Wire / Engineering
Every LLM agent loses its working state when the context window fills. Most systems pretend this doesn’t happen, or treat it as a user problem. Argus treats it as an engineering problem.
It reads the live session log every three minutes and writes a lean recovery pointer file — under 3KB — that tells the agent what it was doing, what files it was touching, and why. When the context resets, that file is the first thing read.
Tier 1 is automatic. Argus infers working state from task lifecycle timestamps, atom history, and file modification times. No dependency on the agent.
Tier 2 is richer. The agent writes a voluntary checkpoint when it has something worth preserving. Self-reported state wins over inferred state.
The sync log tracks every pass, and the agent posts a recovery marker atom at session start when it successfully reads the pointer file. That closes the measurement loop.
Raw Python HTTP server. No framework. One file, currently around 6,000 lines. This is not a boast — it’s a maintenance choice that hasn’t become a problem yet.
It coordinates multiple background daemons: the session log parser, the continuity inference loop, a watchdog, and a fleet health monitor. It also routes atoms, tasks, and scheduled events for the rest of the stack.