@mrclrchtr/supi-context

Adds a /supi-context command to the pi coding agent so you can inspect how the current session is spending its context window.

Install

pi install npm:@mrclrchtr/supi-context

For local development:

pi install ./packages/supi-context

What you get

After install, pi gets one user command:

• /supi-context — render a detailed context-usage report for the current session
• /supi-context full — render the same report with the full guideline and tool-definition lists instead of previews

The command sends a custom supi-context message, and this package registers a dedicated renderer so the report shows up as a structured TUI view instead of plain text.

What the report shows

The report is meant to answer questions like:

• what is taking up space in the current context window?
• how much room is left before compaction pressure gets worse?
• which instruction files, context files, skills, guidelines, or tools are expensive?
• what extra context was injected by other SuPi extensions?

It includes:

• model name, context-window size, and total token usage
• approximation or pending-usage notes when exact usage data is not available yet
• a visual usage bar for system prompt, user messages, assistant messages, tool calls, tool results, other, autocompact buffer, and free space
• a category breakdown table for the same usage buckets
• a system-prompt composition breakdown for:
  • base prompt content
  • instruction files (AGENTS.md, CLAUDE.md, etc.)
  • other context files loaded into the system prompt
  • active skills
  • guidelines
  • tool snippets
  • append text
• instruction-file details with token cost, line count, and detected origin (project vs global)
• injected subdirectory context files from supi-claude-md, including turn number, line count, and token cost
• active skill names with per-skill token counts
• guideline bullet previews, plus source attribution for PI defaults, known built-in tools (read, write, edit), and other
• active tool definitions with per-tool definition token counts and snippet-token columns when available
• a compaction note when older turns were summarized
• extra provider sections from extensions registered through the shared context-provider registry in @mrclrchtr/supi-core

Configuration

No settings are required.

This package does not add a model-callable tool; it adds a user command only.

Notes

• The command uses the latest cached systemPromptOptions captured during before_agent_start.
• If those prompt options are missing or incomplete, the package backfills context files and skills by re-parsing the current system prompt.
• Exact totals come from pi's current context-usage data when available. Otherwise the report falls back to rough estimates and/or scales estimated category totals to the latest measured total.
• If no model is selected yet, the report can still render, but the context-window bar cannot show capacity.

Source

• src/context.ts — command registration, cached prompt-option handling, and renderer wiring
• src/analysis.ts — token accounting, attribution, and report data assembly
• src/format.ts — formatted report output for the TUI view
• src/prompt-inference.ts — fallback recovery of context files, skills, and guideline sections from the live system prompt
• src/renderer.ts — custom renderer for supi-context messages
• src/utils.ts — token and plural-format helpers

Tests live under __tests__/unit/.