[carlrannaberg]

Visualize currently tracks its runtime four times, and installers can select the unresolved template source. Keep skills/visualize/ as the single authored and installable skill, remove the three redundant trees and their transformer, and point development tooling and CI at that directory.

The retained 180-file payload preserves all templates, vendor files, fixtures and executable modes. Its only changes are a short installed-path explanation in SKILL.md and a catalog CLI entrypoint fix: compare both real paths so symlinked, space-containing and case-variant installation paths emit JSON correctly. The isolated package probe exposed the symlink bug; independent review caught and corrected the first fix's case-variant regression.

Catalog navigation links now point at the sole skill directory. Bundled Clean README navigation is rebased during generation so a future catalog repin retains working installed links. The immutable catalog pin, loader/cache behavior, runtime motion ownership and design workflows remain unchanged.

Validation

• Package check retains the former frontmatter, command-hint and placeholder checks; verifies routed references and executes helpers from an isolated package copy.
• All 14 catalog tests pass locally, including symlink/case-variant CLI invocation and a real isolated Git catalog repin with installed-link resolution. The case-variant test skips on case-sensitive filesystems.
• Pinned skills@1.7.4 discovers one skill. Both symlink and copy installs pass for Claude Code, Codex, Cursor, Copilot, OpenCode, Pi, OpenClaw, Gemini CLI and Hermes Agent. The final catalog helper executes through every actual installation destination.
• Existing checks pass: 73 slide/image tests, 72 fixture checks, palette smoke, 35 template contracts, 34 generated fixture detector checks, 103 catalog/preview freshness checks, diagram/Mermaid self-tests, slide/print --all, syntax and both plugin manifest JSONs.
• The generic skill-creator validator rejects the inherited version, argument-hint and user-invocable fields identically on baseline and candidate. Existing repository metadata checks and actual installer parsing remain in place; host metadata was preserved.

Fresh-agent evidence and limits

Fresh Codex agents received only a workspace path, its environment instructions and a realistic approved recap request. Plaintext prompts, input hashes and tool/result transcripts are retained under temp/consolidation/.

• Online trial: selected Editorial retrieval, local artifact creation and desktop/phone/light/dark verification succeeded. Two trials, including the clean-export trial, loaded two full templates; progressive-disclosure compliance was partial. Another trial loaded one template; neither result establishes consistent progressive disclosure.
• Offline trial: calibrated network-denied shell, cold cache, real Editorial retrieval failure, no Clean substitution and no HTML output. Host bootstrap and workspace instruction reads preceded sandbox use; skill reads and retrieval used the sandbox.
• Fresh Claude procedure trial: loaded review/scoring references, ran installed detector/contrast/print checks, rendered the supplied report and left it unchanged. It found a print defect in that historical report fixture. The catalog entrypoint correction followed this trial; that trial did not invoke catalog. Final Codex trials and catalog tests cover the corrected code.

These are procedure tests with an explicitly supplied skill path. They do not establish native automatic activation. Native Codex/Cursor plugin loading, Pi/Hermes/OpenClaw runtimes, Gemini/OpenCode activation and Claude upload surfaces remain unverified. HARNESSES.md distinguishes the installer evidence from these limits.

Configurations pointing directly at removed mount paths must migrate to skills/visualize/; older pinned commits remain available. The Display Claude plugin pin/rename, directory validation and coexistence check remain a separate consumer PR.

The large deletion count is three redundant 180-file trees plus the transformer and sync wrapper. No catalog packages, templates or runtime dependencies are removed from the retained skill.

Final installer checks and fresh Codex repetitions use a clean export of the staged files, excluding ignored Finder files and empty directories left by previous local work. Both clean-export installer modes and online/offline trials completed successfully; raw transcripts were checked again. Independent Opus review found no blockers; its small wording/link clarifications were applied.

CI passed on c8eafc3209cfbe8b67bb4a77f0b6c593a76112fc: CI run 39587353049, attempt 1, both validate and mermaid-windows.