Everything we shipped, dated

Production investigations

A failed deployment now opens an investigation automatically. An agent is dispatched with read-only access by default, reads logs and diagnostics output, writes a summary and a resolution, and can open a follow-up request carrying the fix — linked back to the incident that caused it.

Investigation access levels

Investigations run read-only unless you explicitly raise them to write access. The level is recorded on the investigation, so there is always a record of what an agent was permitted to do.

Design previews

Front-end tasks generate a shareable preview link. Stakeholders can see the built interface in a browser before anything is approved or merged — no deployment required.

Automatic rollback

When a health check fails after a deploy, your rollback command runs on its own and the target is restored to the last commit recorded as healthy. The rollback output is kept on the deployment run alongside everything else.

Health checks and diagnostics

Deployment targets now take a health check and a diagnostics command. A target with no health check cannot be marked healthy, which is what makes rollback trustworthy rather than optimistic.

Deployment phases and merge summaries

Deployment runs report their phase as they go — merge, deploy, health check, rollback — with a summary of every branch that went into the merge.

Three autonomy levels

Projects run at Manual, Supervised or Autonomous. The level controls whether planning starts on its own, whether approved work dispatches itself, and whether healthy work deploys without you. Review gates and escalations still reach a human at every level.

Task review gates

Any role can be marked as requiring human review. When it is, dependent tasks stop until you approve — regardless of the project's autonomy level.

Role contract editor

The Markdown contract governing each of the fourteen roles is now editable in the application. Tighten a role's remit, loosen it, or write your own.

Deployment targets over SSH

Connect a server with a host, port, username and key, then give it a deploy path and a deploy command. Keys are generated and stored encrypted; you install the public half.

Deployment target types

Targets can describe different shapes of deployment rather than assuming one, with a per-target repository path and default-target selection.

Multiple repositories per project

A project can connect several repositories, and a task is scoped to the one it needs. Each repository records when its connection was last verified.

Live run streaming

Watch a task work as it happens — the commands it runs and the output it gets back, streamed to the board rather than collected at the end.

Project chat with the Project Manager

Talk to the manager agent about a project. It has the project's context, and it can turn the conversation into a filed request.

Agents that ask

A specialist that hits genuine ambiguity now stops and asks a question instead of guessing. The task waits in an awaiting-input state until you answer.

Agent profiles and project teams

Assemble a per-project team from the fourteen roles. Each profile carries its own system prompt, rules, model and effort defaults, and its own icon.

Capability substitution

Work planned for a role that is not on the team is reassigned to whichever member can cover it, and the substitution is recorded on the task. A task never disappears for lack of a specialist.

Structured completion reports

Every task now ends in the same shape: summary, files modified, risks, assumptions, recommendations and blockers. Parsed, not just written — the manager reads the risk signal without reading the logs.

Requests, tasks and the board

The first version of the thing: requests you file, tasks the manager plans, and a board that shows both. Typed, prioritised and orderable from the start.

Isolated worktrees per task

Every task runs on its own branch in its own git worktree. Parallel work cannot collide, and an abandoned task is a deleted branch.

Task dependency graphs

Tasks declare what they depend on. Work that is not ready simply waits, and the graph is visible on the board rather than buried in a plan.

Multi-tenant ownership

Projects, repositories, agent accounts and deployment targets are owned, and the boundaries are enforced at the query level rather than in the interface.