Releases
ModelPanel releases — M1 core gateway, M2 streaming + Anthropic + notices + fusion, the ModelTap → ModelPanel rename, and the benchmark evidence trail.
Gateway v0.7.0 — September 25, 2026
shipped- User-curated registry — the gateway ships no provider roster.
registry addpreviews and writes routes with base URL, tier, and context;registry deleterecords a tombstone so catalog refreshes do not resurrect removed routes. - Empty is a valid state — an absent registry file boots with no user routes; a missing registry directory or invalid bytes still fail closed. The
modelpanel/fusionroute remains intrinsic to the gateway. - One adapter, configuration-driven providers — provider URLs and quirks live in the registry, not in a shipped list or adapter code.
- Downgrade hazard — gateways at or below 0.6.0 do not understand the
removedfield; upgrade before pointing an older gateway at a 0.7.0 registry.
Gateway M2 — September 2026
shipped- Anthropic dialect — ingress parser, dialect writer,
/v1/messagesroute (non-streaming + streaming), commit-point semantics — E2E-verified against Claude Code. - Streaming everywhere: chat-completions with fallback-before-first-token, Responses with typed events + dialect-shaped SSE errors, Anthropic SSE.
- Remaining adapters: OpenRouter, Mistral, Gemini, LLM7 (joining Groq, NIM, Cerebras) — plus
store:falserespect,internal_errormapping, null-safe tools filter. - Opportunity notices: NoticeEngine (same-id/tier/keyless/paid-allow gates, TTL dedupe) wired into chat/responses bodies + streamed terminal events; header-safe messages.
- Fusion sidecar (M2 T5): Python stdlib, OpenAI-compatible on the wire, panel recipes, streaming judge, failure masks — plus fused catalog via the sidecar, keyless activation, and hot-reload combo-duplication regression tests.
- Claude Code E2E + quickstart docs (signup links,
pnpm dev,ANTHROPIC_BASE_URLlines).
Gateway M1 — August 2026
shippedNine tasks, spec-traceable:
- Scaffold + config; registry loader with hot-watch (fail-fast boot, combo dedup)
- Key vault with tier flags (
free-permanent,free-credits,discounted,paid), env overrides, chmod 600 - Canonical types + Chat Completions ingress; Responses ingress (priority dialect) with builtin-tool rejection
- OpenAI-compatible egress adapter (config-driven quirks)
- Router: registry order, tier policy, budget skip
- Fastify wiring + fallback loop + TTL response store
- CLI:
serve,add-key,models,status(+doctor)
Pipeline and evidence — August 2026
- Benchmark evidence: matched GSM1K (n=40) — panel synthesize 95.0% vs best single 90.0% vs self-consistency 82.5%; reasoning roster from matched gsm8k+gsm1k evidence (nemotron/minimax/gpt-oss, n=80 pooled).
panel_selector --apply— evidence updates the registry chains the router reads.migrate_vector_identity-style probe work on the tracker side: 30+ providers verified, dead-marking,reasoning_effortcalibration, model-card anchors.
Rename — ModelTap → ModelPanel
2026-08-30The name encodes the differentiator: Mixture-of-Agents is a hand-picked mixture in a paper; ModelPanel is a living, verified, re-selected panel (panels.py, panel_selector, --members). Canonical repo: open-assistants-lab/ModelPanel. Family: HybridDB · CoreMem · ConnectKit · AgentProfile · ModelPanel.
The design spec (2026-08-24, revised 2026-08-30) is the plan of record: registry-driven providers, per-provider discovery, tier taxonomy, fusion promoted to MVP, and the notices model — all shipped across M1/M2 as scoped.
Source of truth for this page: ModelPanel · open-assistants-lab/ModelPanel