The kit and the adapt engine

#The kit

The kit is authored once, in the canonical Agent Skills format, and embedded into the standalone binary. It contains:

• Skills — installed as av:<name>. Most are ported from the kit this project grew out of and are left alone apart from rebranding; skills this project authors meet one bar: a real workflow, an output-format contract, quality-gate self-checks, and a workflow position so they read as one graph.
• Agents — av-explore, av-planner, av-reviewer, av-tester, av-debugger, av-developer, av-git-manager, av-simplifier, av-brainstormer, av-researcher, av-docs-manager, av-project-manager, av-journal-writer: persona, behavioural checklist, status protocol, no external CLI coupling.
• Rules, scripts, and an environment example shared across providers.
• Hooks and a statusline for Claude Code only — session-init, rules-inject, privacy-block, scout-block, session-state, subagent-init — fail-open and covered by tests.
• Workflows — the three canonical graphs av run executes.

ariadnev validate lints all of it — frontmatter shape, sizes, cross-skill av:<slug> references — and prints the live counts. Those counts are deliberately not repeated here; the skill catalog is generated from the release instead.

#The adapt engine

Installation is a projection, not a copy. The adapt engine is a set of pure functions that map each artifact kind onto each provider's verified path and format:

artifact	example target (Claude Code)
skill	.claude/skills/
agent	.claude/agents/*.md
command	.claude/commands/*.md
rules	.claude/rules/*.md
scripts	.claude/scripts/
hook	.claude/hooks/av/*.cjs

Where a provider has no verified target for an artifact, the engine skips it and lists the skip in the install summary. It never invents a path. The full provider × artifact table is the provider reference, and ariadnev contract prints it from the binary.

Skip, do not guess. An artifact kind with no verified target on a provider is refused, reported by name in the install summary, and left absent on disk. A silent "best guess" path would create a file that neither the provider nor doctor recognizes, so the engine chooses a named skip over an unattributable write.

The kit → adapt → projection → receipt/cache pipeline reads left to right: the kit is the authored source, adapt is the pure projection layer, the receipt records what was actually written, and the cache holds the extracted-but-unmodified distribution the binary shipped with.

#The receipt

Every install writes .ariadnev/receipt.json (or ~/.ariadnev/receipt.json for global scope): the exact files written, their hashes, and the version that wrote them. Everything else builds on the receipt:

• ariadnev doctor health-checks the install against it.
• ariadnev audit classifies every installed file as ok, modified, missing, or untracked.
• ariadnev uninstall removes exactly what the receipt lists and preserves files you edited.
• ariadnev adapters regenerate rebuilds a projection other tools can read; nothing in ariadnev reads it back.

#The cache

The binary extracts its embedded kit into a versioned cache under ~/.cache/ariadnev. That cache is the install distribution, not a custom-kit workspace: ariadnev kit refresh discards and re-extracts it, and authoring commands such as ariadnev add-skill are meant to run from a source checkout.