drskill

drskill is a Python CLI that acts as brew doctor for AI coding agent skill loadouts. It scans all coding agents detected on a machine or in a repo, resolves each agent's effective skill set and MCP server configurations, and checks for problems: shadowing, duplicates, description collisions, injection surfaces, broken links, token budget violations, MCP misconfigurations, and tool poisoning. Every finding ends in a fix command or an ack command. The tool is strictly read-only — it never installs, edits, or deletes skills.

The scan runs in three tiers. Static (always runs) reads config files and skill directories with no network or LLM calls. MCP connect (--mcp-connect, opt-in) handshakes with configured MCP servers, enumerates tools, and writes snapshots. Deep mode (--deep, opt-in) sends skill-pair descriptions flagged as overlapping to an LLM (via dspy/LiteLLM) for judgment, budgeted to 25 calls by default.

Key design points include: a fingerprint system (SHA-256 of check + content hashes) so acks survive only while the exact situation persists; a committed verdict cache so one person runs the LLM judgments and the whole team gets them; a self-calibrating lockfile verifier that won't cry wolf when upstream hashing differs; and scope-aware ack routing that separates machine-level from project-level concerns. Seven static injection checks flag suspicious patterns without storing secrets or making LLM calls — the tool quotes evidence and defers intent judgments to the user.

The tool defines 70+ coding agents in data/harnesses.toml, each with search paths, precedence rules, and verified/unverified facets. Uncertainty about harness behaviour is tracked precisely: only findings that actually depend on an unverified facet inherit a ? marker.

Trade-offs include heuristic description checks tuned against public skill corpora (will miss paraphrased conflicts), approximate token counting via tiktoken's o200k_base, and a trust model where both the verdict cache and ack ledger are unsigned — anyone with repo commit access can silence a warning. The tool additionally supports trace auditing (drskill audit) across Claude Code, Codex, Pi, and Copilot session traces to report which skills and MCP tools were actually invoked.