Bug: prompt_async notifications override session's active agent and model

[sjawhar]

Bug Description

When a prompt_async request arrives without explicit agent or model fields (e.g., from Envoy notifications, Legion worker coordination, or any fire-and-forget prompt injection), the session's active agent and model are silently overridden by the default agent's configured defaults.

Steps to Reproduce

1. Start a session and set it to a non-default agent and model (e.g., agent=hephaestus, model=opencode/kimi-k2.5-free)
2. Send a POST /session/:id/prompt_async with only parts (no agent, no model)
3. Observe the next assistant turn uses the default agent's configured model, not the session's active model

Expected Behavior

When prompt_async omits agent and model, the session should preserve whatever agent and model were active from the last user message.

Actual Behavior

The session's agent falls back to the default agent config, and the model falls back to that agent's configured model — discarding whatever the user had actively selected.

Root Cause

In packages/opencode/src/session/prompt.ts, createUserMessage() defaulted omitted agent/model from the default agent configuration instead of reading the current session prompt state (most recent user/assistant message). The synthesized user message then became the new source of truth for ACP session replay, silently redirecting the session.

Impact

This actively affects overnight agent coordination: Envoy notifications delivered to sessions via prompt_async can change the session's agent and model, causing unexpected model switches and errors.

Environment

• OpenCode version: 1.4.0-sami
• Affected endpoint: POST /session/:id/prompt_async
• Also affects: command() (which reuses prompt())

This issue might be a duplicate of existing issues. Please check:

• [Bug] Model parameter ignored when launching subagent with agent name #18615: Model parameter ignored when launching subagent with agent name — similar root cause where agent defaults override explicit model settings in session prompts via the SDK
• [Bug] TUI model selection gets overwritten by agent default model #13456: TUI model selection gets overwritten by agent default model — related issue where agent default model overrides user-selected model in the session

[sjawhar]

Re: potential duplicates flagged by the bot (#18615, #13456):

• [Bug] Model parameter ignored when launching subagent with agent name #18615 shares the same server-side root cause (agent config defaults override explicit model) but describes a different entry point (subagent launch via SDK). The fix in fix(session): preserve active agent and model on prompt_async without explicit fields #21729 addresses the prompt_async / createUserMessage() path and should also fix [Bug] Model parameter ignored when launching subagent with agent name #18615's scenario.
• [Bug] TUI model selection gets overwritten by agent default model #13456 is a separate client-side TUI issue where a createEffect enforces agent defaults — not a duplicate.

PR: #21729

[clarajohan]

I have experienced the opposite of this condition and it resulted in having to reinstall OpenClaw fresh in order to resolve. The main session was picking up its model config from session history, rather than config, and session history model was no longer available (Claude).

We need the ability to reliably configure models from a central location without stale config polluting long-lived sessions.

The selected model should be configured per session, overridable when sending a message. If a model change is intended to be durable, there should be a session model configuration call or parameter that explicitly sets a durable model.