Plugin request hooks don't run for experimental.generate.text #62096

Description

ctx.experimental.generate.text (also ctx.generate.text and POST /api/experimental/generate/text) never runs the model.request, http.request or http.response hooks. Those hooks are the only place a plugin can rewrite outbound model requests, so any provider that depends on them works in sessions and fails in transient generation.

My case: a local plugin registers a claude provider (a copy of the Anthropic catalog) and uses http.request to turn requests into Claude subscription requests (auth headers, the required system block). claude/haiku-5.5 works in sessions. A plugin that builds prompt autocomplete on generate.text with the same model gets the unmodified request, and Anthropic rejects it:

{"_tag":"ProviderRequestError","providerID":"claude","reason":"rate_limit","message":"Error","status":429}

Session hooks also can't change the model, so plugins have no workaround.

Expected: provider-level request hooks also run for transient generation, for example http.request/http.response with kind: "generate" and no session, or a provider-scoped request hook that doesn't depend on a session.

Plugins

A minimal probe plugin (below). Original case: a local Claude subscription provider plugin and opencode-recall autocomplete.

OpenCode version

2.2.5

Steps to reproduce

1. Register this plugin in an otherwise empty config:

   import { appendFileSync } from "node:fs";
   const log = (line) => appendFileSync("/tmp/hooks.log", line + "\n");
   export default {
     id: "probe.http-hooks",
     async setup(ctx) {
       log("setup");
       await ctx.session.hook("model.request", (e) => log(`model.request kind=${e.kind}`));
       await ctx.session.hook("http.request", (e) => log(`http.request kind=${e.kind}`));
       await ctx.session.hook("http.response", (e) => log(`http.response kind=${e.kind} status=${e.response.status}`));
     },
   };

2. Run opencode api --standalone post /api/experimental/generate/text -d '{"prompt":"Reply with exactly: ok","model":{"providerID":"openai","id":"gpt-6-luna-fast"}}'. It returns {"data":{"text":"ok"}}, and /tmp/hooks.log only contains setup.

3. Run opencode run --standalone -m openai/gpt-6-luna-fast "Reply with exactly: ok". The log now shows model.request, http.request and http.response for kind=primary and kind=title.

Operating System

CachyOS Linux (kernel 7.0.8)

Terminal

Not terminal related (reproduced through opencode api)