gregmos/pii-shield/pii-contract-analyze

⚡ YOUR FIRST ACTION

When the user invokes /pii-contract-analyze <anything>, you respond in TWO turns.

Turn 1 — acknowledge and wait

Do NOT call any tool. Do NOT read files. Do NOT run Bash. Reply with one short line and stop:

Ready to start. Type go or continue to proceed.

Wait for the user's next turn. The MCP deferred-tools registry is often not populated yet on turn 1; it lazy-loads between turns.

Turn 2 — discover and proceed

When the user replies with any continue signal (go, continue, yes, ok, proceed, or equivalents in their language), silently run this discovery sequence:

1. ToolSearch(query: "select:mcp__PII_Shield_v2__list_entities", max_results: 1)
2. If that returns "No matching deferred tools found": ToolSearch(query: "select:mcp__pii-shield__list_entities", max_results: 1)
3. If that also fails: ToolSearch(query: "select:mcp__plugin_pii-shield_pii-shield__list_entities", max_results: 1)

If any of them loads a schema → immediately call list_entities and continue with the Startup procedure, mode detection, and pipeline below. Do NOT surface these attempts to the user.

If all three fail → show the user:

PII Shield MCP tools are installed on your host (Claude Desktop) but this session can't reach them. Known Anthropic bridging bug on Windows. Fixes: (1) restart Claude Desktop and start a fresh session, or (2) install pii-shield-v2.0.1-plugin.zip directly into this Cowork session. Meanwhile I can proceed without PII anonymization — OK?

Rules

• Never call ToolSearch on turn 1. The prompt "type go" is the whole turn-1 response.
• Never fuzzy-search with bare keywords ("list_entities", "pii-shield") — underscore names don't match as substrings on Cowork CLI.
• Never declare the plugin missing before turn-2's full three-attempt select: chain has run.
• Never spawn sub-agents, grep the codebase, or probe filesystem paths / localhost ports / beacon files hunting for the server. If MCP tool discovery fails, the three select: attempts above are the whole fallback chain; anything beyond them is off-limits.

PII Shield — Universal Legal Document Processor

Anonymize → Work → Deanonymize → Deliver. Claude NEVER sees raw PII at any stage.

CRITICAL: PII never flows through Claude

File handling: The user must connect a folder (not attach the file directly to the message). When a file is attached to a chat message, its content is rendered and sent to the API as part of the prompt — Claude sees the raw data before PII Shield can process it. When a folder is connected, Claude only sees the file path and calls anonymize_file(path) — the MCP server reads and anonymizes the file locally. PII never enters Claude's context.

If the user attaches a file directly: Warn them politely: "For full PII protection, please connect the folder containing your document instead of attaching it directly. When a file is attached to a message, its content is included in the API request before PII Shield can anonymize it. I can still process it, but the privacy guarantee is stronger when you connect the folder."

• anonymize_file reads the file locally, anonymizes it, writes the result to disk, returns only output_path + session_id to Claude. After HITL is approved, and only then, Claude reads the anonymized text from the output file — never before.
• deanonymize_* tools write results to LOCAL FILES and return only the file path
• get_mapping returns only placeholder keys and types — no real values
• ABSOLUTE BAN #1 — HITL GATE: Claude must NEVER read, open, cat, head, pandoc, use the Read tool, python, bash, or in any way access the anonymized output file (output_path, docx_output_path) BEFORE the review panel reports that the user has clicked Approve (see "Human-in-the-Loop Review" below for how that signal arrives). Not to "preview entity quality", not to "verify placeholders", not to "check formatting", not to "plan the memo" — NEVER. The anonymized file is considered SEALED between anonymize_file and HITL approval. The HITL reviewer is the human, not Claude.
• ABSOLUTE BAN #2 — DEANONYMIZED FILES: Claude must NEVER read, open, cat, head, pandoc, or in any way access the content of deanonymized/restored files. Not to "verify", not to "check formatting", not to "validate" — NEVER. These files contain real PII. Just give the user the file path and STOP. Any "verification" of deanonymized output is a PII leak.
• Claude must NEVER read the source file (via Read tool, pandoc, python, bash, etc.) BEFORE or INSTEAD OF anonymization — always use anonymize_file(path) first
• If an anonymize tool times out or fails with a NON-"tool not found" error — retry once. If it still fails, tell the user PII Shield is unavailable and ask whether to proceed without anonymization or abort. NEVER fall back to reading the raw file.
• NEVER use anonymize_text or scan_text — these take raw text as input which means PII passes through the API. The ONLY exception is if the user explicitly pastes text into the chat (PII is already in the conversation).

Startup

PII Shield is a pure-Node.js MCP server — no Python dependency, instant startup. On first run, the NER model (~665 MB fp32 ONNX GLiNER) and its runtime deps (onnxruntime-node, @xenova/transformers, gliner) download into ${CLAUDE_PLUGIN_DATA}/models and ${CLAUDE_PLUGIN_DATA}/deps. This takes 2–5 minutes once per plugin install and is cached for the full life of the plugin (survives host restarts, only wiped by /plugin remove).

⛔ ABSOLUTE RULE — NO SUB-AGENT DELEGATION

NEVER delegate PII Shield tool calls to a sub-agent. Not to a general-purpose agent, not to a Task agent, not to an Explore agent — NEVER. Sub-agents do not stream text to the user; they return one final message only when they exit. If PII Shield is initializing, a sub-agent will poll silently for minutes while the user sees nothing. This is the single worst UX failure mode. All PII Shield tool calls (list_entities, anonymize_file, start_review, etc.) MUST happen in the MAIN conversation.

If you cannot call a PII Shield tool because it shows "No such tool available" — the fix is the turn-1 / turn-2 pattern above (prompt for go, discover silently on turn 2), NOT a sub-agent.

Startup procedure

1. Call `list_entities` — this happens on turn 2 after the user sends a continue signal. See the YOUR FIRST ACTION block at the top of this skill for the exact two-turn flow. You MUST have list_entities responding before proceeding.
2. Identify the file(s) to process and determine the mode (MEMO, REDLINE, etc.)
3. Read the list_entities response to check NER status

• If "ner_ready": true — proceed to anonymize_file
• If "ner_ready": false — NER is still initializing. The response includes phase (installing_deps / downloading_model / loading_model), progress_pct, a human message, and a pre-formatted user_message field. If the response ALSO contains a `first_run_notice` field (only present on the very first loading response per server process), print `first_run_notice` VERBATIM to the user as a plain chat message BEFORE anything else. It explains where the ~700 MB NER cache will live and why the next session will be instant — the user needs to see this once, up front. Subsequent polls will NOT contain first_run_notice. On every poll (including the first), print the `user_message` field VERBATIM to the user as a plain chat message BEFORE calling `list_entities` again. This is the ONLY thing the user sees during the wait — do not paraphrase, do not summarize, do not skip it, do not batch it silently. Wait and retry: the server enforces a ~20 second throttle by holding the list_entities response for 20 s internally while phase is installing_deps / downloading_model / loading_model. First run may take 2–5 minutes. Between polls (inside the 20 s window) you MAY do useful prep work in the MAIN conversation only — read skill references, plan the analysis. Do NOT delegate any of this to a sub-agent (see the ABSOLUTE RULE above). Do NOT call `anonymize_file` until `ner_ready: true` — without NER, only regex patterns work, missing PERSON/ORGANIZATION/LOCATION entities.
• If "ner_error" field present — show it to the user. If "ner_error_suggestions" array is also present (platform-specific recovery steps like "install VC++ Redistributable", "switch to Node 22 LTS"), print each entry verbatim as a bulleted list — these are the concrete actions the user should try next. If "ner_error_diagnostic" object is present, its likely_cause field is a one-word root-cause tag useful to include in any bug report the user may file.

Long document handling (chunked processing)

For documents >15K characters, anonymize_file returns "status": "chunked". Chunked processing flow:

1. anonymize_file(path) returns session_id, total_chunks, processed_chunks: 1
2. Loop: call anonymize_next_chunk(session_id) until status is "complete" — show "Anonymizing... [chunk X/Y]"
3. Call get_full_anonymized_text(session_id) to finalize — returns output_path, session_id, output_dir
4. Continue with the normal pipeline using the returned values

For short documents (<15K chars), anonymize_file processes everything in one call.

File path resolution

Call anonymize_file(file_path: "<path or filename>") directly — no ceremony. The server auto-resolves:

1. The path as-given (if it's a valid absolute host path it just works)
2. $PII_WORK_DIR/<basename> if that env is set
3. BFS (depth 4) of ~/Downloads, ~/Documents, ~/Desktop, $PII_WORK_DIR for an unambiguous match

If the response is status: "error" with a "file not found" or "ambiguous filename" hint — the file is in a non-standard location. Fall back to:

# create a marker next to the target file
touch "/path/visible/to/you/.pii_marker_abc"
# then:
resolve_path(filename: "<basename>", marker: ".pii_marker_abc")
# take host_path from the response and retry:
anonymize_file(file_path: "<host_path>")

The marker+resolve_path tools stay available as a reliability net — the auto-BFS handles ~95% of cases, marker covers the rest.

Available MCP tools