cel-context

Portable context snapshots for AI agents.

cel-context gives agents, tools, and runtimes a shared way to describe software state. Any source can emit ContextElements, and any consumer can read the resulting ScreenContext without depending on a specific browser, desktop, OCR, accessibility, or monitoring stack.

What It Owns

ContextElement — one UI, app, log, or source-derived fact.
ScreenContext — one snapshot of app/window state and context elements.
ContextSource — source labels such as accessibility, CDP, native API, OCR, vision, merged.
ContentRole — prompt-injection-aware classification.
ContextReference — resilient targeting data that survives ephemeral IDs.
ContextContribution / ContextMerger — generic source contribution merge mechanics.

What It Does Not Own

live cortex operation
runtime freshness / anomaly policy
action dispatch
policy enforcement
monitoring dashboards or compliance exports

Those live in the commercial Cellar/Dilipod runtime unless explicitly opened later.

Example

cargo run -p cel-context --example context_snapshot -- --json

The example constructs a snapshot from caller-provided ContextElements. Live capture belongs in runtime/source crates that emit the same contract shape.

Minimal Element

use cel_context::{ContentRole, ContextElement, ContextSource};
use std::collections::HashMap;

let element = ContextElement {
    id: "app:button:submit".into(),
    label: Some("Submit".into()),
    description: None,
    element_type: "button".into(),
    value: None,
    bounds: None,
    state: Default::default(),
    parent_id: None,
    actions: vec!["click".into()],
    confidence: 0.95,
    source: ContextSource::NativeApi,
    content_role: ContentRole::Interactive,
    properties: HashMap::new(),
};

License