Skip to content

v0.1.51

Choose a tag to compare

@github-actions github-actions released this 15 Nov 01:08
· 73 commits to main since this 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

  • A driver writes to its CLI through a wire of its own, with Claude Code's behind it by @jholhewres in #43
  • Orchestrators and sessions of any CLI message each other through devpit by @jholhewres in #44
  • Settings → Providers finds the six supported CLIs, signs one in from a terminal tab, and keeps a default model and effort per profile by @jholhewres in #45
  • Project agent: starts and learns on its own, a memory page, and Ask no longer stuck on "Reading…" by @jholhewres in #46
  • Codex chats run through codex app-server: streamed answers, commands and edits, approvals, Stop, steer and resume by @jholhewres in #48
  • Project agent: only the person's own settings, the repository's deny rules and instructions, and no messages to other sessions by @jholhewres in #47
  • OpenCode, Grok and Antigravity hold a chat through one ACP driver by @jholhewres in #49
  • An ACP prompt is asked as the command it runs, so Always never allows another command titled alike by @jholhewres in #50
  • Each CLI gets devpit's tools, the orchestrator's brief, its permission mode and resume, without devpit writing the person's config by @jholhewres in #51
  • Cursor holds a chat through the ACP driver, signed in with its own login by @jholhewres in #53
  • Project agent proposals: a count instead of a banner, plain cards, and the orchestrator decides on the person's word by @jholhewres in #52
  • An orchestrator runs on any provider: create one on a non-Claude profile, or switch an existing one by @jholhewres in #54
  • chore: v0.1.51 by @jholhewres in #55

Full Changelog: v0.1.50...v0.1.51