Can a skill trigger proactively compaction?

[gsemet]

🏷️ Discussion Type

Question

💬 Feature/Topic Area

VS Code

Body

It seems like it is not possible to trigger proactively compaction from the LLM (the LLM choosing itself to trigger compaction). Not typing '/compact', I want to trigger compaction at specific points in my workflow. For instance, after a spec has been generated, today i have to say "Please create a new session to type '/compact'". → I want the workflow to do it automatically

[rehuux]

Not possible yet. Auto-compaction is system-triggered only. The LLM cannot decide on its own to compact.

What exists right now:

· Manual trigger — Type /compact in chat to compact whenever you want
· Auto-compaction — Runs automatically when context window fills up

What doesn't exist: LLM-initiated compaction. The LLM can't decide to compact itself at workflow checkpoints.

Workaround for your spec generation use case:

1. Use /compact manually right after the spec is generated
2. Ask the LLM to summarize — "Please summarize our work so far, including key decisions and next steps" — then start a new session with that summary
3. Use a progress file — Maintain a SPEC.md file and ask Copilot to update it at each phase. New sessions can read it

[gsemet]

I have the impression Claude can do it. And why wouldn't be able to expose a tool that trigger the same code that /compact ?

[biruk-arch]

Yeah, I’ve run into this exact wall. Short answer: you can't do it right now.

The LLM doesn't have the system permissions to trigger slash commands like /compact on its own. Compaction only happens if you manually type it, or if the system forces it behind the scenes because you're about to max out your token limit.

Since we can't automate the actual command yet, here is how I handle it to keep things moving:
Write to a file: Have the LLM dump the finished spec directly into a file (like SPEC.md) in your workspace. Then, just open a fresh chat session and reference #SPEC.md. It drops all the chat history bloat but keeps the exact context you need.The quick hand-off: Ask it to "summarize our final decisions into a prompt for a new session." Copy that summary, open a new chat, paste it, and keep rolling.

Hopefully they give the LLM tool access to slash commands soon, but for now, dumping finalized stuff into workspace files is the cleanest workaround.