Metaskill Evolve — Development skill for Claude Code

MetaSkill-Evolve — a $0 reproduction & dissection

An open, Apple-Silicon-friendly reproduction of **MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution** ([arXiv:2607.05297](https://arxiv.org/abs/2607.05297)), plus an honest look at *when* the recursion actually pays off.

What the paper does

Self-improving agents rewrite their own **task skill** (a Markdown file — *what the agent does*) from execution traces. MetaSkill-Evolve makes this **recursive**: every branch also carries a **meta-skill** `m = (ψ, σ, α, π, ε)` — five Markdown files that parameterise the five agents of the improvement pipeline:

symbol	agent	role
ψ	Analyzer	map a failure to a tag + free-form analysis
σ	Retriever	pull same-/cross-branch inspirations for that tag
α	Allocator	set the child budget K ∈ [1, K_max]
π	Proposer	emit a concrete edit to the task skill
ε	Evolver	apply the edit and verify the result

Task skills evolve on a **fast loop**; the meta-skill evolves on a **slow loop** (`every H iterations`) under the *same* pipeline applied to itself — no extra model, no extra objective. All five agents share one **frozen** backbone.

The honest angle (why this repo exists)

The paper's own Table 1 shows most of the win comes from the **fast loop**, and the recursive meta-skill adds a **modest, task-dependent** bonus:

benchmark	no-skill	static	single-level (fast)	+ meta (recursion)
OfficeQA	31.78	36.09	48.94 (+17.16)	55.32 (+6.38)
SealQA	29.17	29.41	37.21 (+8.04)	45.26 (+8.05)
ALFWorld	92.31	90.38	92.31 (+0.00)	94.23 (+1.92)

Recursion helps most on retrieval-heavy QA and is within noise on a near-ceiling embodied benchmark (ALFWorld starts at 92%; *static* even hurts it).

**Headline experiment (not in the paper): does the meta-loop's value shrink as the backbone gets