Skip to content

Releases

re-code 0.1.10

Choose a tag to compare

@github-actions github-actions released this 09 Nov 01:41

Fixed

  • A sandboxed stdio MCP server couldn't read its own files when they live in the home folder (which the
    sandbox empties), so servers such as REA (node ~/.local/share/rea/rea-git/scripts/rea.mjs) failed to
    start. The sandbox now shows a server, read-only, the folders its own command line and settings name:
    a script's project (the nearest folder above it with package.json, .git, pyproject.toml,
    Cargo.toml or go.mod), and folders and files named in its arguments or environment (JAVA_HOME, a
    jar). Never the home folder itself or a broad folder like ~/.config, and secret folders stay hidden.
    Verified live: REA connects inside bubblewrap with its 241 tools.
  • When a server fails to start, re-code waits for its error output to end (up to three seconds) instead
    of a fixed 150 ms, so the explaining line is in the message even when the server is slow to print it.

Install

sudo dnf install ./re-code-0.1.10-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.10-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.9

Choose a tag to compare

@github-actions github-actions released this 09 Nov 01:27

Added

  • A three-step tour on the first run: what re-code found (your local runtime and models, cloud keys),
    what the model can reach (MCP servers, skills, plugins, web search), and how it stays in your hands
    (approvals, the sandbox, undo and rewind, trust). It starts nothing; keys jump straight to the Models
    tab, free cloud models, web search or the skill browser. It opens once, never for anyone with saved
    sessions, /tour shows it again, and --no-onboarding keeps it shut.
  • Clicking a line of a code block in an answer copies that line: wrapped parts joined, line numbers
    left out.
  • Hub fetches report progress: the install, checklist and skill browser windows show "Listing
    owner/repo…", "pdf: 12 of 40 files (80 kB of 260 kB)" and the Hermes index download by the megabyte;
    the command line shows the same on one line that rewrites itself (only on a terminal).

Fixed

  • A stdio MCP server's own error output was drawn straight over the interface (a crashing Node server
    painted its whole stack trace across the screen). The MCP library inherits the terminal's stderr
    unless told otherwise, overriding re-code's setting. A server's error output now goes to the log,
    and when it fails to start, the error names the line that explains it, such as Error: Cannot find module '…/rea.mjs' for a server whose files the sandbox hides.

Install

sudo dnf install ./re-code-0.1.9-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.9-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.8

Choose a tag to compare

@github-actions github-actions released this 08 Nov 19:40

Added

  • The home screen fits the terminal: lists and the detail pane beside them use the window's width and
    height instead of a 116-column, 24-row column. The Settings tab's "Home screen" row keeps the composed
    column for anyone who prefers it, and its "Detail pane" row sets how much of a list's width the
    details take (a third by default, half, or three fifths; never narrower than before).
  • v on a model, provider, MCP server, skill, plugin or tool opens everything the detail pane says in
    the pager, full width and searchable, so nothing is cut short in a narrow pane.

Install

sudo dnf install ./re-code-0.1.8-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.8-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.7

Choose a tag to compare

@github-actions github-actions released this 08 Nov 19:04

Added

  • A dangerous scan verdict can be accepted, knowingly. The scan can't tell a CVE write-up that quotes
    rm -rf / from instructions to run it; the person reading can. o on a dangerous preview (or
    space on a dangerous skill in the Skills tab) lists every finding, says what enabling means, and
    asks; y installs or switches it on by your word. The acceptance is recorded against a fingerprint
    of exactly those findings, so an update that changes them switches the skill off again until you look
    anew. From the shell: re-code skill install --yes --accept-dangerous <what> and
    re-code skill enable <name> --accept-dangerous; --yes alone still keeps a dangerous skill off. A
    plugin's dangerous skill is accepted afterwards, skill by skill, never with the plugin. An accepted
    skill shows ▲ and "dangerous, accepted" in the list.

Install

sudo dnf install ./re-code-0.1.7-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.7-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.6

Choose a tag to compare

@github-actions github-actions released this 08 Nov 18:27

Added

  • The skill search window is now a browser of the whole Hermes index: it lists at once, narrows live as
    you type, and ←/→ step through its sources (Hermes' own skills first, then ClawHub, skills.sh,
    GitHub, LobeHub) with each one's count. Hermes' own skills show their category, and words match
    categories and tags as well as names and descriptions. The header says how many matched past the
    page shown. /skill browse [source] opens within one source; re-code skill browse [--source X] [words] does the same from the shell.

Fixed

  • Piping the command line's output into head wrote a crash file ("failed printing to stdout: Broken
    pipe"); a closed pipe now ends the process quietly with exit code 141, as a pipe death does for other
    tools.

Install

sudo dnf install ./re-code-0.1.6-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.6-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.5

Choose a tag to compare

@github-actions github-actions released this 08 Nov 18:03

Added

  • Tick boxes on every install list. A repository or Hermes category with several skills lists them with
    space to tick and a to tick all; Enter fetches the ticked set into a checklist that shows each
    one's verdict as it lands, and y installs the ticked ones (a dangerous one stays off), k installs
    them off, n discards them all. Skill search hits and plugin listings tick with tab and ^a.
  • The plugin browser groups the Hermes catalog by its categories (tools, desktop, models, memory,
    general, automation, web, platform, voice): ←/→ move along them, and ^a ticks everything shown.
    A checklist takes 25 at a time, since each one is a GitHub listing and GitHub allows 60 an hour
    without a GITHUB_TOKEN.
  • Paste into the install prompt: several links or names (one per line) become a checklist; a pasted
    SKILL.md is staged from the text, scanned and previewed, and recorded as pasted so an update knows
    there is nowhere to fetch from.
  • re-code skill install and re-code plugin install take several things at once and preview each in
    turn; the exit code is the worst of them.

Fixed

  • A repository that is itself a skill but holds more than the file limit (a monorepo with evals and
    tooling around a root SKILL.md, such as conorbronsdon/avoid-ai-writing with 676 files) failed with
    the limit; it now offers its skill folders to choose from, and says why the whole wasn't offered. A
    small repository that is a skill and also holds skill folders lists the whole repository as the first
    choice.

Install

sudo dnf install ./re-code-0.1.5-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.5-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.4

Choose a tag to compare

@github-actions github-actions released this 08 Nov 16:58

Added

  • /untrust (also in the palette) switches a trusted folder's configuration off again after a confirmation
    that lists what goes off; re-code untrust does the same from the shell.
  • A runtime can be started whenever re-code opens: a on the start confirmation remembers the exact
    command; it runs at launch only while the command is still that one (a changed setup asks again), and the
    Settings tab lists it under "Runtimes started when re-code opens" with d to remove.
  • The pager (/diff, /context raw, a tool's output) searches with /, n and N, like the transcript.
  • The Help overlay (F1) scrolls on short terminals.

Fixed

  • A stdio MCP server from a plugin installed during the session, or tested from the command line, couldn't
    see its own files inside the sandbox: the plugin folders are now read when the server starts, not once at
    launch.

Changed

  • The sandbox shows ~/.local/opt (where user-installed toolchains such as Node often live) to commands
    and stdio MCP servers, read-only, like the other toolchain folders.

Install

sudo dnf install ./re-code-0.1.4-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.4-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.3

Choose a tag to compare

@github-actions github-actions released this 08 Nov 06:49

Added

  • A crash writes a file next to the logs (~/.local/state/re-code/crash-<ms>.txt) with the message, the
    place in the code, the version and the system, and nothing from the conversation; its path is named
    once the terminal is restored. re-code doctor --report writes a bug-report bundle: the doctor report,
    the version, recent log lines with secrets hidden, and any crash files.
  • --version names the commit and the build date. Logs rotate daily and keep a week.
  • GitHub Actions: the gate (fmt, clippy, tests, licences) on every push, the rpm built on Fedora 44 and the
    deb built on Ubuntu 22.04 (glibc 2.35, so it installs on Debian 12 and Ubuntu 22.04 or newer); a v* tag
    builds both and publishes the release with checksums and the changelog section as notes.
  • A live cloud test (crates/providers/tests/live_cloud.rs, ignored; key from the environment) that chats,
    calls a tool and sends the result back on any OpenAI-compatible host or on Anthropic.

Fixed

  • Tool calls carry provider data back: Gemini's OpenAI-compatible endpoint attaches a thought_signature
    to each tool call and refuses the next request without it, which broke every tool round trip on Gemini.
    The adapter now keeps the signature with the call and returns it. Verified live on Gemini 3.8 Flash
    (chat, tool call, round trip); the first live confirmation of tool calling on a cloud provider.

Install

sudo dnf install ./re-code-0.1.3-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.3-1_amd64.deb    # Debian 12+, Ubuntu 22.04+

The deb is built on Ubuntu 22.04 (glibc 2.35). bubblewrap is recommended. SHA256SUMS covers both packages.

re-code 0.1.2

Choose a tag to compare

@ZeroDay-Labz ZeroDay-Labz released this 08 Nov 05:58

What's new

  • OpenCode import. o on the MCP or Providers tab, /import opencode, or re-code mcp import opencode reads ~/.config/opencode/opencode.json (and a trusted project's) and brings over its MCP servers and providers. Keys are never copied: {env:VAR} becomes a ${VAR} reference, a literal key is reported. Agents, modes, JavaScript plugins and instruction files are listed as not run here.
  • The command line does what the home screen does. re-code mcp list|add|remove|enable|disable|test|import, re-code skill list|search|install|update|remove|enable|disable, re-code plugin …|marketplace add|list|remove|browse, re-code memory list|show|forget, re-code trust [--yes], re-code untrust. Lists take --json; installs print the same preview as the interface and take --yes or --keep-off (exit 4 when a decision is needed without a terminal, 2 when refused).
  • Stdio MCP servers run in bubblewrap by default. The project is writable, the home folder is emptied except toolchains and package caches, secrets are hidden, network is on, and the server's environment is passed in with --setenv so a key reaches only that server. Each server has a Sandbox setting; the row says unsandboxed when it's off.
  • "Free" means free. A model that is $0 per token but charges per request (OpenRouter's music and image models) is no longer called free, Hugging Face's included-credit models aren't either, and models that don't answer in text aren't listed for chat. Paid models are grouped per provider, so what a Gemini key pays for isn't mixed with what an OpenRouter key pays for.
  • /web gained a Firecrawl preset (self-hosted by address, or cloud with a key). Saved chats can be deleted from the Home list (0.1.1 and later).

Install

sudo dnf install ./re-code-0.1.2-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.2-1_amd64.deb    # Debian, Ubuntu (glibc 2.39 or newer)

Built on Fedora 44. bubblewrap is recommended. SHA256SUMS covers both packages; details in CHANGELOG.md.

re-code 0.1.1

Choose a tag to compare

@ZeroDay-Labz ZeroDay-Labz released this 08 Nov 04:49

Everything in 0.1.0, plus /web: give the model web search through an MCP server of your choice. Your own SearXNG by address, DuckDuckGo and a page reader without any key, or Brave, Tavily and Exa with a key. The pick prefills the MCP form; your address or key stays in your own configuration.

Install

sudo dnf install ./re-code-0.1.1-1.x86_64.rpm   # Fedora
sudo apt install ./re-code_0.1.1-1_amd64.deb    # Debian, Ubuntu

Both packages were built on Fedora 44. The .deb links against that machine's glibc, so it needs a recent Debian or Ubuntu (glibc 2.39 or newer). bubblewrap is recommended: it sandboxes every command the agent runs.

Also in this line

  • An agent loop on local Ollama models: read, search, edit, write, shell in a bubblewrap sandbox, git, with approvals, per-turn undo, saved sessions (now deletable) and compaction.
  • MCP servers, including the ones set up in Claude Code, used live; the MCP registry; per-server tool filters.
  • Skills installed from the Hermes index, GitHub, skills.sh or a link, scanned and previewed before anything lands.
  • Plugins from the Hermes catalog and Claude Code marketplaces; what runs here and what doesn't is stated.
  • Permission modes, hooks, instruction files, memory the program keeps itself, context windows per model, background commands and a helper task tool.
  • Runtime discovery for Ollama, llama.cpp, vLLM and LM Studio.

SHA256SUMS covers both packages. Full detail is in CHANGELOG.md; screenshots are in the README.