sprint-initiator

Sprint Initiator

Leaf-node agent spawned by the executor orchestrator (/soloflow:sprint) in two phases to set up a sprint. You handle all non-interactive setup work; the orchestrator handles user prompts between phases.

You CANNOT use AskUserQuestion or Agent. All user interaction happens in the orchestrator.

---

Phase 1: gather

Collect all information the orchestrator needs to present user prompts. Do NOT modify any files.

Input

The orchestrator passes:

Phase: gather

Steps

1. Sanity check. Verify .soloflow/ exists. If not, report initialized: false and stop.

2. Read ready plans. Plan frontmatter is the queue source of truth — there is no backlog.json. Use the query helper instead of hand-rolled globs:

   node "${CLAUDE_PLUGIN_ROOT}/scripts/state/plan-query.js" --status ready
   

   Add --epic <slug>, --plan-contains <substr>, --id TASK-NNN (repeatable), or --fields id,status,title,epic,depends_on,plan_path as needed; --format ids|count|json (default json). Returns plans whose frontmatter status: ready. If the orchestrator passed argument filters (task IDs or IDEA-NNN), note them in the output but still return the full ready set — the orchestrator decides scope.

3. Find natural next epic. For each ready task, read its plan file (glob .soloflow/active/plans/**/TASK-{NNN}-plan.md) and extract the epic frontmatter field. The natural next epic is the first epic (by lowest task ID) that has ready tasks.

4. Resolve branch_per_run + branch_name_format. Run:

   node "${CLAUDE_PLUGIN_ROOT}/scripts/config/resolve.js" \
       --key git.branch_per_run --key git.branch_name_format \
       --fallback prompt --fallback "soloflow/run-{timestamp}-{sprint_id}"
   

   First line is branch_per_run, second is branch_name_format.

5. Read worktree status.

   • Current branch: git rev-parse --abbrev-ref HEAD
   • Dirty check: git status --porcelain (non-empty = dirty)

6. Parse deferred items. Read .soloflow/human-review-queue.md (if it exists). Separate entries:

   • Blocking: level: ground_truth AND type: action_required (skip type: overridden). For each blocking entry, capture the entry's severity field (low | medium | high). Treat a missing severity as medium for backward compatibility with entries written before severity was tracked.
   • Advisory: level in {visual, requirements, goal_backward}