Skip to main content

Which CLI

:::note Decided 2026-09-20

  • Claude Code through the proxy is the daily agent (px-claude, model brain/auto; the agent tier is deepseek-v4-pro).
  • aider is the cheap editor for targeted changes (architect/editor), not the agent.
  • OpenCode is benchmarked as the open-source third agent.

This replaces "both go forward on equal terms" (2026-09-19). Numbers: agents compared through the proxy. :::

Claude Code — the daily agent​

  • ANTHROPIC_BASE_URL → llm_brain, client key of the dev profile (brain setup claude-code --proxy).
  • Keeps habits, skills, hooks, which always work.
  • Token-hungry (large system prompt, many turns); the proxy's sanitizer removes the Claude-only fields other models reject (client compatibility).
  • Needs a model with a reliable tool loop: that is why its baseline is the agent tier, and brain/auto only moves light sessions down the ladder.

aider — the cheap editor​

  • Native OpenAI-compatible endpoint (brain setup aider --proxy).
  • architect/editor is the original 90/10: --model = reasoning tier (architect), --editor-model = fast tier.
  • Frugal with tokens (repo map, doesn't read everything): the cheapest way to apply a change you already know you want.

OpenCode​

Open source, a TUI similar to Claude Code, native multi-provider. Passes the benchmark through the proxy with far fewer tokens than Claude Code (brain bench run --tool opencode); kept as a measured alternative.

Routing​

Superseded by brain/auto: a decision model picks the tier once per session. Escalation on failure (retry with a higher tier when tests fail) is a design, not built yet (API layer).