BrainHarness Autoresearch v0.3.0

从 brainforge-autoresearch 迁移为 brainharness-autoresearch，统一仓库、技能和插件名称。

• 更新技能入口、元数据和安装命令。
• 修正 README 中过期的 ClawHub 名称说明。
• autoresearch.py 及优化流程保持不变。

安装：

/plugin marketplace add zning1994/brainharness
/plugin install brainharness-autoresearch@brainharness

已有 @brainforge 安装不会因 GitHub 更名而自动变为 @brainharness。本次 GitHub Release 与 ClawHub 审核独立，ClawHub 可下载版本以平台状态为准。

验证：插件清单检查及独立配置目录中的实际安装均通过。

完整变更

v0.2.5 — ClawHub republish under brainforge-autoresearch

Rename-related republish. No functional changes to the optimizer.

Changed

• SKILL.md `name` field: `autoresearch` → `brainforge-autoresearch`
• plugin.json and SKILL.md metadata `homepage` → new repo URL
• Published to ClawHub as `brainforge-autoresearch`; old slug `openclaw-autoresearch` merged in as redirect

Compatibility

• GitHub: old URL `https://github.com/zning1994/openclaw-autoresearch\` auto-redirects
• ClawHub: `openclaw skills install openclaw-autoresearch` still works (redirects to `brainforge-autoresearch`)
• Users on the old ClawHub slug continue receiving updates

Full diff

v0.2.4...v0.2.5

v0.2.4 — Self-optimized SKILL.md

Self-optimization dogfooding

Two rounds of autoresearch run on its own SKILL.md:

Round	Provider	Baseline	Best	Experiments
1	MiniMax M2.7	48.6%	68.1%	15 (3 keep)
2	Claude (improved evals)	80.6%	91.7%	15 (1 keep)

Bug fixes discovered by self-optimization

Both rounds independently found the same documentation bugs:

• LLM eval field names: pass/fail → pass_description/fail_description (docs didn't match API)
• contains/not_contains params: value (string) → values (list)
• Missing fields: LLM eval table now includes required type and name

Other improvements

• Procedure section: added 3-step summary upfront for agent clarity
• Step 4/5 merged into actionable "Review results and apply changes"
• Example restructured: "create eval → run → review" flow
• Added self-eval.json for future dogfooding runs

Full changelog

v0.2.3 — Configurable HTTP Timeout

Added

• --timeout CLI flag to control HTTP timeout per LLM call (default: 180s)

Changed

• Default timeout: 60s → 180s (large prompts were causing mutation failures on Anthropic API)
• Timeout stored on provider instance, fully configurable

Usage

# For large style guides or slow APIs
python autoresearch.py --target SKILL.md --evals eval.json --timeout 300

Full changelog: CHANGELOG.md

v0.2.0 — MiniMax Thinking Support + ClawHub Ready

Fixed

• MiniMax extended thinking support — models with thinking blocks no longer crash
• Fallback to thinking content when text block is missing
• LLM judge max_tokens: 16 → 256
• Multiple field name inconsistencies (rule/check, pass/pass_description, value/values)
• test_inputs: accept both strings and {name, input} objects
• contains rule: support match "all" mode
• Convergence counter double-increment bug
• regex uses re.MULTILINE, word_count is CJK-aware

Added

• --model CLI flag
• ClawHub metadata in SKILL.md (openclaw.requires)
• .clawhubignore

Tested

• brain-search skill: 37.5% → 54.2% pass rate in 5 experiments (MiniMax M2.7)

Full changelog: CHANGELOG.md

v0.1.0 — Initial Release

Autonomous skill prompt optimizer based on Karpathy's autoresearch methodology.

Highlights

• Zero-dependency Python script (stdlib only, Python 3.9+)
• Hybrid eval: rule-based checks + LLM-as-judge
• Supports MiniMax, OpenAI, Anthropic (auto-detect)
• Optional live dashboard (Chart.js)
• Compatible with npx skills add, OpenClaw ClawHub, and standalone use

Install

npx skills add zning1994/openclaw-autoresearch

See CHANGELOG.md for full details.