Model switching for workflow

[MawJPhysics]

I don't understand how this switches between models, if at all.
I am on Claude Pro so I'm limited token wise, from what I understand, it's best to use opus to plan but implementation (aka actual code-writing based on the plan) is more efficiently done on Sonnet. But all your agents are currently set to use opus so when I run something like /orchestrate feature (planner -> tdd-guide -> code-reviewer -> security-reviewer), will it just do it all on opus? Is there a more efficient way? otherwise if i do each command inidivudally, id have to manually sort out the handoffs that orchestrate would have done.

[MawJPhysics]

Reading the performance.md, I think I understand that yes, if you choose a model manually on CLI it will stay constant. But when doing the workflow, performance.md will switch models mid-workflow if appropriate.
I guess that raises another question:
When you use Claude Desktop with the GUI, the moment you switch AI models in a conversation, you clear the entire conversation. I don't know how performance.md would switch AI models without doing the same. Is that why it makes a handoff document instead?

[pangerlkr]

Great questions! You're thinking about this correctly. Let me clarify how model switching works:

Current Default Behavior:
By default, all agents in the /orchestrate workflow are set to Opus. So if you run the full pipeline (planner → tdd-guide → code-reviewer → security-reviewer), everything runs on Opus.

How Model Switching Actually Works:
The performance.md file contains the logic that determines when to switch models mid-workflow. Here's the key insight:

1. Model switching happens at the SKILL level, not within a single conversation
2. Each skill/agent in the orchestration chain can have its own modelConfig
3. When you transition from one skill to the next, performance.md evaluates:
   • What was the output from the previous step?
   • Do we need more reasoning power (use Opus)?
   • Can we optimize with faster/cheaper model (use Sonnet)?

Why Handoff Documents?
You nailed it! The handoff document is crucial because:

• It preserves context from the previous step
• It allows the next agent (with potentially different model) to pick up without needing the full conversation history
• Without it, switching models mid-workflow would indeed lose context

How to Optimize for Your Use Case:
To use Opus for planning and Sonnet for implementation:

1. Create custom agent configurations in settings.json:

   "planner": { "model": "claude-3-5-opus-latest" },
   "code-reviewer": { "model": "claude-3-5-sonnet-latest" },
   "security-reviewer": { "model": "claude-3-5-sonnet-latest" }

2. This way, orchestrate will respect the model assignments per agent

3. The handoff documents will carry the essential context between model switches

This gives you the cost efficiency of Sonnet for implementation while keeping Opus for complex planning and review!

[ccarlosmrz100]

• @claude 

27/log_android_1769350321502.zip)

[affaan-m]

The short answer is: by default, no automatic model switch happens inside a single chat. The workflow uses whatever model/session you are currently running unless a specific tool or adapter creates a separate handoff boundary.

A practical way to think about ECC now:

• planning/review steps can be run in a higher-reasoning session
• implementation/cleanup can be run in a cheaper or faster session
• the handoff is the artifact itself: plan, diff, checklist, eval output, or review notes

So if you want “Opus for planning, Sonnet for implementation,” you should treat those as separate steps or separate sessions rather than expecting one conversation to transparently hot-swap models without losing state.

That is also part of why newer ECC work is moving toward explicit harness controls like audit/routing/eval gates instead of implicit magic model switching.

If useful, I can turn that into a short README section with recommended model-routing patterns for Pro-tier users.