ADR-036 — The coordinator runs on Sonnet; project and user settings choose otherwise

Status: accepted (2026-10-25, spec 100)

Context

A headless session starts on the user’s default model, usually Opus. In the spec-099 benchmarks the coordinator (the main session that runs the slash command) made most of the calls and most of the cost (quick-task: $0.36 of $0.48; build: $1.03 of $1.64), although its work is orchestration: run cplat, brief agents, read their reports. The agents that need judgement (architect, reviewer, product-manager) already pin model: opus in their frontmatter. cplat had no settings file: user configuration was environment variables, per-repo facts were files read ad hoc.

Decision

• cplat headless always passes --model to claude -p. Default sonnet.
• Accepted values: sonnet, opus, haiku. Anything else is refused before a run starts.
• Precedence: --model flag, then the project setting coordinator_model in <repo>/.platform-settings.yml, then the user setting in ${XDG_CONFIG_HOME:-~/.config}/sdlc-foundry/settings.yml, then the default. Each file’s coordinator_model is a string or a map of workflow name to model with an optional default; within a file the workflow entry beats the general one.
• These two files are the platform’s settings files from now on; later settings (budgets, #144) go there, read by scripts/cplat/settings.py.
• Subagents keep the model of their frontmatter; the coordinator setting never changes them.

Consequences

• Runs cost less without any configuration; a project that wants Opus for build writes one line.
• The run record names the model and where it came from, so benchmarks and reports can compare like with like.
• Interactive sessions are not affected: the user picks their model there.