Rig Sprint

Sprint Orchestrator

A sprint is a batch of independent work items that each land on the trunk on their own. Use this when the work decomposes into pieces that don't interleave (no shared runtime contract across siblings, no half-migrated state if you ship one without the others). If children do interleave, use an epic / integration-branch flow instead — same decomposition step, plus a shared branch the children stack on.

Configuration

Reads .rig/config.json (missing keys → defaults):

Tracker modes:

• linear — tickets are Linear issues; use the mcp__claude_ai_Linear__* tools.
• github — tickets are GitHub issues; use gh issue.
• none — ad-hoc mode. There are no ticket IDs. Treat each item in $ARGUMENTS (or each bullet the user gives) as a task description. Skip all create/list/move-ticket steps; the sprint operates directly on the task list, and each task's implementation opens a PR whose title is the task description. plan <feature> still decomposes into a numbered task list — it just prints it instead of creating tickets.

Arguments

The user invoked this with: $ARGUMENTS

What to do

If the user said plan <FEATURE-DESCRIPTION> (decompose a feature)

Create independent work items that can each be picked up and landed on the trunk directly — no parent epic, no integration branch.

1. Read context.

   • Any product/spec docs the project keeps (ask if unsure — don't assume a fixed path).
   • Explore the codebase to understand what already exists. Default the exploration scope to sourceScope[0] (fallback src).
   • Search before creating to avoid duplicates:
     • Linear: mcp__claude_ai_Linear__list_issues with project = tracker.project, a query of the feature's key nouns, limit: 50 — search, don't pull the whole board; page via the returned cursor only if a likely match looks cut off.
     • GitHub: gh issue list --search "<key nouns>" --state all.
     • Ad-hoc (none): grep the codebase / recent branches for prior art.

2. Sanity-check it's actually sprint-shaped (independent), not epic-shaped (interleaved). If at least one child's runtime contract genuinely depends on another being partially complete (the middle state would break the trunk), stop and tell the user to use an epic / integration-branch flow instead.

3. Break the feature into 3–8 discrete items, ordered by dependency. No parent ticket.

4. Materialize the items:

   • Linear: mcp__claude_ai_Linear__save_issue per item — team = tracker.team, project = tracker.project, short action-oriented title ("Add X", "Wire up Y", "Implement Z"), state: "Backlog", priority 2/3/4, a markdown description (Overview, Acceptance Criteria, dependency references). Set blockedBy for every real dependency — not optional; the phaser relies on it to group correctly.
   • GitHub: gh issue create per item with the same title/body, adding the sprint shape label — --label "<L>" where <L> = tracker.shapeLabels.sprint (default sprint) — so a Project board / dispatcher can pick up sprint items directly (skip the label if tracker.shapeLabels is absent; ensure it exists first with gh label create "<L>" --force). Encode dependencies as "Blocked by #N" in the body (GitHub has no native blockedBy).
   • Ad-hoc (none): don't create anything — just produce the numbered task list with an explicit "depends on #k" note per item.

5. Show the user a summary table of what was created (or the task list, in ad-hoc mode):

   ID / #	Title	Priority	Depends On

6. Hand off. End with: "Run /rig-sprint <ID1> <ID2> ... (or paste the task list back) to batch-execute these in phased dependency order." Don't move tickets to Todo or auto-run — the user picks the entry point.

If the user said plan (no description) — or passed no arguments at all

Preview-mode for the existing sprint queue. Neither launches work; they show what's ready to run and wait for the user to pick items or kick off the sprint.

1. Fetch ready work:
   • Linear: mcp__claude_ai_Linear__list_issues with project = tracker.project, state "Todo", limit: 100 (page via cursor if truncated). Also check "In Progress" for in-flight work.
   • GitHub: gh issue list --state open (optionally filtered by a "ready" label the repo uses).
   • Ad-hoc (none): there is no queue — say so and ask the user to supply the task list.
2. Show the user: which items are ready, a suggested phase grouping based on dependencies (blockedBy/blocks relations, or "Blocked by #N" text), and estimated scope per phase.
3. Ask the user to confirm before launching.

If the user specified ticket identifiers (or a task list)

1. Resolve the items.
   • Detect ticket IDs by tracker.ticketPrefix. For Linear, fetch each with mcp__claude_ai_Linear__get_issue; for GitHub, gh issue view <number>. In ad-hoc mode, each argument/bullet is a task description — no fetch.
2. Determine phases:
   • Group items with no inter-dependencies into the same phase.
   • Items that depend on other specified items go in a later phase.
   • Use blockedBy/blocks relations, "Blocked by #N" text, or the description clues (ad-hoc: the "depends on #k" notes) to order them.
3. Show the proposed phases to the user for confirmation.
4. Mark items in-flight (tracker only):
   • Guard — tracker.githubIntegration: if true, skip the bulk move here — each /rig-task sets its own item to "In Progress" at start (Step 1), and GitHub advances In Review / Done from PR events.
   • Otherwise, move every item in the sprint to Todo (Linear: save_issue with state: "Todo").
   • Ad-hoc mode: nothing to move.
5. Execute each phase. Within a phase, items are independent and can run in parallel; whichever PR lands second rebases on the trunk. For each item in the phase, run /rig-task (passing the ticket ID, or the task description in ad-hoc mode) and act on its returned outcome — only a clean review is merge-green; anything else stops for a human. Move to the next phase only after every item in the prior phase has its PR merged.
6. Between phases and at the end, run /rig-tidy to audit the merged commits for dead code, duplicates, and stale comments before building on top of them.

Implementing each item

Each item is implemented by the /rig-task skill — a self-contained spec → worktree → TDD → pre-PR review → open PR → review-bot loop orchestrator. Sprint does not reimplement that loop; it just sequences the items and calls /rig-task per item (Step 5 above), reading the config knobs (tracker.*, vcs.*, test.*, review.*) the same way.

/rig-task never auto-merges and returns an outcome string; sprint uses it to gate phase progression (merge only on clean, run /rig-tidy between phases). See the /rig-task skill for the full per-item cadence.