v0.3.0-beta.1 — first public beta

Effort Autopilot v0.3.0-beta.1 — first public beta

Automatic reasoning-effort selection for the Claude Code CLI: keep using claude exactly as always; each task is classified locally (zero tokens, zero network) and the right /effort level is applied before inference — same model, same session, your manual choice always wins.

What's in this beta

• Hybrid hook + ConPTY broker, live-validated end to end on Windows: block → local classify → /effort with real acknowledgement → exactly one replay. Fail-open on any doubt, always visible, never silent.
• Reversible installer: effort-autopilot install / uninstall / status / policy — user-PATH shim with explicit consent and exact backups; your real Claude binary and settings are never modified.
• Precedence policies: manual-wins (default) and autopilot-wins; per-project .effort-autopilot.json (opt-out or policy override).
• Localized, disclosed decisions (English/Spanish by prompt language), including the CLI's saved-default side effect on every non-max application.
• Mini-AI infrastructure (inactive until calibration ships an artifact): local multilingual embeddings (optional dependency) + trained ordinal head with a deterministic fallback chain; trainer and calibration pipeline included (npm run ml:train, npm run calibrate).
• 162 local non-billable tests; zero-inference diagnostics against the installed CLI.

Install (from source — not on npm yet)

git clone https://github.com/JeronimoRepetto/effort-autopilot.git
cd effort-autopilot
npm install
node bin/effort-autopilot-cli.js install
# open a NEW terminal, cd into any project, run: claude

Full guide: docs/INSTALL.md

Platform support

Platform	Status
Windows 10/11	✅ verified live
Linux	Implemented; verified via WSL
macOS	Implemented; not yet verified — the installer warns

Known limitations (honest by design)

• Claude Code visibly renders UserPromptSubmit operation blocked by hook on each intercepted task (no quiet-block API exists; a native capability proposal is drafted).
• Verified against Claude Code 2.1.238; CLI updates may change acknowledgement texts — the broker then fails open (your prompts always go through unchanged) until updated.
• On 2.1.238, every /effort change except max also becomes your saved default (upstream behavior, disclosed on each application).
• The current classifier's language patterns cover English and Spanish; other languages safely fail open. The multilingual mini-AI activates after the calibration phase.
• Effort is a behavioral signal, not a token cap — no per-prompt savings guarantee is claimed until measured.

Feedback and bug reports welcome — issue templates are set up. If this saves you time or tokens: ko-fi.com/jeronimorepetto ☕