andeerc commented 2026年07月10日 14:01:18 +02:00

Summary

Allow users to specify which LLM model a subagent or task should use directly in the prompt, without having to change global configuration.

Motivation

Currently, all agents/tasks run using the globally configured model. There is no way to mix models within a single session or workflow. Being able to target specific models per task opens up powerful patterns:

• Use a cheaper/faster model for simple subtasks (e.g. writing docs, summarizing)
• Use a more capable model for complex reasoning or analysis
• Experiment with different models side-by-side within the same session

Proposed UX

The user should be able to express model selection naturally in the prompt, for example:

Do task X using model Y. Analyze X using model Y and use model Z to write the documentation.

Nibble should parse the model hint and route the subagent or task to the specified model.

Proposed Behavior

• When a user specifies a model in a task description (e.g. "using model Y"), Nibble (or the coordinator) should honour that hint when spawning the subagent for that task
• The model name/alias should match the same identifiers used in nibble.json / models config (e.g. claude-sonnet-4-5, gpt-4o, gemini-2.5-pro)
• If the specified model is not found or unavailable, fall back to the default model and warn the user
• The feature should work both for:
  • Coordinator-spawned subagents (task agent)
  • Explicit user prompts directing a subtask to a specific model

Examples

`

Single model override

Refactor internal/agent/agent.go using claude-opus-4-5.

Multiple model overrides in one prompt

Analyze the performance bottlenecks using gemini-2.5-pro and use gpt-4o-mini to write the summary.

Named model alias

Write the unit tests for internal/hooks using the fast model.
`

Open Questions

• Should model hints be parsed by the coder agent's LLM (semantic parsing) or via explicit syntax (e.g. @model:name)?
• Should there be a shorthand syntax in addition to natural language (e.g. [model=claude-opus])?
• How should this interact with provider authentication — skip the task if the model's provider is not configured?

## Summary Allow users to specify which LLM model a subagent or task should use directly in the prompt, without having to change global configuration. ## Motivation Currently, all agents/tasks run using the globally configured model. There is no way to mix models within a single session or workflow. Being able to target specific models per task opens up powerful patterns: - Use a cheaper/faster model for simple subtasks (e.g. writing docs, summarizing) - Use a more capable model for complex reasoning or analysis - Experiment with different models side-by-side within the same session ## Proposed UX The user should be able to express model selection naturally in the prompt, for example: ` Do task X using model Y. Analyze X using model Y and use model Z to write the documentation. ` Nibble should parse the model hint and route the subagent or task to the specified model. ## Proposed Behavior - When a user specifies a model in a task description (e.g. "using model Y"), Nibble (or the coordinator) should honour that hint when spawning the subagent for that task - The model name/alias should match the same identifiers used in `nibble.json` / `models` config (e.g. `claude-sonnet-4-5`, `gpt-4o`, `gemini-2.5-pro`) - If the specified model is not found or unavailable, fall back to the default model and warn the user - The feature should work both for: - Coordinator-spawned subagents (task agent) - Explicit user prompts directing a subtask to a specific model ## Examples ` # Single model override Refactor internal/agent/agent.go using claude-opus-4-5. # Multiple model overrides in one prompt Analyze the performance bottlenecks using gemini-2.5-pro and use gpt-4o-mini to write the summary. # Named model alias Write the unit tests for internal/hooks using the fast model. ` ## Open Questions - Should model hints be parsed by the coder agent's LLM (semantic parsing) or via explicit syntax (e.g. `@model:name`)? - Should there be a shorthand syntax in addition to natural language (e.g. `[model=claude-opus]`)? - How should this interact with provider authentication — skip the task if the model's provider is not configured?