| name | cowork-sprint |
| description | Plan and execute multi-feature roadmaps as sequential, concurrent, or mixed sprint clusters. Trigger on sprint planning, run sprints, plan then build, /cowork-sprint, or requests to break an initiative into sprints. Do not use for trivial edits, one-shot fixes, or one feature (use pdca-wf). |
| argument-hint | [goal / feature set / plan-file(s)] [--auto-plan] |
cowork-sprint
Deliver one multi-feature roadmap through the same observable contract on Claude Code and Codex.
Required method
Read ../../shared/references/cowork-method.md completely. It owns roadmap planning, sprint phases, core and risk gates, cluster rules, bounded convergence, the autonomous run-to-completion contract, safety, outputs, and Done.
Then select the runtime from available capabilities:
update_plan + collaboration agents, without Workflow/TodoWrite
-> read references/runtime-codex.md
Workflow + TodoWrite, without update_plan
-> read references/runtime-claude-code.md
both capability sets
-> stop: ambiguous host capability surface
neither capability set
-> stop: unsupported host
part of a set only (e.g. Workflow present, TodoWrite absent)
-> use that host's reference, and SAY SO: name the missing mechanism and what replaces it
before the first dispatch. Never substitute silently.
A host reference maps responsibilities to mechanisms; it does not promise every mechanism exists in
every session. When one is missing or is disallowed by the session's own instructions, the
responsibility does not disappear with it — it is carried by whatever remains, and the substitution
is stated once, up front, and recorded in the sprint report. A run that quietly drops to a different
mechanism looks identical in the report to a run that used the mapped one, which is how a capability
gap becomes invisible.
Do not ask the user to choose a host. Host references map mechanisms only; they cannot waive or redefine the shared contract.
Inputs
| Mode | Input | Planning behavior |
|---|
| Interactive | multi-feature goal | build and review a rolling-wave roadmap before execution |
| Preplanned | roadmap plus sprint artifacts | validate dependencies, reviews, ownership, tests, and cluster schedule |
| Resume | canonical status file | validate revision and resume the first unfinished cluster |
Use scripts/schedule.py to derive clusters. Use scripts/state/state.py for every durable transition. Do not edit status JSON by hand.
Required references
Read specialist references only when their trigger applies: agent-authoring.md, skill-authoring.md, dev-profile.md, refactoring.md, and migration.md.
Outputs and boundary
Produce the shared roadmap, one initiative intent log, reviewed sprint artifacts, targeted and gap evidence, per-sprint provenance commits and checkpoints, final regression, intent audit, documentation sync, completion report, and any required retrospective.
After each sprint gap check, ask exactly: what is here that no WorkList item asked for? This is not a separate agent or review phase.
The leader owns decisions, integration, real command evidence, commits, checkpoints, reports, and approval boundaries. Workers never own leadership, durable state, commits, merges, deploys, or completion claims. Initiative dialogue is recorded once before autonomous execution. A sprint provenance log is artifact-led; it includes a verbatim user-dialogue delta only when the user intervened after the previous checkpoint.
