Skip to content

Releases: sdfdu/evopilot

EvoPilot v0.5.0 — Quality-Gated Skill Compilation

Choose a tag to compare

@sdfdu sdfdu released this 04 Sep 22:03

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-demo

See RELEASE_NOTES.md for the complete history.

EvoPilot v0.1.0 — Adaptive Codex Agent

Choose a tag to compare

@sdfdu sdfdu released this 27 Aug 10:45

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/evopilot

Install 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.