Stella Selfie

Generate persona-consistent selfie images and send to any OpenClaw channel. Supports Gemini, fal, and laozhang.ai providers, multi-reference avatar blending.

Security Scan

The skill appears to implement what it claims (image generation + channel send) but has a few incoherences and a high-privilege setting (always:true) that you should review before installing.

Purpose & Capability

Name/description, required binaries (node), and code files align with an image-generation + channel-sending skill. However the registry metadata/ SKILL.md declare three required API keys (GEMINI_API_KEY, FAL_KEY, LAOZHANG_API_KEY) while runtime logic only requires the API key for the active provider (process.env.Provider or default gemini). Declaring all three as required in metadata is overspecified and inconsistent with the actual runtime needs.

Instruction Scope

Runtime instructions and code stay within the stated purpose: assembling prompts, optionally calling timeline_resolve, reading identity/avatar data from ~/.openclaw/workspace/IDENTITY.md and an avatars directory, calling provider SDKs or REST endpoints, writing generated images to ~/.openclaw/workspace/stella-selfie, and invoking the OpenClaw CLI to send media. This file I/O (reading arbitrary files under the user's OpenClaw workspace and avatars dir) is legitimate for avatar blending but is scope-sensitive — users should be aware the skill will read those local images and identity config. The skill executes the OpenClaw CLI via child_process.exec to send messages (expected for channel delivery).

Install Mechanism

Install entries are Node packages (@google/genai and @fal-ai/client) from standard registries; no arbitrary URL downloads or extraction steps are used. This is proportional to the stated functionality.

Credentials

The skill's metadata lists three provider API keys as required, but runtime only enforces the key for the selected provider. Requiring all three keys up-front is disproportionate and may encourage unnecessary credential exposure. The skill does not request unrelated cloud credentials, but it will read files from the user's home OpenClaw workspace (IDENTITY.md, avatar images), which could contain personal data — this access is functional but should be explicit to the user.

Persistence & Privilege

The skill is published with always: true. That forces the skill to be included by default in every agent run and bypasses usual eligibility gates. There is no clear justification in SKILL.md for needing permanent inclusion; this increases blast radius because the skill can be invoked autonomously and has access to local workspace files and the OpenClaw CLI. This combination (always:true + local file access + ability to call external provider APIs) is the main security concern.

What to consider before installing

What to consider before installing: - always:true is set: ask the author why the skill must be force-included in every agent run; if there's no good reason, avoid installing or ask the registry owner to remove that flag. always:true increases the risk that the skill can be invoked autonomously and access local files. - Credentials: the registry lists GEMINI_API_KEY, FAL_KEY, and LAOZHANG_API_KEY as required. In practice the code only needs the key for the provider you choose. Don't provision all keys unless you intend to use all providers; prefer giving the minimum-permission key for the chosen provider. - Local file access: the skill reads ~/.openclaw/workspace/IDENTITY.md and any configured avatars directory and will base64-embed local images when sending to providers. Make sure you trust the skill and that your IDENTITY.md / avatars dir do not contain private images you don't want processed or transmitted to external APIs. - OpenClaw CLI usage: the skill runs the OpenClaw CLI via exec to send messages. Check that your CLI is up-to-date and that you are comfortable with the skill invoking it on your behalf (sent command arguments may be visible to local process listing while running). - Network calls: the skill calls third-party provider APIs (Gemini via @google/genai SDK, fal via SDK, and laozhang.ai REST). Verify you trust those providers and that you understand where images/metadata will be sent. - Code provenance: the source is marked unknown and homepage points to a GitHub repo. If you plan to use it, review the repo history or prefer a version from a verified publisher. Ask the maintainer to (a) remove the always:true flag if not needed, and (b) update registry metadata to only require provider keys actually used, to reduce unnecessary credential exposure. If you want, I can: (1) list all places the skill reads/writes files and network endpoints it contacts, (2) suggest exact metadata edits to reduce requested env vars and remove always:true, or (3) draft questions to ask the author/maintainer.

Generate persona-consistent selfie images using Google Gemini or fal (xAI Grok Imagine) and send them to messaging channels via OpenClaw. Supports multi-reference avatar blending for strong character consistency.

When to Use

• User says "send a pic", "send me a photo", "send a selfie", "发张照片", "发自拍"
• User says "show me what you look like...", "send a pic of you...", "展示你在..."
• User describes a scene: "send a pic wearing...", "send a pic at...", "穿着...发张图"
• User wants the agent to appear in a specific outfit, location, or situation

Prompt Modes

Mode 1: Mirror Selfie (default)

Best for: outfit showcases, full-body shots, fashion content

A mirror selfie of this person, [user's context], showing full body reflection.

Mode 2: Direct Selfie

Best for: close-up portraits, location shots, emotional expressions

A selfie of this person, [user's context], looking into the lens.

Mode 3: Third-Person Photo

Best for: non-selfie viewpoints, including explicit third-person requests and scenes that should not read as a selfie

A natural third-person photo of this person, [user's context], natural composition, not a selfie.

Mode Selection Logic

Signal	Auto-Select Mode
Strong user keywords: outfit, wearing, clothes, dress, suit, fashion	mirror
Strong user keywords: full-body, mirror, reflection, pose, show the look	mirror
Strong user keywords: selfie, close-up, portrait, face, eyes, smile, looking into the lens	direct
Strong user keywords: third-person, not a selfie, candid shot, 他拍, 路拍, 抓拍	third_person
Legacy keywords: travel photo, tourist photo, 旅拍, 打卡照, 风景合影	third_person

Default policy:

• Interpret explicit user requirements first: camera style, outfit emphasis, body framing, scene, pose, and expression.
• Use mirror by default for outfit / full-body / self-presentation requests, even if the user did not explicitly mention a mirror.
• Use direct by default for selfie requests focused on face, emotion, immediacy, or in-the-moment presence.
• Use third_person only when the user explicitly asks for a non-selfie style or clearly describes a shot that should not read as a selfie.

Default mode when no keywords match and timeline is unavailable: mirror

Resolution Keywords

User says	Resolution
(default)	1K
2k, 2048, medium res, 中等分辨率	2K
4k, high res, ultra, 超清, 高分辨率	4K

Step-by-Step Instructions

Step 1: Collect User Input

Determine from the user's message:

• Explicit context (optional): scene, outfit, location, activity — detect from keywords
• Mode (optional): mirror, direct, or third_person — auto-detect from explicit user intent if not specified
• Target channel: Where to send (e.g., #general, @username, channel ID)
• Channel provider (optional): Which platform (discord, telegram, whatsapp, slack)
• Resolution (optional): 1K / 2K / 4K — default 1K
• Count (optional): How many images — default 1, only increase if explicitly requested
• Has explicit scene?: Does the request contain any specific scene/outfit/location/activity keywords?

Step 2: Enrich with Timeline Context Or Recent Scene Recall

timeline_resolve is an optional enhancement, not a prerequisite.

• If timeline_resolve is unavailable in the current environment, skip this step and proceed with Stella's default behavior.
• If the request is a current-state Sparse prompt — for example "发张自拍", "发张照片", "想看看你", "send a selfie", "send a photo", "show me what you look like" — and timeline_resolve is available, load and follow references/timeline-integration.md.
• If the current request clearly refers back to a single recently resolved timeline scene in the current conversation, load and follow references/timeline-integration.md even if the photo request itself is not Sparse.
• If the user already provided a clear standalone scene, outfit, location, activity, or camera requirement and it is not a callback to a recently resolved timeline scene, do not use timeline enhancement. Follow the default policy directly.
• When you do call timeline_resolve, do not freely rewrite the request into output-slot questions. Use the fixed query rules in references/timeline-integration.md.
• Only enable Nano Banana real-world grounding when the prompt can explicitly include a concrete city plus an exact local date/time anchor from timeline data. If those anchors are missing, do not claim real-world synchronization.
• If timeline returns fact.status === "empty", is missing result.consumption, or any error occurs, immediately fall back to Step 3 without mentioning timeline failure to the user.

Never block image generation on timeline availability. Timeline enrichment is best-effort and should only be used for current-state Sparse prompts or explicit callbacks to a recently resolved timeline scene.

Step 3: Assemble Prompt

Select mode from the default policy first.

If the request is Sparse, and you loaded references/timeline-integration.md and obtained usable timeline context, apply its Sparse-only merge and prompt rules.

When that timeline enrichment includes outdoor real-world grounding, keep the grounding clause as a separate strong instruction sentence rather than a soft atmosphere phrase like Make it feel like....

Otherwise, use the user's explicit context directly and keep Stella's original fallback behavior:

[mirror]  A mirror selfie of this person, [user's explicit context if any], showing full body reflection.
[direct]  A selfie of this person, [user's explicit context if any], looking into the lens.
[third_person] A natural third-person photo of this person, [user's explicit context if any], natural composition, not a selfie.

Step 4: Generate Image

Run the Stella script:

node {baseDir}/dist/scripts/skill.js \
  --prompt "<ASSEMBLED_PROMPT>" \
  --target "<TARGET_CHANNEL>" \
  --channel "<CHANNEL_PROVIDER>" \
  --caption "<CAPTION_TEXT>" \
  --resolution "<1K|2K|4K>" \
  --count <NUMBER>

Step 5: Confirm Result

After the script completes, confirm to the user:

• Image was generated successfully
• Image was sent to the target channel
• If any error occurred, send a concise actionable failure message

Environment Variables

Stella supports multiple providers and a gateway-backed send path, so its sensitive runtime environment variables are explicitly declared in metadata.openclaw.requires.env for OpenClaw's env-injection allowlist. The skill also sets metadata.openclaw.always: true, so these declarations do not become hard load-time gates. Actual credential validation remains runtime-driven inside skill.js, based on the selected provider.

Variable	Required	Description
GEMINI_API_KEY	Required (if Provider=gemini)	Google Gemini API key
FAL_KEY	Required (if Provider=fal)	fal.ai API key
LAOZHANG_API_KEY	Required (if Provider=laozhang)	laozhang.ai API key (sk-xxx); get it at api.laozhang.ai
Provider	Optional	Image provider: gemini, fal, or laozhang
AvatarBlendEnabled	Optional	Enable or disable multi-reference avatar blending
AvatarMaxRefs	Optional	Maximum number of reference images to blend

Credential requirements are provider-specific:

• Default Provider=gemini: requires GEMINI_API_KEY
• Provider=fal: requires FAL_KEY
• Provider=laozhang: requires LAOZHANG_API_KEY

Media File Handling (Gemini)

When Provider=gemini, Stella writes generated files to:

• ~/.openclaw/workspace/stella-selfie/

After successful send, Stella deletes the local file immediately. If send fails, the file is kept for debugging.

Skill Environment Options

Configure in your OpenClaw openclaw.json under skills.entries.stella-selfie.env:

Option	Default	Description
Provider	gemini	Image provider: gemini, fal, or laozhang
AvatarBlendEnabled	true	Enable multi-reference avatar blending
AvatarMaxRefs	3	Maximum number of reference images to blend

Note for Provider=fal users: fal's image editing API only accepts HTTP/HTTPS image URLs. Local file paths (from Avatar / AvatarsDir) are not supported. Configure AvatarsURLs in IDENTITY.md with public URLs of your reference images to enable image editing with fal.

Note for Provider=laozhang users: laozhang.ai uses the Google-native Gemini API format (gemini-3-pro-image-preview). It requires local reference images from Avatar / AvatarsDir and does not use AvatarsURLs. Supports 1K/2K/4K resolution and 10 aspect ratios. Get your API key at api.laozhang.ai — remember to configure a billing mode in the token settings before use.

Delivery Path

• Stella sends via openclaw message send.
• Delivery auth and routing are handled by the local OpenClaw installation, not by skill-level gateway tokens.

External Endpoints And Data Flow

Endpoint / path	When used	Data sent
Google Gemini API	Provider=gemini	Prompt text and selected local reference images from Avatar / AvatarsDir
fal API	Provider=fal	Prompt text and public reference image URLs from AvatarsURLs
laozhang.ai API (api.laozhang.ai)	Provider=laozhang	Prompt text and local reference images (Avatar / AvatarsDir, uploaded as base64)
Local OpenClaw CLI	Always for delivery	Target channel, target id, caption text, and generated media path/URL

Security And Privacy

• Stella reads ~/.openclaw/workspace/IDENTITY.md and local avatar files to build reference context.
• Under Provider=gemini, selected local avatar images are uploaded to Gemini as part of normal image generation.
• Under Provider=fal, only public http/https avatar URLs are sent; local avatar files are not uploaded to fal directly.
• Under Provider=laozhang, local avatar files from Avatar / AvatarsDir are base64-encoded and uploaded to laozhang.ai.
• Generated files (Gemini and laozhang) are written to ~/.openclaw/workspace/stella-selfie/ and deleted after successful send.

User Configuration

Before using this skill, you must configure your OpenClaw workspace. See templates/SOUL.fragment.md for the recommended capability snippet to add to your SOUL.md.

Required: IDENTITY.md

Add the following fields to ~/.openclaw/workspace/IDENTITY.md:

Avatar: ./assets/avatar-main.png
AvatarsDir: ./avatars
AvatarsURLs: https://cdn.example.com/ref1.jpg, https://cdn.example.com/ref2.jpg

• Avatar: Path to your primary reference image (relative to workspace root)
• AvatarsDir: Directory containing multiple reference photos of the same character (different styles, scenes, outfits)
• AvatarsURLs: Comma-separated public URLs of reference images — required for Provider=fal (local files are not supported by fal's API)

Required: avatars/ Directory

Place your reference photos in ~/.openclaw/workspace/avatars/:

• Use jpg, jpeg, png, or webp format
• All photos should be of the same character
• Different styles, scenes, outfits, and expressions work best
• Images are selected by creation time (newest first)

Required: SOUL.md

Add the Stella capability block to ~/.openclaw/workspace/SOUL.md. See README.md ("4. SOUL.md") for the copy/paste snippet.

Installation

clawhub install stella-selfie

After installation, complete the configuration steps above before using the skill.