Sticker Manager

Manage a local sticker / reaction-image library with keyword lookup, quality checks, semantic tags, batch import/collection workflows, and media sending support.

Authors:

• - Wenzhuang Zhu
• TetraClaw (丁蟹)

Repository / backup source:

• - GitHub: https://github.com/TetraClaw/sticker-manager
• If ClawHub is unavailable, use GitHub as the fallback download source.

What this skill does

1. 1. Save the latest inbound image or GIF into a local library
2. Save a specific image from recent chat/media history
3. Search stickers by keyword
4. Rename or delete existing stickers
5. Clean very low-quality files
6. Tag stickers with emotions / scenes / keywords / descriptions
7. Recommend a sticker based on direct text context or recent chat history
8. Batch-collect stickers toward a target count with dedupe and low-quality filtering
9. Batch-import stickers from local directories
10. Prepare vision-model plans for semantic tagging and image understanding
11. Return resolved file paths so the assistant can send matched media with the message tool

Supported formats

• - JPG
• JPEG
• PNG
• WEBP
• GIF

Default paths

• - Sticker library: INLINECODE1
• Inbound media: INLINECODE2

Environment overrides:

• - INLINECODE3
• INLINECODE4
• INLINECODE5
• INLINECODE6

Typical use cases

1. Save a sticker

When the user sends an image/GIF and says things like:

• - "save sticker"
• "store this"
• "save this image"
• "save the previous image"
• "save that image from chat history"

Basic save:

CODEBLOCK0

Save from recent media history:

CODEBLOCK1

Quality-aware save:

CODEBLOCK2

If no name is provided, save_sticker_auto.py exits with code 2 and returns analysis markers so the assistant can ask a model to name the image.

2. Search a sticker

Command:

CODEBLOCK3

List all stickers:

CODEBLOCK4

Matching order:

• - exact filename match
• partial filename match
• fuzzy containment match

3. Rename / delete / clean

Rename:

CODEBLOCK5

Delete:

CODEBLOCK6

Clean very small files:

CODEBLOCK7

4. Tag and recommend

Add tags:

CODEBLOCK8

Suggest for direct context:

CODEBLOCK9

List tag database:

CODEBLOCK10

Prepare model payload only:

CODEBLOCK11

Recommend from chat history:

CODEBLOCK12

5. Batch collect / import / discover

Collect from URLs or local files:

CODEBLOCK13

If the final count is below target, the command exits with code 2 and prints NEED_MORE=....

Import from local directories:

CODEBLOCK14

Discover sources from directories, URLs, or pages:

CODEBLOCK15

Default discovery is lightweight:

• - local directories are scanned immediately
• remote image URLs are returned as pending unless --fetch-urls is used
• page discovery only counts successfully extracted image URLs

Animation rule (mandatory):

• - If the source is originally animated, prefer downloading the animated asset itself.
• Use a generic decision path: file suffix → HTTP Content-Type → downloaded file content validation.
• Do not silently downgrade animated sources to static WEBP/PNG previews.
• If a source looks animated by reference/content-type but the downloaded file validates as static, reject it from the animated batch instead of pretending it is a GIF.
• Before sending or importing a GIF batch, verify the saved file is actually animated-capable rather than a static preview.

6. Auto-tagging and vision planning

Generate a vision plan for a single file:

CODEBLOCK16

Generate plans for a directory:

CODEBLOCK17

Standalone vision fallback plan:

CODEBLOCK18

This returns a JSON plan with:

• - image path
• candidate vision models
• prompt goal
• fallback failure message

Suggested default model order:

1. 1. INLINECODE13
2. INLINECODE14

Sending workflow

After resolving a file path, send it with the message tool.

Example:

CODEBLOCK19

If you need to reply to a specific message, pass replyTo="<message_id>".

Quality rules

The auto-save flow checks media size before saving.

Force-save example:

CODEBLOCK20

Notes for agents

• - Prefer semantic names over timestamp-heavy names
• Avoid platform-specific hardcoding in skill logic
• Use environment variables for library location overrides
• Keep send logic in the assistant/tool layer, not inside raw scripts when possible
• Validate file existence before sending
• For image meaning extraction, try the configured vision-capable model chain in INLINECODE17
• If the primary image model fails, try fallback models before giving up
• If all vision models fail, explicitly tell the user that image meaning, semantic tagging, and quality validation could not be completed reliably
• Treat __MODEL_MATCH__, __AUTO_TAG__, __SEMANTIC_BATCH__, __ANALYZE_HISTORY__, and __CONTEXT_RECOMMEND__ markers as structured handoff payloads for the outer assistant layer

Files

• - scripts/common.py - shared path, i18n, and vision-plan helpers
• INLINECODE24 - keyword lookup and inventory list
• INLINECODE25 - rename, delete, and cleanup
• INLINECODE26 - basic save from inbound or history media
• INLINECODE27 - quality-aware save flow
• INLINECODE28 - tagging, recommendation, auto-tag planning, and context analysis
• INLINECODE29 - batch collection with dedupe and semantic-plan output
• INLINECODE30 - local-directory import with optional auto-tag planning
• INLINECODE31 - discovery from URLs, directories, and static pages
• INLINECODE32 - publish-safety scan for sensitive-looking content
• INLINECODE33 - pytest coverage for CLI and workflow behaviors