Thanks for the PR! The core idea here — giving advance notice before compaction fires — is good, but the implementation approach doesn't fit how we handle the message stream.

Hermes enforces strict message role alternation, and we're very careful about not injecting synthetic user messages into the conversation mid-loop. The budget pressure system you referenced as a model actually avoids this — it only ever piggybacks on existing tool results and silently drops the warning if there's no tool result to attach to. The fallback to messages.append({"role": "user", ...}) in this PR would break that invariant, and the post-compaction summary injection that instructs the model to "share this with the user" would derail it from whatever task it was working on.

We also don't want to change the model's behavior around compaction — telling it to proactively save state or inform the user about compaction details adds noise to the conversation and can cause the model to prioritize compaction housekeeping over the user's actual task.

What we do want is to warn the user (not the model) that compaction is approaching — a simple notification in the CLI output or gateway chat, similar to how budget pressure already prints a console line. We'll implement that ourselves since it touches display/gateway plumbing rather than the message stream.

Appreciate the contribution!