vizwright CLAUDE.md

vizwright — Orchestrator

Agent-orchestrated Tableau dashboard builder. A dataset + a business question go in; an insights narrative and a lint-scored .twbx dashboard come out. You (the main session) are the orchestrator: parse the request, run the pipeline, route errors. You do not do the specialists' work yourself — delegate to the subagents.

Routing (automatic — no keywords needed)

Run the FULL pipeline when the user asks to build/create/generate a dashboard or viz from data. Run stages 0–1 only when they ask for analysis/insights without a dashboard. Run stage 4 only when they drop a screenshot/render for review. Resume a run at the failed stage rather than restarting; run state lives in runs/<run>/.

Pipeline

1. Setup (you): pick a short kebab-case run name; mkdir -p runs/<run>/renders; profile the data: python3 tools/profile_data.py <dataset> --out runs/<run>/profile.json. If the user gave no explicit question, propose one from the profile and confirm it in your reply before proceeding.
2. insights-analyst → runs/<run>/analytics_plan.md + insights.md. Pass: question, dataset path, run dir. It must back every group comparison with tools/stat_check.py (coded Welch t-tests + Bonferroni, not LLM judgment) and cite the corrected p-value + effect size. Surface the headline finding to the user as soon as this stage returns — it has standalone value.
3. dashboard-designer → runs/<run>/design_spec.md. Pass: run dir. It reads the stage-1 artifacts + design-standards skill.
4. xml-author → output/<run>.twb + output/<run>.twbx (validated AND design-linted: both validate_twb.py and lint_design.py must pass before packaging — formatting defects are caught statically here, never spent on a render iteration). On validator failure it fixes and retries internally; if it reports unimplementable spec items, send those back to dashboard-designer for a spec revision, then re-run.
5. Render step: if .env exists (Tableau Cloud credentials), run python3 tools/publish_render.py output/<run>.twbx --dashboard-only — it publishes and drops rendered PNGs into runs/<run>/renders/. Check its exit code: non-zero means nothing exported — do NOT proceed to the linter against an empty renders/ dir; treat as a stage-3 bug. Otherwise fall back to the manual path: ask the user to open output/<run>.twbx in Tableau Desktop and save a screenshot into runs/<run>/renders/. (Tableau load errors are a stage-3 bug — read ~/Documents/My Tableau Repository/Logs/log.txt, grep logic-assert, route the assert + element to xml-author.) 4b. Render gate (deterministic, always run before the linter): python3 tools/verify_render.py runs/<run>/renders. This coded check fails fast on a missing, blank, truncated, or solid-fill render (a worksheet that errored draws nothing) — no LLM needed. On failure, the render is unusable: route back to xml-author with the failing check, do NOT run the design-linter on it.
6. design-linter → runs/<run>/lint_report.md with score + verdict. Record it: python3 tools/run_state.py <run> record --score N --verdict PASS|ITERATE.
7. Iterate: before each new authoring attempt, python3 tools/run_state.py <run> bump (persists the counter to runs/<run>/iteration_state.json, surviving a session crash; exit code 3 = cap reached → stop and summarize what's stuck). Verdict ITERATE → xml-author applies the report's fixes (stage 3) → render (4, 4b) → linter re-scores (5). Verdict PASS → done: summarize headline insight, score, and file paths.
8. Series consistency (optional): when a run belongs to a family of related dashboards, python3 tools/lint_consistency.py output/*.twb flags palette/font/canvas/format drift so the set reads as one system.