1// formatting/clean.js
2// Deterministic post-processing for LLM output.
3
4// Emoji ranges: emoticons, dingbats, symbols, transport, flags, supplemental, etc.
5// Keeps basic punctuation, arrows, math symbols, and currency signs.
6const EMOJI_RE = /[\u{1F600}-\u{1F64F}\u{1F300}-\u{1F5FF}\u{1F680}-\u{1F6FF}\u{1F1E0}-\u{1F1FF}\u{1F900}-\u{1F9FF}\u{1FA00}-\u{1FA6F}\u{1FA70}-\u{1FAFF}\u{2702}-\u{27B0}\u{FE00}-\u{FE0F}\u{200D}\u{20E3}\u{E0020}-\u{E007F}]/gu;
7
8/**
9 * Strip emojis from text.
10 */
11export function stripEmojis(text) {
12  return text.replace(EMOJI_RE, "");
13}
14
15/**
16 * Collapse runs of 3+ newlines into 2.
17 */
18export function normalizeWhitespace(text) {
19  return text.replace(/\n{3,}/g, "\n\n");
20}
21
22/**
23 * Remove trailing filler phrases LLMs like to append.
24 * "Let me know if you need anything else!" etc.
25 */
26const FILLER_RE = /\n{1,2}(?:Let me know if (?:you (?:need|have|want|would like)|there(?:'s| is) anything)|Feel free to (?:ask|reach out|let me know)|Hope (?:this|that) helps|Happy to help|Don't hesitate to)[^\n]*$/i;
27
28export function trimFiller(text) {
29  return text.replace(FILLER_RE, "");
30}
31
32/**
33 * Full cleaning pipeline.
34 */
35export function clean(text) {
36  if (!text || typeof text !== "string") return text;
37  let result = text;
38  result = stripEmojis(result);
39  result = normalizeWhitespace(result);
40  result = trimFiller(result);
41  result = result.trim();
42  return result;
43}
44

1import log from "../../seed/log.js";
2import { clean } from "./clean.js";
3
4export async function init(core) {
5  // Clean AI responses (emojis, whitespace, filler)
6  core.hooks.register("beforeResponse", async (data) => {
7    if (data.content && typeof data.content === "string") {
8      data.content = clean(data.content);
9    }
10  }, "llm-response-formatting");
11
12  // Normalize tool names against registered tools.
13  // Some models generate tool calls with underscores (navigate_tree) instead
14  // of hyphens (navigate-tree), or with slight misspellings. This hook finds
15  // the closest matching registered tool name before it reaches MCP.
16  const { resolveTools } = await import("../../seed/tools.js");
17  const { getToolsForMode } = await import("../../seed/modes/registry.js");
18
19  core.hooks.register("beforeToolCall", async (data) => {
20    if (!data.toolName) return;
21    // Try the name as-is first (fast path)
22    const exact = resolveTools([data.toolName]);
23    if (exact.length > 0) return;
24
25    // Try replacing underscores with hyphens
26    const hyphenated = data.toolName.replace(/_/g, "-");
27    const hyphenMatch = resolveTools([hyphenated]);
28    if (hyphenMatch.length > 0) {
29      log.debug("Formatting", `Tool name normalized: ${data.toolName} -> ${hyphenated}`);
30      data.toolName = hyphenated;
31    }
32  }, "llm-response-formatting");
33
34  log.verbose("Formatting", "Response cleaning + tool name normalization active");
35
36  return {};
37}
38

1export default {
2  name: "llm-response-formatting",
3  version: "1.0.2",
4  builtFor: "TreeOS",
5  description:
6    "Post-processing layer between the LLM and the user. Hooks into two lifecycle events " +
7    "(beforeResponse and beforeToolCall) to clean up the messy, inconsistent output that " +
8    "language models produce. The goal is that every response reaching the user follows the " +
9    "same formatting conventions regardless of which model generated it.\n\n" +
10    "The beforeResponse hook runs a three-stage cleaning pipeline on every AI response. First, " +
11    "emoji stripping: a comprehensive regex covering emoticons, dingbats, transport symbols, " +
12    "flags, supplemental symbols, and variation selectors removes all emoji characters while " +
13    "preserving basic punctuation, arrows, math symbols, and currency signs. Second, whitespace " +
14    "normalization: runs of three or more consecutive newlines collapse to exactly two, preventing " +
15    "the excessive vertical spacing many models produce. Third, filler trimming: a pattern-matched " +
16    "removal of trailing pleasantries that LLMs append reflexively. Phrases like 'Let me know if " +
17    "you need anything else', 'Feel free to ask', 'Hope this helps', and 'Don't hesitate to " +
18    "reach out' are stripped from the end of responses.\n\n" +
19    "The beforeToolCall hook fixes tool name mismatches. Some models generate tool calls with " +
20    "underscores (navigate_tree) when the registered tool uses hyphens (navigate-tree). The hook " +
21    "first checks the exact name against the tool registry. If no match is found, it replaces " +
22    "underscores with hyphens and checks again. A successful match silently rewrites the tool " +
23    "name before it reaches MCP dispatch, preventing tool-not-found errors that would otherwise " +
24    "break the conversation flow. The fast path (exact name match) adds zero overhead to " +
25    "correctly-named tool calls.",
26
27  needs: {
28    services: ["hooks"],
29  },
30
31  optional: {},
32
33  provides: {
34    models: {},
35    routes: false,
36    tools: false,
37    jobs: false,
38    orchestrator: false,
39    energyActions: {},
40    sessionTypes: {},
41
42    hooks: {
43      fires: [],
44      listens: ["beforeResponse", "beforeToolCall"],
45    },
46  },
47};
48