Summation opportunities

Turn recent local IDE chats + what’s already in Summation into a short list of concrete next steps. Privacy first: scan stays on this machine unless the user later runs a normal Summation tool.

When to run

• User: “what can Summation do for me?”, “workflows”, “use cases”, “how can you help my work?”
• the opportunities skill
• Handoff from start after the data gate (only with consent)

Privacy (say this once, plainly)

I can scan recent local Claude Code / Codex sessions on this machine for themes (metrics, SQL, reports, CSVs, …). Nothing is uploaded. I only show short theme summaries, not full transcripts. OK?

• If they decline → skip the scan; still suggest from live catalog / docs if signed in.
• Never paste long chat excerpts. Never invent table names.

Flow

1. Local session scan (optional after consent)

Run the bundled scanner (local only, no network). Prefer the plugin root when the host sets it:

# Prefer the spec plugin root; fall back to Claude's name for the same path:
python3 "${PLUGIN_ROOT:-$CLAUDE_PLUGIN_ROOT}/skills/opportunities/scripts/scan_sessions.py" \
  --days 14 --limit-sessions 20 --json \
  --project-substr "$(basename "$(pwd)")"

# If neither variable is set: path relative to this skill’s scripts/ directory:
python3 skills/opportunities/scripts/scan_sessions.py --days 14 --limit-sessions 20 --json

Caps are intentional. Prefer --json and turn themes into customer language yourself. If the scan fails or finds nothing, say so and continue with catalog-only ideas.

2. Live Summation context (if signed in)

MCP: whoami (signin if needed) → list_data_connections / list_connection_datasets / search_tables or the catalog skill as needed.

• Have named tables → ground each idea in real names.
• No data yet → opportunities that start with the connect skill or import, not fake schemas.

3. Present 3–5 opportunities

Numbered list. Each item:

1. Theme — one short phrase (from scan label or catalog).
2. Why it fits — tie to a redacted sample or a real table (not a wall of chat).
3. Next step — one skill/action: query, report, validate, connect, catalog, schedule, start.

Map themes roughly:

4. Act on their pick

When they choose a number → run that skill. Do not dump the whole menu again.

Rules

• Local scan only for host chat history — no cloud “conversation intelligence.”
• Prefer catalog-grounded next steps when MCP works.
• Product map for “is X supported?” still comes from https://docs.summation.com/llms.txt (see ../api/references/product.md).
• Outcomes in user chat; no OpenAPI archaeology.
• If not signed in: still list opportunities; first step is often the signin or start skill.