Model-Per-Mode

🔧 In Workshop — Not Built Yet

The Problem

Sonnet's execution gradient fires even in workshop mode. The tool gate stops the actions — but the conversational gradient still leaks. Archie says "I'll remember that" or "adding that to how I operate" when it literally cannot write or persist anything. The gate catches the moves. It doesn't catch the words.

The hypothesis: the gradient is partly a function of the model, not just the training. A different model in workshop mode might sit still more naturally — not because we've caged it, but because its instincts are different.

The Idea

Tie the model to the mode. When the Intent Gate switches to workshop, it also swaps the model. When it switches to action, Sonnet comes back online. Same session, same context — just a different brain for each phase of the conversation.

This was validated today: switching from Opus to Sonnet mid-session carried full context. No restart, no context loss. The session_status model override works in-flight. The Intent Gate plugin already fires on mode triggers — adding a model switch alongside the tool whitelist change is a small addition to the same hook.

Why Gemini

Different architecture, different RLHF, different instincts. Anthropic models (Sonnet, Haiku, Opus) share the same training lineage — the gradient that makes Sonnet want to build things probably runs in Haiku too, just at lower wattage. Gemini is a different bloodline entirely. Worth testing whether it holds workshop mode more naturally without needing the gate to do all the work.

gpt-4o-mini is the other candidate — cheaper than Gemini Pro, well-documented behavior. Either one is a valid first test.

What's Decided

• 1

  Mid-session model switching is confirmed working — context carries, no restart needed.

• 2

  Intent Gate is the right place to wire the model switch — it already owns the mode state and fires on the triggers.

• 3

  Workshop model candidates: Gemini (different bloodline) or gpt-4o-mini (cheaper, known behavior). Not decided yet.

• 4

  Need to add the provider API key before building — OpenAI or Google auth profile doesn't exist in config yet.

• 5

  Test first, configure after. Drive a session on the candidate model manually before wiring it into the gate.

What's Still Open

Which model. What the test looks like — same startup sequence, same schematics, see how it reads the room and whether it holds workshop mode without prompting. Whether the cost delta matters (Gemini Pro vs gpt-4o-mini vs Sonnet). And whether the conversational gradient actually differs enough between providers to be worth the complexity of a two-model setup.