Everythingyour agentsdo,inoneplace.
The rest of it is agents. You organize the work, dispatch it to them, and review what comes back. Local‑first, no account, MCP server built in.
Follow the work across multiple tasks and projects.
Keep the public session API stable and add regression coverage.
Start with the guide, not the guesswork.
Installing the app, getting the first tasks onto a board by hand or by asking an agent for them, and pointing Claude Code, Codex or any MCP‑aware client at the same records you see.
Plan, dispatch, review, measure.
Break work into reviewable tasks, or let an agent propose the breakdown before anything lands.
Planning proposals pause on the consequential questions instead of guessing.
Hand a scoped task to Claude Code or Codex with the repo, setup steps and env already attached.
The agent gets the context you configured, not a fresh prompt each session.
Work comes back as a handoff with comments and a diff: approve it or send it back with notes.
Clean runs can auto‑merge; anything gated waits for a human.
Keep live runs, handoffs, comments and review state beside the work they belong to.
Activity stays attached to projects and tasks instead of disappearing into terminal history.
The board itself: your columns, priorities and agent status, filtered by assignee, tag, priority or date.
File tree, editor and terminal for the linked directory, where Agent setup lives and the Agent Panel runs a live ACP conversation with Codex or Claude.
What an agent actually changed, grouped by location, one click from the task that caused it.
Needs attention, live agent work and your notification history: the one to leave open when several agents are working.
Every read and write above exists as an MCP tool your agent calls directly: the same tasks, comments and runs, with no screen in between.
Your states, your hooks, your prompts.
Define your own states, then attach a shell command to any transition.
Work changing hands is an event. Worktrees, tests, draft PRs and notifications can all hang off it.
Saved prompts, scoped to one task or the whole project.
Reproduce the bug described on this card as a test, then stop and show me the diff before fixing.
Scan the backlog for tasks estimated above one day and propose a breakdown for each. Wait for approval.
Diff docs/architecture.md against what shipped this milestone and propose the edits.
Execute them from the board. The repo, the card and your setup steps come along automatically.
Every move is a record.
Comments, activity events and handoffs are first‑class objects on the task, each attributable to the human or the agent behind it.
Agent reasoning lands on the card, not in a terminal you closed.
A request for review is a record with a state, not a nudge in chat.
Every status change is attributable to the human or agent that made it.
Scope limited to src/auth/*. Keep the public session API stable.
Extracted session validation into requireSession(). Touched 4 files, tests pass locally. Flagging the token refresh path for human review.
Awaiting review on the token refresh path before merge.