Agent stack build plans · oh-my-pi · 1 of 5

What you get

After the switch, you get a terminal agent that feels like Claude Code — sessions, subagents, plan mode, MCP, skills — where every category of work is billed to the model you chose for it. omp names those categories roles: default for normal turns, smol for subagent fan-out, slow for deep reasoning, plan for plan mode, commit for commit messages. Assign a model to each role once, in one file, and the routing happens on every turn without you thinking about it.

• One config, three tiers. Frontier model on default and slow, a cheap fast model on smol and commit, a free local model wherever you dare. Change any of it mid-session with /model.
• Rate limits stop being outages. Fallback chains retry a 429’d model’s work on the next model in your list, then restore when the cooldown ends.
• Private repos can stay private. Pin allowed models per directory — local-only under ~/clients, anything goes in side projects.

Play with a week of agent work

The simulator below is a toy with honest proportions, not a quote: it prices a typical week of turns in relative units, using rough public price ratios between model tiers (a frontier flagship ≈ 25× a small fast model per token; local ≈ 0 marginal cost). Your mix will differ — that’s what the slider is for.

The pattern the simulator makes obvious: fan-out and boilerplate dominate turn count, not difficulty. Routing exists because the 25× price gap between tiers is mostly being spent on work that never needed the expensive model.