core: SessionRunner sends forced tool_choice ("required"/"none"/named); providers that only support "auto" fail the whole session

Summary

OpenCode's server-side SessionRunner sends tool_choice values other than "auto" ("none", "required", or named-function choices) on some steps. Providers backed by models/gateways that only implement tool_choice: "auto" (in my case a provider labeled "Console Go") reject these requests with [invalid_request_error], which kills the entire session with "Failed to drain Session" — the subagent/step never completes. OpenCode should either degrade gracefully (retry with "auto"), make forced tool_choice configurable per provider/agent, or at minimum fail a single step instead of the session.

Environment

• opencode version: 0.0.0-beta-18314 (beta channel; from runtime diagnostics snapshot — --version unavailable in this session)
• OS: Linux 6.8.0-138-generic (x64)
• Terminal: Unavailable (no terminal env vars set in this context)
• Shell: /bin/bash
• Install/channel: beta
• Active plugins: none configured (only local/remote MCP servers: github, supabase, vercel, cloudflare)

Reproduction

1. Configure a custom OpenAI-compatible provider whose backend only supports tool_choice: "auto" (observed with a provider labeled "Console Go").
2. Run an agent/subagent turn that causes OpenCode to force a tool choice (observed on subagent runs and structured steps).
3. The provider rejects the request and the session errors out.

Reproduces consistently: the same error appeared across multiple sessions on 2026-08-26 and 2026-08-27 (8+ occurrences in the server log).

Expected Behavior

The session/step should complete. If a provider doesn't support the requested tool_choice mode, OpenCode should fall back to "auto" (possibly with a warning), or the behavior should be configurable per provider/agent rather than hard-forced.

Actual Behavior

Session fails with:

Failed to drain Session
cause="AI.Error: RequestExecutor.execute: Error from provider (Console Go):
Upstream request failed: [invalid_request_error] only `"auto"` is supported for
`tool_choice`. `"none"`, `"required"`, and named function choices are not currently
supported"
at SessionRunner.callModel → SessionRunner.runSteps → SessionRunner.drain

Additional Context

• The request emitter is OpenCode's server session runner; the rejection comes from the upstream model API behind the provider. Sending non-auto tool_choice is valid OpenAI-style behavior, so this is a provider-compat gap — but the current failure mode (whole session dies) is harsh.
• Possibly related: #15226 (tool_choice: "required" incompatible with thinking-enabled models during structured output — closed as not planned) and #37672 (toolChoice plumbing unwired at v2 HEAD).
• No workaround found on the OpenCode side; the practical workaround is to point agents at a provider/model that supports the full tool_choice surface.