Models per station

Different work suits different models. You might want Fable to write specs and plans, Opus to build, and a second vendor to review the diff. Buildline lets you choose a model for each station. It doesn't ship a model list of its own, so new models work the day your tools support them.

#Choose your models

npx buildline models

The picker shows what's available right now:

• Claude Code aliases, read from the claude you have installed (today: fable, opus, sonnet). An alias always means the newest model in that family, so fable moves to the next Fable when Claude Code does.
• Every Claude model id from Anthropic's models API, when ANTHROPIC_API_KEY is set.
• Codex CLI and Gemini CLI, when they're installed.
• Other…: type any model name or id. Buildline saves exactly what you type and passes it to the tool.

You choose either by phase (planning, building, checking, everything else) or for every station. The same picker runs during npx buildline init, and from the interactive menu.

Without prompts:

npx buildline models set planning fable              # spec, design, plan
npx buildline models set building claude-opus-5-5    # build
npx buildline models set checking codex              # verify, review
npx buildline models set review gemini:gemini-2.5-pro
npx buildline models set build inherit               # back to the session model
npx buildline init -y --models planning=fable,building=opus

#Check what your choices resolve to

npx buildline models check

  ✔ fable                    → claude-fable-5-1
  ✔ claude-opus-5-5          → claude-opus-5-5
  ✘ claude-opus-9            There's an issue with the selected model …

For each Claude model this sends one tiny request and reports the model that actually answered. That catches typos, models your account can't use, and aliases that have moved on. For Codex and Gemini it checks the CLI is installed.

#The config

It's one map in .buildline/buildline.json:

"models": {
  "spec": "fable",
  "design": "fable",
  "plan": "fable",
  "build": "claude-opus-5-5",
  "review": "codex:gpt-5-codex",
  "default": "sonnet"
}

Stations: onboard, new, spec, design, plan, build, verify, review, ship. default covers any station you don't list; leave it out to use the session model.

Aliases keep up automatically. Full ids pin a model, so nothing changes under you until you update it. If you edit the file by hand, run npx buildline update to apply it.

#How it's applied

• Claude Code skills get a model: line, so /buildline-plan runs on Fable even if the session started on Sonnet, and the session goes back to its own model afterwards.
• Subagents follow their station: verify uses buildline-verifier, review buildline-reviewer and buildline-finding-checker, spec buildline-spec-critic, design buildline-design-reviewer, and onboard buildline-mapper.
• Stations set to another CLI are handed off. Type /buildline-review 004 in Claude Code with review set to codex, and Claude runs npx buildline step review 004 --headless, waits, and summarises what Codex did. Codex and Gemini follow the same rule through AGENTS.md.
• Autopilot (buildline run) starts Claude Code on the build model. Verify and review inside the loop use their own models, or hand off the same way.

#Opening a station directly

buildline step <station> [id] opens a station in whatever it's set to:

npx buildline step plan 004                  # claude --model fable "/buildline-plan 004"
npx buildline step review 004 --headless     # codex exec … (no chat, edits allowed in the repo)
npx buildline step spec 004 --model opus     # one-off override
npx buildline step build 004 --print         # show the command without running it

Headless runs may edit files in the project but nothing beyond it: Claude uses --permission-mode acceptEdits, Codex --sandbox workspace-write, and Gemini --approval-mode auto_edit.

npx buildline doctor warns if a CLI you chose isn't installed.

Edit this page on GitHub