prd-build-resume

name	prd-build-resume
description	Resume a paused or interrupted PRD build run
argument-hint
allowed-tools	["Bash","Read","Write","Edit","Glob","Grep","Agent"]

PRD Build Resume

Load existing state:

if [ -f .claude/prd-build.state.json ]; then
  cat .claude/prd-build.state.json
else
  echo "NO_STATE_FILE"
fi

If NO_STATE_FILE: tell the user there's no run to resume. Suggest /prd-build <prd-path> to start one.

Otherwise, follow this recovery procedure:

1. Assess State

Read .claude/prd-build.state.json and .claude/prd-build-artifacts/dag.json. Determine:

• What phase was the run in? (decomposing, executing, landing, verifying, paused)
• Which units are landed, active, pending, evicted, failed?
• What pass number are we on?

2. Recover Interrupted Units

Any units with status "active" were interrupted mid-execution. Read dag.json, update those units' status to "pending" and clear their pipeline_stage to null, then write the file back.

3. Resume Execution

Update .claude/prd-build.state.json status to "executing". Append to .claude/prd-build.log.md: "Resumed from checkpoint".

4. Continue from Current Layer

Find the lowest layer number that still has non-landed units. Resume Phase 2 (EXECUTE) from that layer.

Follow the exact same orchestration logic as the main /prd-build command's Phase 2, 3, and 4. The key difference: you already have the work plan — skip decomposition and go straight to execution.

Read the PRD for context:

# The PRD path is stored in the state file — read it from the JSON printed above
cat "$(cat .claude/prd-build.state.json | python3 -c 'import sys,json; print(json.load(sys.stdin)["prd_path"])')"

Then read that file to understand what's being built, and continue orchestrating.