Context compression

Long conversations fill the model's context window, especially when they carry screenshots, file reads and search results. Precursor keeps that in check in two ways. Automatic trimming is free and makes no model call. Compaction summarises the conversation so far so you can keep going in the same thread. Compaction only runs when you ask for it.

After /compact, the older messages stay visible but dimmed; the model works from the summary.

Automatic trimming

Before each turn, Precursor trims what the model sees. The transcript itself is left untouched.

• Inline images are dropped. Tools such as the Playwright browser return screenshots as base64 text inside their result. The model can't see a picture through its base64, so each image becomes a placeholder such as [binary omitted: ~138 KB]. One screenshot alone is about 150,000 characters.
• Large tool results from older turns are shortened. A tool result over 2,000 characters that is older than the last 5 user turns reaches the model as a short preview and a note. The model can call the tool again when it needs the full output. Change the number of turns in Settings → Model → Prompt budgeting → Keep tool results in full for the last N turns (llm_tool_result_keep_turns). Set it to 0 to always send full results.

The per-result cap and the overall input budget (Max tokens per tool result, Max input tokens per request) still apply after these two passes. Past the budget, the oldest turns are dropped first, without a summary. Compaction exists so that doesn't happen.

Compacting a topic or chat

Type /compact in the composer, or click Compact context in the conversation stats panel on the right of the transcript (shown unless turned off in Settings → Chat). The model writes a structured summary of the conversation with these sections:

• the goal
• key facts and decisions, with names, IDs, numbers, URLs and paths copied verbatim
• work done
• open questions
• where things stand

The summary is saved as a compaction marker, and from the next turn on the model sees the summary plus everything after it.

• Say what matters. Add a focus after the command, for example /compact keep the SKU comparison and the pricing figures.
• Nothing is deleted. The messages above the marker stay in the transcript, dimmed. Show summary on the marker shows what the model now works from.
• Undo removes the marker, and the model sees the full history again. The usual undo toast gives you a few seconds to change your mind.
• Compacting again folds the previous summary into the new one.
• Not while a reply is generating. Wait for the reply to finish, or stop it, before you compact.

The summary is written by the conversation's model. Its token usage is counted under /compact in Settings → Usage stats.

Knowing when to compact

The stats panel's context window bar shows how much of the model's window the last turn used. Past 80%, the Compact context button turns amber with a nudge. The collapsed stats rail shows an amber shortcut too.

Past 80% of the window, the stats panel suggests compacting.

After a compaction, until the next turn reports real usage, the bar shows an estimate (marked ≈). The Next turn (estimate) section shows:

• the history the next turn will send
• how much automatic trimming saved
• how many messages the latest marker covers

The estimate leaves out the system prompt and tool definitions.

Compacting an agent

Agents run on the Copilot SDK. The SDK already compacts their history on its own when the window runs low. /compact [focus] in the agent's composer, or Compact context under Usage in the insights sidebar, triggers a compaction now. Only an agent at rest can be compacted: wait for the turn to finish, or stop it, first.

Every compaction shows in the timeline as a context compacted marker, both the automatic ones and yours. The marker shows the tokens before and after; hover it to read the summary. The usage panel's context bar drops right away.

See also

• Attachments: how much of a document the model sees.
• Storage & retention: pruning old tool results from the database, which is a different lever from what the model sees.
• Configuration: the prompt budgeting settings.