Core Concepts
Understand the key concepts behind Intentra CLI.
Events
An event is a single hook invocation from an AI tool. Each time you:
- Send a message in Cursor chat
- Request code completion
- Run a Claude Code command
- Execute a Gemini CLI prompt
...an event is captured.
Scans
A scan groups related events into a logical unit. For example:
- A chat conversation with multiple back-and-forth messages = one scan
- Code completion attempts while typing = one scan
- Separate, unrelated prompts = multiple scans
Scans make it easier to understand your AI usage patterns.
Normalization
Different AI tools report data in different formats. Intentra normalizes all events to a unified schema:
| Tool | Native Format | Normalized |
|---|---|---|
| Cursor | inputTokens, outputTokens | tokens.input, tokens.output |
| Claude | usage.input_tokens | tokens.input, tokens.output |
| Gemini | tokenCount | tokens.input, tokens.output |
This allows consistent analysis across tools.
Anomaly Detection
Intentra detects anomalies that indicate billing-relevant patterns such as:
- Retry loops - Tool retries failed actions repeatedly
- Ignored instructions - Tool diverged from your explicit directives
- Context waste - Excessive tokens loaded but not used
- Duplicate work - Same task performed multiple times
Detection Methods
Intentra uses AI-powered detection (Claude) to analyze context and identify anomalies. This approach understands intent and catches subtle issues that pattern-based rules miss.
Cost Estimation
Intentra estimates costs based on:
- Token counts (input + output)
- Model pricing
- Usage over time
Note: Estimates are approximate and may not match actual provider bills.
Session Deep Dive
The Session Deep Dive provides a detailed inspection view for individual AI coding sessions. Access it by clicking any session row in Explorer search results.
Conversation Thread
The deep dive displays every event in the session as an expandable conversation thread:
- Thinking — AI reasoning steps shown in italicized text
- Tool Calls — Read, Edit, Bash, and MCP calls with rich content inspection:
- Read: Syntax-highlighted code with file name and line numbers
- Edit: Inline diff view showing added (green) and removed (red) lines
- Bash: Terminal-style output with command and exit status
- MCP: Server and tool name with input/output
- Subagents — Color-coded team member entries showing nested tool call trees
Compaction Visualization
A chart showing context window usage over the session. The sawtooth pattern reveals when the AI's context window fills up and compresses (compaction events), with before/after token counts annotated at each drop.
Visible Context Sidebar
A right-side panel showing the current context window composition — how tokens are distributed across system prompts, user messages, tool outputs, thinking, and context files.
Command Palette
Press ⌘K (Mac) or Ctrl+K anywhere in Explorer to search across all sessions. Click a result to jump directly to its deep dive.