[Automated] Draft docs (agentgateway): llm: respect buffer limits on various paths

[github-actions]

agentgateway/agentgateway#3896 — llm: respect buffer limits on various paths

Docs versions: standalone/main, kubernetes/main

What changes for users: LLM conversion paths now bound memory growth. Responses namespace tool descriptions fail with RequestTooLarge when copied descriptions exceed 8 MiB for one request, and Bedrock Runtime streams return an api_error event with Stream processing error when buffered redacted reasoning fragments exceed the stream buffer limit before the block ends.

What the docs now say: The standalone Responses page documents the namespace tool description cap. The standalone Messages page and the Kubernetes Bedrock provider page document the streamed Bedrock redacted-reasoning error.

Not verified: No cluster or live Bedrock provider was available; the configuration and stream behavior are unrun.

How agentgateway-3896 was drafted, and what was not verified

• Code PR: llm: respect buffer limits on various paths agentgateway#3896
• Code issue: none linked
• Product: agentgateway (upstream) → agentgateway/website
• Docs version: standalone/main and kubernetes/main
• Category: new-feature / pr

Plan, and what changed it

• Planned: Cover each affected documentation tree because the planner found LLM conversion targets in both standalone/main and kubernetes/main.
• Learned: The strongest authored homes were the standalone API type pages and the shared Kubernetes Bedrock provider asset; the generated CEL reference already covers guardrails[].action, and the candidate guardrails overview pages were not stale.
• Did: Added the namespace-description limit to content/docs/standalone/main/documentation/llm/api-types/responses.md, added the Bedrock redacted-reasoning stream error to content/docs/standalone/main/documentation/llm/api-types/messages.md, and added the same Bedrock behavior behind a main-only gate in assets/agw-docs/pages/agentgateway/llm/providers/bedrock.md.

Why a documentation change is necessary

The pull request changes behavior that a client can observe when large converted LLM streams or namespace tool requests exceed internal limits. The Responses page already documents namespaced tool conversion, and the Messages and Bedrock provider pages already document Bedrock encrypted or redacted reasoning behavior. Without these edits, those pages describe the converted formats but not the new failure modes that stop the request or stream.

What changed on disk

• content/docs/standalone/main/documentation/llm/api-types/responses.md: Documents the 8 MiB per-request cap for copied namespace tool descriptions.
• content/docs/standalone/main/documentation/llm/api-types/messages.md: Documents the Bedrock Runtime stream error for oversized redacted reasoning fragments.
• assets/agw-docs/pages/agentgateway/llm/providers/bedrock.md: Documents the same Bedrock stream behavior for Kubernetes main and future versions.

How to verify this change

1. Send a Responses request with namespace tools whose copied namespace descriptions exceed 8 MiB while the selected provider requires flattening to function tools.
2. The request fails before reaching the provider with RequestTooLarge.
3. Send a streamed /v1/messages request to Bedrock Runtime that returns redacted reasoning fragments larger than the stream buffer limit before the content block ends.
4. The client receives an api_error stream event with the message Stream processing error, and the stream stops.

What was not verified

• The configuration was not applied to a cluster, and the standalone requests were not run.
• The stream error output was not captured from a live Bedrock Runtime response.
• The exact user-facing field that sets the stream buffer limit was not identified from the dossier, so the pages name the behavior without a configuration path.

How the pages were chosen (triage report)

New configuration: no page mentions it yet

agentgateway/agentgateway#3896 — llm: respect buffer limits on various paths

• Product: agentgateway (upstream) → agentgateway/website
• Docs version: standalone/main, kubernetes/main
• Summary (read from the body): - responses (Responses->Messages): drop the never-read ToolBlock.arguments buffer that kept every tool-call argument delta for the whole stream. - completions (Messages->Completions): only collect streamed tool-call arguments when tool-call logging is enabled. - bedrock (Converse->Messages): cap buffered redacted-reasoning bytes at buffer_limit and send a stream error when it's exceeded; fragm
• Not tested: no cluster was available, so every command in this diff is unrun

Release notes

This pull request also adds its release-note entries:

• content/docs/kubernetes/main/release-notes/release-notes.md: llm: respect buffer limits on various paths (fix, under LLM)
• content/docs/standalone/main/release-notes/release-notes.md: llm: respect buffer limits on various paths (fix, under LLM)

Check these release-note links and names. Each note is on its page, but a link may not reach the drafted content there, or a name may belong to the other mode:

• content/docs/kubernetes/main/release-notes/release-notes.md: llm: respect buffer limits on various paths agentgateway#3896's note names Stream processing error, which content/docs/standalone/main mentions and content/docs/kubernetes/main does not. Check that this tree's readers can set it, or word the note for this mode.
• content/docs/kubernetes/main/release-notes/release-notes.md: llm: respect buffer limits on various paths agentgateway#3896's note names RequestTooLarge, which content/docs/standalone/main mentions and content/docs/kubernetes/main does not. Check that this tree's readers can set it, or word the note for this mode.

Draft branch: pr-tracker-draft-agentgateway-3896

[cloudflare-workers-and-pages]

Deploying agentproxy with    Cloudflare Pages

Latest commit:	baa4c46
Status:	✅  Deploy successful!
Preview URL:	https://d06395a5.agentproxy.pages.dev
Branch Preview URL:	https://pr-tracker-draft-agentgatewa-xg0o.agentproxy.pages.dev

View logs