Skip to content

Task Tracker

Krasnikov Maksim edited this page Oct 22, 2026 · 1 revision

Task tracker: Plane and decomposition discipline

The problem: one task unfolds into fifteen, those into a hundred; a week goes into writing them out, boundaries blur, tasks are forgotten, and real time including manual testnet checks stretches several-fold. A tool does not fix this by itself; rules do, and the tool and the agent help keep them.

12.1 Why Plane

  • Open source (Community Edition, self-hosted), Jira/Linear-style UI: projects, cycles (sprints), modules, task hierarchy, views, pages.
  • Role-based access: members are added to the workspace and to specific projects with roles (admin, member, guest). This covers "show one person the tasks of one project out of fifteen".
  • Plane's stack (API, worker, Postgres, Redis, MinIO) maps onto the Maurice catalogue as an ordinary set of services — it becomes part of the platform's starter kit alongside the registry, Caddy and the notifier.
  • REST API and API keys for automation; for the agent, a tracker driver via Maurice MCP. Whether Plane ships an MCP server is to be checked during implementation; otherwise the driver is written over REST.
  • Jira stays as an external driver for teams that already have it (via the Atlassian MCP); Linear likewise.

12.2 Link to Maurice

  • One Plane workspace per Maurice instance; a Maurice workspace = a Plane project. Creating a workspace in Maurice creates the Plane project, base statuses, labels and cycles; deleting a workspace archives the project (never deletes).
  • Inviting a person to a Maurice workspace as a task viewer adds them to the Plane project as a guest; they log in to Plane with their own account and see only their project.
  • Widgets in the Maurice workspace card: tasks in progress, opened/closed in 7 days, current cycle, backlog growth. Click leads into Plane.
  • Deployments and tasks are linked: a deploy from a branch or commit carrying a task number moves it to in review and attaches the deployment link.
  • Session history (Entire, see Agent Memory) feeds the agent: from the transcript it proposes closing what was done and filing what was found.

12.3 Agent as task author

  • A person says (in chat or by voice in Telegram) what they found; the agent files the task, links it to an epic, estimates, sets status. The person confirms a list instead of typing.
  • The agent can: create an epic with its first layer of tasks, create and fill a cycle, move statuses on events (PR, deploy), tag tasks discovered along the way as found during X.
  • Rights: in dev — create and move freely; delete and change estimates — only with confirmation.

12.4 Decomposition rules (built into the agent, surfaced in the UI)

Rule Essence How Maurice enforces it
Plan in waves Only the next layer is detailed: epic → 5–7 tasks; only the task being picked up now is spelled out. The rest is a title and one line The agent refuses to break down subtasks of tasks not in next / in progress and suggests postponing
At most three levels Epic → task → subtask. A subtask that needs its own subtasks is really a task; the epic was cut wrong Structure validator: warning and a re-slicing suggestion
WIP limit No more than 2–3 tasks in progress per person. New work goes to the backlog tagged found during X Widget turns red when exceeded; the agent will not move a fourth task to in progress without confirmation
Time-box the unknown Research ("poke at the testnet", "size up the contract") is a spike with a time limit and a result of "decision or list of questions" Spike task type with a mandatory deadline; Telegram reminder on expiry
Definition of done Every task has a done criterion verifiable without the author (test, manual check, dev deploy) The agent will not close a task without a filled-in criterion
Overheating signal If more tasks are opened than closed in a week, the epic gets cut or part is postponed Weekly Telegram digest: opened/closed, fastest-growing epics

Stage: Beta, with the dev workspace and MCP. The minimum to start: Plane in the catalogue, auto-project per workspace, the in progress widget and the agent tool create task. The remaining rules are switched on one at a time as usage grows.

Clone this wiki locally