Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/ruvnet/metaharness/evolvenpx skills add ruvnet/metaharness --skill evolvegit clone --depth 1 https://github.com/ruvnet/metaharnessWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00026 | $0.00704 |
| Opus 5 | $0.00013 | $0.00352 |
| Sonnet 5 | $0.00005 | $0.00141 |
| Haiku 4.5 | $0.00003 | $0.00070 |
Grade A, and why
evolve scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 54 lines — stays where its author put it; the contents beside it link to each section on GitHub.
evolve — Darwin Mode self-improvement
kimi-k3-harness ships with Darwin Mode (@metaharness/darwin, ADR-070…146): the model
is frozen; the harness evolves. Each generation mutates ONE of the 7 surface files
(planner, contextBuilder, reviewer, retry/tool/memory/score policy), sandboxes each
child, scores it, and keeps only variants that measurably improve — building an
archive of successful descendants.
Run it
npm run evolve # real substrate: runs your test command per variant (deterministic mutator — no API key, no network)
npm run evolve:dry # mock substrate: fast, fully offline, no test execution
Or directly:
npx metaharness-darwin evolve . --sandbox real --generations 3 --children 4
Safety (secure by default)
- Deterministic mutator is the default — no network, no API key, air-gapped.
- Every mutation passes the
validateGeneratedCodegate: no new imports, network, filesystem, shell, env access, or dependencies — pure refactor/tuning only. - Mutations run in a sandbox; only variants that pass your tests are archived.
- Nothing is promoted without measured improvement (guard against Goodharting).
See @metaharness/darwin for selection strategies (--selection, --crossover,
--curriculum), statistical gates (--fdr, --bench), and the real-LLM mutator (library API).
What the benchmarks taught us (measured, full SWE-bench Lite 300)
Defaults worth carrying into how you evolve and run this harness (full evidence + CIs in
@metaharness/darwin's LEARNINGS.md / bench/results/RESULTS.md):
- Closed-loop repair is the #1 lever (~2×). Feeding test/compiler failure back and retrying took resolve-rate 7.7% → 15.3% on the same cheap model. Iterate against ground truth, don't single-shot.
- Cheap-first + cost-aware routing. Track $/resolve, not just resolve-rate; a cheap model resolved 31× cheaper per fix than a frontier one. Reserve frontier for measured capability gaps.
- Tier the models (Barbarian & Scholar). Cheap sweep + frontier on only the residual = 33.3% at ~6× lower cost than running frontier everywhere.
- Put the output-format contract in a system message + example, and size prompts to the model's real context window — this alone took a weak local model from 0% to ~50% valid output.
- Only trust batch evaluation of the final artifact — in-loop counters drift 1.5–5×.
- The harness multiplies the model; it can't rescue one below the task's reasoning floor. Pick the smallest model above the floor, then let evolution do the rest.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 54 lines · 26 tokens per session scan A 14b28a2866b0
evolve is a skill published in the GitHub repository ruvnet/metaharness (624 stars, last pushed 2d ago), licensed MIT. It adds 26 tokens to every session and 704 once invoked, about $0.0001 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
C31-plan
Turn requirements into validated, executable plans. C31-brainstorm defines WHAT; C31-plan defines HOW — with coverage gates, wave analysis, and threat modeling.
C31-brainstorm
Brainstorming answers WHAT to build. It precedes planning, which answers HOW to build it.
C31-review
Reviews code changes using dynamically selected reviewer personas, then runs a verify-work pass: UAT walkthrough, verifier validation, and decision coverage logging. Writes memory/.planning/phases/XX-VERIFICATION.md.
C31-adopt-project
当用户提供 GitHub 项目链接或说"adopt", "看看这个项目", "研究这个项目", "整合", "学习这个项目"时，自动进行五阶段调研：提取核心哲学 → 差距分析 → 生成报告 → 门控确认 → 执行整合。报告保存到 memory/moc/，执行需用户确认。 MUST trigger when user sends github.com URL or says "adopt", "看看", "研究这个项目", "整合这个项目", "学习这个项目".
C31-compound
Capture a recently solved problem as structured documentation so future occurrences take minutes instead of research cycles.
C31-curator
信息整理, 文档生成, 产出交付 | 将C31-research事实与C31-storm判断组织成可交付文档.