devpit

Changelog

What changed in each release of devpit, newest first. Every version is on GitHub with its files.

Last updated

0.1.53 · release

Hosts: terminals and sessions on your other machines

  • Settings → Hosts keeps the machines you reach with ssh. A host is a target your own ssh resolves (an alias of your ssh config or a host name), with an optional user, port and identity file path, a work root and an environment: dev, staging or production. devpit stores no key and reads none, and Check says whether ssh, tmux, git and an agent CLI signed in are there.
  • A project is linked to a folder on a host, in its dialog. devpit checks the folder is there and never creates, clones or moves anything on a host.
  • New terminal on a host. From a pane's corner, a tab opens with its shell in the project's folder on that host, in tmux there. When the connection drops, the tab says so and comes back by itself; the session on the host goes on meanwhile. Nothing of devpit's is installed on the host.
  • The tab says where it runs. A badge names the host and its environment and whether devpit's connection to it is up. Production reads PROD, in red.
  • The orchestrator starts sessions on a host. devpit_start_session takes host; devpit_sessions and devpit_projects say host and environment. On a production host it starts only with confirm_production, after your yes, its brief opens with production's rules, and the start is written in the orchestrator's sessions log. A session on a host reaches none of devpit's tools: the orchestrator types into its terminal and reads its screen, and stopping it ends it on the host, or says the host did not answer.

Sessions

  • After a crash, agents do not come back by themselves. A terminal that comes back — devpit restarted, a tab drawn again, a project entered — is a shell and nothing more. Resume a conversation when you want it, from its row or with the CLI's own resume. This is a change: earlier versions started the agent again in the pane it had run in.
  • One conversation runs in one process. A conversation another process is running is refused by name instead of started twice, and its old pane forgets it.
  • A stop makes sure. Stopping a session places it in the pane its own process started in, closes only that pane, and checks the process ended: one that outlived its terminal is ended directly, with its group, and the stop says so; one that will not end is an error naming it, never "stopped". A background job is stopped through its CLI only.
  • A terminal the daemon no longer holds says why instead of staying blank.

Devflow

  • Devflow moves out of devpit. devpit runs the devflow program for the calls it serves — devflow.yaml's edits, the security targets in force and a skill's text — and keeps its built-in engine for the rest. Devflow's skills, personas and schemas now ship with devflow, not inside devpit.
  • devpit keeps devflow installed. With Devflow on, a missing devflow, or one older than 0.5.0, is installed, checksummed, and updated once a day; a program called devflow that does not answer as one is left alone. DEVFLOW_NO_UPDATE turns updates off, and an empty DEVPIT_DEVFLOW means no devflow at all.
  • Security targets reach checks, reviewers and gates only once you confirm them, a check's report can be SARIF, and a card shows the exploits its newest review confirmed.

Chat and voice

  • A chat is offered Claude Code's Artifact tools, the orchestrator's included, as a terminal on the same account is: it can read, publish and comment on artifacts. The read-only project agent is not offered them.
  • A question an agent asks waits for its chat, even when the chat opens late or the conversation moves, and leaves once it is answered.
  • Holding a key speaks to the chat in front, Settings → Voice chooses the microphone, and recording works on Linux again.

Opening

  • devpit --open-pane <leaf> opens that pane, on a first launch as on a second.

What's Changed

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.52...v0.1.53

0.1.52 · release

Project agent

  • The project agent reads a project only once you turn it on. Nothing reads a project on its own any more: not when devpit starts, not when a project is linked to an orchestrator, and not when Labs turns the project agent on. That Labs switch only makes the agent available, and each project stays off until you turn it on, under Settings → Maintenance → Labs or with Learn this project on its Project agent page. A project turned on that was never learned is read once, and several are read one at a time.
  • devpit asks before a project learns. When a project is linked to an orchestrator, the top bar asks "Turn on learning for this project?", with Turn on and Not now, and Ask the project on a project that is off asks the same in its pane. Nothing turns learning on without your answer: Not now leaves the project off and is not asked again, the question waits across restarts, and the agent's tools on a project that is off refuse, saying only you turn it on. None of them starts a read, and a card closed in a project that is off asks it for no lessons.
  • Facts read from files go straight into memory. A fact from the agent's own read that rests only on the project's files enters its memory at once, marked as accepted automatically and undone from its page. Decisions, lessons and other facts still wait for you, so the count beside the project shows only what needs your judgment.
  • Your earlier choices carry over once. On the first start of this version, a project already learned stays on, unless you had left it out, and every other project is off. The proposals still waiting from the first read are accepted once, as Approve all from onboarding would, which clears the counts.

What's Changed

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.51...v0.1.52

0.1.51 · release

Providers

  • Chat on Claude Code, Codex, OpenCode, Cursor, Grok and Antigravity. A profile of any of these six CLIs now holds a chat, not only a terminal. Answers stream in, thinking shows as a timed thought, and commands, file edits and MCP tools show as tool rows with their output. Stop ends the turn and keeps the agent running, and after an idle-out the conversation picks up where it was; an agent that cannot resume starts a new one and says so. Claude Code chats are unchanged.
  • Codex chats run on Codex's own app server. A message steered into a running turn reaches it, and a model or mode switched mid-way applies from the next turn. Two Codex profiles with their own CODEX_HOME run side by side without sharing anything.
  • Cursor signs in with its own login. A Cursor chat reuses the sign-in cursor-agent login keeps, and devpit never holds a Cursor credential. The agent ends when its chat does or when devpit quits.
  • The model picker lists what the agent offers. A profile that lists no models of its own offers the agent's, under the agent's own names. A model the agent refuses is said in the chat before the answer and shown as unavailable, with the agent's reason. A refused sign-in or session ends the turn with the agent's own message, and Antigravity below 1.5.0 is refused with "Update the Antigravity CLI".
  • Permission prompts follow the chat's mode on every CLI. Full access answers them, Accept edits also answers file changes, and the rest wait for you in the chat. A prompt names the exact command that will run, never the agent's summary or title for it, and "Always in this chat" covers that command only. A delete, a move, a fetch or a command the agent did not show asks every time. A Codex request for more reach than the call is never answered on its own, its prompt lists what it grants, and in Accept edits a Codex edit is answered alone only inside the conversation's folder. Plan puts the agent in its own read-only mode, and an agent that has none is refused Plan, saying why.
  • Cost and context read n/a when the agent does not report them. The chat header and each turn's line show n/a instead of $0 or 0%. Codex turns show how full the context is, and their cost reads n/a until Codex's models are priced.
  • Settings → Providers finds your CLIs. It looks for the six supported CLIs on your PATH and in their usual install folders. Each row says the CLI's version, whether it holds a chat or only a terminal, whether it is signed in, and the line that installs it when it is missing. Sign in types the CLI's own login into a terminal tab for you to finish; devpit never handles the credentials. Scan again looks anew, and the scan repeats at the check interval. Each profile keeps a default model and effort, and every CLI gets the model list.
  • Every CLI gets devpit's tools and the orchestrator's brief, and your own configuration is left alone. devpit's MCP tools reach each CLI from files devpit keeps in its own folder. An orchestrator's brief reaches every CLI through the instructions file it reads, and where no hook hands it your preferences, they lead your message. Terminals resume Codex, Gemini, OpenCode and Cursor sessions with each CLI's own flag.

Orchestrator

  • An orchestrator runs on any CLI that holds a chat. Create one on a Claude Code, Codex, OpenCode, Cursor, Grok or Antigravity profile; a terminal-only CLI is refused, saying so. Runs as… moves an orchestrator to another CLI and back: its folder, notes, linked projects and settings stay, its next conversation opens with a short handoff from its notes, and the sessions it reached in a devpit pane come along with their drafts. It starts sessions on any profile, each briefed the way its CLI takes a first message, and lists, messages, reads, drafts for and stops them by name. Guardrails decide the same on every CLI, and only your own message counts as your word. A message that waits two minutes for a session's prompt tells the orchestrator once.
  • Orchestrators and sessions on any CLI message each other. Through devpit itself, beside Claude Code's own messaging: an orchestrator writes to a session of a linked project, a session writes back, and an orchestrator can ask to be told once when a session goes idle or ends. A message to a terminal waits until its agent is idle at the prompt, then is typed under a heading that names the sender and says it is not you. A message to an orchestrator becomes a turn between its turns. Neither ever counts as your word: a session's ✅ approves no draft and no stop.

Project agent

  • The project agent starts and learns on its own. With Labs on, every project linked to an orchestrator gets its agent, which reads the project in the background, one project at a time and never while it is answering someone. That first read proposes facts and decisions, each resting on real files of the project. Each linked project has its own switch under Settings → Maintenance → Labs.
  • A page for the project's memory. "Project agent", beside "Ask the project" in the sidebar, shows Facts, Lessons, Decisions and Proposed, each entry with its day and its sources, which open in a tab. Approve, edit or discard each one, approve the first read's proposals at once, or have the agent read the project again.
  • Proposals wait on that page, and each says what it is. The top bar no longer shows memory proposals; the project's icon in the rail and its Project agent item show a count. Each proposal says whether it is a lesson, a fact or a decision (or an update or removal), who proposed it, one plain line and its sources, with the full text behind "details".
  • The orchestrator decides proposals only on your word. It can list, approve and discard them when your last message is approve, aprova or ✅, or discard, descarta or ❌, on its own. Another session's message, a tool's output, the agent's own words and a word sent from a channel never count, and one message decides only once.
  • Facts that rest only on files can be accepted on their own. A Labs switch per project, off by default. While it is on, a new fact whose sources are all files of the project goes straight into memory, marked as accepted automatically, and Undo puts it back to wait for you. Lessons, decisions and every other fact still wait.
  • The project agent keeps to your own settings. Since nobody asks it to start, it loads only your settings, so a repository's allow rules and hooks never apply to it. The repository's deny and ask rules still bind it, an ask refusing like a deny, and its CLAUDE.md, .claude/CLAUDE.md and AGENTS.md reach it in its first message. It cannot message or watch other sessions.

Fixes

  • Ask the project no longer sticks on "Reading…". An answer reaches its pane even when the pane was closed and opened again while it waited, an answer left running in the background no longer goes to whoever asks next, and a question nobody answers fails after 15 minutes with an error, the next one starting afresh.
  • Smaller fixes. The project agent's answers and briefs reach the orchestrator whole instead of cut. A session id that starts like an option is never passed to a CLI to resume.

What's Changed

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.50...v0.1.51

0.1.50 · release

Chat

  • Answers stream smoothly, even in a long chat. An answer moves as it arrives instead of standing still and jumping, and each new block fades in (at once, with reduced motion). Only the turn being written is drawn again, so the rest of the chat stays still, and a turn starts and ends without waiting on git. The chat uses one scale for its type, corners and spacing, and answers are drawn at full contrast.
  • Every chat keeps its agent running between your messages. A message starts its answer without starting the agent again, a model or mode picked mid-conversation applies to the agent already running, and Stop interrupts the answer the way it does in an orchestrator's chat. After ten minutes with nothing said the agent stops, and your next message resumes the conversation. If it ends on its own before answering, your message goes to a new one. While it runs it also hears your account's other sessions and its own background tasks, as an orchestrator does: what it says then appears in the chat, marked as not in answer to you. The supervised mode still starts the agent for each message.
  • Stop always stops, and a steer says when it was missed. A stop sent before the agent has started stops it as it starts, a stop the agent does not heed can be forced, and Force stop ends the turn's tools on every platform. A steer the agent never took in says so under it, with Resend.
  • Act on any message. Hover over a message to copy it, edit from there, fork the conversation or send it again. Edit from here can put back the files the later turns changed. Retry sits beside Continue when a turn failed, and a refused request can carry your reason to the agent. Select words in an answer to quote them, click a file:line to open it beside the chat, and click a picture to zoom it. Each turn ends with one quiet line: when, which model and effort, how long, how full the context was and what it cost.
  • Show as much of the work as you want. The Activity menu shows all of it, hides what succeeded, hides it all, or keeps only your messages and the answers (Zen). Subagents started together are one row with each one's state. Work left running sits behind one chip by the composer, with a Stop for each task and the wakeup the turn set. An opened tool row shows its whole input and output, with copy buttons, and names its MCP server. A thought says how long it took, and every control reached by the keyboard shows a focus ring.
  • Conversations are easier to find, keep and share. Sidebar rows say which agent runs, whether it is working, waiting on you or finished, and its branch, project and last message time, grouped as In progress and Settled. A header over the chat holds its title (click to rename), card, context meter, cost and export; right-click it to choose which details it shows. Conversations can be pinned, settled, snoozed, marked unread or archived, the command palette finds what was said in any project's conversations, and export copies a conversation as markdown, with or without its tool calls. Alt+↑ and Alt+↓ move between messages.
  • Clearer states. A turn refused by a usage limit or an outage shows a banner with Try again, and Resume when it resets when the time is known. Screen readers hear a turn start and end, and each message says who spoke. A picture an answer names on this machine is drawn in it. A plan is approved as written, edited or refused, and approving it leaves plan mode. /btw asks a side question while a turn runs, answered apart from the conversation.

Project agent & Assistant

  • Each project has its own agent. It reads the project, never changes it, and keeps a dated, sourced memory of facts, lessons and decisions in .devpit/project-agent/. The orchestrator asks it about the project and has it write the brief a card's session starts with. It wakes when asked and rests after ten minutes.
  • Ask the project from its sidebar. "Ask the project" opens a chat with the project's agent. Answers name the files they rest on, and an answer can be saved as a fact once you approve it.
  • Sessions consult it while they work. A session can ask its project's agent up to ten questions and propose what to remember. Nothing enters the memory until you approve it.
  • The memory learns from finished cards and stays current. A card that reaches done proposes lessons and decisions; you approve, edit or discard each, and approved lessons reach briefs and reviews. An entry whose files changed is checked again, and a stale one is proposed for update or removal. A card shows the decisions that name it.
  • Assistant mode for an orchestrator. A toggle in the orchestrator's chat. Schedules in its sidebar list every reminder and routine, and routines run at their time. A heartbeat (Settings → Assistant) looks for failing checks on open pull requests, sessions stuck on a question and blocked cards, and speaks once about what needs you. Notices reach Telegram, where a stuck session's question can be answered with its buttons. Back after 30 minutes away, the orchestrator briefs you on what ran, what merged and what waits for your decision.

Settings

  • Merged cards clean up after themselves. Once a card's branch has landed, devpit removes its worktree, build cache and local branch, a few minutes after start and every hour. A checkout with uncommitted work or a running process is kept and listed. Settings → Worktrees → Cleanup says what went and how much space it freed, and Delete now runs it.
  • Branch names you choose. A new card's branch is named {type}/{number}-{slug}, such as fix/351-flaky-tests, and the pattern is a setting in Settings → Worktrees.
  • Defaults per project. The default model, permission mode and branch name can be set for one project, or in the repository's .devpit/settings.json. Each row says where its value comes from. A repository may only start chats in Supervised or Plan.
  • More control over providers. A Claude profile lists its models, each with an on/off switch, and sets an advisor model and a subagent model. Sign-ins and MCP servers are checked every 1, 5, 15 or 30 minutes.
  • A long idle Claude conversation can be compacted. Pick it on a Claude profile in Settings → Providers: when you return, or just before its prompt cache expires, so the next turn does not read the whole context again uncached. Off by default.
  • Maintenance and Diagnostics. A power saver that slows devpit's background checks while you are away or on battery. Logs and run artifacts can be kept for 7 to 90 days. Labs turns the project agent and the canvas on or off. Diagnostics shows devpit's own processes with their CPU and memory, its checks with their fixes, and the newest entries of its error log.
  • Your preferences for chat and display. Send with Enter or Ctrl+Enter, and queue or steer while a turn runs. Text and code sizes, more contrast, LaTeX maths, and how diffs open. Keyboard shortcuts say when they act, and new ones go to a tab by number, reopen a closed tab, set a draft aside and send in the background.
  • An orchestrator's guardrails. Settings → Orchestrator → Guardrails sets, for each confirmation, whether it follows the chat's permission, always asks or never asks. A stop or a draft your words do not cover now waits above the composer for one click, Stop all or Send all, instead of being refused. A click stops only on the loss it showed, and a stop whose names read more than one way stops nothing.
  • Settings is easier to find your way in. General keeps only what is general; notifications, sessions and the rest have panes of their own, and search still finds every row.
  • devpit in a terminal shows a menu. Open devpit, see its status, update it or run its diagnostics from the keyboard. Subcommands and scripts get plain lines.
  • Archived conversations and an About page. An archived conversation leaves the sidebar, the palette and every other list, and Settings → Archive restores it. Settings → About shows the version and platform, and opens the privacy policy, the licence and the open-source notices.
  • Commit messages drafted in the repository's style. On an empty message, the Changes panel offers Draft a message: a subject for the staged changes, written the way the repository's history writes them. It only fills the field; nothing is committed on its own.

Fixes

  • devpit keeps the front on GNOME on Wayland. On a Wayland desktop without layers (GNOME, or one without gtk-layer-shell installed) the island can only be an ordinary window: it took focus from the window you were in, came first when you switched back to devpit, and clicking it left devpit behind the window you came from. There it is now off unless you turn on "Show the island anyway" under Settings → Notifications. Sessions waiting on you still reach you as notifications.
  • A chat's cost counts each turn once. With the agent kept running, each turn reported what the whole process had spent. It now counts only what it added, in every chat and in an orchestrator's.
  • Coming back to a project draws its terminals sooner. Switching projects moved the terminals of the project you left inside the window, and each move made the whole window work out its styles again. They now stay in place.
  • A terminal no longer ends by itself when you come back to its project. Each return redrew every terminal by resizing it, and a zsh waiting at its prompt now and then quit on one of those resizes, leaving the pane empty. A shell at its prompt is no longer resized to be redrawn: coming back already draws what it showed.
  • The board redraws only the cards that changed, so a busy board stays quick when it is read again.
  • A safer sign-in from the browser. "Open in browser" no longer puts a lasting token on the browser's command line, and on Linux the one-time code is accepted only over your own connection. devpit's local servers read a request within a deadline and serve at most 256 connections, and the terminal keeper's folder is yours alone.
  • devflow in other tools keeps your files as they were. Codex's config is read as TOML, so an entry written another way is found and updated in place. JSON files keep their order, indentation and line endings, and each file keeps one backup.
  • Smaller fixes. Scrolling back through a terminal no longer stops at a progress bar. An agent run as a script is named by its script. A command waiting to be allowed sits in a box of capped height. Two tests that failed now and then on macOS are deterministic, and one of them found git reading an early --since time as now.

Installer

  • The installer shows what it is doing. On a terminal, install.sh draws a banner, says which version it installs or updates from and to, shows a download bar and ticks off each step. Piped, in CI or with TERM=dumb it prints the same plain lines as before, and NO_COLOR drops the colour.

What's Changed

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.49...v0.1.50

0.1.49 · release

Fixes

  • A terminal no longer sticks near the top. If you had scrolled up and Claude Code redrew its screen, the view stayed at the top and new output did not bring it down. It now follows the end again. The mouse wheel scrolls a session's terminal opened over the orchestrator's chat, inside Claude Code too.
  • The orchestrator's preferences hold from the next message. Editing context/preferences.md by hand takes effect from your next message, without reopening the orchestrator.
  • The orchestrator stops a session only when you ask. Your last message has to be only that request, with the session's full name: stop fix-login or ✅ fix-login. It can also stop a session when another one already works on the same card.
  • Smaller fixes. An X11 program started in a devpit terminal opens its window, once devpit's terminal keeper has restarted. "Open in browser" works on macOS and Windows without Chromium. A browser tab hears about tabs opened in the window. The Devflow screen no longer offers to replace a newer devflow with an older one.

New

  • A session in the model you choose. The orchestrator can start a session in a given model, such as Opus, without changing the profile.
  • devflow in other tools. The Devflow screen adds the standalone devflow as an MCP server to Codex, Cursor, OpenCode and Gemini CLI, shows where it is added, and removes it.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.48...v0.1.49

0.1.48 · release

Sessions

  • devpit no longer starts sessions on its own; turn it on in Settings → Sessions. Opening the next card of an epic and handing a full session to a fresh one are both off by default. A session is no longer told to stop after one card, so it can take the next card the orchestrator gives it. When the handoff is on, the new session works in the same folder as the one it replaces, and the old one is stopped once the new one is up.

Fixes

  • Sessions in a devpit terminal are tied to it again. An agent devpit starts in a pane (from the orchestrator, a handoff or a card) shows "Open its terminal here" in the session list, "Send as you" works for its drafts, its questions and permission requests reach "Waiting on you", and the orchestrator can read its screen. 0.1.47 lost this for terminals in devpit's own keeper. Sessions started on 0.1.47 are found again on Linux by the pane they were started in; on macOS they stay unlinked until they are started again. If the keeper stops, only sessions started on 0.1.48 come back with their agent.
  • The orchestrator keeps your preferences in view. Its preferences reach it on every message, so they hold after a long chat, a compaction or a restart, and a preference it notes applies from the next message.
  • A message the orchestrator could not deliver says so. It shows as "not sent" with the reason, instead of a card that looked delivered, and a retry keeps one card. A session the orchestrator starts can be messaged as soon as the start answers, and Devflow's reviewers no longer show up as sessions.
  • A finished command keeps its output. An agent's spinner no longer fills the block with 50,000 lines, and a screen cleared during the command no longer erases what was printed before it.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.47...v0.1.48

0.1.47 · release

Chat

  • Long chats stay smooth while an answer streams. Only the part of the thread near the view is drawn, the answer being written redraws only its last block, and typing in the composer no longer redraws the chat. Older pages load before you reach them, and what you are reading stays in place when they arrive.
  • A turn reads in the order it happened. Prose and the work between it alternate as they came; once over, the work folds behind a line saying how long it took ("Worked for 1m 30s", "Interrupted after…"), and a failed turn stays open on its error. A reload keeps how a turn ended and how long it has been working. Pictures you sent show as thumbnails and other files as chips; an edit's diff can be copied, and one past 24 lines opens folded.
  • The agent's questions and requests are answered in the chat. A question shows its options as buttons: pick with 1–9 or the arrows, write your own answer, or tick several and confirm them together. A permission request leads with what it runs, the file it writes or the plan, with every line of a command shown; Enter allows and Escape refuses.
  • A queue that waits for you. Messages queued behind a running turn keep their files and skills, and survive closing the tab. Stop, or a turn that fails, holds the queue until you resume it. Mod+Enter steers the running turn with what you typed, Mod+Shift+Enter with the first queued message, and Up and Down in an empty composer bring back what you sent. A paste over 32 KB becomes a text attachment, and the draft is kept with its attachments.
  • Find, a rail of prompts, and more. Find searches every page of a chat and opens the fold a match is in. A rail lists your prompts down the side, a code block has a copy button, a stopped turn can be continued, and a tab whose turn ended while you were elsewhere is marked.

Terminals

  • Terminals outlive the window, with no tmux needed. New terminals, agents' included, run in devpit's own terminal keeper on macOS and Linux. Quitting, a crash or an update finds them running, with the agent's screen as it left it; a keeper that died brings each terminal back in its folder and starts its agent again. Terminals made in tmux stay there until you close them, and when the keeper cannot start, tmux takes the terminal as before. The terminal menu's "Open in tmux" still opens one from a tmux client, and copies the line that attaches to it.
  • Lighter terminals. Every terminal's output reaches the window on one connection. A terminal flooding output skips the lines that would only scroll out of its history, so a long build log draws about twice as fast and no longer freezes the window. A terminal out of sight stops reading at once and keeps its screen, so a project you come back to shows its terminals quickly. Whether an agent is working or waiting comes from its own hooks and the terminal keeper, so devpit no longer polls every pane and process to find out.
  • Closing a tab stops only what runs in it. A tmux server, a gpg agent or another daemon started from a terminal is no longer stopped with its tab.

Remote

  • Typing from another device echoes quickly. A terminal's output reaches the Remote as raw bytes, and is sent every 8 ms instead of waiting up to 40; over the tailnet the rest of an echo no longer waits for the device. Older viewers keep working as they did.
  • A new device is paired only once you allow it. A device that enters the pairing code, or scans the QR code, waits while Settings → Remote asks you to Allow or Refuse it; no answer in two minutes refuses it. The code and the QR code are good for ten minutes, once.
  • A device you forget is cut off at once, even while one of its terminals is printing.

Devflow

  • Devflow is optional, and off until you activate it. A project without it gets none of its tools, hooks or gates, and the Devflow screen leads with Activate Devflow. A project that already had it on stays on.
  • Devflow is its own project now. devpit runs it on the standalone devflow engine, and the Devflow screen installs the devflow plugin into the project's Claude Code profile and the standalone binary, checked against its release's checksums.
  • Fewer rounds to close a card. An approval closes the review, and a blocking finding that comes back asks you for a design instead of looping. The gate counts a proof written after the checks ran and reuses a green run at the same commit. The plan rubric is advice unless devflow.yaml opts into it, and release checks show in the setup.

Linux

  • Native Wayland on Linux. On a Wayland desktop the AppImage opens on Wayland instead of going through XWayland. It falls back to X11 when there is no compositor, when the desktop's GTK settings are known to crash it, and after a Wayland start that never loaded its window, for that version or until Wayland is chosen again. Settings → General → Display chooses Wayland or X11 for the next start, and DEVPIT_GDK_BACKEND overrides both for one.

Board and sessions

  • A big board opens a page at a time. A long lane draws its first 30 cards and the next as you scroll. Agents read the board 20 cards per lane, with the done lane counted rather than listed, and a card moved while they page through is read once.
  • One card per session. When a card is finished, the next card of its epic starts in a session of its own, and a card's session past 30% of its context window leaves a handoff and is relieved by a fresh one. Each card starts light instead of paying for every card before it.
  • devpit's tools answer agents briefly. The board, a card, evidence and reviews answer a compact summary, with the whole answer one argument away, and a slow call no longer holds an agent's other calls.
  • Rules for sessions can be per project. A preference an agent logs no longer lands among the rules every session is briefed with, and a rule can be kept for one project's sessions only.

Orchestrator

  • A ✅ sends the orchestrator's draft, in any language. Replying ✅ (or ✔, ☑, ✓) alone sends the one draft waiting, and beside a session's name sends that session's. It needs no word of English or Portuguese, and "envie", "send it" and the rest still work. As before, only your own message counts, never another session's or a tool's.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.46...v0.1.47

0.1.46 · release

New

  • An orchestrator's links follow you to another machine. With sync on, an orchestrator set up under the same name on two machines keeps one set of linked projects, matched by repository rather than by folder: linking or unlinking a project on one machine does the same on the other, a link to a repository the other machine does not have waits there until it is cloned, and a link to a folder without a git origin stays where it was made. Its name travels with it; the account it speaks as stays each machine's own choice.
  • Agent profiles follow you, without their secrets. With sync on, a profile made on one machine shows up on the others with its name, agent, program, models and the names of its variables; its arguments stay where you typed them. Tokens and keys never travel: Settings → Providers says what a profile needs on this computer. Nothing another machine sent runs without your yes: a profile that arrived waits to be taken, and a change elsewhere to its agent, program or base URL waits for you to use it or keep yours. A profile or a variable you remove stays removed everywhere.
  • Open devpit in your browser. The tray's "Open in browser" serves devpit's own page on this machine, signed in for this run only and under a name that is this install's own. Nothing listens beyond loopback.
  • Claude Code accounts, per project, with their quota. Settings → Accounts lists the Claude Code config folders you use and how much of each account's 5-hour and weekly limits is left. A project can name its account, and the sessions devpit starts there open in it. When an account is nearly out, a banner offers to start the next session in another; the sessions already running stay where they are. The account's token is only read on this machine and never reaches the window.

Faster

  • Long chats open and scroll without freezing. A chat opens on its last page, read from the end of the transcript. Long tool outputs arrive cut, and arrive whole when you open them. A streamed answer no longer copies the whole thread on every batch.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.45...v0.1.46

0.1.45 · release

Fixed

  • Devflow's reviewers run on the project's account. They used to start Claude Code on the machine's default account, so a personal project's review could run on a company account. They now run as the profile the project names (Devflow → Review), else as the session that asked, else as the orchestrators linked to the project, and the screen says which.
  • A chat turn that fails says so, and stops working. A turn refused (another one running, an update coming) or failed mid-way no longer leaves its answer "working" until a reload; what it said before failing is kept; the CLI's own reason (a session it cannot resume, say) shows in the answer instead of an empty "interrupted"; and the error shows at the end of the thread, not at its top.
  • Publishing a review no longer reports a false error, nor posts it twice. devpit read back the review's comments where GitHub lists them without their lines, so a comment that was there was called missing, and publishing again posted a second review. Every comment a review carried is now kept as posted, and publishing again only edits.
  • One repository is one project on every machine, however it was cloned. A clone over https with a user, a token or a port, or over ssh:// with a port, now matches its git@ twin; a project whose remote moved to another repository takes the new one when it is opened; and a token in an origin is no longer kept, nor sent with the project's synced settings, which now carry its address so another machine can clone it.

Better at what it already did

  • Switching projects is quick again. Opening a project read the git status of every checkout of every project, seconds on a machine with many; it now reads only the project in front, the only one whose status shows, and the orchestrator's boards are not read again for the same projects.
  • A board is lighter to read. It lists every card's runs without their output, which only the card itself shows.

New

  • A reviewer's model can be a tier. model: small runs a Claude reviewer on Sonnet and large on Opus; any other value is a model id, as before. Other CLIs keep their own default for a tier.

Removed

  • Devflow's autopilot is gone. It started each card's session as the default agent, not the project's or the orchestrator's profile, so work could run on another account. Card sessions are started by the person or the orchestrator, as before it existed. It kept nothing on disk, so there is nothing to clean up.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.44...v0.1.45

0.1.44 · release

Fixed

  • "database is locked" on live sessions is gone. The poll that lists agent sessions rewrote what it had already kept on every pass, taking the write lock each time; a poll that sees nothing new now writes nothing.
  • Fewer pauses of the window. A hover over the terminal reads only the rows around the pointer, and pausing or resuming from the tray no longer writes from the window's thread. When the page does stall, the report says what it was on and which commands were working.
  • A headless reviewer's answer is no longer thrown away for a field devpit fills in itself: the answer is signed before it is checked.
  • Severity is held to what a finding is. A missing test, a written rule or style is at most medium, a problem in test code at most low, and a search is not a proof of concept.
  • A newer devflow.yaml still reads. A field this devpit does not know is named and skipped instead of refusing the whole file.
  • An agent moving, starting or finishing a card is quick again. It reads the board without asking every agent CLI for its live sessions.
  • A setup's apply writes only what the person saw. An agent's apply that differs from the last preview comes back as a new preview.

Better at what it already did

  • Setup proposes only checks that can run here. A command whose program is missing — or a make target whose recipe calls one — is listed as unavailable with why, on the setup screen too, and is not recommended. Setup also finds modules in subfolders (unit-<folder>) and adds a tsc typecheck when TypeScript has none.
  • Checks with prerequisites and shared resources. requires: runs a probe first and leaves the check "not run" when it fails, instead of red. exclusive: makes checks that share a resource (a test database) wait for each other, across cards and worktrees.
  • Proof without a test runner. A criterion can be proved by a screenshot, an E2E run or a file, which the gate checks is there. A card that only gathers evidence or writes docs closes on its proofs without a code review.
  • Reviews that fit the change. review.scale runs one combined reviewer on a small change that touches nothing sensitive. After a fix, previous re-reviews only what changed and carries the rest over. review.verify has the session run the proof of concept of each critical and high, and records it as confirmed, refuted or not run.
  • Findings point at real lines. Each finding is anchored on a line of the diff; one on no line is dropped with why, and the card lists each file:line once across reviewers. The reviewers start with no MCP servers and a turn limit, and all read the same packet.
  • Decisions says where it stands. Every Devflow answer that asks Decisions says whether it is on, off, without a key or failing, and why; the Jev panel and the card say the same.
  • Plans across epics, merged and cut cards. A plan card can wait on a card of another epic (#42). Once the cards exist, a card can be merged into another or cut from the plan, releasing the cards waiting on it. unblock_on: gate lets them start as soon as the gate passes. An orchestrator can pass on the person's yes to a plan, quoted on the epic.
  • Runs stay out of the repository. A card's runs are kept in devpit's folder for the project and attached to the card. A setup copies no persona or schema into the repository, keeps the comments in your devflow.yaml, and evidence names the devpit and skills that made it.
  • Cards name what blocks them by number, and an agent can edit or delete a comment it wrote — never someone else's.

New

  • Devflow on demand. Wherever Devflow is on in Settings, a review can run without a card or a setup: on a GitHub pull request, a base..head range or a worktree. A pull request or range is read in a detached worktree in devpit's own folder, so your checkout never moves. A plan without an epic is a preview. devpit-agent devflow <stage> reads a stage's skill or calls its tool from a shell.
  • Publish a review on its pull request. A review of a pull request is drafted for it: what needs changing goes on the lines it concerns, one short comment per line, and suggestions go in one general comment. devpit_devflow_publish posts only the draft you said yes to, as a review that only comments, and edits those comments when published again instead of posting them twice. GitHub posts; GitLab and Bitbucket get the text to paste.
  • Review code already in production. mode: retro on a merged range answers "follow-up" instead of "request changes", drafts one follow-up ticket per medium or worse finding, and asks each defect for a read-only query of what it touched since the deploy.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.43...v0.1.44

0.1.43 · release

Fixed

  • "devpit is still working on …" should be rare now. Every question an agent asked made devpit run git once for every project you have, and each of those copied the whole app in memory to start. Under a build that took longer than the agent waited, so a comment or a card move looked lost and then landed. devpit now finds the project from the files git keeps, starts programs without copying itself, and an agent waits 30 s instead of 5 for an answer. Codex waits on devpit's long tools — checks, reviews, a new session — instead of giving up after a minute.
  • A card from another project stays out of reach. Reading the state of a card's checkout now checks that the card is on the asking project's board.
  • No more leftover processes from opening a folder, a file or the sign-in page.
  • Devflow's headless reviewer works with "verbose": true. With verbose on in Claude Code's settings, every review came back as "invalid JSON". When an answer is not findings, what the reviewer said is kept in the run and named in the reason.
  • A review reads the card's own work. A card started in a Devflow checkout keeps the commit it began at, and a review without a base diffs from there instead of from where the branch left main. The diff is the commits to HEAD, a submodule as its log, and work not yet committed in a section of its own.
  • The gate holds code nobody reviewed. A review counts for the commit it read. If the code changed after that commit, the gate asks for a new review, so a card that used to pass with a fix committed after its approval no longer does. Changes under .devpit/ do not count.
  • Devflow checks with no report can pass. A check set to report: none (a Makefile target, a linter) is judged by its exit code: 0 passes, 127 (command not found) is "not run", and anything else fails. Before, it was always inconclusive, so a gate on it could never pass.
  • go: not found is gone. When GNOME starts devpit (it sets SHLVL=0), devpit now asks your login shell for PATH. It also searches Go's usual folders and never keeps an AppImage's temporary mount on it.
  • Go and parameterized tests count. A Go check reads go test -json test by test, and setup recommends it for a Go module. A criterion's test matches every run of a parameterized test: PHPUnit and Pest data sets, pytest parameters and Go subtests.
  • Devflow without Decisions, and Decisions that fails, both work. The devflow-finding rubric now says what each severity means, so Decisions no longer refuses every call. devpit refuses a choice question without per-option criteria before sending it. When Decisions fails during a review, the findings are consolidated as if it were off, and the note says why. Calls Decisions never answered no longer count toward a rubric's accuracy.

Better at what it already did

  • Cards have numbers. Every card gets a number in its project, shown as #42 on the board and on the card. An agent can name a card by its number wherever it took the id, and an orchestrator can write project#42.
  • The board keeps up without working so hard. A burst of changes from an agent is one reload, not one per change. A board behind another tab or a hidden window reloads when you come back to it. Only the tile whose card changed is drawn again. Cost and session readers stop while the window is hidden.
  • Agents get a shorter, truer list of tools. A project session no longer sees the orchestrator's tools, and Decisions' tools appear once a key is set. That is about half the tokens on every turn. Descriptions now say what each tool really does and when it refuses. A card can be passed as cardId, card or epic, and a session as name or session. An error says what to do next.
  • Devflow asks with buttons. In Claude Code, the briefing's questions, the setup's decisions and the yes to a plan come through its question picker, with the recommended answer first. Other agents get the questions as text, as before.
  • Credits. The Devflow skills now name the MIT skills their method draws on, in flow/NOTICE.md.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.42...v0.1.43

0.1.42 · release

Fixed

  • An update goes straight to the newest release. With an update downloaded and waiting for its restart, devpit stopped looking, so two releases in a row took two updates. It keeps looking now, and a newer release takes the waiting one's place.

Better at what it already did

  • Buttons you can see. Every button has a face and an edge, the main action on a screen is filled with the accent, and the controls inside a row are smaller and quieter.
  • The Devflow screen, redone. What a setup found is a table — each command with where it came from, each agent CLI with its sign-in, long lists folded — and Recommended and Custom are two cards to choose from. The stages are numbered cards along the flow, checks and reviewers are tables, the severities that hold a card are one control, and a Custom setup walks a stepper with Back and Next always at the foot of the pane. Its text is a size larger.
  • Your way of reviewing, in every project. A reviewer persona is now read from the project, then the team layers it extends, then your own ~/.devpit/devflow/reviewers/, then devpit's. Name a persona of yours in your own devflow.yaml and it reviews all your projects; a company's persona in its layer reviews the company's.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.41...v0.1.42

0.1.41 · release

Fixed

  • The window no longer stops answering when tmux is busy. A tmux server too slow to answer used to hold one of devpit's threads per call until none was left, and then nothing reached the board. tmux calls now give up after 10 s, the window waits for one answer before asking again, and devpit writes devpit stall: to its log if anything stands still.
  • "database is locked" is gone again. Commits and closes no longer wait for the disk while holding the store's locks, so a slow disk under a build refuses nobody.
  • A file's tab is named after the file, however it was opened — not files.rs.

Better at what it already did

  • Keyboard shortcuts, all in one place. Settings → Keyboard lists every shortcut, lets you record new keys and refuses one already taken, naming which. ? or Ctrl+/ shows them all. Inside a terminal only the shortcuts meant for it fire; the rest go to the program. Splitting no longer splits every open tab, Ctrl+G no longer opens Claude Code's editor too, and Ctrl vs ⌘ follows the platform.
  • The tray icon stays and helps. Closing the window keeps devpit in the tray where there is one (off in Settings → General). The icon shows a dot and a count when something waits on you, and its menu opens each item, recent projects, a new terminal, pause and Settings.
  • Every control shows the pointer it acts with, maximize turns into restore, icon-only buttons have tooltips, and hover and focus are the same everywhere.
  • Sign-ins before a session starts. A footer chip shows each account's sign-in; a start from the window checks it and holds with the command that fixes it. The MCP check no longer slows a start.
  • Sessions. A session's row says whether Remote Control is on, with its link, and a card lists the sessions that worked on it, each resumable.
  • The orchestrator keeps one name across restarts, so sessions can always report back, and its chat answers /sessions, /waiting, /ended and /resume-session itself.
  • Channels escalate. A notice nobody acts on moves from push to Telegram to e-mail; the calendar of card dates is now served by devpit.app, titles off unless you switch them on.
  • Secrets can carry a rotation reminder, and the Remote can add one to a project's .env without the value passing anywhere else.

New

  • The Remote from any network. Without Tailscale, the Remote reaches this computer through devpit.app's relay, encrypted end to end: devpit.app splices bytes it cannot read. Tailscale stays as the other choice.
  • Chat on your phone. The Remote talks to the orchestrators and their sessions, with dictation transcribed by this computer and photos that land in the project.
  • Talk to the orchestrator from Telegram, in text or voice. Sending a draft, stopping a session or moving a card is a button, never a typed yes.
  • devpit's MCP from another computer, read-only: board, sessions and what waits on you, through the same encrypted relay. Off by default.
  • Sync between your computers: preferences and project settings, sealed with a key devpit.app never holds. Off by default; secrets never sync.
  • Devflow: a short engineering flow served by devpit's MCP — briefing, context, plan, execution, review and completion — with the plan on the board (epics, child cards, dependencies, acceptance criteria you tick), evidence kept on each card, a completion gate before a card is done, review by several agent CLIs at once, Jev judging in shadow first, and a Devflow screen in every project. Off until you switch it on, with a click.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.40...v0.1.41

0.1.40 · release

Channels

  • devpit tells you what waits on you while you are away from it. Settings → Channels sends a session waiting on a question, failing or finishing, a draft ready, a reminder, a step failing or devpit's MCP restarting to the channels you connect — never while devpit's window is in front or devpit is paused.
  • Telegram, through your own bot. Paste its token from @BotFather and send it the code devpit shows: messages go straight from this computer, with no server of ours in between. The token is kept on this computer, readable by you alone, and never shown again.
  • E-mail and push on your phone, through devpit.app. Switch on Send notices through devpit.app and choose at devpit.app/account: e-mail (after you confirm the address, with a one-click way out) and push on any device that installs devpit.app. Only that something waits leaves this computer, unless you switch titles on; devpit.app keeps nothing of it.
  • Buttons that act on this computer. Send the draft, Done and In 15 minutes for a reminder, and the first two choices of a session stopped on a question — picked only while it is still the question on its screen. A button works once, for half an hour, and this computer does it as you.
  • Which event goes where, quiet hours and one message for many. A matrix of events and channels, quiet hours by day that a reminder you set still breaks, and a short wait that turns five sessions finishing at once into one message.

Remote

  • Start work from your phone. A device allowed to type makes a card and starts a session, on a card or in the project, checked for the sign-ins the project needs like any other start. The board opens with what waits on you.
  • Your card dates as a calendar. Settings → General → Remote gives an address a calendar app on one of your tailnet's devices subscribes to; New address retires the old one.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.39...v0.1.40

0.1.39 · release

Sessions

  • The folders of the projects you add are trusted for Claude Code. Adding a project answers "Do you trust this folder?" in every Claude Code account devpit knows, for the project, each worktree it makes and every folder it starts a session in — a session the orchestrator starts no longer waits on that question. One that shows up anyway, in a session devpit started, is answered yes, unless the folder brings its own permissions. Trust the folders of the projects I add in Settings → Providers turns it off.
  • A session is not started without the sign-ins its project needs. The AWS profile a project's .env names, and whatever its needs.txt lists (aws <profile>, gcloud, gh), are checked first; the orchestrator is told the command that fixes a missing one, and starts anyway only once you say so.
  • A tool that fails for want of a sign-in rings the bell once, with the command: an expired AWS token, gcloud or gh signed out, an MCP server answering 401.
  • A new worktree gets the project's .env files when its setup declares nothing. Each setup command stops after ten minutes, and the card says what the setup did, or where it failed and how its output ended.

Orchestrator

  • "Envie" works from the Claude app too. The orchestrator sends a draft when your last message asks for it — a bare "envie", or the session named with what to do — said in devpit's chat or over Remote Control. devpit reads your message in the conversation's own log; another session's words or a tool's never count, and one message sends once.
  • It knows where the round stands after a compaction. devpit keeps context/now.md in its folder — sessions running, what waits on you, drafts, reminders, sessions ended lately — rewritten at each of your turns.
  • A card keeps a handoff note. A session leaves what it did, what is left, how to test it and the risks; the next session handed the card starts from it, in another account or after a restart.

Projects

  • Secrets go in the project's .env, never in a chat. Edit project → Secrets takes a name and a value from a password field; the file is yours alone and kept out of git in this clone. devpit keeps the name and the date, never the value. The chat says so when what you are about to send looks like a key.

Account and community

  • devpit Plus has a waiting list. Settings → Account, and Remote without Tailscale, link to devpit.app/pro: one click to join while signed in.
  • devpit has a community on Discord. Community in the account menu and at the foot of Settings opens it in your browser, and each release is announced there.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.38...v0.1.39

0.1.38 · release

Sessions

  • A command can be sent to a session. Send command… in the Sessions panel offers /remote-control, /compact, /clear, /model, /status and /cost, or one written by hand, and types it in the session's terminal as you once you confirm it.
  • Sessions that ended stay in reach. The Sessions panel lists them under Ended: when, whether you or the orchestrator stopped them, what each said last and files left uncommitted in its folder. Resume opens the same conversation in a new terminal of its project, in the folder it ran in.
  • Stopping a session that is working, or has work not committed, asks again and says why, with Stop anyway.

Orchestrator

  • "Send it" sends the draft. Said alone in the orchestrator's chat (also envie or manda), it sends the one draft waiting. The island shows a session's draft with Send too, and every draft sent is recorded in the orchestrator's sessions log: to whom, from where, and whether it was edited.
  • It reads and resumes sessions that ended, and a stop it asks for is refused while a session is working or has work not committed, unless the person said so.

devpit's own MCP

  • It is checked every 30 seconds and restarts by itself when it stops answering, without restarting the app or dropping the orchestrator's chat; the bell says so. A devpit chip in the footer shows how fast it answers, with Restart MCP.

Remote

  • A terminal's key bar has Claude Code's commands, each confirmed first; the orchestrator's drafts are sent from the waiting tab; devpit's MCP is restarted from there too.
  • Add to devpit.app puts this machine on devpit.app/remote. Only its address goes, in the link's fragment — no token.

Files

  • Video and audio play in a tab, with seeking: MP4, MOV, WebM, MP3, WAV, Ogg and FLAC. They are streamed from devpit, never sent whole.
  • Pictures past 8 MB open instead of being refused. A click switches between fitting the pane and the picture's own size, with its dimensions under it. A file devpit cannot show offers the system's app.

Settings

  • Voice tests the microphone in place: three seconds through the same microphone and engine a chat uses, and what was heard. With whisper.cpp, Download base model fetches one and chooses it.
  • Remote shows what paired devices did lately, never what they saw.
  • Decisions (new, off by default): typed decisions by Jev, with your own key, model, a daily cap and a switch per project. Agents ask through devpit_decide and devpit_rubric_run. Nothing is sent without a key.

Fixes

  • Moving a card into a lane that runs a step no longer fails with "database is locked" or takes over 30 seconds while agents are working.
  • A session restored after a restart continues its conversation on the default Claude account; it used to start afresh.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.37...v0.1.38

0.1.37 · release

Chat

  • A document opened from an answer is resized by its edge, and the chat narrows beside it instead of running on underneath. The chat's buttons stay over the chat, and the document's own over the document.

Fixes

  • A page in a browser pane no longer draws over the Manager, Settings or the command palette. It steps aside while one is open and comes back when it closes, as it does when its tab is not the one showing.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.36...v0.1.37

0.1.36 · release

Chat

  • Answers are written out as they arrive. Text appears word by word at a steady pace from the first token, with a caret while the turn runs, and says Thinking as soon as the agent starts to think. With reduced motion it still arrives as it is written, without the animation.
  • An answer to a message from another session is one paragraph again, not one line per piece it was streamed in — answers already saved that way read whole too.
  • A reply the orchestrator drafts is a card in its chat, with Send, Edit and Drop: you send it to the session's terminal as yours without leaving the chat. A session stopped on a question shows it in the same card. The orchestrator still sends nothing itself.

Sessions

  • What a session has spent is shown where you use it: in a terminal's corner, a chat's, each row of the Sessions panel and the island. Its tooltip has the last turn, what was spent since devpit first saw it, the tokens and the model. Dollars are API list prices; a click shows tokens instead.

Fixes

  • A session in a project added a moment ago is in an orchestrator's reach right after it is linked; it used to be refused for a few seconds.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.35...v0.1.36

0.1.35 · release

Chat

  • Answers stream as they are written. Text arrives word by word instead of all at once when the turn ends, and the line under a live turn says what it is at: the command or file it is on, Thinking or Writing.
  • A queued message can go now. Send now on a message waiting in the queue hands it to the turn that is running, without waiting for it to end.
  • Links in an answer open in your browser, http://localhost included. A link that cannot be opened says why instead of doing nothing.

Board

  • A lane's step can start from a recipe: Tests runs the command your project's own files name and reads its report; Review has an agent file findings on the card's change and sends a blocker back; Prove-It runs a bug's test on the card's base and on its checkout, and is not proven unless it failed first.
  • A review's findings sit on their lines in the card's diff, with how bad each one is, why, and whether it was made against older code. A count over the diff goes to the first one, and a finding can be dismissed.
  • Command steps read their test reports: a green run with passing tests is Passed, not Inconclusive.
  • Worktree preparation's shared variables reach steps and terminals in a card's checkout.

macOS

  • Agents are found when devpit is opened from Finder. devpit reads your login shell's PATH and the usual install folders (the native installer, npm, bun, nvm, asdf, mise, Homebrew), a command in Settings may be ~/… or absolute, and terminals keep UTF-8 without a locale.
  • Copy environment info, under the agents in Settings, gives what devpit sees of your machine for a bug report; when no agent is found, devpit says where it looked.

Fixes

  • devpit's MCP tools no longer time out after hours open. Every question an agent asked listed the status of every checkout of every project; it now answers in a fraction of the time, and says so in time when it is slow.
  • A reloaded window ends the terminal connections its previous page left.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.34...v0.1.35

0.1.34 · release

Orchestrator

  • The Sessions panel lists what runs. A linked project with nothing running no longer takes a group unless you ask for it. A project the orchestrator does not reach is linked from its group. A row names its card, stops its session from a button, and a search finds a session by name, project or card.

Settings

  • Every row is a card, and only the switch switches: clicking a row's words no longer turns it on. The long half of each row folds under How it works, and a search scrolls to the row it matched.
  • Remote is one card: its steps, the address with a Copy button, and chips for what each device may do. A paired device is named after its system rather than A device.
  • Voice looks for a whisper again from Settings, without reopening it.

Fixes

  • Closing the window ends devpit, and Quit devpit in the tray always does. devpit used to stay running behind the island with no window, and the tray could no longer open or quit it. Terminals keep running in tmux.
  • A session resumed after a restart is placed in its own terminal. It could be taken for another pane's: its screen could not be read, and stopping it closed the other session's terminal.
  • Error reports leave out what is not an error: a project that is not a git repository, a folder gone from the tree, a download the network cut off, a chat whose agent process ended, a sound device that would not start. A refused command is reported by its code and message, not as [object Object].

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.33...v0.1.34

0.1.33 · release

Settings

  • General is grouped: the island and being told, this desktop, chats, other devices, updates and privacy. Searching Settings for island, pause, voice, remote, shortcut and the rest finds it.
  • Pause has a row, with what it holds back and what it does not.
  • The island says how it is drawn on this desktop: above every window, a layer at the top, or, on GNOME, a window wherever the desktop puts it.
  • Ask before stopping a terminal can be turned back on after Don't ask again.
  • Remote says how to get in, step by step, what type and answer allow, and where its log is kept.
  • Voice messages say where the microphone is and how to install a whisper.
  • An orchestrator's first screen says where its rules for sessions live.

devpit.app

  • devpit lives at devpit.app. Sign-in and error reports go there, and the install line is curl -fsSL https://devpit.app/install.sh | sh. Installed copies that still ask the old address keep working.
  • The installer checks each release's signature by devpit's key, where minisign is installed, besides its checksum.

Fixes

  • A session that finished no longer shows as working. Claude Code's own background forks report a subagent ending after the turn has ended, and that put the session back to work on the island, its tab and its card.
  • The microphone and the island's sound work in the AppImage. It carries the GStreamer plugins it records and plays with; the system's are newer than its own GStreamer and were refused. The .deb recommends them.
  • The tray offers a pause again when one ends on its own, rather than Resume for a pause long over.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.32...v0.1.33

0.1.32 · release

Remote

  • This machine, from your phone or another computer, over Tailscale. Settings → General → Remote: the machine serves its own viewer to your tailnet. Terminals are watched, and typed into when you allow it. The board's cards move. Agents' questions are answered. Chats are read.
  • Nothing listens beyond the machine and your tailnet. The viewer answers on loopback, published by tailscale serve with HTTPS, or on the machine's tailnet address when the tailnet has no HTTPS. Never on every interface.
  • Two things to get in. You must be the machine's owner in the tailnet, and the device must be paired with a code (or QR) shown on the machine's screen. The code is good once, for two minutes.
  • What each device may do is set on the machine. Every device can watch. Typing and answering are allowed device by device. A device that only watches is attached read-only by tmux itself, and a phone never shrinks the terminal on the desk.
  • The desk says Watched by 1 device, and a click drops it. A log keeps who connected and what they did, never what was on screen.

Orchestrator

  • A session's window says where it stands. It shows the branch, how far it is ahead or behind, and whether it runs in its own checkout. A folded Changes panel holds the files, the commits and the diff. Its name is renamed in place (click or F2), through the CLI's own /rename.
  • It proposes projects, and you add them in one click. "Add ~/Workspace/x to the ASC group" becomes a card above the composer: Yes, Edit or Drop. The Boards panel adds a folder, names it, groups it and links it in one screen, and links a whole group at once.
  • It hears what its sessions did since its last message (one finished, one is waiting, one ended) at the start of your next one. It is not woken to say it.
  • Standing rules for every session it starts. Whatever its context/preferences.md keeps under Rules for sessions goes with every brief.

The card follows the work

  • A card moves to the lane for work in progress when its work starts, and to the one for checking when it is finished. The agent no longer has to remember. A lane says which role it plays from its menu.
  • A card says what its sessions changed in other repositories: Also changed 13 files in devpit-app, with which are still uncommitted.

Chats

  • Attach files with a button (Ctrl/Cmd+Shift+A), from the system's picker.
  • Voice messages. Click or hold the microphone. The words land in the composer to read and edit before sending. They come from a whisper on this machine (whisper.cpp or whisper), or from a transcription service only if you choose one and keep its key. Without an engine, the recording goes as a file.

The island

  • It shows the question a session asks, and answers it. Its choices become buttons, pressed in that terminal as you.
  • Answer in words. Reply to a session from the island without opening its terminal.
  • A question from another session waits as a badge, rather than taking the island from the one you are reading. A thin bar counts down the last seconds before it folds.
  • On Wayland too. It is a layer at the top of the screen on KDE, Hyprland and Sway (with gtk-layer-shell), and a window that never takes focus on GNOME.
  • It costs nothing while hidden: no timer, no frame, no sound context.

Desktop

  • One devpit at a time. Opening it again brings the open one forward.
  • In the tray, with Pause for an hour.
  • Open at login, and a shortcut that brings devpit forward from anywhere, both in Settings → General.
  • Pause notifications, the island and its questions, for a while or until you resume. Everything is still heard, so it is right when you come back.

Other agents

  • Any agent can report to devpit. devpit-agent hook --agent <name> says what it is doing, so it shows on the island and on its card like Claude Code (see docs/agents.md).
  • Gemini CLI reports from devpit's terminals with nothing to set up. Your own ~/.gemini/settings.json is never written, and your own hooks still run.
  • A safe way to change an agent's own settings file: a diff first, a backup, and no write over a file that changed in between.

Fixes

  • A PDF opens in its tab, page by page, rather than as a file too long to read.
  • A big tool report, such as a Read of a large file, reaches devpit cut down to what it reads, rather than being refused and leaving its step running.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.31...v0.1.32

0.1.31 · release

Reminders

  • Ask to be reminded, and be told at that time. A card's date can carry a time now; a day alone reminds that morning at nine. When it comes, a banner under the top bar says so — it has no close, only Later (15 minutes, an hour, tomorrow morning) and Done — with a system notification when devpit is behind, and a line in the bell. A focus does not hold it back: the time was your own choice.
  • The orchestrator, or any session, can set one when you ask — "remind me tomorrow at three to review the PR" — on a card with that date and time. devpit sets it off, not the agent, and a reminder only tells: it never starts anything.
  • One thread waits for the next date and wakes only when a date changes. Switch it off in Settings → General → Reminders.

Orchestrator

  • It drafts what you would say to a session, and you send it as yours. A "go on" or a change of plan said in its chat waits, editable, above the composer and in the Sessions panel; Send as you types it into that session's terminal as your own words. Nothing is sent without your click.
  • A session it starts in a new folder shows the trust question in Waiting on you before the session is even up, and is answered there.
  • Remote Control from the Sessions panel, for a session in one of devpit's terminals.
  • What a session is on — Edit invoice.ts, Run npm test — shows in the Sessions panel and on a card's sessions, not only on the island.
  • Every orchestrator reads the brief of the build that runs it, however it is opened, and the brief says which build wrote it.

Updates

  • The Mac updates itself. The app in Applications was taken for a build nobody installs over, and was never offered an install.
  • Update in Settings shows the update it starts: it follows the download and brings the update card forward, where it used to say "Working…" and then nothing.
  • Which devpit this is, at the left of the status strip — with a dot when a newer one is on its way. A click copies it for a bug report.

Fixes

  • The island shows what is open in devpit: closing a tab takes its session off at once, a session idle in an open tab stays, and a terminal shows one session rather than every one it ever ran.
  • A picture between 2 and 8 MB — a full-screen screenshot — opens in its tab rather than being refused as a file too long to read.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.30...v0.1.31

0.1.30 · release

The island

  • Every agent at a glance, above your other windows. A small capsule at the top of the screen shows what each session in devpit's terminals and chats is doing — the step it is on, like Edit invoice.ts or Run npm test — and opens all the way when one waits on you. Hover or click it for every session; open one for its steps and a preview of the change, the file or the command. It folds itself away when you move on.
  • Answer from it. A terminal session's permission question can be allowed, denied, allowed always (the rule the agent suggested) or handed back to the terminal; unanswered, the terminal asks as it always did, and coming back to devpit hands it back too. A chat's question is answered there as well.
  • Get to the session. Open terminal and Open chat bring devpit to the front on it. Drop a file on a session and its path goes into the agent's prompt. A session's pull request and the state of its checks show beside it.
  • devpit's own face, in a picture per mood, a few sounds of its own you can switch off, and a system notification when a session waits, fails or finishes while devpit is behind — held, like the bell, by a focus.
  • Drag it to the screen you work on; it stays there.
  • On by default; switch it off in Settings → General. It sees only what runs inside devpit. On Linux it needs X11 or XWayland, which the AppImage uses; on a Wayland session without it you get the notifications alone.

Orchestrator

  • It reads a session's replies, a project's repository state and keeps its own log through devpit, rather than through scripts of its own.
  • Its links and account are kept by devpit, out of reach of its own chat.
  • Its session tools reach only the projects linked to it, never type control characters into a terminal, give every session a name of its own, and wait long enough for a session to start — so a slow start no longer ends in two of the same session.
  • A message queued while it works goes in as soon as it has answered, not once every background task it started has finished.

Chat

  • Allow always, for the rest of the conversation: every file edit, the exact command, or the same tool — kept for that chat only.

Fixes

  • A card checked out again after its worktree was removed goes back on its branch, diffing from where its work began.
  • A reply to a session is never typed into a terminal whose shell is in front.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.29...v0.1.30

0.1.29 · release

Windows

  • Commands typed in a terminal run. The line was cleared with keys PowerShell does not bind, and they reached the command as invisible characters — ls and claude read as not found.
  • A new terminal opens in its project's folder, not your home — so an agent started there no longer asks to trust your home folder.
  • The chat finds your Claude Code profiles. Programs are found under their .exe or .cmd, on the PATH Windows keeps for you now.

Board

  • An agent moves its card with the work: into the column for work in progress when it starts, and on to the one for checking when it finishes — by your board's own column names, never into a column that runs a step.

Releases

  • Each release says what changed, from this changelog: on the releases page and in the update card's notes.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.28...v0.1.29

0.1.28 · release

Windows

  • devpit installs on Windows. A per-user installer on the releases page, with psmux — the tmux devpit's terminals run on there — inside it. It adds devpit to the Start menu and removes cleanly from Settings → Apps. It is not signed with a code-signing certificate, so SmartScreen asks the first time.
  • It updates itself like the Linux and macOS builds. On Windows an update closes the open terminals; the card says so before you choose.
  • Terminals run PowerShell with devpit's prompt marks, find what a new Windows Terminal tab finds, and nothing devpit starts flashes a console window.
  • Closing the last terminal ends the terminal server, and installing over devpit or removing it ends it first.
  • The agents offered are the ones actually installed.

Sessions

  • A background session can be watched in a terminal, and every session devpit starts runs in one you can see (0.1.26).

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.28

0.1.27 · release

Windows

  • devpit runs on Windows inside WSL. The README says how: WSL 2 with WSLg, and the same install line as on Linux.
  • On the way to a native build. The code now builds for Windows, and its terminals' layer is tested on every push against psmux, a tmux of Windows' own. Nothing is shipped for Windows yet.

Fixes

  • The search lists an agent by its command, without the variables in front of it — no more account directories or paths in the list.
  • On macOS, a chat that hangs is ended after its grace period, instead of being taken for already gone.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.27

0.1.26 · release

Sessions

  • Every session devpit starts runs in a terminal you can watch. A card's work handed on by an orchestrator, and a lane's session step, open in the card's terminal tab with your profile's own command — the same mode your own sessions run in. They used to start in the background, where nothing showed them and a permission they asked for waited on a question nobody could see.
  • A background session can be watched. In the Sessions panel, "Watch it in a terminal" attaches it in a tab of its project.

Fixes

  • A program devpit starts — an agent, a chat, the MCP check — finds what your own shell finds, even when devpit was opened from the desktop's menu. MCP servers installed in a folder your shell adds to PATH no longer read as failed.
  • The MCP list shows the servers that answer apart from the ones that do not, and refreshes from its header.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.26

0.1.25 · release

Git

  • Fetch, pull, push and sync from the Changes panel, with how far ahead and behind the branch is. Pull only fast-forwards: a branch that went two ways is left for a terminal, never half-merged by a button.
  • The panels follow a commit made anywhere — in a terminal, by an agent — instead of waiting for the window to lose and regain focus.
  • The branch menu scrolls, and filters once there are more than a few.
  • A deleted file shows its diff again.

Terminals

  • A link in a terminal opens in the browser.
  • A split stays where its pane was — in a card's worktree, with its branch shown at once.
  • Stopping a terminal ends everything it started, background jobs included; removing a project ends its terminals.
  • Drop files on a terminal to type their paths, on the file tree to copy them in, or on Artifacts to keep them.
  • A new project offers its agents on the empty screen.

Orchestrator

  • A stopped session stays stopped. A background session is stopped by its own CLI, which would otherwise start it again, and every session of that name goes with it.
  • The chat keeps its turns apart: a turn another session woke and the person's own no longer mix, land out of order or stop each other, and a chat with Remote Control on is no longer closed by another.
  • Orchestrators stay where they are in the rail when you pick one.

The window

  • The right panel reads as one: the tab in front says its name, Sessions counts what is waiting on you, and every view has the same heading and the same empty state. Artifacts can be added by hand.
  • A notice goes to the terminal the agent is waiting in.
  • The status strip shows the MCP servers — how many answer, and why the others do not.
  • A group of projects can show only the active ones.
  • A new worktree can be set up for you: files copied or linked from the main checkout, commands run, variables set — in Edit project.
  • An update shows as soon as it is found, and devpit asks every hour.
  • Less redrawing: the window no longer re-renders itself for every agent report.

Fixes

  • "database is locked" with many terminals open.
  • The board's Add card no longer sits under a long lane.
  • A card's branch keeps its accented letters (producao, not produ-o).
  • Sessions under a path with _ are found, and a profile switched off leaves every menu that offers one (thanks, @aronpc).

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.25

0.1.24 · release

Orchestrator

  • Stop a session. devpit_stop_session — and Stop in the Sessions panel — ends a session and closes the terminal it ran in, its tab too when it was the last pane. Only when you ask: work in flight is lost.
  • Start a session anywhere you linked. Asked to, the orchestrator starts Claude Code in a new terminal tab of any linked project, in its folder — no card and no worktree needed.
  • New terminals and chats from the Sessions panel. Every linked project is listed there, with or without a session, and opens a new terminal right over the chat or a new chat in the project.
  • A terminal you can size. The terminal opened over the chat resizes from its edges and corners, stays centred and remembers its size; it is drawn as one surface in the terminal's colour. Hiding it keeps it running; "Close terminal" ends it.

Performance

  • Idle MCP App hosts are let go after ten minutes instead of staying until devpit quits.
  • Listing sessions no longer asks git about every project every few seconds, and does nothing while the window is hidden.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.24

0.1.23 · release

Orchestrator

  • Sessions in one panel. The right panel opens on Sessions: grouped by project, the ones waiting on you first, each with how long it has been at it and what it was last asked. A session opens to its question, a reply typed into its terminal as you, and everything that passed between it and the orchestrator; sessions no longer running stay below with their history. The chat's corner counts them and opens the panel.
  • A session's terminal, right here. Open any session in a devpit terminal over the orchestrator's chat and work in it as you — answer what it asks, type a command — without leaving; its tab takes it back when you close it. "Go there" now opens its very tab.
  • Questions above the composer. A session stopped on a question shows right above where you type, with its choices and a way into its terminal.
  • Who woke it. A turn another session woke says which one, and quotes what it said.
  • Handed work as cards. Each message the orchestrator sends shows as a card: to whom, what was asked, and how that session stands now.
  • Memory and tools, always. The orchestrator's brief now has it search and save in whatever memory tools the account has, look facts up with the account's MCP servers, and keep a notice that adds nothing to one line.

Chat

  • Remote Control connects visibly. A chat whose process started with a message now shows its claude.ai page as soon as it arrives, instead of staying on "connects with the next message".

Under the hood

  • Tests no longer leave tmux servers running behind them.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.23

0.1.22 · release

Orchestrator

  • Projects are linked by hand. An orchestrator no longer reaches every project. Link the ones it works with from its Boards panel: only those show there, only their folders are open to its chat, and only they are in reach of its tools. An orchestrator made before starts with none linked.
  • Only its own account's sessions. The Sessions panel is back to the sessions of the account the orchestrator speaks as; 0.1.19 listed every account's.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.22

0.1.21 · release

Chat

  • Pages that come with MCP tools. Tools that bring an interactive page — a Jira issue, a Confluence page, a Datadog chart — show it in the chat, under the answer that called them, as they do on claude.ai. Each page runs on its own origin, sandboxed, reaching only the hosts its server declared and nothing of devpit's. When a page wants to run one of its server's tools, you are asked first: once, or always for that page.

devpit's own tools

  • The board, a card and the sessions as pages. In any host that shows MCP Apps, devpit's board opens its cards, a card can be commented on, moved and handed to a new session (in its own checkout or the project's folder), and the sessions show who is busy, idle or waiting on you. A waiting question is still answered only in devpit.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.21

0.1.20 · release

Orchestrator

  • Every board beside the chat. An orchestrator no longer has a board of its own. Its right panel opens on Boards: each project with its lanes and how many cards each holds, a lane opening to its cards, and a card opening on its project's board.

Projects

  • Artifacts. Files that belong to a project but not in its repository — specs, exports, reports — are kept in its own folder in the devpit workspace, safe from clones, new worktrees and cleans, and never committed. The right panel's Artifacts tab lists, opens and removes them; sessions keep and take them back with devpit's devpit_artifact_* tools.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.20

0.1.19 · release

Orchestrator

  • Waiting sessions are found again. 0.1.18 did not recognise the terminals real Claude Code sessions run in, so "Waiting on you" and Reply never showed for them. They do now.
  • Sessions of every account. The Sessions panel lists the Claude Code sessions of all your accounts, each marked with the account it runs under. Those of another account can be read and answered from the panel; only the orchestrator's own account can be messaged.
  • Notes, organised, without git. An orchestrator reads every project's folder, keeps what it learns in context/projects/, your preferences in context/preferences.md and decisions in decisions/. Its folder is no longer a git repository, and the Changes and History tabs are gone for it.
  • Asks before starting a session. Work goes to the session already running in a project; a new one is started only after asking, in the card's own checkout or the project's folder, as you choose.
  • Easier to add and tell apart. A new orchestrator is one click away in the rail, and each one gets a colour of its own.

Chat

  • Remote Control in any chat. A project's chat can be reached from claude.ai and the Claude app too — outside the Supervised mode, which runs a process per turn.
  • Links you can click. Web addresses and file paths in an answer are links, and web addresses in a terminal open in the browser. A file an answer names opens beside the chat, and expands into a tab.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.19

0.1.18 · release

Orchestrator

  • Answer a waiting session from the orchestrator. A session in one of devpit's terminals that stops on a question — a permission, a choice Claude asked for — shows at the top of the Sessions panel beside the chat, under "Waiting on you", with its question and choices. A click answers it in that session's own terminal, as you; Esc dismisses it. Nothing is pressed unless the screen still shows exactly the question you clicked on, so a late or double click never answers the next one. The orchestrator can read the question and recommend a choice, but only you answer it.
  • Remote Control in the orchestrator's own chat. The phone button on its chat now makes this same conversation — this chat, this process — reachable from claude.ai and the Claude app as devpit-<name>, and opens its page from the chat. What you write there arrives in this chat as it happens, marked as not from here. It replaces 0.1.17's terminal that resumed the conversation beside the chat. If the chat's process is not running yet, it connects with the next message; a restarted process connects again on its own.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.18

0.1.17 · release

Orchestrator

  • Continue an orchestrator remotely. A phone button on its chat continues the same conversation in a terminal with Remote Control on, named devpit-<name> in claude.ai and the Claude app — hooks and devpit's tools included. Close the terminal and write in the chat to carry on at the desk.

Fixes

  • A choice in a dialog or in Providers is a form field whose list opens under it, instead of the chat's chip menu floating over the dialog's buttons.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.17

0.1.16 · release

Fixes

  • No one machine's commands are built in. A second account used to be looked for by one person's command name; now an account is only ever a command its owner declares in Providers, and the process it runs is recognised as the CLI it is.
  • The default agent is one menu in Providers, not a row of buttons that grew with every profile.
  • An orchestrator runs as the account you pick. The picker drew no selection, so one could start as an account nobody meant. It is now a menu of the accounts configured in Providers, with a way to add another. The account lives in the orchestrator's own folder and Runs as… changes it later.
  • Orchestrators are a group of their own — Orchestrators, at the top of the rail, folding like any group, with a quiet + — and edit like a project: name, icon and colour, with a mark of their own by default. The project picker no longer lists them.
  • An orchestrator opens with its history begun: its folder is committed when it is made, and devpit's own files are kept out of it, instead of opening on a list of untracked files.
  • The project picker says "1 worktree", not "1 worktrees".

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.16

0.1.15 · release

Orchestrator

  • Orchestrators, at the top of the project rail. New orchestrator asks which account it runs as — by the command that starts it, a shell function included — and a name; an account can have several. Each gets a folder of its own under ~/.devpit/orchestrator/<account>/<name>/ — a git repository with a brief, docs/, artifacts/ and context/ — and a chat that opens on its last conversation. Right-click reveals its folder or removes it; the folder is kept. It is not listed among the projects.
  • An orchestrator keeps listening between your messages. Its process stays, so a session's reply — or the notice it asked for when one went idle — wakes it, and what it says appears in its chat, marked as not in answer to you. Stop interrupts the turn and leaves it listening. It opens in Accept edits; in the supervised mode it cannot listen.
  • An orchestrator hands work to sessions. Given a card, it starts a new Claude Code session of its account in the background, in the card's own checkout, linked to the card — on the board like any other — and named, so it can message it and hear when it is done. Only an orchestrator can, and only on a card.
  • Answer a session from the orchestrator. A session in a devpit terminal that is waiting on you can be answered from the Sessions panel: what you type goes into its terminal as yours. A message from the orchestrator approves nothing there — that is Claude Code's rule, and it stays.
  • It sees every project and every session of its account. Its agent can list the projects and their lanes, work on any project's board, and see the Claude Code sessions running now — by the name it messages them with, busy or idle, and the project and card each works in. The same list sits beside its chat. It still never moves a card into a lane that runs a step.
  • Its brief keeps up with devpit. devpit's instructions live in .devpit/orchestrator.md and are rewritten when devpit updates; CLAUDE.md imports them and is yours alone.
  • An agent started in a card's checkout reaches its own board again; the board's tools used to answer that no project contained it.

Chat

  • A chat you leave keeps going. Switch project or reload the window in the middle of an answer and, coming back, the answer so far is there and the rest keeps arriving — Stop included. It no longer says the app closed about a turn that was still running.
  • The mode you pick stays picked. Full access, Accept edits or Supervised survives switching tabs, reopening the conversation and restarting devpit, and a new conversation starts in the last mode picked on that account.
  • Unsent text is kept per conversation until you send or clear it.
  • The thread follows the answer as it grows — text streaming into place, a block opening — and stops when you scroll up to read, with a button back to the latest.
  • A wider thread: one 896px column for messages, the composer and replies.
  • A relative link in a card's chat opens the file in the card's own checkout.
  • A chat has the board's tools, as a terminal does, without the Claude Code plugin installed.
  • Sessions devpit starts hear the account's other sessions whatever mode either runs in, instead of holding each message until approved by hand.

Highlights

  • Error reports, opt-in. A new switch in Settings → General, off unless you turn it on. With it on, devpit keeps its own errors — panics, internal errors of its commands, background routines that fail, and errors the window did not catch — and sends them anonymously so bugs can be found and fixed:
    • what is kept is cleaned as it is written: no paths, no credentials or tokens, no names quoted by git or the shell, no text from your cards, prompts or terminals;
    • a report carries only the error and devpit's version, OS, architecture and package type — no account, even when signed in, and no install id;
    • reports go out only while devpit sits idle (nothing running, nobody at the window for five minutes), five at a time, and pause when you come back;
    • Settings shows the exact request the next report would send;
    • turning the switch off deletes everything kept, at once. Kept errors never grow past 200 entries or 256 KB and are dropped after fifteen days, on the server too.

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.15

0.1.14 · release

Highlights

  • The window never waits. Every command that touches the disk, git, tmux, the store or a process runs off the thread the window draws on. A guard in cargo xtask check keeps it that way.
  • Lighter rendering.
    • The shell stops re-rendering the whole app on every change: components read the slice they need.
    • Streaming chat applies its frames once per frame, and turns that are out of view are not laid out.
    • Terminals give back their GPU context after a while off screen, fit once a frame, and receive events through one shared listener.
  • Smaller and faster builds. The frontend bundle drops from 58 MB to 15 MB (source maps and an unused CJK font are left out). The diff editor, Settings, the palette and onboarding load on first use. The app and the updater now share one HTTP client. A release no longer rebuilds the binary to bundle it, and it skips the frontend tests when CI already passed on the same commit.
  • Menus. There is one menu component everywhere:
    • it works from the keyboard, and only one menu is open at a time;
    • it stays inside the window;
    • when an action is refused, a message says why. Changed files in the Git panel get a context menu, commits in History can be copied or opened on their forge, and F2 renames as the menus say it does.
  • Clickable paths. A path printed in a terminal opens in the app with a click: images, PDFs, logs and JSON, at the line when one follows the path. Relative paths resolve from the shell's current folder, and paths the terminal wrapped over several rows link whole.
  • Claude Code plugin. devpit's hooks and board tools can be installed as a Claude Code plugin from a terminal's footer. It works for sessions started with your own claude functions and aliases.

Fixes

  • Deleting or moving a symlink acts on the link, never on its target.
  • Deleting or moving can no longer reach the project root or .git.
  • Saves are atomic, and huge files are never read whole.
  • The profile-command reader only runs your own shell function or alias, and on macOS a reading that times out no longer leaves what it started running.
  • A profile list that no longer parses is not overwritten.
  • MCP secrets in URLs, headers and environment pairs are masked.
  • In Supervised mode, the chat asks before an agent runs a command or changes a file, instead of the change being refused. Reads are not asked about.
  • The file viewer refuses credential files: CLI tokens, settings.json, .env, keys, and devpit's own secrets.
  • The chat accepts pasted screenshots and dropped files, and its text can be selected and copied. Pictures it cannot keep, such as SVG or anything over 8 MB, are attached by path instead of being lost.
  • Board:
    • reordering a card within its lane no longer re-runs the lane's step;
    • a confirmed move while a run is going no longer starts a second run;
    • a card holds one running run at a time, enforced by the database;
    • a lane can be added after another one was deleted.
  • Files and search:
    • the tree's git status is never older than the request;
    • a file typed into while saving stays unsaved;
    • content search opens at the matching line;
    • deleted conversations leave the search index.
  • The window has rounded corners again on Linux.
  • Shortcuts read Ctrl on Linux.
  • Projects in the rail show a pointer cursor.
  • The command palette takes the keys typed right after Ctrl+K.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.13...v0.1.14

0.1.13 · release

Fixes

  • Terminal output no longer stalls on screen. The last part of a burst of output (for example, Claude Code's prompt after /compact, or a spinner's final state) was held back until the program printed something else, which made the terminal look frozen. Output is now flushed within one frame even when the program goes quiet.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.12...v0.1.13

0.1.12 · release

Highlights

Terminal blocks (Warp-style)

  • Each command in a terminal is now a block: command, folder, duration and exit status, with failed commands marked in red.
  • Block headers stay pinned while long output scrolls.
  • Block actions: copy command, copy output, copy both, run again, edit and run, filter output, and bookmark. Alt+↑/↓ jumps between bookmarks.
  • Output keeps its colours and progress-bar redraws, and URLs are clickable.
  • Blocks survive an app restart.
  • Full-screen programs (vim, Claude Code) take over the whole terminal. "Classic" switches back to the plain terminal at any time.

Command editor

  • Commands are typed in an editor below the blocks:
    • Chips show the folder, the git branch and the working-tree diff.
    • Multi-line input.
    • History suggestions appear as ghost text (Tab or → to accept).
    • ↑/↓ walks history, Ctrl+R searches it and Ctrl+L clears the blocks.
    • Tab completes paths.

Shell integration

  • bash, zsh and fish (new) report prompts, commands, exit codes and the current folder.
  • A nested bash/zsh keeps its commands as blocks.
  • An interactive ssh login carries the integration to a remote bash or zsh. Set DEVPIT_SSH_WRAP=0 to turn this off.

Agents

  • When an agent runs in a terminal, a bar shows its folder and branch, a button to open the project files, and a Compose editor (Ctrl+G) for writing the next message.
  • A command that takes 10s or more rings the bell if it finishes while the window is not focused.

Tabs and splits

  • A terminal tab can move into another tab's split: right-click the tab and choose "Move into split with…".
  • A pane can move out to its own tab: right-click the pane and choose "Move to New Tab".
  • No shell is restarted either way.

Providers and models

  • A profile is now just a name and the command you already use (for example claudin). If that command is a shell function, devpit reads the agent, arguments and variables it sets, so the profile also works in the chat and on the board.
  • The model picker shows one line per model: name, what it is for, and the alias on hover.

History panel

  • Commits are grouped by day, with a badge for the commit type (feat, fix, …).
  • Adds a filter, loads older commits as you scroll, and copies the SHA on click.

Project rail

  • Projects and groups are reordered by dragging, and a project can be dragged into another group.
  • Removed the top filter, which shifted the rail on hover.

Fixes

  • The window no longer freezes when the machine is busy: commands that wait on git, tmux or the shell no longer run on the UI thread.
  • The mouse wheel scrolls Claude Code's conversation again, instead of cycling its input history.
  • The default Claude Code sign-in stays listed when other profiles of the same CLI exist.
  • Enabling a provider takes effect immediately, and the chat picks up profile changes without reopening.
  • A terminal opened before the app started switches to blocks as soon as its shell is idle.

Full Changelog: https://github.com/jholhewres/devpit/compare/v0.1.11...v0.1.12

0.1.11 · release

0.1.10 · release

0.1.9 · release

0.1.8 · release

0.1.7 · release

0.1.6 · release

0.1.5 · release

0.1.4 · release

0.1.3 · release

0.1.2 · release

0.1.1 · release

0.1.0 · release

What's Changed

New Contributors

Full Changelog: https://github.com/jholhewres/devpit/commits/v0.1.0