Changelog

Mid-turn steering (foundation). An engine steering channel can deliver a message to a *running* turn at a safe boundary — demonstrated end-to-end with the mock backend. The Claude Code live path is opt-in (LECTERN_STEER) and its components are in place (message format, gating); the full stream-json stdin wiring is deferred pending live verification against the CLI (its input format is an open documentation gap). See docs/mid-turn-steering-design.md.

Skills are scanned on import. Importing a skill — including an external SKILL.md from the ecosystem — now runs the static red-flag scan (destructive shell, secret exfiltration, prompt-injection markers) and surfaces its findings. By default it warns and still imports (your machine, your call); LECTERN_SKILL_STRICT=1 refuses a hard-flagged skill. The token-spending model audit stays publish-only. lectern skills import prints the result.

Opt-in run sandbox (Linux). Set LECTERN_SANDBOX=1 to run an agent's backend inside a [bubblewrap](https://github.com/containers/bubblewrap) sandbox that confines writes to the workspace — the rest of the filesystem is read-only — so an injected instruction that lands becomes a bad transcript rather than files written outside the repo. LECTERN_SANDBOX_NET=off additionally isolates the network. It's off by default, degrades with a clear error if bubblewrap isn't installed, and lectern doctor reports its availability.

Local Agent2Agent (A2A) interop. Lectern can now exchange work with other A2A agents on your machine — A2A (v1.0) is the inter-agent standard that complements MCP. It is off by default, loopback-only, and opt-in:

Cloud login, sync, and accounts. Lectern is now fully local — there's no sign-in, and nothing leaves your machine. Your provider logins (Claude Code, Antigravity, OpenCode) are unchanged, and encrypted session export/import still works.

Checkpoints & rewind. Lectern snapshots your workspace before an agent writes to it, so you can undo a run you don't like and try a different prompt. Snapshots use a private, per-workspace git store that is completely separate from your project's own .git, so it works on non-git folders and never touches your history. Rewind reverts edits and removes files the agent added, and is itself reversible.

In-app auto-updates (Linux). The desktop app checks for newer signed releases and offers a one-click Restart & update — download, install, relaunch. Every update is verified against Lectern's signing key before it installs.

"What's new" on update. After updating, the app shows what changed in the new version, drawn straight from this changelog.

Smarter, leaner memory. Recall now applies a relevance floor, so small talk recalls nothing while genuine matches still surface. When memory content is needed, only the most relevant window of a file enters context (roughly 9× less recalled context on a typical task). The agentic path passes recalled paths, not file contents.

One-click provider setup. Each provider in Settings (Claude Code, Antigravity, OpenCode, OpenRouter, Ollama) expands an OS-aware panel with the exact install command, a copy button, an auth next-step, and links. Vetted user-space installers (OpenCode, Ollama) have a one-click install that streams its output.

Feature-level documentation at [getlectern.vercel.app/docs](https://getlectern.vercel.app/docs) covering chat commands, the Conductor, the brain, scheduling, and the Hub.

First public release. Local-first agent orchestration with per-task model routing, a persistent brain (memory, learned skills, code graph), and multi-provider support (Claude Code, Antigravity, OpenCode, OpenRouter, Ollama).

Terminal stack — CLI, a full TUI, and a background daemon, installable with one command; also a Nix profile install.

Desktop app — a native cockpit (Tauri) with cross-platform installers for Linux (AppImage, .deb), Windows (.exe), and macOS (.dmg); tiled sessions, an embedded terminal, live streaming, and a work-panel preview rail.

The Conductor — plans a task, routes each step to the model that fits, and cross-reviews the result.

MCP + the Hub — a catalog of MCP servers with cross-harness registration, and a community skills hub with an audit gate.

Sessions are shared across the desktop, TUI, and CLI through one engine store.