Releases: sdfdu/evopilot
Releases · sdfdu/evopilot
Release list
EvoPilot v0.5.0 — Quality-Gated Skill Compilation
Highlights
- Separate structural validation from semantic Skill quality assessment.
- Generate human-readable and machine-readable quality annotations for every compiled bundle.
- Block generic, repetitive, low-evidence, or underspecified Skills before installation.
- Bind installation approval to the reviewed bundle contents and destination.
- Expose 23 dependency-free MCP tools with 24 passing behavioral tests.
Try it
git clone https://github.com/sdfdu/evopilot.git
cd evopilot
python plugins/evopilot/scripts/evopilot.py doctor
python plugins/evopilot/scripts/evopilot.py demo --destination ./evopilot-demoSee RELEASE_NOTES.md for the complete history.
EvoPilot v0.1.0 — Adaptive Codex Agent
The first public release of EvoPilot: a local-first Codex copilot that learns proven work habits and grows through safe, versioned extensions.
Highlights
- Local SQLite memory with confidence, evidence, scope, risk, and timestamps.
- Privacy-minimized Codex hooks that record tool categories and outcomes—not raw inputs or outputs.
- Transparent habit detection and workflow promotion thresholds.
- Dependency-free stdio MCP server with seven inspectable tools.
- Human-in-the-loop policy for destructive, external, financial, credential, system-wide, publication, and permission-expanding actions.
- Six focused Skills: Idea Lab, Dev Flow, Tool Operator, Knowledge Coach, Automation Watcher, and Extension Foundry.
- Explicit memory deletion and portable JSON export.
- Windows-tested database lifecycle and MCP protocol tests.
Install
codex plugin marketplace add sdfdu/evopilotInstall EvoPilot from the Codex plugin interface, enable its MCP server under your preferred approval policy, and start a new task.
Known limitations
- v0.1 analyzes action frequency and outcomes, not complete cross-app sequences.
- Hooks observe Codex tool activity, not every interaction inside arbitrary desktop applications.
- Generated extensions remain a guided promotion workflow rather than unrestricted runtime self-modification.