DEV Community

Cover image for Klepai v0.1.0: regenerate keeps your old answer
Crystal Studio
Crystal Studio

Posted on

Klepai v0.1.0: regenerate keeps your old answer

Klepai is our free desktop chat for the model servers you host: Ollama and any OpenAI-compatible endpoint on your network, in one window on macOS, Windows and Linux. The v0.0.2 post covered the collision guard. This one is about v0.1.0, our first beta, which shipped today: regenerate no longer throws your previous answer away.

GitHub logo CrystalPlatforms / klepai

Klepai - OSS Self-Hosted LLM Manager for Every OS.

Klepai by CrystalPlatforms

Klepai is a private, self-hosted LLM manager: a desktop chat app for macOS, Windows and Linux that connects to your own model servers over the LAN.

  • Ollama first-class + any OpenAI-compatible endpoint
  • Rich chat: streaming, markdown, code blocks, thinking blocks
  • Agent mode: your server models code in a project folder through built-in tools, with Claude Code and Codex CLI as ready-made extra brains
  • Everything local: SQLite, no accounts, no telemetry

Status: Klepai Alpha v.0.0.1 (released) · see PRD.md

Updates

Update bundles are signature-verified against GitHub Releases: the app ships with the public key in app/src-tauri/tauri.conf.json (plugins.updater.pubkey), and release artifacts are signed with the matching private key at build time.

Maintainer setup (done once, 2026-10-24):

  1. Keypair generated with npx tauri signer generate -w ~/.tauri/klepai.key (no password). The private key lives OUTSIDE the repository, so never commit it.

  2. If CI ever builds releases, store the contents…

Quick answer: v0.1.0 adds variant navigation. Regenerate keeps the old answer as a variant, the last message grows < and > controls with a "2/3" readout, and each variant is labelled with the model that produced it. Stopping a generation midway keeps the partial answer as a variant marked "aborted". Sending a follow-up prunes the variants and keeps the one you displayed.

Why regenerate was the wrong verb

The old flow deleted the previous answer the moment a retry started. Comparing attempts meant memorising them. The PRD (plans/betas/v0.1.0.md in the repo) states the problem plainly: a weak answer can only be overwritten, not kept beside its retry.

How variants work

Regenerate now appends to a ring instead of replacing. Each variant stores its content, its model label, and a visible marker — exactly one variant per answer is on screen. The controls exist on the last assistant answer only; the rest of the thread stays fixed, because a conversation-tree engine was rejected as overkill.

Two guards bound the ring. It caps at ten variants, and creating the eleventh drops the oldest. And the arrows sleep while a stream runs, so navigation cannot race a live generation. Everything lives in SQLite, so the ring and your visible pick survive a restart.

Stopping halfway

Stop keeps its evidence. A generation you halt midway persists as a variant with an "aborted" marker next to the model label. An abort with zero content creates nothing, because there is nothing worth keeping.

What the model sees

This is the decision the feature stands on: the message row always mirrors the visible variant. The model reads ordinary conversation history, so it reads the variant you chose to display — no special context logic anywhere. When you send the next message, the other variants get deleted in one SQL statement and the displayed text simply stays as the message.

The update banner

The second story in this release: the app now announces updates instead of hiding them in Settings. One check at app start — no background polling. On a new release the window shows a banner with Install, which runs the signature-verified update flow from Alpha, and See what's new, which opens the GitHub release page. Dismiss it and it returns next launch, until you update. Quiet states stay quiet.

Also in this build

Conversations with Claude Code or Codex always run in Agent mode now. The Natural Mode toggle is gone for CLI brains; picking a CLI brain flips the mode in the same write, and older conversations normalise on reopen.

What's next

The roadmap grew on the planning side: v1.2.0 is a read-only Diff view, v1.3.0 a per-conversation sandbox for agent isolation. Both PRDs live under plans/ in the repo. Repo and downloads: github.com/CrystalPlatforms/klepai. Feedback is fastest as GitHub issues.

Landing: https://klepai.pages.dev

Top comments (0)