context-compactor

import { createConfigSchematics } from "@lmstudio/sdk";

export const configSchematics = createConfigSchematics()
  .field(
    "triggerPercent",
    "numeric",
    {
      min: 0.05,
      max: 0.95,
      displayName: "Compaction trigger",
      subtitle:
        "Compact once this fraction of the context window is used. Lower keeps the context short, " +
        "which keeps generation fast, at the cost of compacting more often. Tool definitions are " +
        "not counted by the measurement, so leave some headroom.",
      slider: { min: 0.05, max: 0.95, step: 0.05 },
    },
    0.6,
  )
  .field(
    "keepRecentTokens",
    "numeric",
    {
      int: true,
      min: 500,
      max: 32000,
      displayName: "Recent context kept verbatim",
      subtitle:
        "How many tokens at the end of the conversation are never summarized. Measured in tokens, " +
        "not messages: one tool result can be worth forty ordinary replies, and a tail counted in " +
        "messages can overflow the window on its own.",
      slider: { min: 500, max: 32000, step: 500 },
    },
    6000,
  )
  .field(
    "chunkTokens",
    "numeric",
    {
      int: true,
      min: 2000,
      max: 32000,
      displayName: "Summarize in chunks of",
      subtitle:
        "Tokens of history per summary. Each chunk costs one model call whatever its size, and " +
        "reading is far cheaper than writing — so smaller chunks make compaction slower, not " +
        "faster. Lower it only if summaries are losing detail.",
      slider: { min: 2000, max: 32000, step: 1000 },
    },
    16000,
  )
  .field(
    "maxToolResultTokens",
    "numeric",
    {
      int: true,
      min: 500,
      max: 32000,
      displayName: "Maximum size of a tool result",
      subtitle:
        "Tool results are truncated to this many tokens before the model sees them. Servers set no " +
        "such limit of their own: reading one large file can return more than the context holds and " +
        "kill the reply outright. Compaction cannot help — it runs before the reply, not during it.",
      slider: { min: 500, max: 32000, step: 500 },
    },
    4000,
  )
  .field(
    "vaultPath",
    "string",
    {
      displayName: "Archive folder (optional)",
      subtitle:
        "Folder where the verbatim transcript and the consolidated state note are written on every " +
        "compaction — an Obsidian vault works well. Leave empty to keep everything in memory and " +
        "write nothing to disk.",
    },
    "",
  )
  .field(
    "showStatus",
    "boolean",
    {
      displayName: "Announce compaction in chat",
      subtitle: "Show a one-line notice with the before/after token counts when compaction runs.",
    },
    true,
  )
  .build();