devtu

DevTU — ToolUniverse Development Toolkit

One consolidated skill for the entire ToolUniverse (mims-harvard/ToolUniverse) development lifecycle. Read the mode from the user's intent and open that mode's playbook under modes/<mode>/README.md (progressive disclosure — the full former per-skill content, references, and templates are preserved there unchanged).

Sub-command modes

Mode	Use when	Playbook
self-evolve	"run the self-improvement loop", "do a debug/test round", evolve ToolUniverse end-to-end (orchestrates the other modes)	modes/self-evolve/README.md
auto-discover-apis	expand coverage — gap analysis, web-search for life-science APIs, batch discover + create + validate + PR	modes/auto-discover-apis/README.md
create-tool	add a new tool: Python tool class + JSON config + registration + real test examples	modes/create-tool/README.md
fix-tool	a tool fails tests / returns errors / has schema-validation issues — diagnose root cause, fix, validate	modes/fix-tool/README.md
code-optimization	writing/refactoring tool Python: reusable code patterns, per-API bug fixes, schema/pre-commit checklist	modes/code-optimization/README.md
optimize-descriptions	make tool JSON descriptions clear (prerequisites, abbreviations, required-vs-optional params, usage)	modes/optimize-descriptions/README.md
optimize-skills	improve research-skill quality: tool verification, evidence grading (T1-T4), completeness, synthesis	modes/optimize-skills/README.md
github	push code safely, activate pre-commit hooks, clean staged files, fix CI	modes/github/README.md

The cycle (self-evolve orchestration)

Discover → Create → Test → Fix → Optimize → Ship → Repeat

• Discover → auto-discover-apis — gap analysis + API search + scoring
• Create → create-tool — tool class + JSON config + test examples
• Test → self-evolve — launch researcher-persona agents to find issues (verify EVERY report via CLI first — 50%+ are false positives)
• Fix → fix-tool (code-level patterns in code-optimization)
• Optimize → optimize-descriptions + optimize-skills
• Ship → github

Shared rules (all modes) — CRITICAL

• Verify before fixing: python3 -m tooluniverse.cli run <ToolName> '<json_args>' — many agent-reported bugs are MCP-interface false positives.
• Tool contract: return_schema MUST have oneOf (success + error); test_examples MUST use real IDs (no TEST/DUMMY/PLACEHOLDER); tool name at most 55 chars; NEVER raise in run() (return an error dict); set a 30s timeout on all HTTP.
• Registration (most-missed): class @register_tool + entry in default_config.py + wrapper generation on tu.load_tools().
• Code hygiene: always run Skill(skill="simplify") after writing or modifying code; ruff check before commit.
• Git: never push to main; never keep multiple open fix PRs; always git fetch origin && git rebase origin/main before push; commit messages use "Feature"/"Fix" (never "Bug"); no AI attribution; repo is mims-harvard/ToolUniverse (confirm with git remote -v).

Consolidation note: this skill supersedes the former eight devtu-* micro-skills (devtu-auto-discover-apis, devtu-code-optimization, devtu-create-tool, devtu-fix-tool, devtu-github, devtu-optimize-descriptions, devtu-optimize-skills, devtu-self-evolve). Their complete bodies, reference files, and templates are preserved verbatim under modes/<mode>/. Any older Skill(skill="devtu-<mode>") invocation is now this skill in that mode.