Session Monitor¶

When your coding agent runs autonomously, you're flying blind — tokens burn silently, context fills up, and tool errors pile up without warning. The session monitor gives you a real-time dashboard so you can catch problems early and understand where your budget is going.

Monitor your coding agent sessions in real-time with a comprehensive analytics dashboard. Supports Claude Code, OpenCode, and Codex CLI.

Accessing the Dashboard¶

Click the Agent Hub icon in the activity bar (left sidebar) to access all monitoring views.

Session Analytics Dashboard¶

The main dashboard panel provides:

• Token Usage — real-time input/output token tracking with model-specific pricing
• Cost Tracking — per-model cost breakdown with accurate pricing
• Context Token Attribution — stacked bar chart showing where your context budget goes (system prompt, CLAUDE.md, user messages, assistant responses, tool I/O, thinking)
• Token Usage Tooltips — hover for quota projections and estimated time to exhaustion
• Context Window Gauge — input/output token usage vs. limits, with theme-aware colors that adapt to light, dark, and high-contrast themes
• Compaction Detection — timeline markers showing when context was compressed and how much was lost
• Context Health — real-time fidelity score showing how much context degradation has occurred from compactions, with a color-coded gauge (green/yellow/red)
• Truncation Tracking — detects when tool outputs are truncated by the agent, with per-tool breakdown and total count
• Cycle Detection — alerts when your agent enters repetitive loops (e.g., Read → Edit → Read → Edit on the same file)
• Activity Timeline — user prompts, tool calls, errors, and subagent spawns with full-text search
• Timeline Search & Filtering — filter by event type, noise classification (system reminders, sidechains)
• Session Navigator — collapsible panel to switch between active and recent sessions
• Tool Analytics — categorized tool usage with drill-down to individual calls
• Toast Notifications — dismissable feedback toasts (e.g. "Copied to clipboard") with aria-live for screen readers
• Session Summary — AI narrative generation with progress notification
• Session Dump — export the current session as text, markdown, JSON, or HTML via Sidekick: Dump Session Report (Command Palette, status bar menu, or Session Analytics toolbar)
• HTML Session Report — self-contained HTML report with full transcript, token/cost stats, model breakdown, and tool-use summary via Sidekick: Generate HTML Report (Command Palette or Session Analytics toolbar). Also available in the CLI as sidekick report or by pressing r in the TUI dashboard

Dashboard Sections¶

The dashboard organizes information into three collapsible groups:

• Session Activity — Activity Timeline, File Changes, Errors
• Performance & Cost — Model Breakdown, Tool Analytics, Tool Efficiency, Cache Effectiveness, Advanced Burn Rate

- Analytics — Tool Frequency chart, Event Distribution doughnut, Activity Heatmap grid, and Event Patterns with template clustering

• Tasks & Recovery — Task Performance, Recovery Patterns

HTML Session Report¶

Generate a self-contained HTML report for any session — full transcript with collapsible thinking blocks and tool detail, token/cost stats, model breakdown, and tool-use summary. The output file has zero external dependencies and can be shared or archived as a single .html file.

• VS Code: Run Sidekick: Generate HTML Report from the Command Palette or the Session Analytics toolbar. The report opens in a webview panel. Also available as the "HTML Report" format option in Sidekick: Dump Session Report.
• CLI: Run sidekick report to generate and open in the default browser, or press r in the TUI dashboard. See the CLI HTML Report docs for flags and examples.

Subagent Tree¶

The subagent tree view displays spawned subagents in a hierarchical parent/child structure. When an agent spawns other agents, they appear as nested children in the tree, with collapsible nodes showing the agent count. The tree uses trace-based parsing from session logs to reconstruct the spawn hierarchy, falling back to a flat list for providers that use different directory structures. Running agents are tracked in real-time and merged into the tree as they complete.

Event Stream¶

A sidebar tree view showing live session events as they happen. Each event displays a color-coded icon by type (user prompt, assistant response, tool use, tool result), a summary, and a timestamp. The view maintains a ring buffer of the 200 most recent events and updates in real-time as your agent works.

Session Intelligence¶

Three detection systems run continuously during a session to surface problems that are otherwise invisible.

Context Health¶

As your agent's context window fills up, the runtime compresses (compacts) the conversation to make room. Each compaction loses information. The context health gauge tracks this degradation as a percentage — starting at 100% and dropping with each compaction event.

• Green (70-100%) — context is mostly intact, decisions and earlier work are reliable
• Yellow (40-69%) — noticeable degradation, earlier decisions may have been lost
• Red (below 40%) — significant information loss, consider starting a fresh session

The gauge appears in the dashboard's Session Activity section. If context health drops below 50%, session handoffs automatically include a warning so the next session knows to re-verify earlier decisions.

Truncation Detection¶

When tool outputs exceed size limits, the agent runtime silently truncates them — your agent sees partial results but keeps going as if nothing happened. Sidekick scans every tool result for truncation markers and tracks which tools are affected.

The dashboard shows the total truncation count and a per-tool breakdown. Files that produce repeated truncated output are automatically surfaced as knowledge note candidates, since they likely contain gotchas worth capturing.

Cycle Detection¶

Sometimes agents get stuck in loops — reading a file, editing it, reading it again, editing it again — burning tokens without making progress. Sidekick uses a sliding-window algorithm to detect repeating patterns of tool calls.

When a cycle is detected, a VS Code notification fires with the affected files so you can intervene. The mind map also highlights cycling files with a distinct visual indicator.

Provider Status¶

The dashboard polls API health for both Claude (status.claude.com) and OpenAI (status.openai.com), then shows the banner only for the monitored provider. When the relevant API is degraded or experiencing an outage, a banner appears in the gauge row showing:

• Color-coded indicator (yellow for minor, red for major/critical)
• Affected components and their status
• Active incident name with link to the status page

The relevant status page is shown based on the monitored provider — Claude status for Claude Code, OpenAI status for Codex, and no provider-status banner for OpenCode. The banner is hidden when all systems are operational. Polls every 60 seconds, pausing when the dashboard is not visible. Also available as a standalone CLI command: sidekick status, which checks both endpoints directly.

Quota & Rate Limits Display¶

The dashboard shows quota or rate-limit data depending on the active provider:

• Claude Max ("Subscription Quota"): 5-hour and 7-day utilization from the Anthropic OAuth usage API, with color-coded gauges, elapsed-time projections (e.g., 40% → 100%), countdown timers, and auto-refresh every 5 minutes.
• Codex ("Rate Limits"): Primary and secondary rate-limit windows extracted from token_count events in the Codex event stream. Data arrives automatically during active sessions — no separate polling needed.
• OpenCode: No quota display — OpenCode does not provide rate-limit data.

When quota cannot be fetched, the dashboard keeps the section visible and shows a structured unavailable state for missing credentials, expired sign-in, rate limits, network/server failures, or unexpected API responses. Quota failure transitions appear as lightweight dashboard toasts and are recorded in notification history, without triggering native VS Code popup notifications.

Project Timeline¶

See every session your agent has run in the current project at a glance. Each session appears as a card showing its label, duration, token usage, task and error counts, and which model was used. Filter by time range (24h, 7d, 30d, or all) and click any card to expand tool breakdowns, task lists, and error summaries. See the Project Timeline page for full details.

Knowledge Notes¶

Capture reusable knowledge — gotchas, patterns, guidelines, and tips — attached to specific files in your codebase. Notes appear as gutter icons in the editor and in a dedicated tree view. Right-click notes to edit, confirm, or delete them. Use "Inject Knowledge Notes" to append them to your instruction file so your agent benefits from what you've learned. See the Knowledge Notes page for full details.

Session Discovery¶

The monitor automatically discovers sessions based on your configured provider. If the session is in a different directory:

• Use "Sidekick: Browse Session Folders..." to manually select a session folder
• Selection persists across VS Code restarts
• "Sidekick: Reset to Auto-Detect Session" clears manual selection

Configuration¶

Accessibility¶

All webview panels include:

• prefers-reduced-motion support — animations and transitions disabled when the OS-level setting is enabled
• Focus-visible outlines (2px) on all interactive elements for keyboard navigation
• Custom scrollbar styling (6px, themed) and text selection colors matching the VS Code editor
• Responsive layout adjustments for narrow sidebar panels (under 260px)
• Card entrance animations with stagger delay in Task Board, Plan Board, and Project Timeline (respects reduced motion)