assistant-setup

Set up the assistant for the current user.

Read the procedure in agents/assistant-onboarding.md and execute it inline as the main assistant — do not try to spawn assistant-onboarding as a subagent (custom agents in this plugin are procedure documents, not registered subagent types). The steps:

0. Run the connection check first. Before anything else, run ./scripts/setup-connections.sh --check via mcp__Control_your_Mac__osascript (file and bash tools are sandboxed in Cowork and cannot reach the plugin directory). Surface the full output in chat. If the Salesforce CLI (sf) is missing, tell the user to install it (npm install -g @salesforce/cli, then sf org login web, then claude mcp add salesforce -- npx -y @salesforce/mcp). Do not skip this step.
1. Resolve the persistent data directory. Run the following via mcp__Control_your_Mac__osascript (Cowork) or the Bash tool (Claude Code CLI):

   PLUGIN_DATA_DIR=$(cat "$HOME/.claude/aise-assistant.datadir" 2>/dev/null)
   if [ -z "$PLUGIN_DATA_DIR" ]; then
     PLUGIN_DATA_DIR=$(ls -d "$HOME/.claude/plugins/data/aise-assistant"* 2>/dev/null | head -1)
     PLUGIN_DATA_DIR="${PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/aise-assistant}"
   fi
   echo "PLUGIN_DATA_DIR=$PLUGIN_DATA_DIR"
   ls "$PLUGIN_DATA_DIR/about/" 2>/dev/null || echo "(empty — fresh install)"
   

   Capture the printed PLUGIN_DATA_DIR value. Use it as the literal path for all file reads and writes in this session. Never use $CLAUDE_PLUGIN_DATA for writing — it resolves to a volatile temp path, not the persistent directory.
2. Detect the Notion connection and resolve the current user via notion-get-users self → auto-fills identity.md with the Notion user ID.
3. Ask HITL questions covering identity, voice, and workspace preferences in one combined elicitation form (call read_me with modules: ["elicitation"] first, then render a single card — no sequential question-by-question flow). Reserve AskUserQuestion only for a single ad-hoc clarification that arises after the form is submitted.
4. Auto-discover the AISE team roster from the Customer Tracker (Step 2.5 in the agent) and present it for confirmation in the combined HITL form.
5. Optionally (--scrape-voice or when the user opts in via the form): read 5–10 recent sent emails from Gmail and 5–10 recent Slack messages, distinguishing internal vs client-facing tone, and draft a voice.md from the user's actual writing style. For Slack, read the slack_search_public_and_private tool description to find the Current logged in user's user_id is <ID> line — use from:<@USER_ID> as the query (not the email address).
6. Write files to <PLUGIN_DATA_DIR>/about/ (the literal path from step 1) — identity.md, voice.md, workspace.md, tracker-memory.md, team-roster.md. Create the directory if it doesn't exist. Present computer:// links to each written file so the user can open them.
7. Confirm setup in chat. These files live at <PLUGIN_DATA_DIR>/about/ and persist across plugin updates. They are deleted on uninstall and are machine-specific — re-run /assistant-setup after a full reinstall or on a new machine.

Modes (mutually exclusive):

• Default (no flag) — fill gaps only. Preserves existing values in <PLUGIN_DATA_DIR>/about/; asks only about fields still set to <TBD>. Safe to re-run whenever.
• --update — drift check. Re-resolves Notion identity (catches user ID changes, role changes), then walks each section asking the user to confirm or update values that may have drifted. Useful after a role change, team move, or workflow shift.
• --reset — wipe everything. Deletes the three files from <PLUGIN_DATA_DIR>/about/ and runs the full onboarding flow from scratch. Use when handing the assistant off to a teammate, or starting clean after a major shift.

Modifier flag (combinable with any mode):

• --scrape-voice — skip the opt-in question and go straight to Gmail+Slack scraping for the voice draft. Distinguishes internal vs client-facing tone.

Don't ask the user for values that are retrievable. Notion user ID, primary email, time zone — pull from the connected account, don't ask. Reserve HITL questions for genuine preferences (sign-offs, em-dash rule, Slack register, etc.).

Save scraped raw email/Slack samples to a temp file the user can reference if they want to tweak the inferred voice.md by hand.

Cowork: reading about/ files

In Cowork, Read is sandboxed to session outputs. Use mcp__Control_your_Mac__osascript to read files. Always derive the path from the pointer file — not from $CLAUDE_PLUGIN_DATA which is a volatile runtime var.

Pattern:

do shell script "PLUGIN_DATA_DIR=$(cat \"$HOME/.claude/aise-assistant.datadir\" 2>/dev/null || ls -d \"$HOME/.claude/plugins/data/aise-assistant\"* 2>/dev/null | head -1); cat \"$PLUGIN_DATA_DIR/about/identity.md\" 2>/dev/null || echo NOT_FOUND"

Run one call per file (identity.md, voice.md, workspace.md). If a file returns NOT_FOUND, the user hasn't run /assistant-setup yet — prompt them to do so before continuing.

Cowork: writing about/ files

In Cowork, Write, Edit, and mcp__workspace__bash are sandboxed. Use mcp__Control_your_Mac__osascript to write to <PLUGIN_DATA_DIR>/about/. Always derive the path from the pointer file — not from $CLAUDE_PLUGIN_DATA.

Pattern: Python script via heredoc

1. Destination path — read from the pointer file inside the Python script:

   import os, pathlib, glob, subprocess
   result = subprocess.run(['sh', '-c', 'cat "$HOME/.claude/aise-assistant.datadir" 2>/dev/null'],
                          capture_output=True, text=True)
   data_dir = result.stdout.strip()
   if not data_dir:
       matches = glob.glob(os.path.expanduser('~/.claude/plugins/data/aise-assistant*'))
       data_dir = matches[0] if matches else os.path.expanduser('~/.claude/plugins/data/aise-assistant')
   about_dir = pathlib.Path(data_dir) / 'about'
   about_dir.mkdir(parents=True, exist_ok=True)
   

2. Write all three files in one osascript call. The pattern:

   • do shell script wraps a bash heredoc using << 'PYEOF' (single-quoted delimiter — passes content through literally with no substitution)
   • Python creates the directory then writes each file using .write_text(content, encoding='utf-8')
   • File content goes in Python triple-single-quoted strings, which handle apostrophes and backticks without restriction

3. Critical constraint: The outer do shell script command is an AppleScript string delimited by double quotes. Any double-quote character anywhere inside the heredoc content will terminate this outer string early and cause a syntax error. Keep all file content free of double-quote characters. If a double-quote is unavoidable, use the Python hex escape \x22 in the Python source.

4. Verify after writing. In Cowork, Read is sandboxed — use osascript cat calls (same pattern as the read section above) to confirm content landed correctly before reporting success.