Seshat-BI: Skill for Claude Code

retail-scaffold

The authoring front door for a NEW seshat check rule. Adding a rule means wiring the same five places every time; doing it by hand silently under-governs a rule when a place is missed (a shipped rule once had no glossary row). This skill runs the retail scaffold CLI verb, which automates the mechanical part with a strict write/print split, and interprets its output.

Boundary vs retail-govern (read first)

Both wrap a governance CLI verb; they do opposite jobs:

• retail-govern -- INTERPRET existing seshat check findings: map a reported rule id (D8, C2, S2, G1, …) to its meaning and fix. It reasons about rules that already exist.
• retail-scaffold (this skill) -- AUTHOR a NEW rule + verify its wiring. It creates a rule that does not yet exist, or doctors an existing rule's five-place wiring for drift.

Use retail-govern when a check FAILED and you need the fix; use retail-scaffold when you are ADDING a rule or checking whether one is fully wired.

Scope + non-negotiables

• Invoke-and-interpret only. This skill runs retail scaffold and reads its output; it does not hand-edit the five places itself.
• The CLI (not this skill) is the source of the five wiring places. They are declared in src/seshat/scaffold.py (FIVE_PLACES); this skill points at that authority and never re-enumerates a competing list (anti-fork).
• Never invents rule intent (DEC-1). The author supplies the id, the title, and the real check logic. The generated stub yields no findings until the author fills it in; its test stub fails on purpose (honest red).
• Never self-grants a wiring pass (Principle I). Whether a rule is correctly wired is disposed of by the test suite + the gate exit code, not by this skill or the helper.
• Prose + golden records are PRINT-only (Principle V). The helper writes exactly three targets and PRINTS the rest (golden-regen commands, a suggested glossary row, the import/__all__ edit) for a human to apply by hand. It never writes the glossary or a golden record.