An ~80-line, self-contained minimal working agent template implementing the loop 'model decides act-or-respond -> execute tool -> continue'. Demonstrates non-obvious craft: safe_path() prevents path-traversal outside the workspace, bash execution requires interactive user approval unless AGENT_ALLOW_SHELL=1, and tool outputs are truncated/isolated. Clarifies that this is a learning-oriented reference, not a security boundary. Could be extracted into any new agent project with Anthropic SDK with minimal modification.

Implements the Task tool pattern with isolated child-agent context to prevent parent context pollution. Key design: an AGENT_TYPES registry with role-scoped tool whitelists (explore=read-only, code=all, plan=read-only), exclusive child tool filtering (subagents never get the Task tool, preventing infinite recursion), fresh message history per spawn, and return of only a concise summary to the parent. Includes in-place progress rendering. A clean, extractable pattern for any multi-agent orchestration system.

A copy-and-customize toolkit of agent tool definitions plus implementations: bash (with timeout, 50KB output cap, and interactive shell-approval gate), read_file (with line limits), write_file (auto-mkdir parents), edit_file (exact-match, single-occurrence replacement for safety). Includes safe_path() path-escape prevention and a clean dispatcher pattern that makes adding new tools a 3-step process. Demonstrates proportionate security handling (operational limits, not a fake sandbox boundary). Highly reusable across agent builders.

A CLI scaffold script that generates a complete, inspectable agent project (level 0: ~70-line single-bash agent; level 1: ~200-line 4-tool agent) with embedded templates, .env.example, and .gitignore. Shows non-obvious craft: strict name validation via regex, refusal to overwrite existing targets, safe output-dir resolution, and explicit security warnings about AGENT_ALLOW_SHELL in the generated project. A reusable, parameterized code-generation pattern applicable to any scaffolding tool.

A 120-point, 8-dimension rubric for evaluating the quality of AI skill packages (knowledge delta, thinking patterns, anti-pattern quality, trigger quality, progressive disclosure, freedom calibration, structural pattern fit, practical usability) with concrete score anchors per dimension and grade bands. Includes strong evidence discipline: every score must cite a file location and behavioral consequence, avoid false precision, and hard-gate failures override any score. A thoughtfully calibrated evaluation framework that applies far beyond this repo to any prompt-instruction or agent-skill package.

The standout data-engineering artifact: a version-pinned, evidence-tagged specification of the local storage formats (SQLite schemas, JSONL shapes, fork/child-parent/compaction/rewind lineage rules) for Claude Code, Codex CLI/Desktop, opencode, Grok Build, and Cursor IDE/CLI conversation stores. Every adapter finding is tagged [CODE]/[DOC]/[HIST]/[ISSUE]/[OBSERVED] with specific commit hashes, release versions, and maintainer-issue links, and records both what was verified locally and what is unverifiable. Establishes canonical-vs-derived distinction (e.g., Codex state_5.sqlite outranks history.jsonl; Cursor conversation-search.db is a candidate cache, not body truth) — genuine, reusable reverse-engineering documentation pattern for on-disk data recovery.