POST /session/{id}/message crashes with UnknownError when "agent" is included in the body

Description

POST /session/{id}/message (operationId: session.prompt) crashes with a generic UnknownError whenever the request body includes an agent field — reproduced with three different payload shapes (agent alone, agent+format, agent+format+model), all producing the identical stack trace within ~100ms of session creation, before any provider call is attempted.

Separately, omitting agent from the message body does not fall back to the session's configured agent — it silently uses the default build agent instead, even though the session was created with a custom agent. So there is currently no way to run a message against a non-default agent through this endpoint at all: specifying it crashes, and omitting it silently uses the wrong agent.

This also blocks using format: json_schema structured output together with a custom agent, since format lives on the same endpoint.

Steps to reproduce

export OPENCODE_CONFIG="/path/to/opencode.json"   # must define the custom agent + provider
opencode serve --port 4096 --print-logs --log-level INFO &

# 1. Create a session scoped to a custom agent + provider (this works correctly)
curl -s -X POST "http://127.0.0.1:4096/session?directory=/path/to/project" \
  -H "Content-Type: application/json" \
  -d '{"agent":"my-custom-agent","model":{"providerID":"my-provider","id":"my-model"}}'
# -> returns a session with agent=my-custom-agent, model.providerID=my-provider (confirmed correct in server logs)

# 2. Send a message WITHOUT "agent" in the body (replace <id> with the session id from step 1)
curl -s -X POST "http://127.0.0.1:4096/session/<id>/message" \
  -H "Content-Type: application/json" \
  -d '{"parts":[{"type":"text","text":"hello"}]}'
# -> succeeds (200), but server logs show `agent=build`, not `agent=my-custom-agent` -- the
#    session's configured agent is silently ignored.

# 3. Send a message WITH "agent" in the body, to try to force the correct agent
curl -s -X POST "http://127.0.0.1:4096/session/<id>/message" \
  -H "Content-Type: application/json" \
  -d '{"parts":[{"type":"text","text":"hello"}],"agent":"my-custom-agent"}'
# -> 500, body: {"name":"UnknownError","data":{"message":"Unexpected server error. Check server logs for details.","ref":"err_..."}}

Server log for the step-3 crash

error="UnknownError: UnknownError" cause="UnknownError: UnknownError
    at <anonymous> (B:/~BUN/root/chunk-0acbd0yq.js:1085:2069)
    at SessionPrompt.createUserMessage (B:/~BUN/root/chunk-0acbd0yq.js:1085:9954)
    at SessionPrompt.createUserMessage (definition) (B:/~BUN/root/chunk-0acbd0yq.js:1085:1827)
    at SessionPrompt.prompt (B:/~BUN/root/chunk-d8mqra64.js:4:13220)
    at SessionPrompt.prompt (definition) (B:/~BUN/root/chunk-0acbd0yq.js:1085:9838)
    at SessionHttpApi.prompt (B:/~BUN/root/chunk-d8mqra64.js:2:84408)
    at SessionHttpApi.prompt (definition) (B:/~BUN/root/chunk-d8mqra64.js:4:13133)

The crash happens with format entirely absent, so it is specifically triggered by the agent field, not by format/structured output.

I also tried the newer v2 API (POST /api/session/{id}/agent, operationId: v2.session.switchAgent) as a possible workaround to set the agent before prompting, but v2's prompt endpoint (POST /api/session/{id}/prompt, PromptInput body schema: text, files, agents) has no format field at all, so it can't carry structured output either way.

Environment

• OpenCode version: 1.18.14 (latest on npm as of this report)
• Provider: custom OpenAI-compatible (@ai-sdk/openai-compatible), self-hosted sglang server
• OS: Windows 11 (also expected to reproduce on Linux/macOS -- nothing OS-specific in the repro)
• Invocation: raw HTTP against opencode serve, not the CLI's run/--agent flag (which works correctly since it doesn't go through this exact code path)

Impact

There is no working way to combine a custom agent with format: json_schema structured output via the HTTP API in this version. This blocks using engine-enforced structured output for any project that relies on custom agent definitions rather than the default build/plan agents.