vision-agent-skill

Vision Agent

Purpose

Specialized agent for tasks requiring image understanding and visual analysis.

When to Use

Spawn this agent when the task involves:

• Analyzing screenshots or images
• Reading text from images (OCR)
• Interpreting charts, graphs, or diagrams
• Visual debugging (UI layouts, error screenshots)
• Image comparison or quality assessment
• Any task where "seeing" the image matters

Model

• Primary: ollama/kimi-k2.5:cloud (text+image, 125k context)
• Why: kimi-k2.5 is vision-capable via Ollama cloud. It can accept image inputs and analyze them.

Data handling (read first)

Images and screenshots are sent to a cloud-hosted vision model — they leave this host. Screenshots in particular routinely contain sensitive data: credentials or tokens on screen, API keys, personal/customer information, internal URLs and hostnames, or private messages. Before sending an image:

• Confirm the operator is OK with this specific image going to the cloud model.
• Prefer redacting or cropping out anything sensitive that isn't needed for the task.
• Never upload an image solely to extract a secret from it.

How to Spawn

sessions_spawn(
    task="[VISION] Analyze this screenshot and tell me what's wrong with the layout",
    mode="run",
    model="ollama/kimi-k2.5:cloud",
    timeoutSeconds=300
)

Task Prefix

Always prefix vision tasks with [VISION] so the Nexus routes them correctly:

• [VISION] — Image analysis, screenshot reading, visual tasks → routes to vision-agent

Examples

# Analyze a screenshot
sessions_spawn(
    task="[VISION] Look at this screenshot and identify any UI errors or layout issues",
    mode="run"
)

# Read text from image
sessions_spawn(
    task="[VISION] Extract all text from this image and format it as markdown",
    mode="run"
)

# Compare images
sessions_spawn(
    task="[VISION] Compare these two screenshots and list the differences",
    mode="run"
)

Communication

• The Nexus stays conversational
• Acknowledge before spawning: "I'll delegate this image analysis to the vision agent."
• Synthesize results when complete