Elpis is a terminal environment for coding agents, built on OpenAI's Codex CLI. The core problem: long agent sessions bury useful state under the history of how the agent got there. Every request pays for all of it. Elpis separates working context from durable evidence. After each turn, a meaning-aware pruner (Ace) evicts completed tool outputs, resolved dead ends, and low-signal history from the live context to disk. The next request starts lean. Measured against the Codex CLI baseline on the same task: Elpis finishes with 93% context free; Codex finishes at 73%. Features: • Context Ledger — always-visible working set with ctrl+clickable paths • Portable continuity — goal and checkpoint state survive restarts • Bounded memory — selective, size-capped, attributable • Explicit provider choice — OpenAI, Anthropic, Gemini, OpenRouter • No analytics by default v0.1.1 · Linux x86_64 · Open source (MIT)

The working-context vs durable-evidence split is a smart way to cut token spend. Since Elpis wraps Codex CLI, does it let users point at a custom base URL/model endpoint, or is it tied to OpenAI's API directly? Curious how Ace's pruning behaves across different upstream models.

The meaning-aware pruner is a clever move. Skipping the rerun-token tax on resolved dead ends is the kind of detail that separates a polished tool from a clever demo.

Elpis

Terminal environment that keeps your AI agent's context lean

Terminal environment that keeps your AI agent's context lean