arbeitsplan

Decides whether a compiled arbeitsplan run executes in-session, as a headless matrix of fresh claude -p processes, or inside the Workflow tool, and emits the checked backend object workflow.json requires. Use when the user asks whether to run something as a workflow or a matrix, whether a job should run in parallel agents or in this session, or why a run had to be forced into one. Recommend-only. For which pattern fits, hand off to arbeitsplan-patterns; for which of the four approved workflows applies, to arbeitsplan-start.

Turns a stated problem into an executable, budgeted workflow spec plus the takt beats that enforce its order. Use when the user describes a change and wants it run as an agentic workflow, asks to plan or set up a swarm, says the advice they have is too vague to act on, or asks what shape a piece of work should be run in. Refuses and points at zirkel when the problem turns out to be a question rather than a change. Writes the run's workflow.json under analysis/arbeitsplan/ and, behind an approval gate, .claude/takt.local.md.

Compiles and runs a headless sweep — cases x models x plugin-states x repeats, one fresh `claude -p` process per cell — after proving the environment can authenticate. Use when the user wants genuinely independent candidates, a real per-candidate --model, a plugin-present versus plugin-absent comparison, or asks to benchmark or ablate a plugin. Also use afterwards to read results back. A cell that could not be fairly measured is UNMEASURED, never FAIL.

Says which agentic pattern fits a problem and with what numbers — fan-out width, model tier, stop rule, cost — grounded in a frozen catalog of measured configurations. Use when the user asks what shape work should be run in, how many agents to use, whether to run something in parallel, or whether a named pattern is a good idea. Also use to explain why a pattern is rejected here. Read-only; it proposes, it never compiles.

Read-only inventory of what a swarm run can and cannot do in this repository before any tokens are spent. Use before arbeitsplan-compile or arbeitsplan-run, when the user asks whether this repo is ready for a swarm, or when a run failed and the cause might be the environment rather than the work. Reports git/worktree support, test commands, other plugins' live guards, and any open run lock. Writes nothing.

Executes a compiled workflow as a redundant swarm — N candidates over the same scope in their own git worktrees, judged blind, exactly one landed. Use after arbeitsplan-compile, or when the user asks to run the workflow, start the swarm, or build several versions and pick one. Halts and surfaces rather than retrying when candidates fail; convergence comes from widening, never from repeating. Writes only inside candidate worktrees until the landing step.

Names which of four approved workflows fits a stated task, the plugins to install for it, and the permission mode to run it in — and says plainly when none of the four fits. Use when the user asks where to start, which workflow or plugin combination applies, whether to use plan mode or auto mode for something, what to install before beginning, or how to combine werkstoff's plugins for a build, an unfamiliar-repo read, a bug fix, or a UI/plugin review. Recommend-only — it never installs anything, writes workflow.json, or changes permission mode itself.

Read-only report on an arbeitsplan run that has already started — phases done, candidates built and refuted, referee verdicts, open doubts, halts with their reasons, dispatch budget used, whether a lock is open, and the single next command. Use when the user asks how the run is going, what the swarm decided, why it stopped, how to resume it, why something was denied, or what is still holding a lock. Advances nothing and writes nothing.