Self Evolve
Methodology and deterministic evaluation tools for iterative improvement of skills, repositories and agent workflows, with independent review and regression checks.
Install
git clone https://github.com/DaizeDong/self-evolveInstall command taken from the project's README or registry entry. Always review code before running it.
README excerpt
Read full README on GitHub →self-evolve
Improve a skill or repository through proposed changes, evaluation in a Git worktree, and evidence-based acceptance.
⭐ Design Philosophy
A self-improvement loop can raise its score by changing the task, dropping hard cases, or grading its own output. Open-ended work makes that especially difficult to detect because there may be no single ground truth. self-evolve therefore starts with an observable improvement contract and keeps the comparison stable.
- Evidence sources. The loop is
profile → reflect → propose → patch → evaluate → accept or review. A uses executed tests, B uses independently verified facts, and C needs supplied measurements. Missing measurements remain unavailable; changing a label cannot create evidence. - Fixed evaluation contract. Freeze the policy, obligations and holdout before editing, and compare the parent and candidate by the same identities. This limits task removal and grader changes as routes to a higher score, at the cost of starting a new evaluation contract when the goal changes.
- Proposal and acceptance responsibilities. Models propose changes; deterministic gates handle evidence and acceptance. Retain actual provider metadata when judging independence; different requested aliases cannot establish it. A gate still depends on the quality of its inputs and the assumptions behind its statistics.
- Recovery and auditability. Failures, rejection, and review are useful outcomes. Preserve measured populations and exact candidate bytes, with the final documentation and source revision ready before review. This costs storage and review effort but makes a decision auditable and a rollback possible.
- Data and authorization boundaries. Real evidence stays in a PRIVATE Git companion. Worktrees isolate edits, but do not prove operating-system isolation. Accepted candidates are archived; publishing, merging, and sending follow the caller's authorization as separate actions.
The full rationale and tradeoffs explain these choices. Current limits define what the shipped loop can execute.
Implemented A/B loop
Diagram source · render script
ACCEPT archives a candidate snapshot; merging requires its own authorization.
CONTINUE and queued review can start a fresh round, bounded by --max-rounds
and circuit breakers. Checks can reject before evaluation; unavailable required
evidence, insufficient review independence, or failed restoration can end the
invocation early. See runtime and recovery.
Install
Clone the repository with its pinned guard and style submodules:
git clone --recursive https://github.com/DaizeDong/self-evolve.git
The skill entry is SKILL.md; command adapters live in commands/.
Use the absolute tools/sie_cli.py entrypoint from another directory or an
installation alias.
Config
Set SELF_EVOLVE_CONFIG (alias SELF_EVOLVE_CONFIG_DIR) to an existing PRIVATE
Git companion, or SELF_EVOLVE_DATA_DIR to exactly its absolute data/ path.
An inherited DATA_DIR takes precedence over CONFIG. The data/ directory may be
absent; runtime writes are path-checked against that data/ layout, with no
runtime visibility proof and no companion-root or public-worktree fallback.
Similar plugs
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local…
Why use many token when few token do trick. Viral skill + proxy for coding agents that cuts 65% of tokens by talking like a caveman.
Production-grade engineering skills for AI coding agents.
The 1st Collaborative AI Design Agent Workspace. Local-first and open source. Create prototypes, websites, apps, dashboards, presentations, images, and videos with your team…