claudeframework: Skill for Claude Code

Orchestrate parallel /ship runs for decomposed PRDs. Analyze file conflicts across features, group non-conflicting features into parallel batches, execute each in an isolated git worktree, and merge results back to main conflict-free.

$ARGUMENTS is optional:

• First positional arg: decomposition name (matches .claude/specs/decomposition-<name>.md)
• --dry-run — generate specs and show batch plan, but don't implement
• --max-parallel N — limit concurrent worktrees (default: 5)
• --batch-only — run batch 1 only, then stop for review

Architecture

Fleet is a thin orchestrator — it never reads code files or accumulates implementation context. It reads manifests, spec metadata, and delegates all per-feature work to /ship subagents running in isolated worktrees.

Fleet orchestrator (this skill)
  ├── Step 0: load decomposition manifest
  ├── Step 1: collect file claims from specs
  ├── Step 2: conflict analysis → batch plan
  ├── Step 3: execution plan + user approval
  ├── Step 4: batch execution loop
  │     ├── Launch /ship agents in isolated worktrees (background)
  │     ├── Wait for completion
  │     ├── Merge completed branches to main
  │     └── Advance to next batch
  ├── Step 5: final integration verify
  ├── Step 6: failure triage
  └── Step 7: fleet report

Step 0 — Load decomposition manifest

1. Parse $ARGUMENTS for the decomposition name and flags (--dry-run, --max-parallel N, --batch-only)
2. If a decomposition name is provided, read .claude/specs/decomposition-<name>.md
3. If no name provided, find the most recently modified decomposition-*.md in .claude/specs/
4. If no decomposition manifest exists: stop — "No decomposition found. Run /decompose first to break your concept into PRDs."
5. Parse the manifest's features table: extract feature name, PRD path, dependencies, status
6. Filter to features with status pending only (skip completed, in-progress, skipped)
7. If no pending features: "All features are already completed or in-progress."