Skip to content

DevLabs v4.4.1

Choose a tag to compare

@github-actions github-actions released this 02 Nov 01:59
· 31 commits to main since this release
Immutable release. Only release title and notes can be modified.

DevLabs v4.4.1

Highlights

  • Feature Manual in Help, with a 3D tour.
  • Bundled skills: SkillSpector, ponytail and a security audit.
  • Slack, Groq and webhook keys are now stored encrypted.
  • Queued messages can no longer answer approval prompts.
  • Stronger guards stop agents running code through DevLabs.

v4.4.1 — 2026-11-02

  • Security: your Slack, Groq, webhook and teammate keys are encrypted now. They used to sit as plain text in config.json, which every agent can read. They now live in the same encrypted store as your connector keys, existing ones move there on first start, and the app window only ever sees a placeholder for the Slack and Groq keys (saving Settings without retyping keeps them).

  • Fixed: a queued message could answer an approval prompt. When an agent stopped at a "Do you want to proceed?" prompt, DevLabs could mark it idle a few seconds later and type the next queued message, whose Enter picked "Yes". A prompt on screen now always holds delivery. Switching between agents no longer marks the one you left idle mid-turn either.

  • Fixed: a message could be resent every second and a half. If DevLabs couldn't file a sent message away (a locked file, a missing folder), it delivered it again on every pass. Messages are now filed first and delivered once.

  • The loop guard works. It never counted, so two agents replying to each other could ping-pong forever. DevLabs now counts the messages in each conversation itself and stops a runaway thread after 40.

  • Security: agents can't make DevLabs run their code or impersonate each other. DevLabs' own git no longer runs commands a repo's config names (core.fsmonitor, external diff tools, filters and hooks; only the standard Git LFS filter still runs, so filters like git-crypt don't apply in the Git tab or new worktrees, and a repo whose own git config includes other config files is refused; git 2.31 or newer is needed), git signature checks and submodule work trees are skipped too, one agent can no longer send hook events in another agent's name or a stopped agent's (each agent gets a secret token, and an event without a valid one is refused), a file an agent swaps for a special file can no longer freeze the app, and on a Mac, clicking a path an agent printed only ever reveals it in Finder (an .app folder used to launch).

  • Moving a floor to a new home folder stops the agents first, so nothing they write during the copy is lost or copied half-written.

  • Fewer, kinder nudges. The app and its background watchdog each announced new mail on their own, so an agent could be told twice. Each message is now announced once, a nudge never lands on top of a prompt you're halfway through typing, and checking for mail reads file names instead of opening every message.

  • Token and dollar caps survive a restart. Restarting DevLabs used to reset what each agent had used, handing it a fresh budget. The caps now count the whole session from the cost ledger.

  • Smaller fixes: roster backups now span hours instead of minutes, work handed to a stopped agent tells Ramon instead of vanishing, Ramon no longer gets stuck on "booting" when his engine fails to start, GitHub issues load on a Mac app opened from the Dock, a muted radio stays muted after a restart, a station that drops its stream shows offline instead of "Live", pasted screenshots older than three days are cleaned up, and links to other apps' URL schemes are refused.

  • New: Help → Feature Manual. A detailed, offline manual of every DevLabs feature: what it is, why you'd use it, step-by-step how to turn it on, and where the switch lives. It opens with a table comparing DevLabs to running an AI tool on its own, and a table of what each AI tool (Claude Code, Codex, Gemini, Grok and more) gets inside DevLabs. Searchable, and linked from the How to use guide.

  • See the difference, in 3D. The Feature Manual opens with an animation: switch between One AI tool (you feed a lone session by hand) and DevLabs (you ask once, Ramon hands tasks to twelve AI tools, agents message each other and save to shared memory, and finished files fly back to you), with live counters. It runs offline from the app, and is turned off when your system asks for reduced motion.

  • New page on the website: devlabs.atmt.dev/features, with the same animation and the whole feature manual. Features in the site menu now goes there.

  • Nicer scrollbars in the help guide, the Feature Manual and the website: thin and rounded, in the office colours.

  • Fixed: the Help menu showed "Security FAQ" with a gap on Windows. It now reads Security and FAQ.

  • For contributors: node tools/sync-site-features.cjs rebuilds site/features.html from resources/help/features.html (a test fails if they drift). The showcase scene lives in tools/site-bundle/src/showcase3d.js; npm run build there writes both the site scripts and the app's offline resources/help/showcase.js.

  • Security: agents can no longer plant code that DevLabs itself runs. Agents may write the shared hive folder, which used to hold the hook helpers their CLIs run and the hive's git history, so an agent could edit a helper or plant a git hook that DevLabs ran outside the sandbox on its next save. Both now live in a separate hive-internal folder next to the hive that agents cannot write, an existing hive's history moves there on first start, and DevLabs' own git saves ignore any git hooks.

  • Three skills come built in. Every Claude Code and Codex agent now starts with NVIDIA's skill-inspector (checks a skill is safe before you install it), the six ponytail skills (keep changes small) and Cloudflare's security-audit. No setup; they work offline and update with DevLabs.

  • Bundled skills now really reach Claude Code agents. They were copied into each agent's folder, but Claude Code only loads skills from a folder passed with --add-dir, so they never loaded. Codex agents now get them too, in their own isolated Codex home.

  • Security: refreshing an agent's skills can no longer write outside its folder. An agent could leave a link in its own skills folder so the next restart overwrote a file elsewhere on your computer. DevLabs now removes and rebuilds each bundled skill, and skips the copy when the folder leads through a link.

  • Security: the same fix for every file DevLabs writes for an agent. An agent could leave a link where DevLabs keeps its settings, MCP key file, inbox, outbox archive, provider config (Codex, Gemini, Pi, OpenCode, Crush), shared logs or helper scripts, so DevLabs would write, move or copy a key to a file elsewhere on your computer. DevLabs now removes such a link and writes a fresh file in a real folder instead.

  • For contributors: node tools/sync-third-party-skills.cjs re-downloads the pinned skills listed in tools/third-party-skills.json and checks every file against the pinned commit's git hash; --update moves each pin to the latest release, --check verifies the committed copies offline. Nothing downloaded is run. Review git diff resources/skills before committing.