Technical Constraints
The platform had to ingest heterogeneous local signals, preserve ordering, and stay auditable without introducing opaque automation or high operational overhead.
Engineering Case Study
Erebus
Event-driven Linux coordination layer that records system context, infers higher-level state, and makes troubleshooting replayable.
Problem
Operational debugging on a personal workstation is usually reactive and ephemeral. Erebus captures low-level events and inferred state so system behavior can be searched, replayed, and explained after the fact.
Constraints
Environment
The system targets a Linux desktop with systemd, D-Bus, Wayland session signals, and local SQLite storage.
Tradeoffs
Erebus favors append-only event history and explicit confidence scores over brittle single-state assumptions. That increases storage and modeling work, but keeps inference inspectable.
Architecture
- Emitter processes collect focused slices of system activity such as GPU state, network changes, focus transitions, and session events.
- Events are written to an append-only store that acts as the canonical timeline for replay and debugging.
- A belief engine derives higher-level state from those events, assigning confidence rather than collapsing everything into binary conclusions.
- Search and inspection layers expose both raw history and inferred state so operational questions can be answered from evidence.
Key Technical Decisions
Append-only event log
Reconstructing system behavior requires a durable timeline. Immutable history makes replay and audit possible even when inference logic changes.
Confidence-based belief engine
Desktop state is often ambiguous. Confidence scores let the system reason under uncertainty without hiding that uncertainty from the operator.
SQLite with local full-text search
The workload is local-first and read-heavy. SQLite keeps deployment small while still supporting indexed investigation of historical context.
Emitter isolation by domain
Separate collectors make failures easier to contain and allow each integration point to evolve without destabilizing the full pipeline.
Operational Capabilities
Cross-substrate session tracking
The system correlates activity that spans focus changes, locks, network transitions, and other state boundaries into a usable operational story.
Replayable inference
Because the event history is preserved, newer inference logic can be tested against older timelines without losing the original evidence.
Searchable operational context
FTS-backed queries make it possible to inspect incidents as sequences instead of isolated log lines.
Current Outcome
- Erebus replaces ad hoc workstation debugging with a durable model of what happened and why the system believed it was happening.
- The project is still in active development, but the architecture already establishes a path for auditable local automation.
Evidence
Private operator artifact / Private
Private operator artifact
Local project evidence includes append-only event history, emitter output, and inference traces from a personal Linux environment.
Note: Public repository is not currently available; private machine identifiers, socket paths, policies, and actuator details are not published.
Sanitized architecture summary / Sanitized
Sanitized architecture summary
The public case study summarizes the event log, belief engine, and inspection workflow without exposing host-specific automation details.
Links
- Public Repository: Public repository not available.
- Related: Aether
- Related: Arachne