agentrigs
+ submit rig
~ / rigs / whilehq / whileai-sdk

whilehq/whileai-sdk

Scientific RL and SFT post-training for AI agents: build evals that can fail, simulate situations, judge checked against people, pass@1 with a 95% interval, train with GRPO/SFT/DPO, prove every gain on a held-out set. uv add whileai

↗ GitHub ★ 21 Apache-2.0 updated 3d ago project Claude Code
share on X
ARCHETYPE
Skill Collector
Ten-plus skills loaded on demand. A procedural-knowledge library.
CONTEXT TAX · EVERY TURN
~1.0k tokens
Featherweight · median rig: 2.3k · breakdown
GUARDRAILS
0/5
No committed guardrails · details

Copy this rig

# review before running: this installs third-party code
$ npx degit whilehq/whileai-sdk/.claude ./rig-whileai-sdk  # inspect, then merge into .claude/

MCP servers are added to Claude Code at local scope; env vars are shown as YOUR_… placeholders — we never store values. Files are fetched with degit into a separate folder so you can review before merging.

$ npx degit whilehq/whileai-sdk/skills/audit-your-judge .claude/skills/audit-your-judge
$ npx degit whilehq/whileai-sdk/skills/character .claude/skills/character
$ npx degit whilehq/whileai-sdk/skills/dpo-pairs .claude/skills/dpo-pairs
$ npx degit whilehq/whileai-sdk/skills/grpo-verifier .claude/skills/grpo-verifier
$ npx degit whilehq/whileai-sdk/skills/harness-search .claude/skills/harness-search
$ npx degit whilehq/whileai-sdk/skills/manage-experiments .claude/skills/manage-experiments
$ npx degit whilehq/whileai-sdk/skills/pick-a-method .claude/skills/pick-a-method
$ npx degit whilehq/whileai-sdk/skills/sft-from-traces .claude/skills/sft-from-traces
$ npx degit whilehq/whileai-sdk/skills/strengthen-your-evals .claude/skills/strengthen-your-evals
$ npx degit whilehq/whileai-sdk/skills/tool-call-efficiency .claude/skills/tool-call-efficiency
$ npx degit whilehq/whileai-sdk/skills/watch .claude/skills/watch
$ npx degit whilehq/whileai-sdk/skills/whileai-simulations .claude/skills/whileai-simulations
$ curl -fsSL --create-dirs -o .claude/agents/docs-designer.md https://raw.githubusercontent.com/whilehq/whileai-sdk/main/.claude/agents/docs-designer.md
$ curl -fsSL --create-dirs -o .claude/agents/style-guide.md https://raw.githubusercontent.com/whilehq/whileai-sdk/main/.claude/agents/style-guide.md
$ curl -fsSL --create-dirs -o .claude/agents/user-sim.md https://raw.githubusercontent.com/whilehq/whileai-sdk/main/.claude/agents/user-sim.md

This rig commits no guardrails. Here is the community baseline instead — the deny/ask rules most often found across all 7,204 rigs:

{
  "permissions": {
    "deny": [
      "Read(./.env)",
      "Read(**/.env)",
      "Read(~/.ssh/**)",
      "Bash(rm -rf *)",
      "Read(**/*.pem)",
      "Bash(rm -rf /)",
      "Bash(git push --force:*)",
      "Bash(sudo *)",
      "Read(.env)",
      "Bash(rm -rf /*)",
      "Read(./.env.*)",
      "Read(~/.aws/**)",
      "Bash(git push --force*)",
      "Bash(rm -rf:*)",
      "Read(**/*.key)",
      "Read(**/.env.*)",
      "Bash(sudo:*)",
      "Bash(git reset --hard*)",
      "Bash(git reset --hard:*)",
      "Read(.env.*)"
    ],
    "ask": [
      "Bash(git push:*)",
      "Bash(git push *)",
      "Bash(git commit:*)",
      "Bash(rm *)",
      "Bash(rm:*)",
      "Bash(git rebase *)",
      "Bash(wget *)",
      "Bash(npm publish:*)",
      "Bash(git commit *)",
      "Bash(gh pr merge *)"
    ]
  }
}

Skills (12)

audit-your-judgecharacterdpo-pairsgrpo-verifierharness-searchmanage-experimentspick-a-methodsft-from-tracesstrengthen-your-evalstool-call-efficiencywatchwhileai-simulations

Subagents (3)

docs-designer
model: opus
Reviews the Mintlify docs (docs/) and the while.ai website for the scientific style in docs/reference/constitution.md and docs/reference/style.md with a UI/UX designer's eye - information architecture
style-guide
model: opus
Reviews the whileai SDK's public surface for developer ergonomics against docs/reference/style.md and docs/reference/constitution.md (PyTorch / DSPy / Unsloth shape, research-backed defaults from rlhf
user-sim
model: opus
Plays a researcher or applied-AI engineer using whileai for real - their own keys, their own compute (Modal, Prime Intellect), a scientific RL/SFT recipe with a feedback loop - by following the docs a

Similar rigs

bradautomates/claude-video
Give Claude the ability to watch any video. /watch downloads, extracts frames, transcribes, hands it all to Claude.
★ 18.1k project Claude CodeCodex 1 skills 14d ago
Minimalist 1.1k tok ·
nozomiishii/dotfiles
🧙🏿‍♂️ Doting Dotfiles
★ 7 setup Claude CodeCodex 12 skills 1 hooks 3d ago
Skill Collector 899 tok ·
tance-mang/chinese-webnovel-skills
中文网文写作工具包 · Chinese web-novel writing toolkit (23 skills) — a Claude Code plugin, also works in ChatGPT / DeepSeek / Gemini
★ 61 collection Claude Code 33 skills 4mo ago
Skill Collector 1.3k tok ·
SethGammon/Citadel
The operating layer for Claude Code + OpenAI Codex: persistent project memory, intent routing, safety hooks, cost telemetry, and parallel agent fleets.
★ 924 project Claude CodeCodex 5 mcp 48 skills 7 agents 13d ago
Orchestrator 38.0k tok ·
copied ✓