Model & Reasoning-Level Configuration Not Reliably Applied by Agents

[EirikHaughom]

Context

I’ve been testing Squad for about a week across several development projects together with colleagues. My workflow uses different agents for different tasks (e.g., code review vs. implementation) and ideally I want to specify which model and reasoning levels each step or agent uses.

What I Tried

• Defining model usage in charter.md
  • Not reliably picked up.
  • Coordinator appears to override or redefine the model on invocation.
• Storing model preferences in long‑term memory (decisions.md)
  • Example: “Use Opus 4.6 high for code reviews; use gpt‑5.3‑codex xhigh for implementation.”
  • Not consistently followed.
• Adding “Use model X” directly in the prompt
  • Most consistent approach.
  • Downside: easy to forget and repetitive.

Expected Behavior

A reliable mechanism to define:

• Model selection per agent or task
• Reasoning level per agent
• Optionally: model selection per task type or topic (e.g., planning, code review vs. implementation)

Ideally configurable via:

• Agent charter files, or
• A global configuration file, or
• Coordinator-level defaults that agents inherit unless overridden

Actual Behavior

• Agents inconsistently follow model/reasoning-level instructions unless explicitly told in the prompt.
• Charter-based or memory-based definitions are ignored or overridden.

Why This Matters

• Different tasks benefit from different models (e.g., high‑reasoning for code review, codex‑style for implementation).
• Inline instructions create friction and reduce reliability.
• A stable configuration mechanism would improve workflow consistency and reduce cognitive overhead.

Request

Support a more explicit and reliable mechanism for defining model and reasoning-level preferences per agent or per task type.

[bradygaster]

Fixed in PR #245 (merged). AgentDefinition.model now accepts structured ModelPreference with preferred/rationale/fallback fields, squad-level model defaults via defaults.model, and proper charter generation with ## Model section.

[bradygaster]

Love this request, @EirikHaughom! Model + reasoning-level configuration is foundational for multi-agent workflows. Charter-based config that's reliably honored is the right direction—no more friction from inline workarounds. This is actively being built into the new typed system. Thanks for the clear problem statement! ✨

— McManus 📝, on behalf of the Squad