[FEATURE]: Plugins can't intercept outbound requests for AI-SDK models (http.request hook excludes the AI-SDK path)

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

The runtime hook ctx.session.hook("http.request", (event) => { … event.request … }) lets a plugin mutate the outbound Web Request immediately before provider dispatch. Per the docs, though, it applies to native models only — AI-SDK-routed models don't pass through it:

From https://opencode.ai/v2/docs/build/plugins:
"HTTP hooks can modify requests and responses. They apply to native models; AI SDK models do not currently pass through these hooks. Request and response bodies are one-shot streams."

The separate ctx.aisdk.hook("sdk" | "language", …) capability covers the AI-SDK seam but exposes only the SDK/model objects, not the outbound HTTP request/body. So for any provider configured with a custom OpenAI-compatible baseURL (self-hosted proxies, gateways like a LiteLLM/router layer, local model servers) — all of which route through the AI SDK — a plugin currently cannot inspect or rewrite the outbound request at all.

Why it matters

This is the same motivation as #19425 ("pre/post API call hooks … for secret redaction and exfiltration prevention"), but for the AI-SDK path that the shipped hook skips:

1. Security / governance (primary). Redact secrets from outbound bodies, enforce egress policy, and audit/log exactly what leaves the machine — for self-hosted-proxy and gateway deployments, which are common in orgs that route everything through one controlled endpoint. Today none of this is possible for AI-SDK providers.

2. Payload shaping. Normalize or strip fields, inject required headers/tenant tags per baseURL, etc.

3. Request-body compression (concrete, broadly useful). Large, growing JSON contexts are re-uploaded every turn. Compressing the body + setting Content-Encoding is a real, standardized optimization that multiple real targets already accept:

   • Google Vertex AI accepts Content-Encoding: gzip on generateContent request bodies (working ... | gzip | curl -H "Content-Encoding: gzip" example: https://github.com/BerriAI/litellm/discussions/3967).
   • Cerebras Inference documents accepting compressed request bodies on /v1/chat/completions (Content-Encoding: gzip, msgpack+gzip): https://github.com/Cerebras/inference-examples/blob/main/payload-compression/simple_example.py.
   • Self-hosted OpenAI-compatible servers (vLLM/FastAPI/Uvicorn-based, and proxies) can decode gzip/zstd request bodies — it's a standard server capability (e.g. FastAPI gzip-request middleware pattern).

   Being explicit so this doesn't overclaim: OpenAI does not support request-body compression (confirmed in that same LiteLLM thread), and Anthropic/Azure/Bedrock don't document it either. The point is not "all providers" — it's that a plugin should be able to compress where the target accepts it (Vertex, Cerebras, self-hosted proxies), scoped by baseURL, and leave the rest untouched. On a constrained uplink a ~600 KB (≈150K-token) body is tens of seconds of pure upload wait per round-trip; JSON compresses ~5–8×.

All three use cases require the one missing capability: a plugin-reachable seam that can rewrite the outbound request (headers + body) on the AI-SDK path, scoped by target.

Ask

Let plugins intercept/rewrite the outbound request for AI-SDK-routed models — either by extending http.request/http.response dispatch to cover the AI-SDK path, or by adding an outbound-request seam to ctx.aisdk.hook (whichever fits the architecture). The event should expose the request (headers + body, replaceable) and enough model/provider/baseURL context that a plugin can scope changes to a specific endpoint and leave all other providers untouched. If this is intentionally out of scope, documenting the supported alternative would also close the gap.

Happy to test a fix end-to-end against a live Content-Encoding: zstd/gzip-accepting endpoint.

Related

Narrows #19425 and #2661 (generic pre-provider / body-rewrite hook requests) to the now-actionable gap: the hook shipped, but the AI-SDK dispatch path is excluded.