[FEATURE]: @agent:provider/model syntax to override subagent model on the fly

[adambiggs]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

I’d like OpenCode to support an explicit per-subtask model override when invoking subagents, via @agent:provider/model syntax (example: @explore:openai/gpt-5.2).

What I want to add:

• Allow users to specify the model for a specific subagent invocation inline, without changing global defaults.
• Ideally this is supported anywhere @agent invocations are supported (at least TUI first), with autocomplete for available providers/models.
• The chosen model should be stored on the message/subtask metadata so it’s visible/auditable and can be surfaced in the UI.

Benefits:

• Lets users pick a faster/cheaper model for simple subtasks and a stronger model for harder ones, per call.
• Avoids needing multiple nearly-identical agents just to switch models.
• Makes model selection explicit and reproducible when reviewing a session history.

[github-actions]

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

• [FEATURE]: Dynamic model selection for subagents via Task tool #6651: Dynamic model selection for subagents via Task tool (proposes model_tier parameter with global/agent-level mappings)
• [FEATURE]: Commands Model Variant Specification #7713: Commands Model Variant Specification (adding variant field to commands for per-command model overrides)
• [FEATURE]: Tool that lets the model switch models #8278: Tool that lets the model switch models (automation of model picking with instructions)

These issues all address the core need for per-task/per-command model selection without changing global defaults. The proposed @agent:provider/model syntax in this issue provides an alternative approach to the tiered model system proposed in #6651.

Feel free to ignore if this specific syntax approach addresses a gap not covered by those issues.

[farukg]

Love this idea. Related use case that goes a step further: shell-specialist → summarizer pipeline.

I want to delegate shell command execution (fd, rg, eza pipelines) to a cheap code-specialist model (opencode/grok-code), then have the main expensive model (Opus/GPT-5) interpret and summarize the raw results. Today I achieve this with a custom shell-exec subagent pinned to a specific model, but the @agent:provider/model syntax would make ad-hoc overrides much more ergonomic — especially when I want to temporarily test whether a different model handles a specific subtask better, without editing agent config files.