Send a message that starts with the trigger (default /compact, optionally followed by a note):

• The prior conversation is written to compact-<timestamp>.md (the /compact message itself is excluded).
• If "write a seed summary" is on, the current model summarizes the conversation into compact-<timestamp>.seed.md.
• Your message is replaced with a status note pointing at both files and showing the summary.

Then open New Chat and paste the summary to continue with less context.

/compact ship notes
→ [compact] Exported this conversation.
  - Transcript: …/compact-2026-06-21-1432.md
  - Seed:       …/compact-2026-06-21-1432.seed.md

  Summary for the next chat:
  We were building plugin-compact; chose a preprocessor because tools
  can't pullHistory… next: wire docs + ship.

Configuration

• Compact trigger — the word that fires an export (default /compact). Must be the start of the message; /compacting and please /compact do not match.
• Export directory — folder for the .md + .seed.md files. Blank = the plugin's per-prediction working directory.
• Timezone — IANA name (e.g. America/Mexico_City) for the filename timestamp. Blank = this machine's timezone.
• Enable /compact — watch for the trigger (on by default).
• Write a seed summary — also summarize via the current model (one extra model call per /compact; on by default). Turn off for a raw export with no model call.
• Max tokens per summary call (default 4000) — the token budget for each summarization call. A long conversation is summarized in chunks of this size, then merged. Set it below your model's loaded context length. It can't be auto-detected: LM Studio's SDK reports the model's maximum context, not the window you loaded, so the plugin would over-size and overflow. Default 4000 is safe on essentially any model; raise it toward your context length (e.g. 24000 for a 32k load) for fewer, faster chunks.

Notes

• The export is preprocessor-only — there is no tool, because only the preprocessor's controller exposes pullHistory().
• Long conversations are summarized in chunks (map-reduce). A conversation that won't fit the budget is split into Max tokens per summary call-sized parts (measured with the model's tokenizer), each part summarized, then the part-notes merged into one seed. This is slower (several model calls) but means compaction works on exactly the big conversations that need it, instead of overflowing the context and producing nothing.
• The summary is best-effort: if the model errors or returns nothing, the full transcript is still written and you're told why (the failure reason, or that the model returned nothing).
• After a /compact, the model still generates a brief reply to the status note — a preprocessor can't suppress the follow-on turn. It's harmless; the useful artifacts are the two files.

MIT licensed.