Releases: MasihMoafi/Elpis
Release list
Elpis v0.1.2
Elpis is a terminal environment for coding agents. The agent runs the model loop;
Elpis owns context, memory, continuity, retrieval, permissions, and provider choice.
Install (Linux x86_64 or macOS Apple Silicon):
curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/main/scripts/install-elpis.sh | bash && ~/.local/bin/elpisAlready running Elpis? Use elpis --update.
What's new since 0.1.1
- Durable Ace pruning — selective steady and pressure passes, configurable reclaim targets,
/pruneand/force-prune, animated savings, a/contextbreakdown, and saved totals that survive session resume. Luna runs the pruning pass at High reasoning effort. - Provider routing — explicit Anthropic, Gemini, and OpenRouter routing with corrected authentication boundaries and live model/status reporting.
- Accountable agent work graphs — persisted task graphs, bounded roles and scopes, dependent verification, and an
/agentview of task state and evidence. - Permission and interface fixes — permissions can cycle while work is running; hidden commands no longer intercept typed input; obsolete usage-limit reset UI was removed.
- Documentation and evaluation refresh — the README now includes context-retention and tool-call evidence, with reproducible continuity and work-graph evals.
Known limits
- Durable memory extraction works, but promotion still requires recall evidence that has not been reached on Masih's install.
- Work-graph and continuity evals are automated; Masih's final functional acceptance remains outstanding.
- Windows,
/auto, voice input, and LSP integration remain out of scope.
Verify each download with its matching .sha256 asset. Full docs are in the README.
Elpis v0.1.1
Elpis is a terminal environment for coding agents. The agent runs the model loop; Elpis owns everything around it — what goes into each request, what stays on disk, and what carries across sessions.
Install (Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/main/scripts/install-elpis.sh | bash && ~/.local/bin/elpisInstall (macOS, Apple Silicon):
curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/main/scripts/install-elpis.sh | bash && ~/.local/bin/elpisThe macOS build is CI-proven — it's tested headlessly in the release pipeline on real Apple Silicon runners — but not yet verified on a developer's own machine. Report anything unexpected.
Already running Elpis? Use elpis --update.
What's new since 0.1.0
- macOS (Apple Silicon) — first release with a native
arm64build. Intel Macs are not supported; the installer will tell you to use Rosetta if you try. - RAG removed — the bundled retrieval engine (a Python/PyTorch sidecar) and the
/ragcommand are gone. Retrieval is now a capability you register like any other tool, via MCP; ask in plain language and the agent calls it and cites sources. - Durable memory on by default — was off by default in 0.1.0.
- Dev-rule files now load on any machine — previously the Context Ledger only found a project's dev-rule files (
skills/dev/*.md) if they happened to sit in a folder next to the project. Fixed: it now also checks a fixed location under Elpis's home directory, so the rules travel with the machine, not the project's disk layout. - Context Ledger: durable memory row is toggleable — the keyboard toggle silently excluded the memory row from selection; fixed.
- Startup cost measured and reduced; debug builds now flag themselves instead of pretending to be a release build.
What it does
- Context pruning on three levels — optional RTK shell-output filtering before the agent sees a command's output, Codex's inherited safety cap for exceptionally large output, then Ace's meaning-aware post-turn pass. Useful results become a compact conclusion plus an evidence pointer; dead ends leave working context. In a controlled same-task comparison, Elpis ended with 93% free context against Codex's 73%.
- Context Ledger —
/contextlists exactly which files are in the working set and why, with ctrl+clickable paths. Add files, exclude them, or remove them outright. - Durable evidence — full conversations, terminal events, and artifacts stay exact on disk after they leave working context, and can be retrieved later.
- Portable session continuity — older history compacts into goal and checkpoint state, so a session resumes without replaying the transcript.
- Its own state — configuration, sessions, history, logs, and hooks live under Elpis's own directory, so Elpis and Codex no longer silently alter each other.
- Self-update —
elpis --updateverifies the release checksum and replaces the binary atomically, leaving the existing one untouched on any failure. - Bounded local memory — selective and size-capped, backed by provenance, and now on by default.
- Provider choice — OpenAI, Anthropic, Gemini, and OpenRouter, chosen explicitly rather than routed silently.
- No telemetry by default — no analytics are uploaded, and every OpenTelemetry exporter is off unless you configure one.
Not in this release: Windows, /auto routing, multi-agent control, voice input, LSP integration.
Verify the download with the matching .sha256 asset. Full docs in the readme.
Elpis v0.1.0
Elpis is a terminal environment for coding agents. The agent runs the model loop; Elpis owns everything around it — what goes into each request, what stays on disk, and what carries across sessions.
This is the first stable release. It replaces the two earlier preview builds, which have been withdrawn: they predate the pruning correctness work, Elpis's own state directory, and the updater, and should not be installed.
Install (Linux x86_64):
curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/main/scripts/install-elpis.sh | bash && ~/.local/bin/elpisAlready running Elpis? Use elpis --update.
What it does
- Context pruning on three levels — optional RTK shell-output filtering before the agent sees a command's output, Codex's inherited safety cap for exceptionally large output, then Ace's meaning-aware post-turn pass. Useful results become a compact conclusion plus an evidence pointer; dead ends leave working context. In a controlled same-task comparison, Elpis ended with 93% free context against Codex's 73%.
- Context Ledger —
/contextlists exactly which files are in the working set and why, with ctrl+clickable paths. Add files, exclude them, or remove them outright. - Durable evidence — full conversations, terminal events, and artifacts stay exact on disk after they leave working context, and can be retrieved later.
- Portable session continuity — older history compacts into goal and checkpoint state, so a session resumes without replaying the transcript.
- Its own state — configuration, sessions, history, logs, and hooks live under Elpis's own directory, so Elpis and Codex no longer silently alter each other.
- Self-update —
elpis --updateverifies the release checksum and replaces the binary atomically, leaving the existing one untouched on any failure. - Bounded local memory — selective and size-capped, backed by provenance.
- Provider choice — OpenAI, Anthropic, Gemini, and OpenRouter, chosen explicitly rather than routed silently.
- No telemetry by default — no analytics are uploaded, and every OpenTelemetry exporter is off unless you configure one.
Not in this release: macOS, Windows, /auto routing, multi-agent control, voice input, LSP integration. Local RAG works only from a source checkout.
Verify the download with the matching .sha256 asset. Full docs in the readme.