Phase 2.5 — Goals

Track A — Multi-Provider Adapters

Goal 2.5.A: Prove the Provider interface generalizes beyond OpenAI

Milestones:

• 2.5.G1.M1 — Anthropic provider adapter: Implement packages/provider/anthropic/ proving provider.Request/provider.Response are vendor-neutral. Handles system-field extraction, Anthropic SSE event parsing, and error mapping to the shared apierror.Error envelope.
• 2.5.G1.M2 — Model capability registry: ModelCapability struct with context window, max output tokens, tool/vision/streaming support, and tokenizer family. Registry.Capability(model) alongside Registry.For(model).
• 2.5.G1.M3 — Self-hosted LLM adapter (vLLM-first): Generalized openaicompatible package reusing OpenAI client logic with configurable base_url. Health-check at bootstrap, circuit breaker for self-hosted backends, explicit 503/queue-full mapping.

Acceptance criteria:

• Second provider added without duplicating streaming/auth logic
• ModelCapability queryable for all registered models
• Self-hosted vLLM endpoint registered and health-checked at bootstrap

---

Track B — Tokenizer Registry

Goal 2.5.B: Model-accurate token counting for every registered model

Milestones:

• 2.5.G2.M1 — Tokenizer registry: packages/tokenizer Go package with Tokenizer interface (Count(model, text) (int, error)). tiktoken for OpenAI models, Hugging Face tokenizers via CGo bindings for open-weight models. Tokenizers loaded once at startup, keyed by TokenizerFamily from the capability registry.

Acceptance criteria:

• Token counts match ground truth for OpenAI and at least one open-weight model family
• No network dependency on Hugging Face at runtime (tokenizer files bundled or cached locally)

---

Track C — Response-Side Processing Pipeline

Goal 2.5.C: Create the extension point for Phase 3 response stages

Milestones:

• 2.5.G3.M1 — Response middleware hook (non-streaming): ResponsePipeline with Stage interface in packages/responsepipeline. Typed decode of OpenAI chat response JSON, pipeline execution, re-encode. Fail-open by default. p99 overhead < 2ms benchmarked.
• 2.5.G3.M2 — Streaming response pipeline design (design doc + prototype only): ADR evaluating token-window buffering vs. post-hoc correction vs. disable-for-flagged-agents. Throwaway prototype measuring buffering-vs-flush interaction.

Acceptance criteria:

• Byte-for-byte identical output vs. current passthrough when pipeline has zero stages
• ADR published with streaming transformation decision

---

Track D — Multi-Backend Pluggable Embedding Service

Goal 2.5.D: Embedding backends selectable per deployment profile

Milestones:

• 2.5.G4.M1 — Embedder interface + registry: Python ABC (EmbeddingBackend), selected by EMBEDDER_PROFILE env (cpu/gpu/hosted). Profile is a deployment-time choice, not per-request.
• 2.5.G4.M2 — TEI backend (GPU): HTTP client to Hugging Face TEI sidecar running BAAI/bge-m3. No hand-rolled batching — TEI handles it.
• 2.5.G4.M3 — Hosted-API backend: OpenAI text-embedding-3-large as default hosted option.
• 2.5.G4.M4 — Content-hash embedding cache: Redis SHA-256(text) -> vector cache, backend-agnostic, >80% hit-rate target.

Acceptance criteria:

• At least two real EmbeddingBackend implementations (TEI + hosted OpenAI or Cohere) passing the same contract test suite; CPU MiniLM profile remains stub (ADR-0046)
• Org config carries embedding_profile + embedding_dim, validated against pgvector column at migration time
• Re-embedding migration runbook deferred to Phase 3 (non-goal for 2.5; exit criteria updated accordingly)

---

Track E — Schema Pre-Work

Goal 2.5.E: Prepare Postgres schema for Phase 3 memory store

Milestones:

• 2.5.G5.M1 — Temporal validity columns: valid_from, valid_until, observed_at on ibex_core.memories. Migration-only, no application code.
• 2.5.G5.M2 — Multi-label category support: ibex_core.memory_categories join table replacing single category enum exclusivity.
• 2.5.G5.M3 — Relationship graph readiness: Graph-traversal index on memory_relationships, helper view for recursive CTE traversal.

Acceptance criteria:

• All three migrations apply cleanly on top of existing Phase 3 draft schema with zero application code depending on new columns
• Verified via make db-migrate on a seeded dataset

---

Track F — MCP Server Skeleton

Goal 2.5.F: Prove MCP auth and tool-call audit pipeline before memory exists

Milestones:

• 2.5.G6.M1 — MCP server with stub tools: Python MCP server using official SDK. OAuth via AuthService.ValidateToken gRPC. Stub write_memory/search_memory handlers. Every tool invocation traced to ClickHouse.

Acceptance criteria:

• MCP server passes official conformance/inspector tool against stub handlers
• Auth reuses AuthService.ValidateToken via gRPC (no second identity system)
• Tool invocations written to ClickHouse ibex.mcp_tool_calls (not llm_traces; ADR-0050)

---

Track G — Exit / Ops Pack-Up

Goal 2.5.G: Prove Phase 2.5 seams operate together with local observability

Milestones:

• 2.5.G7 — Phase 2.5 exit gate: Local LGTM stack (ADR-0051), make verify-phase25, multi-service e2e, thin K8s observability chart, roadmap sign-off.

Acceptance criteria:

• make observability-up / make observability-smoke green with Proxy Critical Path dashboard
• make verify-phase25 maps exit criteria to runnable checks
• Phase roadmap status completed; Phase 3 entry clear