local-budget: Skill for Claude Code

Budget Visualizer — the shared visual discipline

This is the shared visual discipline any budget-* skill references when the user wants a chart or full visual report. The individual skill still decides WHEN to offer one.

The primary path: render_report

Visual reports are rendered by the deterministic renderer (src/local_budget/report/), not hand-authored per request. The skill that owns the request (normally budget-monthly-brief) calls render_report(period, narrative?) after user confirmation. The renderer:

• builds the fixed page — masthead → stat strip → spend-vs-budget chart → monthly trend → flags — directly from reports.py/detect.py data, so the over/floor classification, the shared bar scale, the anomaly month-filter, the recurring cross-reference (aliases, bill-like allowlist), and money formatting are all computed once, server-side, and covered by tests (tests/test_report_charts.py, tests/test_report_flags.py);
• renders in PRESS — the warm-paper editorial brand shared with local-fitness, city-report, devlog, ghostwriter and the résumé. Owned by src/local_budget/report/brand.py (paper #F5F0E6, ink #181510, dim #6E675C, accent #E8501F), overridable via BUDGET_BRAND_FILE, and guarded by tests/test_report_brand.py. web/static/palette.css is the DASHBOARD's colors only — the PDF no longer reads it;
• writes reports/budget-report-<period>.pdf (gitignored dir, 0700/0600) via headless Chrome discovered automatically (LOCAL_BUDGET_CHROME overrides).

The only LLM-authored content on the page is the optional narrative — 1-3 plain-text sentences grounded in rendered blocks already printed this conversation. It is HTML-escaped into a fixed slot; markup is not rendered.

Chart-content changes (row rules, colors, new sections) are code changes to report/charts.py + its golden tests — relay user requests for them as feedback, don't try to override per request.

An ad-hoc "just one chart" ask outside the monthly report is still served by the full render_report PDF — point the user at it rather than hand-building a one-off (a lone stat row is the only exception worth hand-writing, and even then prefer the tool).