Changelog

All notable changes to Baton. Versions follow semver; the control protocol (baton protocol) and the event log format (BEP) carry their own version numbers.

#Unreleased

#0.3.9 — 2026-10-27

Fixes

• The event log no longer grows with Baton's own agent or ACP agents (Gemini, Qwen, …) either: their running commands' output is shown live and stored once, like Claude's and Codex's. Logs are compacted once more, and a command whose result doesn't repeat its output keeps its last output.
• baton -c "task" (or --resume with a task) could start a second lead next to the resumed one; the task now goes to the resumed lead. Whatever asks for a lead at the same moment gets the same one.
• --continue resumes the last run that had agents (not the empty run baton history starts).
• Compacting the log waits for a daemon that is still shutting down (on macOS and Linux it could lose the run's last events).
• ctrl+c twice quits only right after its "press again" warning: a ctrl+c that closed a menu, cleared what you typed or left a text box no longer counts.
• Windows: a client that stops reading (a suspended laptop, a stalled phone) is disconnected instead of making the daemon hold everything for it, and lagging clients skip live output.
• HTML exports and replays show agents that stopped in full again.
• A long-running daemon checks for updates every day, not just when it starts.
• /web in the app: the other addresses it lists carry the access key.
• A worktree that can't be checked says why in one line.

#0.3.8 — 2026-10-27

A new web view (baton web, /web)

• Redesigned from scratch: black, quiet, with a faint touch of depth (controls that look pressable, status lights, recessed code). A sidebar with the run, the autonomy switch and every agent on a desktop; tabs on a phone.
• Agents' work is readable: tool calls grouped and described in words (Run npm test, Edit src/app.ts) with their output and diffs a click away; thinking folded; Markdown with code blocks, tables and task lists; long messages shortened until you open them; worktree paths shortened to the part that matters.
• What agents send each other shows as cards: who to whom, done or failed, lines changed, acceptance checks, work waiting to be merged — without the instructions Baton adds for the agents.
• New views: Activity (the plan and the conversation between agents) and Changes (all unmerged work, with Merge). An inspector shows an agent's details, files and conflicts.
• The inbox lets you dismiss old problems; the tab title shows how many things need you.
• A command menu (Ctrl K), single-key navigation, @ and / suggestions and message history in the input.
• Live updates no longer redraw the page: what you've opened, selected or scrolled to stays put, and only the agents whose work changed are sent again.
• Names in any script work (the page used to stay on "Connecting…" for a name like Łukasz or 山田).
• Agent text: a heading followed by many spaces no longer hangs the page or an HTML export, and file names and URLs keep their underscores (__init__.py isn't bold any more).

#0.3.7 — 2026-10-27

The event log stays small (.baton/state.db)

• A running command's output was stored again with every new chunk, all of it each time: one long command could add hundreds of megabytes, and a busy project's log reached gigabytes. Output is now shown live (a few times a second) and stored once, with the command's result. Big raw tool results that repeat what the card shows aren't stored either.
• Logs from earlier versions are compacted once, the next time you start baton in the project (a 3.6 GB log became 170 MB). baton compact does it on demand.
• After updating, baton notices a daemon still running the older version and offers to restart it (--yes to just do it): the log gets compacted and the agents that were running come back.

Knows when there's a new version

• Once a day the daemon looks for a newer Baton release and says so in the orchestrator's feed, with the command that updates your install (baton update, npm or Homebrew). Leaving the app repeats it. Starting Baton never waits on this. BATON_NO_UPDATE_CHECK=1 turns it off.

Resume brings back what was running, quickly

• Resuming a run starts only the agents that were still running when it ended. Agents you killed, or that failed or finished earlier, stay gone, and an agent restarted under the same name comes back once. (It used to bring back every agent the run ever had, including every agent an earlier resume had brought back.)
• The run list (/resume, /history, baton history) opens right away: it no longer reads every past run in full. Resuming reads only the agents' lifecycle and latest events.
• baton --continue (-c) resumes the last run, baton --resume <run> a given one, and baton --resume opens the list to pick from. baton resume [run] does the same.

Several Baton windows at once

• Starting Baton while another one was busy could crash it ("Failed to start HTTP Client thread") when the machine was short on memory. Baton now uses much less memory (bounded tool output and history per agent, nothing kept for agents that are gone, slow clients can't make the daemon queue without limit), checks agents a few at a time instead of all at once, and only contacts a local Ollama / LM Studio when something is listening there.

Keys and copying

• ctrl+c twice quits and stops all agents, from anywhere in the app. /quit still leaves them running in the background.
• Selecting text with the mouse copies it to the clipboard when you let go.
• /web shows the whole link (and the other addresses it's reachable at), and c copies it.
• Copying on Windows keeps non-English characters intact.

#0.3.6 — 2026-10-11

Resume brings the conversation back

• Resuming a past run (/history, now also /resume) shows each agent's earlier conversation on its card, above an "earlier, from the resumed run" marker. Agents that support it (Claude, Codex, OpenCode, Gemini, …) also continue their own session, so they remember it too.
• The orchestrator is resumed as well: a lead that hasn't done anything yet makes way for the resumed one; a busy lead gets the earlier conversation shown above its own.

Select the orchestrator like any agent

• tab from the input (or ↑ from the top row of cards) selects the orchestrator: z zooms it, and the other card keys work on it too. esc goes back to typing.

npm

• Packages carry the repository link in the form npm expects (no more warnings on publish).

#0.3.5 — 2026-10-11

Pick models from a list

• Worker model (defaults.model): the model for workers on your default worker agent, in /settings → Agents. ←→ cycles through the models that agent offers, enter types any name.
• Lead model works the same way, and both show what the agent offers.
• /model <agent> suggests that agent's models as you type.
• The lists come from the agents themselves: Codex's model cache, opencode models, Claude's model names. An agent that can't say still takes any model you type.

#0.3.4 — 2026-10-11

Approvals with several Codex workers

• Every Codex worker numbered its approval requests the same way, so when two of them were waiting, the newer request replaced the older one: the inbox showed only the last worker's request (or none, once that one was answered) while the other worker sat at "Waiting for approval". Requests are now unique across agents, so the inbox lists every one and approves the one you select.

#0.3.3 — 2026-10-11

Phone & web view

• Closing the QR code screen and opening it again no longer fails with "port already in use". Reopening /web shows the running web view as it is (phone access stays on, same link), and esc says that it keeps running.
• When the web view does restart (switching phone access, HTTPS or Tailscale), it waits for the old port to free up, then picks a free one if it has to. A port you set yourself (remote.port) gets a clear message instead.

Releases

• npm publishing works without a token (trusted publishing), skips versions already up, and can be re-run for a released version (only: npm).

#0.3.2 — 2026-10-11

Windows: no more flashing console windows

• Baton's background process has no console, so every command it started (git, agent CLIs, checks) used to pop up a console window that closed right away, sometimes dozens in a row. Every process Baton starts is now hidden, and a test keeps it that way.
• The verify command before a merge, cmd: secrets and notification commands now run in Git Bash or PowerShell on Windows instead of sh, which isn't there.

#0.3.1 — 2026-10-11

Every team feature has its own switch

• [features] in the config: roles, skills, acceptance and notes. Switched off, the agents never hear of it: no tools, no text in their instructions, and the matching spawn_agent fields go away.
• Team notes are now opt-in (they write a file into your project). baton init asks.
• baton features [name on|off], /features [name on|off] in the app, and a Team features page in /settings.
• ui.markdown: turn Markdown rendering off to see agents' raw text (/settings → Layout).
• Your own playbooks can follow the switches with {{#notes}}…{{/notes}} and {{^notes}}…{{/notes}}.
• notes.enabled is replaced by features.notes.

#0.3.0 — 2026-10-11

Roles, skills and team memory

• Roles: tester, reviewer, debugger, refactorer, docs, security, researcher and frontend are built in, plus your own in .baton/roles/*.md (baton role new). A role brings instructions, an access level, a model size, an agent, an acceptance check and skills. /spawn tester, /roles.
• Worker specs: the lead's spawn_agent takes instructions, skills, files (claimed for the worker), acceptance and access. Its instructions now include a brief template with examples.
• Acceptance checks: Baton runs the command when the worker finishes. The work counts only when it passes, with two retries on failure, and the check goes through your approval rules.
• Read-only workers can't edit files, even in YOLO. Reviews now use the reviewer role.
• Skills (Agent Skills layout, from .baton/skills and .claude/skills) for every agent, not just Claude: a menu in their instructions and load_skill to read one. baton skill new, /skills.
• Team notes in .baton/notes.md: agents note and recall facts, new workers start with them, and review findings are kept. /notes, baton notes.
• Saved teams: [teams] plus /team <name> [task].
• pick_agent suggests a role. Cards show the worker's role, and r/o for read-only workers.
• Reviews prefer agents confirmed to be signed in.

Markdown everywhere

• Agent messages render as Markdown in the TUI (bold, italics, code, lists, task lists, quotes, tables, code blocks, links) instead of showing ** and backticks.
• The phone/web view and HTML exports render it too, with a renderer that escapes everything first, so agent text can't inject markup.

#0.2.0 — 2026-10-11

Windows

• Baton runs natively on Windows 10 and 11: irm https://k6w.github.io/baton/install.ps1 | iex, or npm i -g batonsh. CI tests it on Windows alongside macOS and Linux.
• The daemon listens on a named pipe. Processes are cleaned up with taskkill. npm-installed agent CLIs (codex.cmd, …) are found and started. The Baton agent runs commands with Git Bash, or PowerShell without it.
• Worktrees are byte-exact copies even with core.autocrlf (no more CRLF churn on merges).
• baton update replaces a running baton.exe safely. baton doctor checks Git for Windows.

Fixes

• The lead's pane no longer shows "approval pending" after you've approved.
• In the agent grid, Tab moves to the next agent and Esc goes back to the input (the hint said so; the keys didn't).
• Plain text symbols instead of emoji everywhere in the app (↯ auto, ▲ YOLO, ‖ paused, …).
• Inbox entries read "codex-1: Bash(…)".
• Relative paths are shown with "/" on every platform.

Docs

• A new README and website with full docs, a plain-words guide to how Baton works, a command-line reference, and a re-recorded demo.

#0.1.0 — first release

The first public version: one terminal for a team of coding agents.

Team

• An orchestrator (Claude Code, Codex, OpenCode or Baton's own agent) plans, spawns workers, routes messages and lands their work. Every agent gets its own git worktree and a 3-way merge queue that runs your checks first.
• Cross-vendor reviews (Codex reviews Claude and vice versa). Merges wait for an unresolved review unless forced, and a reviewer that fails is reported as "no review".
• Track record per agent and task kind (pick_agent, /scores), model routing tiers, best-of-N compare.
• Agents: Claude Code, Codex, OpenCode, every ACP agent (Gemini, Cursor, Copilot, Goose, Qwen, Kimi, Cline…), any prompt-taking CLI in a PTY, A2A agents, and Baton's own agent on any provider.

Control

• One autonomy dial for every agent (normal · ask · auto · yolo, shift+tab), deny rules that hold in yolo, project trust for anything executable in a repo's config.
• Plan approval and editing, diff review with inline comments, undo merge, budgets, a watchdog for stuck and looping agents. Acknowledgments and agent-to-agent ping-pong no longer burn turns.

Everywhere

• TUI with 5 layouts, 11 styles and 40+ themes, a guided welcome, lessons and recipes.
• Phone and web view with one-tap approvals (LAN, HTTPS, Tailscale), ntfy/Slack/webhook notifications, shared runs with presence.
• baton run for scripts and CI, baton mcp (Baton as an MCP server), baton a2a, GitHub issues → runs → pull requests, replay, HTML/Markdown export, OpenTelemetry, TypeScript and Python SDKs.

Safety

• Agents that aren't signed in, or whose plan ran out of quota, are refused up front with the reason, and the orchestrator is told which agents are ready.
• The Baton agent's commands run sandboxed: seatbelt on macOS; bubblewrap, firejail or Landlock on Linux; or a docker/podman container on any OS. It also has web_search and web_fetch.
• DNS-rebinding protection on every server Baton runs, and on outgoing fetches and webhooks.

Install

• One binary for macOS and Linux (x64 and arm64, plus musl), from install.sh, npm or Homebrew, verified against SHA256SUMS. baton update updates in place. A Python SDK is on PyPI.