Verify Loop

The outer loop. Runs autopilot skill iteratively, gated by verification-before-completion, with self-heal recovery. Stops only when 100% verified or stop condition triggers.

The loop

init: bash .claude/scripts/loop-iteration.sh start --until <ISO> --max-iter <N>
      (machine-enforces the wall-clock/iteration budgets below)

while not stop_condition:
    1. run `bash .claude/scripts/loop-iteration.sh` — deterministic gate + picker:
       exit 2 → HALT (abort cap / run budget / external TASKS.md edit — reason printed);
       exit 1 → stop (empty or fully-blocked backlog, or verify pre-flight red);
       exit 0 → it names the nominated task (canonical next-task.sh grammar)
    2. invoke autopilot skill on that task    (5-phase execution)
    3. run verification-before-completion gate
       a. if PASS  → `task-status.sh T-<id> done`, commit, continue
       b. if FAIL  → invoke self-heal (3 attempts max)
          - on heal success → continue
          - on heal failure → `task-status.sh T-<id> failed --note "<reason>"`,
            log to OVERNIGHT_REPORT, continue
       (task-status.sh feeds the abort/progress recorder — never flip by hand)
    4. WIP checkpoint
    5. evaluate stop conditions
end

at end:
    6. run /dream skill (consolidate memory)
    7. write OVERNIGHT_REPORT.md
    8. open PRs for completed tasks
    9. exit cleanly

Stop conditions (any one stops)

• No more unblocked tasks in tasks/TASKS.md
• Wall-clock budget exceeded (--until <time> or default 4 hours)
• Token budget exceeded (--budget <N> or session limit)
• 3 consecutive task ESCALATIONs (self-heal exhausted on 3 different tasks in a row)
• Any security check fails on a diff (semgrep critical, gitleaks hit)
• User sends /stop, Ctrl+C, or any interrupt
• tasks/TASKS.md modified externally — loop-iteration.sh compares its hash to the loop's snapshot (.claude/state/tasks-md.snapshot, refreshed by every sanctioned mutation) and exits 2 on mismatch

Invocation patterns

From a Cloud Routine (the headline scenario)

Run /verify-loop until all unblocked tasks pass verification OR 04:30 wall clock.
Use /implement, /verify, /review per task. Use playwright-skill for UI.
Use semgrep for security. WIP-checkpoint every 10 minutes.
On completion, run /dream then write OVERNIGHT_REPORT.md.
Open one PR per completed task on claude/overnight-YYYY-MM-DD-<task> branches.

From /loop (in-session)

/loop --budget 500000 --until 02:00

From the command line (headless)

claude -p --max-budget-usd 20 --max-turns 200 \
  --output-format stream-json \
  /verify-loop

Verification gate (the heart)

Wraps obra/superpowers:verification-before-completion:

For each task, after Phase 5 of autopilot, before marking [x]:

1. Run task's accept: command → exit 0 required
2. Run bash .claude/scripts/verify.sh → exit 0 required (lint + typecheck + unit tests)
3. Run integration tests if defined → exit 0 required
4. For UI changes: webapp-testing skill → screenshots captured, no console errors, no failed network
5. For security-touched diffs: semgrep + codeql + dep audit → no high or critical findings

Only when ALL FIVE pass is the task [x]. Otherwise [!] and self-heal kicks in.

Hard rules

• Never skip the gate. Even at 4:25 AM with 5 minutes left.
• Evidence required. A passed gate writes evidence to verify/<date>-<task>/ — the gate itself is the proof.
• Auto-PR per task. One task = one PR. Don't batch.
• Don't auto-merge. Verify loop opens PRs; merge requires human or branch-protection automation (claude-code-action + auto-merge for green CI).
• Self-heal max 3. Per task. After 3, escalate and move on.

What it produces

• N completed PRs on claude/overnight-<date>-<task> branches
• M handoff files in .swarms/streams/<run-id>/handoff-*.yaml (coordinator schema; the autopilot skill writes here too)
• 1 OVERNIGHT_REPORT.md at repo root
• 1 /dream consolidation pass at the end
• Updated tasks/TASKS.md with [x] and [!] marks

References

• obra/superpowers/skills/verification-before-completion (gate logic)
• obra/superpowers/skills/executing-plans (loop structure)
• .claude/skills/autopilot (per-task execution)
• .claude/skills/self-heal (recovery)
• .claude/skills/handoff (NEXUS report format)
• Native /loop command: https://code.claude.com/docs/en/scheduled-tasks