A Claude Code mod (desktop app and terminal, version 2.1.321 or later) that picks the model and the effort level for each prompt: stronger on hard tasks, lighter on your quota for simple ones.
/plugin install automode --marketplace FloGV01/automode
Then run /reload-plugins.
- Sorts each prompt into six tiers: Haiku low, Sonnet medium, Sonnet high, Opus high, Opus max, then Fable only on evidence of failure or on request (
@fable). - Local rules first; a fast Haiku judge settles unclear cases; falls back to a safe default (Sonnet high) if the judge does not answer.
- Escalates on its own when there is evidence of failure: repeated failing tests or a correction from the user.
- Protects the prompt cache: no model switch in the middle of a task while the context is warm.
- Risk floors: private data, financial models, secrets, migrations, requested reviews.
- Never touches subagents.
AutoMode [On|Off] Sonnet 5.5 medium Ultracode [On|Off]
- AutoMode in green: active. In orange: shadow mode. In red: off.
- Ultracode in green: on. In grey: off.
By default AutoMode starts in shadow mode: it decides and logs without applying anything. After 7 days, /automode report reports the simulated savings and how choices were spread, so you can decide whether to turn it on.
/automode [active | shadow | off | auto | ultracode on|off | report | status]
Prompt prefixes: @haiku, @sonnet, @opus, @max, @fable, @auto, effort low|medium|high|very high|max.
Prompts are understood in English and in French.
The log stays local in ~/.claude/automode/ and never contains prompt text: only the choice, the reason, token counts and cost.
MIT