v1.1.2

[1.1.2] — 2026-06-06

Release hardening patch for the external model catalog cutover, catalog pricing correctness, and response-body lifecycle fixes. No public API breaks. Fixes every issue labelled release-1.1.2: #132, #133, #134, and #185.

Fixed

• Azure OpenAI, Azure Foundry, and Vertex AI catalog pricing aliases (issue #132): gateway provider IDs such as azure-openai, azure-foundry, and vertex-ai now resolve against the catalog's canonical provider prefixes (azure_openai, azure_foundry, azure, and vertex_ai). Cost calculation walks the provider-specific fallback chain, skips unpriced preferred chat entries when a priced fallback exists, and includes explicit regression coverage for Azure and Vertex pricing.
• OpenAI response-body lifecycle (issue #185): non-streaming OpenAI-compatible responses now drain the HTTP body after decoding so transports can reuse connections, and streaming responses close the OpenAI stream when the gateway stream goroutine exits.
• Provider and circuit-breaker lookup race (PR #170): routing strategy provider lookup now snapshots the provider and circuit-breaker maps under the gateway lock before dispatch, avoiding concurrent map access during runtime discovery or config reload.
• Admin API key mutation leak (PR #190): key-management APIs now return copied in-memory key records so callers cannot mutate the store's internal state through returned pointers.
• Shutdown hook regression coverage (PR #171): added unit-level tests around hook shutdown behavior so future edits keep close/drain semantics intact.

Changed

• Model catalog loading now consumes the external release artifact (issue #133): the default catalog source is the latest ferro-labs/model-catalog release artifact. Gateway startup and refresh use remote-first loading with the embedded catalog_backup.json as fallback, preserving offline startup while allowing catalog updates without an ai-gateway release.
• Catalog lookup uses a reverse model-ID index (issue #134): bare model-ID lookups avoid scanning the full catalog, while preserving the previous behavior for arbitrary caller-constructed Catalog values through validation and fallback scanning.
• Streaming content matching precompiles regular expressions (PR #189): repeated streaming content checks no longer compile regexes on the hot path, and config validation fails fast on invalid patterns.
• CI and release workflows use Go 1.25.11: vulnerability scanning now runs against the patched Go 1.25 toolchain so standard-library govulncheck findings match the release environment rather than a stale host toolchain.

Added

• Catalog coverage guardrail: added a provider/catalog coverage test that verifies registered providers either have priced catalog entries for representative models or are explicitly documented as dynamic/no-prefix exclusions.
• Catalog backup refresh guardrails: added the scripts/refresh_catalog_backup.sh helper and release workflow checks to keep the embedded fallback catalog aligned with the external catalog artifact.
• Catalog load observability: added gateway_catalog_loads_total{source,result} metrics and structured remote/fallback logging with catalog URLs sanitized for credentials and query strings.

Notes

• models/catalog.json was removed from the repository. Runtime catalog loading now uses the remote release artifact plus the embedded models/catalog_backup.json fallback.
• The release notes generator reads this 1.1.2 section directly when publishing the v1.1.2 GitHub release.

Contributors

Thanks to everyone who shipped this release:

• Release/v1.1.2 (#188) — @Rachit-Gandhi @toller892 @ibobgunardi @MitulShah1

---

Full changelog: https://github.com/ferro-labs/ai-gateway/blob/v1.1.2/CHANGELOG.md