fix(opencode): preserve prompt tool enables with empty agent permissions

[jquense]

Issue for this PR

Closes #17063

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

When an agent has permission: {}, prompt-level tool enables can be dropped before the model request is built.

This PR applies session/prompt permission overrides in the same pre-request filtering path that decides which tools are exposed to the model, and adds a regression test for the tools: { question: true } case.

How did you verify your code works?

• added regression coverage in packages/opencode/test/session/llm.test.ts
• reproduced the bug with a local SDK repro script
• confirmed the repro succeeds against the local dev server with this change

Repro script used for testing

packages/sdk/js/example/question-permission-repro.ts

import { createOpencodeClient } from "@opencode-ai/sdk/v2"
import { spawn } from "node:child_process"

const key = process.env.ANTHROPIC_API_KEY

if (!key) {
  throw new Error("Set ANTHROPIC_API_KEY before running this repro")
}

const port = 4100 + Math.floor(Math.random() * 1000)
const proc = spawn("bun", ["run", "dev", "--", "serve", `--hostname=127.0.0.1`, `--port=${port}`], {
  cwd: new URL("../../../../", import.meta.url),
  env: {
    ...process.env,
    OPENCODE_CONFIG_CONTENT: JSON.stringify({
      enabled_providers: ["anthropic"],
      provider: {
        anthropic: {
          options: {
            apiKey: key,
          },
        },
      },
      agent: {
        repro: {
          mode: "primary",
          model: "anthropic/claude-sonnet-4-5-20250929",
          permission: {},
        },
      },
    }),
  },
})

proc.stdout?.pipe(process.stdout)
proc.stderr?.pipe(process.stderr)

const url = await new Promise<string>((resolve, reject) => {
  let out = ""
  const timeout = setTimeout(() => done(new Error(`timed out waiting for dev server\n${out}`)), 15000)
  const done = (err?: Error) => {
    clearTimeout(timeout)
    proc.stdout?.off("data", onStdout)
    proc.stderr?.off("data", onStderr)
    proc.off("error", onError)
    proc.off("exit", onExit)
    if (err) reject(err)
  }
  const onStdout = (chunk: Buffer) => {
    out += chunk.toString()
    for (const line of out.split("\n")) {
      const match = line.match(/opencode server listening on\s+(https?:\/\/\S+)/)
      if (!match) continue
      done()
      resolve(match[1]!)
      return
    }
  }
  const onStderr = (chunk: Buffer) => {
    out += chunk.toString()
  }
  const onError = (err: Error) => done(err)
  const onExit = (code: number | null) => done(new Error(`dev server exited early: ${code}\n${out}`))
  proc.stdout?.on("data", onStdout)
  proc.stderr?.on("data", onStderr)
  proc.on("error", onError)
  proc.on("exit", onExit)
})

const client = createOpencodeClient({ baseUrl: url })
const session = await client.session.create()
const sessionID = session.data!.id

console.log("session:", sessionID)
console.log("sending prompt")

const result = await client.session.prompt({
  sessionID,
  agent: "repro",
  tools: {
    question: true,
  },
  parts: [
    {
      type: "text",
      text: "do you have a question tool?",
    },
  ],
})

console.log(result.data?.parts?.find((part) => part.type === "text")?.text)

proc.kill()

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍

[rekram1-node]

/review

[github-actions]

lgtm