hermes-persona-stickers

hermes-persona-stickers 🐱🎨

Instant, Character-Consistent Emotion Reaction Stickers & Memes for Hermes Agent.
Built for the Nous Research Hermes Agent Plugin Hackathon (Nov 2026).

hermes-persona-stickers gives Hermes Agent an expressive visual voice. Instead of replying with dry text, Hermes can stamp bespoke chibi stickers on demand—pouting, boasting, sighing, or celebrating alongside the user with sub-30ms instant response—via the generate_persona_sticker tool or /sticker command.

⚡ The Breakthrough: Why Sub-30ms Matters

Traditional AI image diffusion takes 10–30 seconds per generation—far too slow for real-time instant messaging. Nobody wants to wait 15 seconds in the middle of a lively chat for a reaction meme.

hermes-persona-stickers solves this through a Dual-Engine Architecture:

                       ┌─────────────────────────────────────────┐
                       │   Hermes Dialogue / Emotion Extractor   │
                       └───────────────────┬─────────────────────┘
                                           │
                        ┌──────────────────┴──────────────────┐
                        │ Mode Auto-Resolution & Asset Check │
                        └──────────┬───────────────────┬──────┘
                                   │                   │
         [Pre-baked Pool Exists]   ▼                   ▼   [Zero GPU / Headless VPS]
      ┌───────────────────────────────┐     ┌───────────────────────────────────┐
      │   Engine A: Instant Stamper   │     │    Engine B: Fast Pure Vector     │
      │ • Pre-baked transparent base  │     │ • Pure Python + Pillow drawing    │
      │ • 3-layer comic caption text  │     │ • 0 GPU, 0 models, 0 MB download │
      │ • 32-bit RGBA PNG / WebP      │     │ • Sub-50ms instant generation     │
      │ ⚡ 12.88 ms | ~18 KB           │     │ ⚡ 35.00 ms | ~15 KB               │
      └───────────────────────────────┘     └───────────────────────────────────┘

📊 Real-world Benchmarks (Tested on RTX 3070 Ti Laptop & Linux Cloud VPS)

| Pipeline Stage | Traditional Diffusion | Our Instant Stamper (Engine A) | Pure Vector (Engine B) |
|---|---|---|---|
| Turnaround Latency | 12,000 – 25,000 ms | 12.88 ms (0.013s) | 35.12 ms (0.035s) |
| User Wait Time | Noticeable disruption | Zero perceptible delay | Zero perceptible delay |
| Output File Size | 400 – 800 KB (raw PNG) | 7.98 – 19.4 KB (WebP/PNG) | 14.2 KB (PNG) |
| Hardware Requirement | High-end GPU required | Any CPU / 512MB RAM | Any CPU / 512MB RAM |
| Character Consistency | Random face drift | 100% Identical Persona | 100% Identical Vector |

✨ Key Technical Highlights

1. 🎯 Persona Consistency

Pre-baked persona bases keep the same face and colors across turns. Asset prompts used (offline, outside this plugin) locked:

Hair Lock: (silver hair:1.2), very long hair, (complete twintails:1.3), ahoge
Eyes Lock: (aqua eyes:1.2), (cyan eyes:1.1) (strictly avoiding muddy green/brown drift)
Features Lock: cat ears, cat tail, fangs
360° No-Crop Margin: Generous negative boundary penalties (cropped:1.5), (cropped hair:1.5), touching border ensure 40px+ margin on all sides, completely eliminating severed hair edges.

2. 🛡️ Floodfill Alpha Decoupling & 3-Layer Captioning

Corner Floodfill: Connected-component background removal starting exclusively from the 4 outer image corners, preserving internal white hair highlights and eye reflections.
Smooth 6px Sticker Border: Euclidean Distance Transform (EDT) generates a clean white die-cut rim.
3-Layer Comic Typography: Dark plum drop shadow $\rightarrow$ 4-8px crisp white stroke $\rightarrow$ vibrant hot pink fill.
Telegram Native Compatibility: Enforces 32-bit RGBA PNG and WebP output, preventing Telegram's Image_process_failed document-fallback bug.

🚀 Quick Start

Installation

Install directly via Hermes Plugin Manager:

hermes plugins install hermes-persona-stickers

Slash Command

Generate an on-demand reaction sticker at any time:

/sticker tsundere 哼！才没有在等你呢！
/sticker smug 哼哼~小菜一碟！
/sticker sigh 真是服了你了...
/sticker shock 骗人的吧？！
/sticker happy 好耶超开心！

Native Agent Tool Calling

Hermes automatically calls generate_persona_sticker whenever strong emotional banter occurs:

{
  "name": "generate_persona_sticker",
  "arguments": {
    "emotion": "tsundere",
    "text": "才没有被切到头发呢！",
    "size": 256,
    "mode": "auto"
  }
}

The tool returns a platform-native MEDIA:<path> tag that renders natively as an image bubble across Telegram, Discord, and Slack!

⚙️ Usage notes

No plugin config schema is required. Stickers are written under the active Hermes profile:

<HERMES_HOME>/plugin-data/hermes-persona-stickers/

Pass emotion, text, size (128 / 256 / 512), and mode (auto | pool | fast) as tool arguments or via /sticker.

🧪 Automated Testing

Run the full verification suite:

pytest test_plugin.py -v

Core suites:

test_emotion_detection: Keyword & conversational tone extraction
test_pool_sticker_rendering: Sub-50ms pre-baked asset stamping
test_fast_vector_fallback: Pure Pillow chibi drawing
test_unified_render_sticker: Graceful multi-tier fallback
test_tool_handler: Hermes Tool Schema & media tag output
test_input_validation: Emotion whitelist + size clamp
test_profile_output_dir: Profile-aware plugin-data path
test_hook_removed: No post_llm_call registration

🏆 Hackathon Submission Metadata

Competition: Nous Research Hermes Agent Plugin Hackathon (Nov 2026)
Target Category: System & Personality Enhancement / Multimedia Experience
Author: Eric Black & Xuewan (雪丸)
License: MIT

README from the reviewed commit 7db126a; it updates when the author re-pins.

About