Claude-Pipeline: Command for Claude Code

Execute the implementation plan with context isolation per step.

---

Prerequisites

Check that plan exists and optionally verify drift:

SESSION_DIR=$(cat .claude/artifacts/current.txt 2>/dev/null)
if [ -z "$SESSION_DIR" ] || [ ! -f "$SESSION_DIR/plan.md" ]; then
  echo "ERROR: No plan.md found. Run /plan first."
  exit 1
fi

If --force is NOT provided and no drift-report.md exists, warn:

"Warning: /pmatch has not been run. Consider running /pmatch first to verify plan-design alignment."

If drift-report.md exists and verdict is DRIFT_DETECTED, require --force:

"ERROR: Drift detected. Run /pmatch to see issues, then either fix them or use /build --force to proceed anyway."

---

Build Execution

Use the builder agent.

Input: Read plan.md from the current session directory.

Process:

For each step in the plan (sequentially):

1. Fresh Read — Read only the files mentioned in that step
2. Verify BEFORE — Check that file content matches BEFORE code (for MODIFY)
3. Apply AFTER — Use Edit/Write to apply the changes
4. Verify Criteria — Check each acceptance criterion
5. Log Result — Add to build-report.md

After all steps:

1. Run build verification: npm run build
2. Run type verification: npx tsc --noEmit
3. Output final build-report.md

---

Context Isolation

Each step starts with a fresh context:

• Only read files explicitly mentioned in that step
• Don't assume state from previous steps
• This prevents context bleed and ensures deterministic execution

---

Error Handling

If a step fails:

1. BEFORE mismatch: Stop immediately. The plan is stale.
2. File not found: Stop immediately. The plan has a bug.
3. Build fails: Continue logging but flag in report.

Do NOT try to fix issues — that's improvisation. Report and stop.

---

Output

After build, report:

## Build Complete

**Session:** {session-directory}
**Verdict:** [SUCCESS | PARTIAL | FAILED]

### Summary
- Steps completed: [N/M]
- Files changed: [N]

### Verification
- Build: [PASS | FAIL]
- Types: [PASS | FAIL]

### Files Changed
[List of files with actions]

### Next Steps
1. Run `/denoise` to clean debug code
2. Run `/qf` for code quality check
3. Run `/qb` for behavior validation
4. Run `/security-review` for security audit