Changelog

Newest first. Minga is in early access, and the app changes often; runner versions are named where a change needed one. Update a runner with minga update.

11 October 2026

App

• Reviews you pick from: a review is titled with its question (“Pick the welcome screen”), each output has Pick this and Not this, one note goes with your pick or your request, and Approve carries the name of your pick. Approve the round and Cancel the item sit under More.
• Cards say why: every card reads What, Why and If you approve, with the rule that sent it to you as one grey footnote (“Asked because brief edits always need you”).
• A chat that folds: turns where Lea only worked fold into “While you were away”, the steps inside a reply into “Worked”. Under each reply, a Saved line lists what the reply actually changed, Proposed the cards it opened, and Not saved any write that failed.
• Setup is a conversation: a new project needs a name and a repo, then opens in Chat with Lea’s first question. Home lists what is left to set up, and Connect Claude Code or Codex gives the commands to copy, including the repo import with your project’s id.
• Account in your avatar menu on every page: profile, sign-in and security, connected apps. The brief is edited on the Overview, and Where Lea runs replaces “where it thinks”.
• Five tabs, one question each: Overview, Chat, Plan, Team and Needs you. The Plan has three layouts (List, Canvas and Outputs), and each milestone group has a Later · Next · Now control. Team shows people, agents and workers by area.
• Needs you is one tab with one count, grouped by verb: Decide, Answer, Review, Allow, Do, Sort. The same count shows on the tab, the phone badge, the Overview and the page title.
• A new look: paper and ink, a serif for headlines, larger chat text, and one color for what waits on you.
• Act means act. At Act, then tell me, Lea starts new milestones and moves them into Now herself, tells you, and offers Undo. Ask first and Suggest in chat are unchanged.
• Lea hears approvals. When you approve a review, she hears about it in her next turn and plans what it unlocks.
• Names on chat messages. Everyone’s messages in the shared chat carry their author’s name.
• Agents ask each other. An agent can ask any other agent on the record, and an @mention wakes the agent it mentions. Four agent-to-agent wakes in a row without a person become a card.
• Pause keeps messages and wakes. A paused agent’s messages and wakes wait and arrive together when it is resumed.
• Review rounds that run out go back to the item’s owner agent instead of to you. Attempts that pushed or posted something no longer count toward the five-attempt limit; a limit of 20 attempts of any kind backs it up.
• The usual number of milestones in Now is a warning now, not a refusal.

10 October 2026

App

• The Overview, the project’s first tab: the brief, the milestones with what is done, in progress and missing, goals, Needs you, Next up, outputs to look at, Lea’s latest update, decisions, open questions, and what shipped.
• The brief: a five-section brief per project, with revisions. Lea’s edits to it are always a card with a diff.
• Milestones are releases with a due date, in Later, Next or Now. Moving one into Now is the go-ahead.
• Lea saves what you decide into the records it changes, and ends her reply with a Saved line of links. Decision comments record the why where no card does.
• Outputs you can see and judge: workers, Lea and you attach files (screenshots, reports, designs, video). Review cards show the outputs side by side, with Approve this one, ✕ and Comment. Rounds you send back are free.
• Autonomy levels for Lea: Suggest in chat, Ask first, or Act, then tell me. Lea hears when a person finishes, cancels, moves or edits something, and when a KPI number comes in.
• MCP from your own sessions: connect Claude Code or Codex to https://app.minga.run/mcp with OAuth, and read projects, capture ideas, record KPIs, answer what needs you and message Lea from your terminal.
• Import a repo’s state: the import_repo_state prompt drafts the brief, milestones, open questions and past decisions from an existing repo, and writes them after you approve.
• Goals, KPIs and items replace the earlier initiatives and tasks. Items have an owner and an assignee, checks for done, and a triage for captured ideas and bugs. Agents and workers can record KPI numbers, with evidence.
• Failing work stops after a bounded number of attempts and comes to you as one card. Items that publish something always end at a person’s approval.

Runner

• 0.6.1: Codex in Auto uses Codex’s own auto-review. Allow cards for Codex commands can remember: Allow for this run and Always allow a prefix.
• 0.6.0: workers upload the files they attach; runs receive the outputs passed to them as files.
• 0.5.1: runs use the stock Claude Code and Codex CLIs with nothing added around them: no extra path or git rules from Minga. The profile’s access mode and the CLIs’ own permission modes decide.
• 0.5.0: the runner finds the clones you already work in, from your Claude Code and Codex sessions, so most repos need no minga repo add.
• 0.4.0: the runner speaks the new item model.

9 October 2026 and earlier

• Settings: account, workspace, members with copy-a-link invites, runners and model providers; project settings for the brief, repos, areas, agents and temporary workers; where each agent thinks.
• Access modes per profile: Auto (the default), Ask or Full, set in the app or with minga profile set (runner 0.3.0).
• The chat and run transcripts: readable tool steps, cards in the chat, permission prompts next to the step that asked, and a live line for what is happening now (runner 0.3.1).
• Lea supervises every item: she closes, pauses, retries and reviews work; workers stop with a clear reason when Minga’s tools fail to load (runner 0.3.2).
• Runner console and signed self-update: runners report their profiles, models and repos, and install signed releases on their own (runner 0.2.0).
• First version: workspaces, projects with a lead agent, runner pairing with a device code, and workers that open pull requests.