Inspect context usage

The context window holds the material available to the model for its next response: instructions, conversation history, and tool results. Open /context to see how much space is used and what contributes to it. This inspection does not send a model request.

Read the report

The compact report shows used, free, and reserved space, a category breakdown, and loaded instruction files. Large tool results are called out so you can spot an unusually expensive file read or command output.

Use /context all to expand individual messages and tools. A used total without breakdown rows means that snapshot has no category detail. You do not need to send another message to make the report valid.

Understand the measurement

Tool results update the used count as they arrive. Rewinding updates it for the remaining conversation. Between replies the count is an estimate that Vera corrects with the provider’s count for the last request, including reasoning the provider carries forward, so it stays close to what the next request uses. Each attached image counts as about 1,600 tokens until the provider reports its real size. After resume, the loaded-file list describes the last measured request.

The maximum comes from the model’s known context window. If a local server does not publish a window, Vera leaves it unknown rather than estimating one. Used, free, and reserved space remain distinguishable without color.

The composer shows ctx ?% when the model’s window is known but the current conversation has no measured context yet. It replaces the question mark with the measured percentage after a turn. Saved sample conversations may have no measurement even when they contain messages. Any measured use fills at least one cell of the meter. In a narrow terminal the composer drops the token counts and keeps the meter, the percentage, the model, and the effort.

Large instruction files

Everything in the INSTRUCTIONS section is sent with every request: home and project rules, AGENTS.md, AGENTS.local.md, memory, the selected agent’s instructions, and the skill catalog. When that total reaches 8,000 estimated tokens, /context shows a red line under INSTRUCTIONS with the total and the biggest source, and Vera shows one short notice after the first reply in a conversation, plus the same sentence as a brief overlay that types on quickly, then a tick. Press Escape to close it when no dialog is open; otherwise it clears after a few seconds. The numbers match the Instructions row and the section heading. While the red line shows, the per-file size warnings are hidden. Trim or split the biggest source to clear it. The notice appears again in a new or resumed conversation while the total stays over.

Change the instruction budget

Find vera.context in /customize under Extensions and note its absolute path. Add an explicit entry to the home’s config.json and set instruction_budget_tokens:

{
    "extensions": [{
        "path": "/absolute/path/to/src/core-extensions/context",
        "enabled": true,
        "config": {
            "instruction_budget_tokens": 8000
        }
    }]
}

Set it to 0 to turn the warning off; /context still shows the total. A value that is not a whole number of 0 or more prevents the extension from activating.

Inspect an instruction file

1. Press Ctrl+O in the report, or run /context sources.
2. Choose a source and press Enter to preview it in Customize.
3. For a writable source, open it in your editor from the preview.

Previews show the file’s current saved contents. They are not historical copies of what the model received.

Copy or close the report

Drag over text to copy a section, or press Enter to copy the full Markdown report. Escape closes it without adding the report to your conversation.

To change when Vera summarizes a long conversation, see Context limits and compaction.