Docs / How it works / Choosing modelsVIEW .MD

Choosing models

Published 02/11/2026

Each role (pilot, builder, reviewer) in ade.toml has one engine and, per engine, a model and an optional reasoning effort. Models are family aliases, so each engine resolves its latest release. Change a role's model or effort under [roles.<role>]:

TOML

[roles.builder]
provider = "claude"
models = { claude = "sonnet", codex = "gpt-6.1-sol", opencode = "openrouter/deepseek/deepseek-v4.1-flash" }
effort = { claude = "high", codex = "high", opencode = "high" }

Effort is off until you set it. An unset effort comes from the engine: codex takes [providers.codex] default_effort; claude and opencode fall to their own user config, so set effort per role to keep lanes independent of your setup. ade config --roles shows every engine's model and effort for each role, the active engine marked, and the codex compaction source.

claude#


Model key	`[roles.<role>].models.claude`
Effort key	`[roles.<role>].effort.claude`: low, medium, high, xhigh or max
How ADE passes effort	`claude --effort <value>`
Kit default	pilot `opus`; builder and reviewer `sonnet`; effort unset

codex#


Model key	`[roles.<role>].models.codex`
Effort key	`[roles.<role>].effort.codex`: low, medium, high or xhigh
How ADE passes effort	`codex -c model_reasoning_effort=<value>`
Compaction keys	`[providers.codex] auto_compact_token_limit` and `context_window`, passed as `codex -c model_auto_compact_token_limit` / `model_context_window`; unset means the lane inherits `~/.codex/config.toml`
Kit default	`gpt-6.1-sol` for every role; effort and compaction unset

opencode#


Model key	`[roles.<role>].models.opencode`, named provider/model (openrouter/..., ollama/..., anthropic/...)
Effort key	`[roles.<role>].effort.opencode`: the model's own variant names
How ADE passes effort	`opencode --variant <value>`
Kit default	`openrouter/deepseek/deepseek-v4.1-flash` for every role; effort unset

Run one card on another engine#

Set ADE_PROVIDER for the one command, and ade.toml stays as it is:

ADE_PROVIDER=codex ade conduct

ADE_PROVIDER forces every role onto that engine, using each role's model from the table above. fake is a built-in engine that is deterministic and spends no usage, handy for a rehearsal.

Change a model or effort for one run#

ade loop and ade conduct take --model ROLE=ENGINE:MODEL and --effort ROLE=LEVEL, each repeatable. They hold for that run only (lanes included); ade.toml is not edited, and the next run without the flags uses it:

ade loop c12 --model builder=claude:haiku --effort builder=max

Look any key up with ade config WORD.