kunsanglee/ai-ready

Audit codebases for AI-readiness using a 7-category 100-point rubric, then apply ROI-prioritized improvements (CLAUDE.md scaffolds, ANTIPATTERNS.md seeds, INDEX.md, ARCHITECTURE.md Mermaid diagrams, freshness Stop hook, trend sparkline). Multi-module and single-module (package-as-module) layouts are scored on parallel rubrics. v0.2.0+: optional .ai-ready/config.json drives frontmatter-aware INDEX grouping and preserves user-edited lazy-load triggers. v0.3.0+: the config rubric section makes scoring respect project reality — docs/design counts as ADR, springdoc as API contract, project-level .claude/settings.json lint/test hooks as a verification gate. v0.4.0+: apply guards human-adopted docs — generators refuse to overwrite files lacking the auto-generated signature (--force to override) and apply runs a per-doc diff-confirm loop. v0.4.2+: INDEX generation respects .gitignore — only tracked or unignored .md is indexed, so generated artifacts and .claude/commands no longer pollute INDEX; non-git dirs fall back to a full walk. v0.5.0+: apply maintains docs surgically — the doc-touching scripts (gen_index, gen_arch_diagram, extract_section, inject_module_map, inject_lazy_load_index) gain a --json facts-only mode and the AI adds/updates only what changed while preserving human curation, instead of mechanically overwriting whole docs. v0.6.0+: the autonomous verification loop (formerly the separate loop-engine plugin) is folded in — /loop-run (maker→checker→deterministic-rubric→decide→stall loop until pass or brake), /loop-review (one-shot adversarial check), /loop-lessons (human-approved harvest of caught mistakes into the knowledge layer), the loop-checker and loop-lesson-synthesizer subagents, and the _loop-engine scoring shell (severity assigned by the shell, not the LLM; BASE rubric + optional project-local .loop/rubric.md). Build/test/lint commands, ticket pattern and base branch are runtime-detected (detect_build.py) so no adapter file is persisted; the loop's knowledge layer is the docs/ANTIPATTERNS.md that ai-ready writes, which the loop reads and appends to. v0.8.0+: the loop-build skill adds multi-phase autonomous build-out — loop-run's verdict engine wrapped in an outer phase loop with a per-phase subagent maker (SendMessage-continued within a phase) so many phases build out unattended. v0.8.1+: loop-build fail-loud-validates phases.json before the unattended traversal — a jq assertion rejects a malformed decomposition (empty/non-array .phases, missing name/steps/ac_cmd, or an out-of-enum status) with exit 65 so a corrupt spec halts for a human instead of silently mis-traversing. v0.8.2+: the loop skills (loop-run/loop-build/loop-review) and loop-checker recover findings via a deterministic orchestrator-named output file rather than scraping the checker's final chat message — a background/unattended session never delivers a subagent's final text inline, so file recovery is the only mode-agnostic path; loop-checker's no-write Bash rule gains one narrow carve-out (write only that single findings JSON path, refuse if it points at a source file), the orchestrator clears the file before each spin so a stale prior-cycle result can't masquerade as a pass, and a fail-loud [ -s ] guard rejects an empty/missing findings file instead of silently passing.