Skip to content

v0.1.50

Choose a tag to compare

@github-actions github-actions released this 14 Nov 14:10
· 116 commits to main since this 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

  • Stops and drafts the person's words do not cover wait for one click instead of a refusal by @jholhewres in #5
  • v0.1.49 leftovers: Codex config read as TOML, tool files edited in place, paging and stop-guard fixes by @jholhewres in #6
  • A board reload redraws only the tiles whose card changed by @jholhewres in #7
  • [leveza L] Fix the medium and low findings left by the [leveza] cards by @jholhewres in #8
  • #315 follow-ups: a stop's click re-reads the loss, colliding names stop nothing, one press on the Remote and Telegram, Send all for drafts by @jholhewres in #10
  • Kept terminals keep their place in the page across a project switch by @jholhewres in #9
  • #286 leftovers: the sign-in code off other users' reach, R1's limits on the tailnet host, the relay on done, a capped command box by @jholhewres in #11
  • A shell at its prompt is not resized to be redrawn, and the switch harness ends what it starts by @jholhewres in #12
  • Settings: split General into Notifications, Remote, Sessions and Voice by @jholhewres in #13
  • Chat: an answer that moves while it streams, a Remote that follows every turn, steer confirmation and a stop that always stops by @jholhewres in #14
  • install.sh shows a banner, a progress bar and each step on a terminal by @jholhewres in #16
  • The island stays off on GNOME on Wayland unless asked for, so it no longer takes devpit's focus by @jholhewres in #15
  • Chat light render: only the live turn redraws, no typing tick, git off the turn path, a streaming ruler by @jholhewres in #17
  • Merged cards' worktrees, build caches and branches are removed on their own by @jholhewres in #18
  • Project agent v1: a per-project agent with a dated, sourced memory, asked by the orchestrator by @jholhewres in #19
  • Chat base visual: one type, radius and spacing scale, one chat stylesheet, text at AA by @jholhewres in #20
  • A long idle Claude conversation can be compacted when you return or just before its prompt cache expires by @jholhewres in #21
  • Project agent: approved lessons from closed cards, and a read-only agent process by @jholhewres in #22
  • Chat activity and background: activity filter, grouped subagents, background panel, full tool details, timed thoughts by @jholhewres in #23
  • Memory entries are re-checked when their files change, stale ones are proposed, and decisions name their cards by @jholhewres in #24
  • Two tests that flaked on macOS are deterministic, and touched_since reads 0 as the epoch by @jholhewres in #25
  • Chat message actions: hover bars, edit from here, fork, retry, refuse with a reason, turn line, quote, clickable paths, image zoom by @jholhewres in #26
  • Card branches are named {type}/{number}-{slug}, and the pattern is a setting by @jholhewres in #27
  • Chat states: usage-limit banner, screen-reader announcements, local pictures in answers, plan card, /btw by @jholhewres in #29
  • Schedules in the orchestrator sidebar, and an Assistant mode toggle with its brief and tools by @jholhewres in #30
  • devpit in a terminal shows a menu; subcommands and scripts get plain lines by @jholhewres in #33
  • Assistant mode: routines run, a heartbeat that speaks once, Telegram both ways and the comeback briefing by @jholhewres in #32
  • Conversations: sidebar rows, a chat header, conversation actions, search across projects, export and reading keys by @jholhewres in #28
  • Sessions consult their project agent while they work, within a budget, and propose what to remember by @jholhewres in #34
  • Every chat keeps its agent process between turns by @jholhewres in #35
  • Settings has Maintenance and Diagnostics: power saver, log and run retention, Labs flags, and devpit's processes, checks and error log by @jholhewres in #36
  • Each project's sidebar asks its project agent, and an answer becomes a fact only once approved by @jholhewres in #37
  • An orchestrator's guardrails follow its chat's permission, each set in Settings by @jholhewres in #38
  • Settings keep the chat's send key and follow-up, sizes, contrast, LaTeX and diff preferences, and keybindings say when they act by @jholhewres in #39
  • Providers: switchable models, advisor and subagent model per profile, check interval; model, permission mode and branch name per project or in .devpit/settings.json by @jholhewres in #40
  • Conversations can be archived and restored, Settings has an About page, commit messages are drafted in the repository's voice, and the chat header's details are chosen by @jholhewres in #42
  • Release prep v0.1.50: changelog, release notes and version bump by @jholhewres in #41

Full Changelog: v0.1.49...v0.1.50