bubble-shield-anonymizelisted

# Bubble Shield — anonymise before reading The `bubble-shield` hook blocks reads of protected client folders because raw identifying data must never enter the model context in clear. This skill is the sanctioned path: anonymise locally first, then work on the cloaked copy. ## When you were just blocked If a tool call was denied with a `🔒 Bubble Shield` message, do NOT try to bypass it (no `cat`, no copying the file elsewhere, no reading via Bash). ### Fastest path — read it through `bubble_shield_read` (one file) The plugin ships an MCP tool **`bubble_shield_read`** (namespaced, e.g. `mcp__plugin_bubble-shield_bubble_shield__bubble_shield_read`, or just call it `bubble_shield_read`). Give it the blocked file's path and it returns the file's contents with client PII replaced by `⟦NOM_0001⟧`-style tokens. This is the preferred way to read a single protected file, **especially in Cowork**, where it's the mechanism that actually works (a normal Read of a protected file is blocked by design; `bubble_shield_read` is the sanctioned read). ``` bubble_shield_read(path="~/Dossiers-clients/dossier-dupont/contrat.pdf") → returns the cloaked text; work on THAT. ``` It handles .pdf/.docx/.txt/.md/.csv/.json, and uses the same vault as the rest of Bubble Shield (so tokens are consistent and reversible). ### How `bubble_shield_read` works in this version — READ THIS Since v1.23.0 the read path is **fast by design (zero AI models at read time)**. It serves a **pre-computed masked co