Crosstabs for Codex and Claude Code

This repository contains source release 0.4.0, identified by GitHub tag v0.4.0. It pins crosstabs==1.4.0 for local execution. Source publication does not imply directory submission, approval, or listing.

The plugin declares two on-device stdio MCP servers:

• crosstabs-headless: exactly 35 deterministic project-workflow tools for local files, project handoff, approved dictionary edits, guarded transformations, design-aware methods, canonical aggregate widgets, tabulation, reports, and audit history.
• crosstabs-statistics: 40 registered tools (39 analytical tools plus runtime status) and two local evidence resources.

Together the two servers declare 75 tool registrations. The 35-tool headless catalog is recorded exactly in plugins/crosstabs/parity.json. Package publication and Registry status are recorded separately from source publication in release-state.json. Both provider-directory submission, approval, and listing states remain false.

The historical source release 0.3.0 pins package 1.3.0; its complete evidence and older 0.2.6/1.2.2 evidence remain recorded separately. The shared get_runtime_status name appears on both servers, so 75 registrations represent 74 distinct names.

Local execution boundary

• Plugin version: 0.4.0.
• Exact package requirement: crosstabs==1.4.0.
• Transport: local stdio only.
• Runtime: Python 3.10 or newer and Node.js 22 or newer.
• Project files, respondent rows, project state, calculations, and generated reports remain in the user-controlled local process and filesystem.
• The declared Registry identity is io.github.crosstabs/crosstabs; the exact 1.4.0 record was verified active at publication timestamp 2026-09-06T01:01:07.667448Z. Registry publication does not establish directory approval.
• The historical public plugin 0.2.3 remains recorded separately in plugins/crosstabs/release-state.json.

The exact uvx commands are present in plugins/crosstabs/.mcp.json. Installation requires the exact package to be published and independently verified; check sourceRelease.packagePublished before use.

Install the source release

After the tag and exact package are public, add the GitHub-backed Codex marketplace:

codex plugin marketplace add barangaroo/crosstabs-codex-plugin --ref v0.4.0
codex plugin add crosstabs@crosstabs

For Claude Code, load the tagged source directly without a directory listing:

git clone --branch v0.4.0 --depth 1 https://github.com/barangaroo/crosstabs-codex-plugin.git
claude --plugin-dir ./crosstabs-codex-plugin/plugins/crosstabs

These are opt-in local loading paths, not an official directory listing. The OpenAI submission guide currently does not support local stdio-only MCP plugins; do not move private project workflows to a public endpoint to bypass that restriction. A separate Claude directory submission requires an authorized Console or organization session and a retained submission receipt.

The local JSON store enforces 8 MiB per serialized project (including audit and undo) and 32 MiB per combined database (including replay). Capacity rejection preserves the original database. A disposable four-project fixture with 3,200 rows, 12 columns, and one full undo snapshot per project exercised import, save, reload, and analysis at 33,308,653 database bytes (99.3% of the file cap). This is a bounded fixture, not a maximum row count or a memory or compute guarantee. Parser admission ceilings do not override persistence limits. Release verification must use the exact packaged artifact.

Runtime and reproducible plans

Call get_runtime_status on each server with expectedPackageVersion: "1.4.0" before relying on its tools. Check versionMatch and the returned inventory. This reports the executing package, not the installed plugin version. Restart stale MCP processes after updating the plugin; do not infer a newer runtime from a manifest pin.

Create a portable plan with create_analysis_plan for explicit saved table IDs and the inspected revision. Review its readable settings, call validate_analysis_plan, and use run_analysis_plan only when valid. Stale data, definitions, or revisions require a newly created and reviewed plan. Plans and runs are read-only; artifact saving is a separate explicit local action. Workload limits and unsupported complex-survey plans must remain visible.

Guided sampling weights

Use preview_raking_weights with explicit population targets for the active saved wave. Review aggregate convergence, margins, weight diagnostics and warnings; no rows are silently dropped and preview does not mutate the project. apply_raking_weights requires the matching reviewed fingerprint, explicit human approval, acknowledgement of every warning, revision and replay guards, a new weight column/definition, and explicit table assignments. It creates sampling weights, not permission for ordinary or complex-survey inference.

Local validation

Run the source-only verifier:

python3 scripts/verify_plugin.py

Validate the shared Claude Code package structure when the Claude CLI is available:

npx --yes @anthropic-ai/claude-code@2.1.201 plugin validate --strict plugins/crosstabs

Run the exact packaged servers after PyPI publication:

uv run --isolated --python 3.12 --with "mcp==1.29.0" --with "packaging==26.2" python scripts/mcp_smoke.py

For a source-only headless check without resolving the package:

uv run --isolated --python 3.12 --with "mcp==1.29.0" --with "packaging==26.2" \
  python scripts/mcp_smoke.py \
  --headless-bundle ../crosstabs-lite/mcp-server-python/crosstabs_mcp/headless-mcp.mjs

The Codex and Claude manifests reuse the same skills/ tree and .mcp.json; there is no second behavior fork.