 Dynamic OG
- d2.5 Command Palette
- d2.6 Enhanced Code
- d2.7 Sidebar Breadcrumbs
- d2.8 NOT Found
- d3.1 Information Architecture
- d3.2 Seed Pages
- d3.3 Quickstart Five
- d3.4 Api Reference
- d4.1 Vercel Preview
- d4.2 Github Actions
- d5.1 Visual QA
- d5.2 Verify Phase15
- d6.1 Cloudflare Domain
- d6.2 DNS Production
- d6.3 Cross Site
Single Provider E2E
Phase 2: Single Provider E2E — complete. Authenticated OpenAI-compatible chat, auth cache, directives, sessions, and ClickHouse traces ship on main.
- 2.1.1 Provider Interface
- 2.1.2 Openai NON
- 2.1.3 Openai Streaming
- 2.1.4 Provider Routing
- 2.1.5 Provider Error
- 2.1.6 Idempotency-Key for provider retries
- 2.2.1 Auth Cache
- 2.2.2 Token Revocation
- 2.3.1 Directive Migrations
- 2.3.2 Directive Resolver
- 2.3.3 System Prompt
- 2.4.1 Sessions Checkpoints
- 2.4.2 Session Store
- 2.4.3 Proxy Session
- 2.4.4 Session Idle-Timeout Sweeper
- 2.5.1 Clickhouse Schema
- 2.5.2 Clickhouse Client
- 2.5.3 Async Trace
- 2.6.1 Latency Benchmark
- 2.6.2 Phase2 Exit
Provider Generalization & Foundation
Generalize the single-provider proxy into a pluggable, multi-backend platform before building the memory engine.
- 2.5.G1.M1 Anthropic provider adapter
- 2.5.G1.M2 Model capability registry
- 2.5.G1.M3 Self-hosted LLM adapter (vLLM-first)
- 2.5.G2.M1 Tokenizer registry
- 2.5.G3.M1 Response middleware hook (non-streaming)
- 2.5.G3.M2 Streaming response pipeline design
- 2.5.G4.M1 Embedder interface and registry
- 2.5.G4.M2 TEI backend (GPU profile)
- 2.5.G4.M3 Hosted-API embedding backend
- 2.5.G4.M4 Content-hash embedding cache
- 2.5.G5.M1 Temporal validity columns
- 2.5.G5.M2 Multi-label category support
- 2.5.G5.M3 Relationship graph readiness
- 2.5.G6.M1 MCP server skeleton
- 2.5.G7 Phase 2.5 Exit
Phase 3
Phase 3 proves memory can be written, deduplicated, conflict-checked, and retrieved correctly and fast — in isolation, testable via direct API calls.
- 3.1.1 Memory schema v2
- 3.2.1 VectorStore interface and pgvector-HNSW implementation
- 3.C.1 PII detection and redaction stage
- 3.C.2 Content and near-duplicate detection
- 3.C.3 Temporal-interval-aware conflict detection
- 3.C.4 Multi-label memory classification
- 3.C.5 Write-path persistence and orchestration
- 3.D.1 Semantic search read path
- 3.D.2 Composite scoring in search ranking
- 3.D.3 Hot cache (Redis sorted set)
- 3.E.1 Memory security integration test suite
- 3.E.2 End-to-end memory lifecycle test
- 3.E.3 Retrieval-quality benchmark sign-off
- 3.E.4 ADR consolidation and index update
- 3.E.5 Phase 3 exit audit document
Phase 3.5
Phase 3.5 closes the learning loop: conversations become memories (extraction) and memories become context on the next request (assembly), wired into the real proxy hot path.
- 3.5.A.1 Celery Worker Skeleton
- 3.5.A.2 Task Observability and Dead-Letter Handling
- 3.5.B.1 Extraction Prompt v2 and Structured Output Contract
- 3.5.B.2 Cost-Tiered, Provider-Agnostic Extraction Execution
- 3.5.B.3 Incremental Turn Tracking and Idempotency
- 3.5.B.4 Extraction Quality Evaluation Harness
- 3.5.C.1 Token Budget Calculator
- 3.5.C.2 Parallel Retrieval Orchestration
- 3.5.C.3 Composite Scorer v2 — Relevance-Gated
- 3.5.C.4 Packer v2 — Bounded DP Knapsack
- 3.5.C.5 Context Formatter
- 3.5.C.6 gRPC service skeleton and degradation contract
- 3.5.D.1 Go gRPC client for Context Assembly
- 3.5.D.2 Wire Assemble into dispatchProviderCompletion
- 3.5.D.3 Response headers and metadata
- 3.5.D.4 Extraction task enqueue
- 3.5.D.5 Config, feature flags, and env vars
- 3.5.D.6 Integration tests
- 3.5.E.1 MCP server skeleton and transport
- 3.5.E.2 search_memory and write_memory tools
- 3.5.E.3 record_feedback tool
- 3.5.E.4 MCP observability and rate limiting
- 3.5.F.1 End-to-end learning-loop test suite
- 3.5.F.2 Cross-tenant isolation matrix
- 3.5.F.3 Benchmark sign-off and ADR index
Operator Platform & Multi-Provider
Phase 4 is where IBEX stops being a proxy with an API-only backend and becomes an operable product: humans get a management surface, a second LLM provider proves the abstraction, and rate limiting becomes real cost control.
Phase 4.5
Phase 4.5 makes IBEX observably trustworthy over time: behavioral fingerprinting, drift detection with calibrated statistical methodology, and directive regression testing with LLM-judge evaluation.
Advanced Retrieval & Graph Memory
Phase 5 upgrades retrieval from single-vector ANN search to hybrid dense+sparse+reranking with graph-aware traversal.
Phase 0 → Phase 1 → Phase 1.5 → Phase 2 → Phase 3 → Phase 4 → Phase 5
See the phases overview for dependencies and exit criteria.