Weekly Update — May 11 – 18, 2026

[missBerg]

Quieter week post-v0.6 — two of last week's "for review" PRs landed (the Anthropic → AWS Bedrock translator and the AIGatewayRoute 15-rule cap), a handful of new design threads opened across shadow-traffic mirroring and governance hooks, and the GA target moved up from end-of-June to mid-June. Thanks to everyone who shipped, reviewed, and weighed in.

🛣️ Road to v1.0 GA

From the maintainer call: v1.0 has been pulled forward to mid-June. Holding the GA through July 4 + the start of European summer would have buried the release, so the RC-to-release window is being compressed to ~1 week to land it ahead of the holiday. If anything slips, the team would rather wait until after summer than ship a low-visibility GA. Tracking issue: #2083.

A 0.6 patch release is also under consideration this week — the open security and compatibility fixes called out in 👀 below are the leading candidates. Separately, quota policy work for 0.7 (#1869) is getting a new push from a contributor at Bloomberg — comments on the existing PR around unit transformation between the API and the rate-limit service still need to be resolved before merge.

✨ What's new

Post-v0.6 it was a quieter week. Two PRs surfaced in last week's 👀 / 💡 lists landed:

• Anthropic /v1/messages → AWS Bedrock translator — #2103 by @Killusions. Lets you point an Anthropic-format client at any Bedrock model — not just Claude — covering messages, system prompts, tools, thinking config, images, streaming (EventStream → SSE), and error translation. The Anthropic-shaped counterpart to today's OpenAI → Bedrock path. Relates to #2102.
• AIGatewayRoute.spec.rules capped at 15 — #2123 by @siddharth1036. The Gateway API enforces maxItems: 16 on HTTPRoute.spec.rules, and the controller adds a catch-all "route-not-found" rule on top — so a 16+ rule AIGatewayRoute silently produced an HTTPRoute the API server rejected, with the AIGatewayRoute still reporting Accepted: True and the data plane on a stale snapshot. Now the failure surfaces at admission time on the resource the operator actually wrote. Closes #2120.

💡 New design proposals & feature requests

A few new threads worth a read if any of these are in your wheelhouse:

• Shadow traffic mirroring for AIGatewayRoute — #2137 by @mtparet. Adds a mirrors field on AIGatewayRouteRule that sends a fire-and-forget copy of each request to one or more shadow backends — with the upstream ExtProc pipeline (model rewrite, header/body mutation) applied to the mirror leg, and LLMRequestCosts suppressed so shadow traffic doesn't double-count in billing. The standard pattern for qualifying a new model or region against live production traffic before flipping routing. Author has a working PoC on a fork.
• Governance policy hook in the LLM ExtProc pipeline — #2128 by @imran-siddique. A route-level gRPC governance hook that runs in upstreamProcessor.ProcessRequestHeaders() after translation and before upstream auth — for model access control ("team-X can only use gpt-4o-mini"), content scanning, and audit. Phased: pre-request allow/deny first, async response audit next, buffered response enforcement later. Open question on integration shape — ExtProc-native hook vs. ext_authz vs. hybrid.
• MCP backend API keys are embedded as plaintext in the generated HTTPRoute — #2141 by @aishwaryaraimule21. MCPRoute backendRefs with a securityPolicy.apiKey (via secretRef) end up writing the resolved plaintext into an HTTPRouteFilter (header) or URLRewrite (query param), so anyone with HTTPRoute read access can recover the token without ever touching Secrets. The proposal points at adopting the BackendSecurityPolicy pattern already used by AIServiceBackend. Initial fix for the header case is up in #2134.

👀 PRs looking for review

If you have time to help review, these are open and waiting:

• #2092 — OAuth 2.0 Token Exchange for MCP upstream auth by @nacx — still the next significant landing, flagged again in this week's maintainer call.
• #2134 — Stop embedding plaintext API keys in MCP backend HTTPRoutes by @aishwaryaraimule21 — security fix for #2141; a strong candidate for a 0.6 patch.
• #2138 — Handle per-backend HTTPRoute cluster names by @mturac — restores upstream ExtProc injection for the 7-part cluster names Envoy Gateway 1.7 emits with priority-based failover, so cross-schema retry (e.g. OpenAI → AzureOpenAI) actually re-translates the request path on the fallback leg. Fixes #2135.
• #1982 — OpenAI Files API by @sivanantha321 — prerequisite for the Batch Processing API proposal.
• #2099 — Reasoning and image support in Anthropic ↔ OpenAI translation by @anurags25.
• #2129 — Media data input for Gemini by @hustxiayang — tightens the input-type mapping so PDFs, video, and audio stop being smuggled in as InlineData on the image path.

🙏 Thanks to this week's contributors

@missBerg, @nacx, @Killusions, @siddharth1036, @aishwaryaraimule21, @mturac, @anurags25, @sivanantha321, @hustxiayang, @cjackal, @ajac-zero, @immanuwell, @taiman724, @arpitjain099, @mtparet, @ismael-dm, @brumhard, @imran-siddique, @azaretsk — and everyone who showed up to triage, review, and discuss.

See you next week!