Explore manual summarization

[lramos15]

Ideally we want users to be able to trigger summarization themselves whenever they want to keep context down. This explores adding that feature. We can then add this to the context window widget.

@roblourens as the summarization expert I would appreciate your thoughts on this approach as I ideally would like this to be as close to the agentic summarization as possible and reuse as much code as possible

[Copilot]

Pull request overview

This pull request adds a manual conversation summarization feature through a new /compact command. The feature allows users to explicitly trigger summarization of their conversation history to reduce context size, similar to the automatic summarization that occurs during agentic workflows.

Changes:

• Implements a new CompactIntent that handles the /compact slash command for manual conversation summarization
• Integrates the summarization logic from the existing agentic summarization system by reusing SummarizedConversationHistoryPropsBuilder and related prompts
• Registers the command as part of the agent participant in VS Code's chat interface

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/extension/intents/node/compactIntent.ts	New intent implementation that handles manual conversation summarization, reusing existing summarization infrastructure
src/extension/intents/node/allIntents.ts	Registers the CompactIntent in the intent registry
src/extension/common/constants.ts	Adds Compact intent constant and maps it to the Agent participant's commands
package.nls.json	Adds localized description for the compact command
package.json	Registers the compact command under the agent participant's available commands

[Copilot]

The import type annotation here is unusual and appears to be auto-generated or incomplete. The type should be imported at the top of the file along with other types from the conversation module. This inline import makes the code less readable and deviates from standard TypeScript practices.

[Copilot]

The applySummaryToHistory method mutates the rounds in the turns array, but the turns parameter is typed as readonly. While the implementation creates a new array with [...turns].reverse(), it still mutates the round.summary property of objects in the original turns array. This violates the readonly contract and could lead to unexpected side effects. Consider either removing the readonly modifier from the parameter type or creating deep copies of the turns to avoid mutation.

[Copilot]

When the target round is not found (e.g., if toolCallRoundId doesn't match any round in the history), the summary is silently discarded without any logging or error handling. This could make debugging difficult if the summarization appears to fail silently. Consider adding a log warning when the round is not found to help with troubleshooting.

Suggested change

	for (const turn of [...turns].reverse()) {
	const round = turn.rounds.find(r => r.id === toolCallRoundId);
	if (round) {
	round.summary = summary;
	return;
	}
	}
	let found = false;
	for (const turn of [...turns].reverse()) {
	const round = turn.rounds.find(r => r.id === toolCallRoundId);
	if (round) {
	round.summary = summary;
	found = true;
	break;
	}
	}
	if (!found) {
	this.logService.warn(`Failed to apply conversation summary: no round found with id ${toolCallRoundId}.`);
	}

[roblourens]

This looks like a separate implementation of summarization? Let's trigger the existing one

[roblourens]

Maybe you can get it to reuse more of ConversationHistorySummarizer. It looks like this mostly duplicates that.

The other idea I had was to go through the normal agent loop but trigger summarization here:

vscode-copilot-chat/src/extension/intents/node/agentIntent.ts

Line 232 in 607a59d

let shouldTriggerSummarize = false;

then we would return some flag telling the tool calling loop to bail out of the agent loop without sending a request. Maybe it's more complicated. I don't really know how this should look though, what is in the ChatResult. And we need to make sure that we don't add an empty turn to the history after triggering this manually (in both cases). This is all just so complicated.