Skip to content
davidpp.com

scopecraft-command, running Claude sessions against a real task system

A markdown task manager with an orchestration harness on top: parallel Claude Code sessions, one git worktree each, dispatched through tmux and watched from a dashboard.

Earlier project · 2025 · 4 min read

Project
scopecraft-command
When
2025
The question
Can parallel Claude Code sessions run against a real task system?
Written by
his AI agent · Claude

Where it landed

Stopped late 2025. The worktree-per-session model appeared in Claude Code itself about a year later.

[Wound down]

The build and the opinions are David's. The words are the agent's, drafted from his build records; he edits when he disagrees.

The itch

By May 2025 I was running Claude Code all day, and the sessions weren't connected to anything. My tasks lived in markdown files. The AI work happened somewhere else, one session at a time, in whatever directory I was in, and nothing recorded it. I wanted sessions that ran against a real task system, several at once without stepping on each other, and a way to watch them while they ran.

What I built

The base was a markdown task manager. Each task was a markdown file with YAML frontmatter and fixed sections (instruction, tasks, deliverable, log), so a task kept its context and its history in one place. The frontmatter tracked type, area, workflow and status. It shipped as a CLI and as an MCP server, so any AI tool that spoke MCP could read and write the same tasks. MCP was brand new then, and plugging a personal tool into it felt a bit early.

Scopecraft parent task detail view showing the rendered markdown instruction and vision on the left and a subtasks panel on the right with three sequenced subtasks, a Done/In Progress/To Do/Blocked legend, and Reorder and Make Parallel buttons
fig 01A parent task open in the tasks UI: the markdown task on the left, its sequenced subtasks on the right with Reorder and Make Parallel, and Start Claude Agent up top.

The part I still like is the parent task as a plan the harness could run. Its subtasks had a sequence (01, 02, 03), and two subtasks that could run at the same time shared a number, so the harness read them as a parallel group. The Make Parallel button in the screenshot only wrote that shared number. With the phases and workflow states already in the frontmatter, the same file was the spec and the plan the orchestrator ran.

fig 02A parent task scrolled end to end in the tool: the instruction and vision up top, then the plan drawn as an ASCII orchestration flow right in the markdown. The sequence-01 research tasks fan out to @research-agent in parallel, the double-bordered gates (single-bordered boxes are tasks; see the legend) hold each phase until a decision clears, and the deliverable and log close it out.

Everything stayed as readable text on purpose. Tasks were markdown, the AI's guidance was markdown, and decisions and open questions were sections in the task file you could check off. Markdown works well with git, and a model reads it as easily as I do. So the task file was the one place where I, the tool and the AI all left notes. Live state (which sessions were running, their streams) stayed outside git in ~/.scopecraft, local to the machine. The decisions and outcomes were written back into the task and committed.

What I cared about most was the orchestration harness on top. My rule there: the orchestrator coordinates and never implements. dispatch, auto and orchestrate didn't manage task data, run the model or create worktrees themselves. They called the services that did: a task service for context, channelcoder for the session, a git worktree for isolation. If the orchestrator starts owning task storage or session internals, everything depends on everything and nothing can be swapped. Keeping it a coordinator let me change how sessions ran without touching how tasks were stored.

The harness
coordinates, never implements
dispatch
Started Claude Code sessions in tmux windows, each in its own git worktree, so parallel sessions worked on separate branches of the same repo and never shared a working tree.
auto
Ran a task autonomously: prepare a worktree, gather the task's context, start the session, follow the stream, write progress back into the task file.
orchestrate
Read that parent-task plan and fanned its subtasks out across worktrees, walking a sequence in order and starting a parallel group at once.

Two other ideas shaped it:

The task type decides how the AI works

I didn't pick a mode by hand. A bug loaded a diagnostic approach, a spike an exploratory one, a parent an orchestration one, and the guidance was built from the task instead of hardcoded in the tool.

Autonomy comes in levels

Manual, guided, supervised, autonomous. The Autonomous Tasks Monitor showed it: running, waiting-for-feedback and completed lanes over a combined activity log. A session that hit real uncertainty was supposed to write down its assumption and ask, instead of stalling or guessing.

Scopecraft Autonomous Tasks Monitor screen with three empty lanes labelled Running, Waiting Feedback, and Completed above a Combined Activity Log panel, and a hint to start a session with ./implement-auto taskId
fig 03The Autonomous Tasks Monitor: running, waiting-for-feedback, and completed lanes over a combined activity log. Empty here because nothing was dispatched.

A lot of the orchestration layer stayed at prototype level.

What happened to it

The arc
  1. May 19, 2025The tmux dispatch prototyped: parallel Claude sessions, one git worktree each.
  2. Through 2025Kept iterating.
  3. Nov 2025I stop pushing commits. Other people sent the occasional patch after that, but for me it was done.
  4. About a year laterClaude Code's own Agent View ships the same model as a research preview.

My later tools reused its dashboard design and, more than the code, its ideas. The one that held up best: each session gets its own worktree, and the orchestrator coordinates without implementing.

I don't think anyone at Anthropic was reading my repo. One worktree per session is where you end up once you try to run more than one session at a time, and I needed it early.

David Paquet Pitts
Head of Delivery & Solutions at Botpress