Insights

Sophon's insights engine watches what you do — sessions, messages, tokens, tool invocations, workflow runs, cron jobs, memory growth — and turns it into something you can glance at. The Dashboard home shows an AI-written summary of your week; mobile's home tab streams a feed of cards ("3-day streak", "you used web.search 14 times Monday"); the agent can query insights directly and bring them into a conversation.

What it computes

The engine computes 15+ metrics rolled up across time buckets (day / week / month):

• Sessions — total, active, longest, shortest
• Messages — sent, received, by agent, by channel
• Tokens — consumed, by provider, by agent, cost estimate
• Tool invocations — top tools, tools by risk, tools per agent
• Workflows — runs, success rate, average duration
• Cron jobs — fires, success rate
• Memory — entries added, entries forgotten, top entities
• Approvals — requested, approved, rejected, timed out
• Documents — uploaded, indexed size, top tags
• Streaks — days with at least one session

Metrics are cached 30 minutes. The cache is user-scoped — your metrics never see anyone else's data.

Insight cards

Raw metrics are dull. The engine produces cards — rule-based contextual callouts:

Cards surface in the Dashboard home section and mobile home feed. Each card shows the metric it's based on; you can dismiss cards you don't want to see again.

AI-generated summary

The engine periodically composes a short narrative of your activity using the configured LLM:

"This week you spent most of your time in code-review sessions with the reviewer agent — 47 messages total. Your morning-briefing workflow fired every weekday and you approved all 5 summary emails. Token spend is down 12% from last week, tracking closer to your budget."

The summary is cached 2 hours (to save tokens), with a fallback to a rule-based template if the LLM call fails or times out. It's user-scoped, so your summary reflects only your activity.

Home feed (mobile)

The mobile app's Home tab is built around insights:

• Hero card — greeting ("Good morning, Enes"), current streak, quick stats
• Daily thought — rotating curated quote (one of 12 built-ins)
• AI summary — the narrative from above, truncated for mobile
• Insight cards — scrolling list
• On this day — LLM-generated reflection from historical activity one year / one month / one week ago
• Active tasks — live list from the task ledger
• Recent sessions — jump back in

Pull-to-refresh recomputes metrics immediately (bypasses the 30-min cache).

Dashboard

Home on the Dashboard mirrors the mobile feed with more detail:

• Hero greeting + headline stats
• AI summary (full length)
• Insight cards
• Charts — session timeline, token usage, tool usage breakdown, workflow success rate

Settings → Insights lets you:

• Enable / disable specific card categories
• Adjust sensitivity of anomaly detection
• Bypass the AI summary (keep only rule-based cards)
• Clear card dismissals

The insights.query tool

Agents can query insights during a conversation:

insights.query({
  category: "tokens",       // or "sessions", "tools", "workflows", "memory", …
  window: "week",           // day / week / month / all
  format: "summary"         // summary (LLM), raw (JSON), cards
})

Useful for questions like:

"How often did we use the gmail.send_email tool this month?" "What's my session streak?" "Which workflow has failed most often?"

The agent pulls the metric and answers without needing its own analytics implementation.

CLI

sophon insights metrics                 # all metrics for the current week
sophon insights metrics --window month
sophon insights summary                 # AI narrative
sophon insights cards                   # active insight cards
sophon insights query --category tokens --window day

REST API

GET  /api/insights/metrics?window=week
GET  /api/insights/summary
GET  /api/insights/cards
POST /api/insights/cards/{id}/dismiss

All endpoints are user-scoped — they return only the calling user's insights.

Configuration

{
  "Sophon": {
    "Insights": {
      "MetricsCacheDuration": "00:30:00",
      "SummaryCacheDuration": "02:00:00",
      "AiSummaryEnabled": true,
      "AnomalyDetectionEnabled": true,
      "CardCategories": ["activity", "usage", "productivity", "milestone", "anomaly"]
    }
  }
}

Limits and gotchas

• User-scoped only. There's no tenant-level aggregation for admins today. Enterprise admins can query usage reports via the Admin → Audit pipeline instead.
• AI summary costs tokens — the summary uses a fast-tier model with a 2-hour cache, so costs are small, but they're not zero. Disable AiSummaryEnabled if you don't want it.
• Anomaly detection is rule-based, not ML. It catches obvious spikes (3× vs prior) but won't catch subtle drift.
• On this day needs at least 7 days of history to start being interesting.
• Metrics are eventually consistent — events (tool invocations, token usage) can take up to 30 minutes to appear in metrics due to the cache.

Where to go next

• Mobile App — the home feed that Insights powers
• Tasks — task-level history that feeds metrics
• Audit Logging — admin-level aggregation for compliance