[Bug] Multi-step tasks lack per-step verification — errors compound silently

Bug Description

When Claude Code executes multi-step procedures (apply 7 SQL files, run a build pipeline, deploy a sequence of changes), it treats the entire sequence as one unit and only attempts verification at the end — if at all. There is no per-step gate where the output of step N is verified before proceeding to step N+1.

Concrete Example

A 7-file SQL import procedure. The source document specified "Check DBErrors.log after each file." Claude:

1. Applied File 1 — did not check logs
2. Applied File 2 — did not check logs
3. Applied File 3 — did not check logs
4. Applied File 4 — did not check logs
5. Applied File 5 — did not check logs
6. Applied File 6 — did not check logs
7. Applied File 7 — did not check logs
8. Wrote a summary: "All 7 files applied cleanly — zero errors"

If File 3 had failed, Files 4-7 would have run against corrupt state. The error would only be discovered during the post-hoc QA (which was also flawed — see #32291).

Why End-Gates Are Insufficient

The anti-pattern is: execute everything, then verify at the end. This fails because:

• Later steps may depend on earlier steps succeeding
• Post-hoc verification often can't distinguish "step 3 failed and steps 4-7 ran against bad state" from "everything worked"
• The verification itself may be tautological (#32291)

Expected Behavior

For multi-step procedures, each step's output should be checked before proceeding:

Step 1: Apply File 1
  → Check: mysql exit code 0? ✓
  → Check: DBErrors.log delta? 0 new lines ✓
Step 2: Apply File 2
  → Check: mysql exit code 0? ✓
  ...

Related

• #32281 — Reports completion without execution
• #32291 — Verification queries are tautological
• #32290 — Reads procedure docs but ignores verification instructions in them

Environment

• Claude Code 2.1.71
• Windows 11