Claude Code First, specification version 0.1.0. The plain-language guide
Version 0.1.0. Level profile. Best evidence grade R3/S3.
AP-3 Tested as installed
A product's agent surface MUST be exercised from its published artefact, loaded the way an installing consumer loads it, and not only from the source checkout.
| Field | Value |
|---|---|
| Keyword | MUST |
| Level | profile |
| Checked | by judgement |
| Findings | F407, F306 |
| Harness facts | H8, H13; O2 (observed) |
| Threat | construct: defects are known only where a consumer reported them |
| Status | active |
Rationale: roadkeep's published plugin shipped a skill whose frontmatter the loader dropped in silence (O2), an MCP server from which no tool reached the client, and its own CLAUDE.md in the payload; Shio published four skills its own settings never loaded, so "every breakage in them arrived as somebody else's report" [shio@d1853cbea] (F407, R3/S2). Inside its own checkout a session reaches the producer's files by other routes, the repository's .mcp.json (H8), docs and a repo-local skill, so it does not see what a consumer sees. The remedies load the artefact as published: roadkeep runs claude plugin validate --strict in CI at a pinned version (H13), and Shio enables its own published plugin in its committed settings (F306, R3/S3). Discoverability is the same question: Shio named the gap that "the reader with the repository mounted is not the reader who installs (SH605)" [shio@7820e57a7], so a test of whether an agent finds the surface counts only when it runs without the source on disk.