v1.0.0¶

Release date: 2026-08-23 Tag: v1.0.0 (annotated tag object 77b27fa, pointing at commit 7e0d24f) Architecture freeze: SGW-V3-DESIGN-FREEZE-2026-08-20-01 Routing policy freeze: SGW-BLUEPRINT-2026-08-18-01

The first production release of the Smart LLM Supervisor Gateway.

Highlights¶

OpenAI-compatible surface. POST /v1/chat/completions with streaming, POST /v1/route/decision for dry runs, GET /healthz, GET /metrics.
Four routing aliases. smart-router, smart-router-flash, smart-router-pro, smart-router-free. See User guide → Model aliases.
Frozen model registry. Nemotron 3 Ultra (Scout), DeepSeek V4 Flash (worker), DeepSeek V4 Pro (senior), GLM-5.2 (emergency fallback). See Architecture → Model roles.
Privacy gate. Sensitive by default; free models refused on sensitive content; secret-pattern redaction with automatic sensitive upgrade. See Security → Sensitive data.
Budget ledger. Per-goal and per-day soft / hard caps with PostgreSQL-backed atomic reservation. Fail-closed on hard cap. See Architecture → Budget.
Model health registry. Per-model state machine with HEALTHY / DEGRADED / UNHEALTHY / HALF_OPEN / DISABLED / STALE / QUARANTINE and substitution into a senior fallback chain. See Reference → Model registry.
Production deployment. smart-openrounter.bee1x.one, 46.250.252.0, nginx + certbot, isolated Docker network, separate PostgreSQL volume. The full production topology is documented in docs-internal/operations/production-topology.md (operator-only).
Evidence. 60 real OpenRouter calls across four systems and 15 tasks (POST-RELEASE-E4). Real coding client (Aider 0.86.2, 27 chat-completions, 0 privacy violations, fixture diff 4,617 bytes — POST-RELEASE-E3). Independent review: PASS, 0 BLOCKER, 0 HIGH.

Evidence¶

Artifact	Path
Final completion report	`PROJECT-COMPLETION-REPORT.md`
Production deployment evidence	`artifacts/WAVE-H/`
Post-production validation	`artifacts/WAVE-I/`
Real coding client (Aider)	`artifacts/POST-RELEASE-E3/`
Real evaluation (60 calls)	`artifacts/POST-RELEASE-E4/`
Final independent review	`artifacts/FINAL-EVIDENCE-REVIEW/`
Roadmap	`AUTONOMOUS-ROADMAP.md`

What's intentionally absent¶

The OpenAI Responses API. Use Chat Completions.
The Anthropic /v1/messages shape.
Multimodal / vision input.
Multi-tenant RBAC.

Known issues and limitations¶

Architectural finding: API key value not verified¶

The current implementation checks that the bearer token is non-trivial in length, but does not verify the value against a server-side allowlist. Rotation is run for hygiene; it does not strengthen authentication by itself. This is tracked as an open architectural decision in PROJECT-COMPLETION-REPORT.md §"Unresolved Items" and in the v1.0.0 doc note at User guide → Authentication.

Streaming UX with DeepSeek extended thinking¶

The streaming normalizer folds reasoning-only deltas into the first content delta so that OpenAI-compatible clients (notably Aider) see a non-empty response. For very long reasoning passages the first content chunk may be large; if you need byte-perfect reasoning passthrough, use --no-stream.

Nemotron free endpoint terms¶

Nemotron 3 Ultra's free endpoint on OpenRouter explicitly warns against confidential or personal data. The gateway treats this as a hard boundary: smart-router-free on sensitive=true returns 403.

Upgrade path¶

v1.0.0 is the first production release. There is no upgrade path from a prior version. Future changes will be issued as v1.x or v2.x releases with documented migration steps.