Mathodology Project Orientation

Repository Contract

This branch is a skills-only GitHub tree. Treat it as an AI-coding knowledge pack, not as the original runnable Mathodology application.

Current work should normally edit only:

• any file under .claude/skills/<skill-name>/ — its SKILL.md,

agents/openai.yaml, and a scripts/ directory when the skill ships one (e.g. .claude/skills/mathodology-award-gates/** and .claude/skills/mathodology-dev-test-release/scripts/validate_repo.py)

• .claude/agents/<agent-name>.md (including .claude/agents/mathodology-award-judge.md)
• .claude/workflows/<workflow-name>.md
• .claude/skills/mathodology-whole-project/scripts/create-source-backup.sh
• .claude/skills/mathodology-whole-project/scripts/update-project.py
• AGENTS.md
• README.md
• README_en.md
• docs/SKILLS.md
• docs/SKILLS_zh.md
• docs/INSTALL.md
• docs/INSTALL_zh.md
• docs/WORKFLOWS.md
• docs/WORKFLOWS_zh.md
• docs/BACKUP.md
• docs/BACKUP_zh.md
• .gitignore
• .mcp.json

.mcp.json is a deliberate exception to the "no project config" rule: it registers the keyless search MCP server (uvx free-search-mcp) that mathodology-evidence-search depends on, so a fresh clone has working evidence search with no per-user setup. It must stay a client-configuration file — no secrets, no API keys, no local absolute paths. A user who wants a different server (a local checkout, a keyed engine set) overrides it at local scope, which takes precedence over the project file.

LICENSE is retained but should not change unless the license changes.

What Is Absent

The current GitHub tree should not contain application source, generated clients, CI, deployment, package-manager, installer, data, or test trees. .mcp.json is the one configuration file that is present, and it is whitelisted for the reason given above.

If a task requires historical implementation detail, use Git history or another branch in a separate worktree. Do not add those files back to this branch as part of ordinary skills maintenance.

Boundary Check

Before publishing cleanup or skill maintenance, verify tracked files stay inside the skills-repository whitelist:

python3 .claude/skills/mathodology-dev-test-release/scripts/validate_repo.py whitelist

Run validate_repo.py all for the full set of maintenance gates (skills, metadata, links, whitelist, agents, sync, evidence, updater). The whitelist logic lives only in that script; do not re-inline it here.

Text Check

Search for stale current-path claims before committing (requires ripgrep; the grep -E fallback is shown second):

rg -n "apps/|crates/|packages/|docs/skills|Dockerfile|docker-compose|just |cargo|pnpm|uv run|\.github|installer" README.md README_en.md AGENTS.md docs .claude/skills
grep -REn "apps/|crates/|packages/|docs/skills|Dockerfile|docker-compose|just |cargo|pnpm|uv run|\.github|installer" README.md README_en.md AGENTS.md docs .claude/skills

Hits are acceptable only when they describe removed historical material or an absence check. They must not instruct agents to edit or run missing current files. The current skills, scripts, and award-gate content do not match this pattern, so any hit is a stale claim to review.

Choosing Skills

• Whole repository backup, transfer, or orchestration: use mathodology-whole-project.
• Skill text or metadata changes: use mathodology-skill-authoring.
• Award-level Codex or Claude Code phase workflow: use docs/WORKFLOWS.md.
• Award-run gate schemas (handoff/gate/scorecard/decision_memo), judge-panel protocol, and figure/PDF QA scripts: use mathodology-award-gates.
• Former subsystem knowledge: use the matching archived subsystem skill.
• Validation and publishing checks: use mathodology-dev-test-release.