Implements symlink-safe, atomic flag file writes using O_NOFOLLOW and temp+rename patterns, with ownership verification for symlinked parent directories. Also provides a layered config resolution (env var → repo-local .caveman/config.json walking up to root → user XDG config → default). The readFlag function enforces size caps and whitelist validation to prevent injection attacks. This is genuinely reusable defensive filesystem I/O code that solves the common problem of safely writing predictable-path config/flag files in shared or symlinked home directories.

A JSONC-tolerant settings.json reader/writer with a hand-rolled state-machine comment stripper that correctly handles comment-looking sequences inside quoted strings. Includes atomic writes (temp + rename), defensive hook field validation that prevents a single malformed entry from poisoning an entire settings file, idempotent add/remove helpers using substring markers, and orphaned-hook pruning. The stripJsonComments function alone is a reusable utility for any tool that needs to read JSONC config files without pulling in a heavy dependency.

A pure-Node prose compressor that mirrors the caveman style rules — dropping articles, filler, pleasantries, hedging, and leaders while preserving protected segments (fenced code, inline code, URLs, paths, CONST_CASE identifiers, dotted.method calls, version numbers). Uses a sentinel-based approach to protect segments during transformation. The compress + compressDescriptionsInPlace API is reusable for any tool that needs to shrink MCP tool descriptions or similar prose fields without breaking technical content.

Orchestrates LLM-based file compression with multiple safety layers: a sensitive-file denylist (credentials, keys, .ssh/.aws paths) that refuses to send content to third-party APIs, backup verification via readback comparison before overwriting the original, YAML frontmatter preservation by surgical split/reassemble, empty/identical output rejection, and a validate-then-targeted-fix retry loop. The safety guard pattern — refusing to process files matching secret/PII heuristics before any API call — is a reusable defensive pattern for any tool that ships file content to external services.

Validates compressed markdown against originals by checking heading count/order, exact code block preservation (with nested fence support), URL sets, path sets, inline code occurrence counts, and bullet count drift. Returns structured ValidationResult with errors vs warnings. This is a reusable pattern for any lossy text transformation tool that needs to verify structural and content integrity of the output.

A three-arm eval harness that compares baseline (no system prompt), terse control ('Answer concisely.'), and terse+skill to isolate the skill's actual contribution from the generic terseness effect. Runs real LLM calls per (prompt, arm) and snapshots outputs with metadata. The design insight — measuring skill delta against a terse control rather than baseline — is a genuinely useful pattern for anyone building prompt skills who wants honest efficacy numbers.

An ESM Bun plugin that handles dynamic caveman mode tracking for opencode, with a creative cross-runtime module loading solution: it reads the CommonJS caveman-config.js source and evaluates it via new Function() with injected module/exports/require because opencode's compiled Bun binary rejects both require() of on-disk files and await import() of CJS. Also handles opencode's command template expansion (recovering the level argument from expanded template text) and quoted-message unwrapping for non-interactive mode.