Who contributed to Anthropic on May 20, 2026?

1 developer shipped this update, including tobinsouth.

What were the notable Anthropic updates?

interceptor: streaming RPC support via Stream-shaped intercept_streaming (#121), payload: AnyMessage and Payload — type-erased, lazily-decoded message bodies (#113), and dispatcher: pass Payload to call_unary so handlers reuse the interceptor decode (#119).

@anthropics

Claude SDKs and developer tooling

github ↗

Pick a date

Topics: AI / ML Full archive →

The Wire · Showcase

CONNECT-RUST CLOSES STREAMING RPC SECURITY GAP, INTERCEPTOR CHAIN NOW COVERS ALL CALL SHAPES

By RepoJournal · Filed 06:02 UTC on May 20, 2026 · About Anthropic

1 person shipped this

tobinsouth @tobinsouth 1 cited

Streaming RPCs have been bypassing the entire interceptor chain in connect-rust - a security hole that's fixed as of today with full Stream-shaped intercept support.

The interceptor chain in connect-rust covered only unary RPCs [1], leaving server-streaming, client-streaming, and bidirectional streaming calls to skip authorization entirely. That's not a gap, that's a vulnerability. PR #121 adds Interceptor::intercept_streaming, a Stream-shaped interceptor that covers all three streaming shapes and closes the chain. In parallel, the stack is tightening around message decoding efficiency: Payload now type-erases and lazily decodes message bodies [2], so when an interceptor reads the request body, that decode is cached and reused by handlers instead of decoding twice [3]. HttpClientBuilder gets connect_timeout control [4], letting callers on flaky network paths (NAT gateways under churn, silent SYN drops) fail fast instead of waiting 130 seconds on kernel defaults. And Server now proxies with_interceptor and with_interceptor_arc directly [5], so standalone Server users don't have to drop down to ConnectRpcService constructors. On the Claude plugins side, first-party integrations from Apollo, Appwrite, Atlassian, and 21 other major-brand orgs are now promoted and SHA-pinned [6]. A separate sweep bumps 25 plugin pins to upstream HEAD across the h-r range [7]. Over on claude-code-action, the allowed_bots check now fires after actor account type resolution [8], keeping the allowlist correctly scoped to bot/app accounts. Three point releases shipped: v1.0.127 with the actor resolution refactor [9], plus v1.0.126 and v1.0.125 [10], [11]. One critical fix: claude-plugins-official now quotes CLAUDE_PLUGIN_ROOT in all five hook commands [12], unblocking Windows paths with spaces that were erroring on every tool call.

Action items

→ Merge and deploy interceptor streaming support (#121) before adding any new streaming RPC handlers anthropics/connect-rust [immediate]
→ Pin the CLAUDE_PLUGIN_ROOT quote fix if running plugins on Windows anthropics/claude-plugins-official [immediate]
→ Review Payload caching behavior in your request body inspection interceptors anthropics/connect-rust [plan]
→ Upgrade claude-code-action to v1.0.127 for correct allowed_bots scoping anthropics/claude-code-action [plan]

References

[1] interceptor: streaming RPC support via Stream-shaped intercept_streaming (#121) anthropics/connect-rust
[2] payload: AnyMessage and Payload — type-erased, lazily-decoded message bodies (#113) anthropics/connect-rust
[3] dispatcher: pass Payload to call_unary so handlers reuse the interceptor decode (#119) anthropics/connect-rust
[4] client: HttpClientBuilder with connect_timeout (#117) anthropics/connect-rust
[5] server: proxy with_interceptor / with_interceptor_arc to ConnectRpcService (#123) anthropics/connect-rust
[6] Add 24 first-party plugins from major-brand orgs (#1919) anthropics/claude-plugins-official
[7] Bump 25 plugin SHA pins to upstream HEAD (huggingface–railway) ↗ anthropics/claude-plugins-official
[8] Resolve actor account type before applying allowed_bots (#1330) anthropics/claude-code-action
[9] v1.0.127 ↗ anthropics/claude-code-action
[10] v1.0.126 ↗ anthropics/claude-code-action
[11] v1.0.125 ↗ anthropics/claude-code-action
[12] fix: quote \${CLAUDE_PLUGIN_ROOT} in hookify and security-guidance hook commands anthropics/claude-plugins-official

Quick answers

What shipped in Anthropic on May 20, 2026?: Streaming RPCs have been bypassing the entire interceptor chain in connect-rust - a security hole that's fixed as of today with full Stream-shaped intercept support. In total, 17 commits, 15 pull requests, and 4 releases landed.
Who contributed to Anthropic on May 20, 2026?: 1 developer shipped this update, including tobinsouth.
What were the notable Anthropic updates?: interceptor: streaming RPC support via Stream-shaped intercept_streaming (#121), payload: AnyMessage and Payload — type-erased, lazily-decoded message bodies (#113), and dispatcher: pass Payload to call_unary so handlers reuse the interceptor decode (#119).

CLAUDE AGENT SDK FIXES STDIN CLOSURE BUG IN BACKGROUND TASKS

The Python SDK's query method was prematurely closing stdin when background subagents were still in flight, breaking tool calls and silently bypassing hooks - now fixed in v0.2.127.

ai 16 shipped 1-min read

@anthropics 1 day ago

BUFFA v0.9.1 PATCHES REGRESSION THAT BROKE ORDINARY SCHEMAS

Buffa shipped an emergency patch overnight after v0.9.0 broke codegen on schemas with a few hundred proto files, small enough to hit in production.

ai 113 shipped 1-min read

@anthropics 2 days ago

CLAUDE FOUNDATION MODELS ADDS APP ATTEST, PLUGIN ECOSYSTEM STABILIZES

iOS developers can now ship Claude calls without API keys in the binary, backed by Apple's Secure Enclave.

ai 111 shipped 1-min read

@anthropics 3 days ago

BUFFA FIXES V0.9.0 REGRESSION, CLAUDE PLUGINS ECOSYSTEM GAINS 9 NEW INTEGRATIONS

Buffa's codegen was silently rejecting moderately large schemas due to a memory bound meant only for untrusted input, but a fix shipped overnight restores throughput and clears the path for googleapis-scale proto sets.

ai 156 shipped 2-min read

Elsewhere on the wire

AI Agents about 10 hours ago

CLAUDE OPUS 5 LANDS ACROSS THE STACK

The newest Anthropic model is now live in langchain, Cline, and llama-index, with native support for extended reasoning and 1M context windows.

ai-agents 28 shipped 1-min read

Local LLMs about 10 hours ago

OLLAMA LANDS LAGUNA SUPPORT AND CRUSHES MEMORY LEAKS WHILE SGLANG HITS V0.5.16 WITH CONFIDENCE-DRIVEN SPECULATIVE DECODING

Ollama shipped three critical performance and reliability fixes for Metal residency and concurrent access patterns, while SGL-Lang released 0.5.16 with a new speculative algorithm hitting 383.7 tok/s on DeepSeek-V4.