File size: 5,229 Bytes
ea8c728 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 | # Agentic Usage Guide
This guide shows how to run Shinka with coding agents using the project skills:
- `shinka-setup`: scaffold task files (`evaluate.py`, `initial.<ext>`, optional run config)
- `shinka-convert`: snapshot an existing repo into a Shinka task directory
- `shinka-run`: launch and iterate evolution batches via `shinka_run`
- `shinka-inspect`: load top-performing programs into a compact context bundle
It covers:
- installing Shinka
- installing Claude Code and/or Codex CLI
- installing the skills from this GitHub repo with `npx skills add`
- running a practical setup -> run -> inspect loop
## 1) Install Shinka
From a clean machine:
```bash
pip install shinka-evolve
# or
uv pip install shinka-evolve
```
Set API keys (example):
```bash
cp .env.example .env 2>/dev/null || true
# Edit .env and add OPENAI_API_KEY / ANTHROPIC_API_KEY as needed
```
## 2) Install Agent CLI(s)
Install one or both.
### Claude Code
```bash
npm install -g @anthropic-ai/claude-code
claude --version
```
### Codex CLI
```bash
npm install -g @openai/codex
codex --version
```
## 3) Install Skills from the Repo with `npx skills add`
The Shinka skills live directly in this repo under `skills/`. You do not need to copy files by hand or publish a separate npm package.
Install all current Shinka skills globally for Claude Code and Codex:
```bash
npx skills add SakanaAI/ShinkaEvolve --skill '*' -g -a claude-code -a codex -y
```
This installs from the GitHub repo source. The explicit `--skill '*'` makes "install all skills" unambiguous and avoids interactive prompts.
Installed skills currently include:
- `shinka-setup`
- `shinka-convert`
- `shinka-run`
- `shinka-inspect`
### Project-local install
Use this if you want the skills installed only for the current repo:
```bash
npx skills add SakanaAI/ShinkaEvolve --skill '*' -a claude-code -a codex -y
```
Typical project paths:
- Claude Code: `.claude/skills/`
- Codex: `.agents/skills/`
### Global install paths
For the global install command above, the relevant skill roots are:
- Claude Code: `~/.claude/skills/`
- Codex: `~/.codex/skills/`
### Install one skill only
For a narrower install:
```bash
npx skills add SakanaAI/ShinkaEvolve --skill shinka-setup -g -a claude-code -a codex -y
```
## 4) Setup Skill Walkthrough (`shinka-setup`)
Ask the agent to scaffold a new task directory and evaluator contract.
Example prompt:
```text
Use shinka-setup to scaffold a new task in examples/my_task.
Language: python.
Goal: maximize <metric>.
```
Illustration (setup flow):
Expected output:
- `initial.<ext>` with evolve block
- `evaluate.py` producing `metrics.json` + `correct.json`
- optional `run_evo.py` / `shinka.yaml` scaffolds when requested
## 5) Run Skill Walkthrough (`shinka-run`)
Use `shinka_run` for agent-driven evolution loops.
Minimal batch:
```bash
shinka_run \
--task-dir examples/my_task \
--results_dir results/my_task_agent \
--num_generations 10
```
With core knobs via `--set`:
```bash
shinka_run \
--task-dir examples/my_task \
--results_dir results/my_task_agent \
--num_generations 20 \
--set evo.max_api_costs=0.5 \
--set evo.llm_models='["gpt-5-mini","gemini-3-flash-preview"]' \
--set db.num_islands=2 \
--set db.parent_selection_strategy=weighted
```
Illustration (run flow):
## 6) Inspect Skill Walkthrough (`shinka-inspect`)
Use `shinka-inspect` after one or more batches to generate an agent-ready context file.
Minimal:
```bash
python skills/shinka-inspect/scripts/inspect_best_programs.py \
--results-dir results/my_task_agent \
--k 5
```
With filters and explicit output:
```bash
python skills/shinka-inspect/scripts/inspect_best_programs.py \
--results-dir results/my_task_agent \
--k 8 \
--min-generation 10 \
--max-code-chars 5000 \
--out results/my_task_agent/inspect/top_programs.md
```
Output:
- default file: `results/my_task_agent/shinka_inspect_context.md`
- contains ranking + code snippets for top programs
- designed to be loaded directly into coding-agent context
## 7) Batch Iteration Rules (Important)
When using `shinka-run` skill:
- unless user explicitly requests fully autonomous execution, ask for config confirmation between batches
- keep `--results_dir` the same across continuation batches so prior state can reload
- change `--results_dir` only when intentionally forking a new run
## 8) Quick Validation Checklist
Before first run:
- `shinka_run --help` works
- task dir has `evaluate.py` + `initial.<ext>`
- API keys are available in environment
- `npx skills list` shows the installed Shinka skills
- for global installs, skills appear under `~/.claude/skills/` and/or `~/.codex/skills/`
- for project installs, skills appear under `.claude/skills/` and/or `.agents/skills/`
After each batch:
- check run artifacts/logs under the chosen `results_dir`
- review score and correctness trend
- run `shinka-inspect` and review the generated context markdown
- choose next batch config (budget, models, islands, attempts, generations)
|