dsh-skill-trace

When an Agent silently selects a Skill and delivers output, what you actually see is only the result. You have no readable record of what it loaded, which steps it declared, what it depended on, or whether you could replicate the work yourself. dsh-skill-trace gathers every observable Skill-loading event in a session, the declared steps, and your own notes into a single local receipt, turning a fleeting side effect into something you can review, learn from, and manually extend.

Three views anchor the experience. The Skill receipt lays out each turn, step, request, result, and empty state in chronological order, explicitly distinguishing "not observed" from "not used" and never filling gaps with sample data. The flow map reorganises the same fact set into nodes and dependency edges for users who think in system structure, reusing the same source of truth without inventing an extra layer of AI inference. The My Skill workspace is a read-only learning index: search discoverable Skills in the current scope, write down your understanding, improvement intent, and verification plan for each session, then log the outcome of a manual re-test and the next action. Everything stays on your machine. The plugin never stores full prompts, full Skill bodies, tokens, cookies, credentials, or project files; it does not discover, install, sync, route, rewrite, or publish any Skill, and it implements no telemetry or cloud sync.

A clear evidence line runs through every view: a successful skill(name) call only proves the tool was loaded; it does not prove the Agent followed every instruction, nor that the Skill caused the correct outcome. That boundary is stated explicitly in the receipt, the map, and the learning loop. If you work in DeepSeek Harness Desktop and want to move from "I saw it happen" to "I can do it next time," if you refuse to equate "loaded" with "effective," or if you need a reviewable record before touching a Skill, this local receipt puts the facts, the path, and your judgment in one place you can keep reopening.