Compress your prompts before they hit the LLM. Remove polite filler, redundant phrases, and verbosity automatically — saving tokens and reducing latency.
Troglodyte sits between you and the LLM, compressing prompts in real-time:
| Removes | Examples |
|---|---|
| Polite fillers | "please", "thank you", "I would appreciate" |
| Redundant phrases | "in order to" → "to", "due to the fact that" → "because" |
| Excessive verbosity | "I was wondering if you could" → "" |
| Articles & pronouns | (Balanced/Aggressive modes) — Pronouns preserved in Balanced mode ✅ |
Issue: Short technical prompts like const config = { ... } failed to trigger Smart Mode because the threshold was too high (> 0.25). A prompt with 8 tokens and 1 keyword scored only 0.125, bypassing Smart Mode entirely.
Fix: Lowered threshold from 0.25 to 0.15:
Impact: Short code snippets now correctly trigger Smart Mode, reducing synonym replacement ratio from 100% → 30% as intended.
Issue: phrases.ts contained ~75 build-log/MSVC/compiler output phrases (e.g., "Build started", "Fehler") that had zero relevance to prompt compression but inflated regex alternation size.
Fix: Removed all build-log entries from src/dictionaries/phrases.ts.
Impact: Regex alternation reduced by ~35% (~125 → ~85 entries). Faster matching, less memory.
The word tokenization regex character class now properly escapes the hyphen: [^\w\u00C0-\u024F\u1E00-\u1EFF\-] — preventing unexpected behavior with certain Unicode characters.
Test Group 2 now includes real assertions for English/German output validity and explicit language overrides (previously a no-op).
Deeply nested JSON/XML structures (>10 levels) are now safely aborted to prevent potential ReDoS attacks.
Sliding window rate limiter added: 10 requests/second per session. Exceeded limits return original text without compression.
Issue: All pronouns (he, she, it, er, ihn, etc.) were stripped in balanced mode, breaking reference tracking across sentences.
Fix: Added effectiveBlacklist that filters out core pronouns only when level === 'aggressive'. In balanced mode, essential pronouns are preserved:
Impact: Multi-sentence context tracking now works correctly. LLMs receive prompts where "John said he would fix it" isn't corrupted to "John said would fix."
Issue: The phrase "step by step": "steps" created broken grammar ("debug this steps").
Fix: Changed replacement to preserve semantic intent without breaking syntax:
!,? → Clean Ending)Issue: Orphaned punctuation like ?,! or !,? survived filtering, creating visual noise.
Fix: Robust cleanup that strips non-alphanumeric trailing symbols and preserves question/exclamation intent:
Issue: baseUrl and custom path aliases (@dictionaries/*, @engine) triggered TS5101 deprecation warnings in TypeScript 6.x.
Fix: Reverted to relative imports (./troglodyte, ../dictionaries/phrases). No functional change — cleaner, future-proof configuration.
Ensures LM Studio's bundled SDK takes precedence over local node_modules.
Recent updates introduce significant performance improvements:
Recent updates introduce significant performance improvements:
incremental and isolatedModules in tsconfig.json for faster development builds.Over the LM Studio Hub:
Or manually:
That's it! The plugin is now permanently installed. No terminal needs to stay open.
Simply run the same command again:
LM Studio will automatically reload the updated plugin.
For active development with hot-reload:
Leave this terminal open. The plugin will appear in LM Studio's plugins list.
Click the gear icon next to the plugin to access settings:
| Setting | Options | Default |
|---|---|---|
| Compression Level | Gentle / Balanced / Aggressive | Balanced |
| Smart Mode (Context-Aware) | On/Off | On |
| Protect URLs & Links | On/Off | On |
| Protect Version Numbers & IDs | On/Off | On |
| Protect Markdown Headers | On/Off | On |
| Protect File Paths | On/Off | On |
| Protect JSON/XML Structures | On/Off | On |
| Language Mode | Auto-Detect (EN/DE) / English / German | Auto-Detect |
| Show Statistics in Console |
Note: Smart Mode threshold was lowered to
0.15(v1.3.1) for better detection of short code snippets.
Only removes polite filler and courtesy words:
Removes articles, basic prepositions. Pronouns preserved for context tracking:
Maximum compression — removes almost all function words including pronouns:
Send this verbose prompt:
"Hello there! I was wondering if you could possibly help me out? I would really appreciate it if you could explain how to install Node.js on Windows step by step. Thank you so much for your time and assistance!"
Expected compressed output (Balanced):
(Note: "step by step" → "sequential", not the broken "steps")
Send this verbose prompt:
"Hallo! Ich würde mich sehr freuen, wenn du mir bei dieser Aufgabe helfen könntest, bitte und danke!"
Expected compressed output (Balanced):
Send this prompt to verify file paths are preserved:
"check C:\Source Code\ServiceMonitor for issues."
Expected: Path stays as C:\Source Code\... (not corrupted to C:\src Code\...)
Send this prompt to verify both path types work:
"Bitte analysiere den Code in /home/user/project/src/main.ts und ./lib/utils.py."
Expected: Both paths preserved intact:
Send this prompt to verify German compression works correctly and version numbers aren't fragmented:
"Hallo! Ich würde mich sehr freuen, wenn du mir bei dieser Aufgabe helfen könntest, bitte und danke! Erkläre mir bitte Schritt für Schritt, wie man Node.js unter Windows installiert. Vielen Dank!"
Expected: Clean German output with "Node.js" intact and proper phrase replacement:
❌ Wrong (old buggy behavior):
These elements are never modified, regardless of compression level:
| Element | Example | Protection Method |
|---|---|---|
| Code blocks | `code` and blocks``` ```` | PUA placeholder (min 15 chars) |
| URLs | https://example.com/path?id=123 | PUA placeholder (min 20 chars) |
| Version numbers | v1.0.0, 2.1.3-beta | Regex match + protect |
| Software names with dots | Node.js, TypeScript | Regex match + protect |
| Issue/PR refs | #456 | Regex match + protect |
| UUIDs | 550e8400-e29b-41d4-a716-446655440000 | Regex match + protect |
Runs tsc --noEmit for fast validation without generating output files.
Current Test Results: ✅ 12/12 assertions pass (v1.3.1)
lms dev --install completed successfullydist/index.js exists after buildCheck LM Studio's console (F12 → Console tab) for error messages.
Try Gentle mode or disable specific protections in settings.
This was a known bug fixed on May 16, 2026. Ensure you have the latest version with path protection enabled.
For detailed architecture and technical documentation, see:
| Issue | Status |
|---|---|
| Placeholder overhead causing negative compression | ✅ Fixed - compact Unicode placeholders (\uE000+) now used |
| "source" → "src" synonym corrupting file paths | ✅ Fixed - Windows path protection added before synonym phase |
| Phrase replacement order causing partial matches | ✅ Fixed - phrases sorted by length (longest first) |
Punctuation spacing broken (out?explain, Node. js) | ✅ Fixed - smart cleanup chain with CAPITAL-aware spacing |
Logic symbols breaking natural language (and → &&) | ✅ Fixed - all logic symbol replacements removed |
protectFilePaths config field unused | ✅ Fixed - now wired up and functional (May 18, 2026) |
| Placeholder overflow risk (~1M items) | ✅ Fixed - MAX_PLACEHOLDERS check added (May 18, 2026) |
| Duplicate entries in de-filler.ts | ✅ Fixed - ~90 duplicates removed (May 18, 2026) |
| Error handling lacking user feedback | ✅ Fixed - stack trace + truncated error message shown (May 18, 2026) |
Path extensions stripped (main.ts → main) | ✅ Fixed - regex includes / and uses lookahead boundaries (May 18, 2026) |
| Placeholder counter duplication (50% waste) | ✅ Fixed - removed unused generatePlaceholder() function (May 18, 2026) |
| Language detection O(n²) performance | ✅ Fixed - converted indicator arrays to Sets for O(1) lookup (May 18, 2026) |
| Regex compilation per compression call | ✅ Fixed - pre-compiled all phrase regexes in constructor (May 18, 2026) |
| String concatenation O(n²) memory | ✅ Fixed - use array join pattern for reconstruction (May 18, 2026) |
MIT
Last Updated: June 24, 2026 — v1.3.1 Release
.match() + .split() calls with unified .matchAll(/([^\s\w]+)|(\w+)/gu) loop. ~50% reduction in memory allocations during word filtering./\\b(?:const|...)\b|[{}]/g. ~30% faster technical context detection./\b[a-zäöüß]{3,}\b/g to module scope so V8 caches compiled bytecode. Eliminates redundant compilation overhead per call.Record<string, string> to Map<string, string> for O(1) direct access without prototype chain traversal. ~20% faster synonym lookups..trim() calls and branch mispredictions in the hot loop.| On/Off |
| On |
| Markdown headers |
## My Header |
| PUA placeholder (min 12 chars) |
| Windows file paths | C:\Source Code\... | Regex match + protect |
| Placeholder restoration O(n²) time | ✅ Fixed - single-pass Map-based replacement (May 18, 2026) |
| "Node.js" fragmented into parts | ✅ Fixed - added . to word pattern regex (May 18, 2026) |
Orphaned punctuation scattered (!, .!) | ✅ Fixed - cleanup step removes standalone punctuation (May 18, 2026) |
| No input validation (security risk) | ✅ Fixed - added null/empty check + 1MB limit warning (May 18, 2026) |
| PUA restoration regex broken | ✅ Fixed - proper \uE000-\uF8FF Unicode escapes used (May 24, 2026) |
| Dead synonym entries (no-ops) | ✅ Fixed - 17 no-op entries removed (May 24, 2026) |
detectTechnicalContext double-call | ✅ Fixed - result cached in variable (May 24, 2026) |
| escapeRegex cascading double-escaping | ✅ Fixed - single-pass regex replacement used (May 31, 2026) |
| Word filtering reconstruction misalignment | ✅ Fixed - separate filtered word array prevents delimiter shift (May 31, 2026) |
| XML depth tracking matching unrelated tags | ✅ Fixed - proper searchPos skip logic added (May 31, 2026) |
| Technical context over-counting overlapping patterns | ✅ Fixed - simplified keyword/brace counting eliminates overlap (May 31, 2026) |
return totalTokens > 0 && (codeScore / totalTokens) > 0.15; // was 0.25
const protectedPronouns = new Set([
// English
'he', 'him', 'his', 'she', 'her', 'it', 'they', 'them', 'their',
// German
'er', 'ihn', 'ihm', 'sein', 'sie', 'ihr', 'es', 'wir', 'uns', 'euch', 'mein', 'dein',
]);
'step by step': 'sequential', // instead of 'steps'
'Schritt für Schritt': 'sequenziell', // German equivalent
.replace(/[^\p{L}\p{N}\p{So}\p{Sk}\p{Sc}\p{Sm}\uE000-\uF8FF]+$/, ''); // Strip mixed punctuation with emoji support
if (['?', '!'].includes(lastChar)) { text = text + lastChar; } // Re-add if original ended in ? or !
"peerDependencies": { "@lmstudio/sdk": "^1.5.0" }
lms get crunch3r/troglodyte
cd "C:\Source Code\LM Studio Plugins\troglodyte"
lms dev --install
lms dev --install
cd "C:\Source Code\LM Studio Plugins\troglodyte"
npm run dev
Input: "Hello! Could you please help me? Thank you so much!"
Output: "Help me"
Input: "John said he would fix it. He needs to call me tomorrow."
Output: "John said he fix it. He needs tomorrow." ✅ (pronouns kept!)
Input: "Please explain step by step how to solve this issue."
Output: "sequential how solve issue." ✅ (grammar intact!)
Input: "I would really appreciate it if you could provide a detailed
explanation of the implementation process."
Output: "Provide detailed explanation implementation process"
possibly help me out? explain sequential install Node.js Windows. ✅
würde freuen mir Aufgabe helfen, könntest ✅
analysiere Code /home/user/project/src/main.ts ./lib/utils.py. ✅
würde freuen mir Aufgabe helfen, könntest sequenziell Node.js Windows installiert. ✅
! würde freuen Aufgabe helfen könntest, Erkläre steps Windows installiert,!, Node.js.! ❌
npm install
npm run build
npm run typecheck
npx ts-node src/tests/compression.test.ts
# or via npm script:
npm test
npm run dev
troglodyte/
├── src/
│ ├── index.ts # Entry point
│ ├── promptPreprocessor.ts # Pipeline orchestrator + rate limiting
│ ├── troglodyte.ts # Compression engine (main logic)
│ ├── config.ts # UI configuration
│ └── dictionaries/
│ ├── en-filler.ts # English blacklists
│ ├── de-filler.ts # German blacklists
│ ├── phrases.ts # Multi-word replacements (~125 entries, v1.3.1)
│ └── synonyms.ts # Single-word abbreviations
├── dist/ # Compiled output
├── package.json # v1.3.0: peerDependencies added, typecheck script
├── tsconfig.json # v1.3.0: deprecated baseUrl removed
├── PROJECT_SUMMARY.md # Comprehensive documentation
├── CHANGELOG.md # Version history
└── README.md # This file