This repo is still early and engine-first.
The bar for contributions is simple:
- keep the Aperture core SDK package small
- do not couple source-specific logic into core
- prefer deterministic, testable behavior over speculative abstraction
- keep adapter logic thin and reversible
The most valuable contributions are grounded in real multi-agent human attention workloads.
If you use Aperture, please try it against:
- multiple agent sources competing for attention
- approval-heavy workflows
- failure + blocked-task mixes
- noisy status streams where low-value work should stay ambient
The goal is not just to make the code pass. The goal is to pressure-test whether the engine is actually making good attention decisions.
Additional adapters are welcome.
Good adapter contributions:
- preserve the
ApertureEvent -> ApertureCore -> FrameResponseboundary cleanly - keep source-specific transport and mapping outside the Aperture core SDK package
- make ingress and egress both explicit when the upstream system supports them
Install dependencies:
pnpm installRun the main checks:
pnpm boundary:check
pnpm test
pnpm typecheckFor a full release-shaped pass, run:
pnpm release:checkRun the companion surface:
pnpm demo:tuiBefore adding code, ask:
- does this strengthen the attention engine?
- does this belong in an adapter instead of core?
- can this be expressed with fewer public constructs?
Good contributions:
- better deterministic judgment
- better signal quality
- better tests around attention behavior
- thin ingress or egress adapters
- real-world workload reports and scenario traces
- clearer docs and examples
- refactors that preserve the public CLI and SDK behavior while reducing hidden coupling
Bad contributions:
- source-specific conditionals inside core
- broad UI framework work not tied to the engine
- speculative anticipation layers without real signal grounding
- unnecessary new public abstractions
- new non-test imports from
packages/*/srcintopackages/core/src/*instead of using a supported contract