fix: preserve explicit model selection in command subtasks

[Haohao-end]

Issue for this PR

Closes #18615

Type of change

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

What does this PR do?

This fixes model precedence in the command/subtask delegation path.

The bug was that an explicit model input could still be overridden by an agent-configured fallback model when resolving a command/subtask request.

I changed the precedence in packages/opencode/src/session/prompt.ts so that it is now:

1. explicitly provided model
2. command-configured model
3. command agent fallback model
4. previous / inherited model fallback

Why this works:

• the explicit model is now preserved when a command/subtask is created
• existing fallback behavior is still preserved when no explicit model is provided

How did you verify your code works?

I tested locally with:

• cd packages/opencode && bun test test/session/prompt.test.ts --timeout 30000
• cd packages/opencode && bun run typecheck

I also added focused regression coverage to verify that explicit model selection is preserved in the command/subtask path.

Screenshots / recordings

N/A (not a UI change)

Checklist

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

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Based on my search, I found one potentially related PR:

PR #17788: "fix: prioritize CLI --model argument over command markdown model"

• fix: prioritize CLI --model argument over command markdown model #17788
• This addresses a similar model precedence issue, prioritizing explicit model arguments over fallback models in commands

This PR is related because both deal with model precedence and ensuring explicit model selections aren't overridden by fallback models. However, PR #17788 focuses on the CLI --model argument path, while PR #18752 addresses the subtask/command delegation path.

[gmnstr]

I noticed this PR while researching subagent model-routing behavior and ended up opening #26117 before I found it. The two overlap heavily.\n\nOne difference in #26117 is that it keeps the internal delegated model scoped to command-triggered subtasks only, so ordinary/manual subtasks keep the existing fallback path. It also explicitly avoids pre-seeding child session agent/model after I hit a switch-event/history regression during QA, and adds regression coverage around fallback behavior and non-authoritative session metadata.\n\nIf it’s useful, I’m happy to help fold those guardrails/tests into this PR instead of splitting reviewer attention across two overlapping fixes.