Elpis v0.2.0

Elpis v0.2.0

Elpis v0.2.0 is a Linux-first coding-agent TUI centered on explicit context control
and auditable tool-output admission. The maintainer accepted this version's local
candidate before release. Smart Prune remains experimental and off by default.

Install

Linux x86_64. The installer checks the downloaded binary against its SHA-256
sidecar; review the script before running it.

curl -fsSL https://raw.githubusercontent.com/MasihMoafi/Elpis/v0.2.0/scripts/install-elpis.sh | bash && ~/.local/bin/elpis

September 6 maintenance refresh

The v0.2.0 downloads have been refreshed with these corrections:

• Ledger, /context, /usage, and dashboard evidence links open private readable reports with Markdown downloads. The maintainer confirmed the Ledger link opens in normal browser use.
• Background dashboard polling preserves existing evidence links and avoids flashing refresh controls. Unavailable subscription-cost text is smaller and wraps within its column.
• Context and dashboard model labels follow the active model selection. Smart Prune's Luna optimizer uses the selected Max reasoning effort without changing the main model's effort.

Existing v0.2.0 users should rerun the installer above to get the refreshed build. Restart Elpis after updating, then open newly generated evidence links. Existing file URLs in old terminal output are not rewritten. Smart Prune remains experimental; this refresh makes no new quality, latency, or savings claim.

What changed

• Smart Prune (Experimental, off by default) can optimize eligible textual tool results before the main model first sees them. Once admitted, Smart Prune does not revisit that history.
• Clear recovery boundaries: /prune enables Smart Prune for subsequent turns without rewriting sent history. /force-prune <1-100> remains an explicit emergency action that may reduce prompt-cache reuse; /compact remains native Codex compaction.
• Context observability: the Context Ledger, /context, and /dashboard distinguish current usage, estimated attribution, history rewrites, Smart Prune admissions, and optimizer overhead. Ledger and /context share category colors and a full-window scale; individual categories remain estimates. The local candidate's visual acceptance is recorded.
• Explicit continuity and memory: GOAL.md, ES.md, development rules, and the user-maintained MEMORY.md are visible admission choices. Elpis does not claim automatic memory extraction or promotion.
• Safer customization: ordinary skills and plugins do not participate until deliberately enabled. Configured MCP servers remain independent of the plugin gate.
• Accountable work graphs (under development, off by default): persisted DAGs validate dependencies, write scopes, evidence, and required verifier tasks before dispatch.
• Provider and activity improvements: model/reasoning selection, live turn timing, cost availability, dashboard updates, and failure accounting are more explicit.

Evidence boundary

Focused tests establish Smart Prune's first-exposure placement, exact fail-open
behavior, and append-only tested request prefix. An older normal-work ON pilot
observed 95.85% cached input overall. A separate September 5 ON-only smoke admitted
one smaller output, retained a planted fact, and observed 91.57% cached input on
its follow-up response. Neither is an OFF/ON comparison; total cost, latency and
general task-quality effects remain unproven. See the
experiment log.

Known limits

• Smart Prune uses an extra optimizer request and can add latency. It is Experimental and disabled by default.
• Manual MEMORY.md admission works; automatic memory promotion does not exist.
• Work graphs and native Anthropic/Gemini routes remain under live user acceptance.
• This release targets Linux x86_64 only. macOS and Windows are not included.

Read the technical preprint, evaluation results, and full README.

Verified source: f0e68084caf1c385f56f2748be1a10066bfe77b3. Release checks.