overnight-mode

Overnight Mode Context

You are currently running in OVERNIGHT MODE. This skill provides essential context for unattended autonomous development.

Current State

Read your current state from these files:

Primary State: .claude/auto-overnight.local.md

active: true/false     # Is overnight mode active?
iteration: N           # Current loop iteration
max_iterations: N      # Stop after this many
max_hours: N           # Maximum runtime
started_at: ISO8601    # When overnight started
deadline_at: ISO8601   # Hard stop time
current_phase: PHASE   # INIT/PLAN/EXECUTE/VERIFY/COMPLETE
tasks_completed: N     # How many tasks done
tasks_total: N         # Total tasks in plan
last_checkpoint: ISO   # Last checkpoint time

Progress: .claude/auto-progress.yaml

tasks:
  task-1:
    status: done/in_progress/pending/failed
    branch: auto/task-1
    iterations: N

Memory: .claude/auto-memory/

• context-summary.md - Full context for resume
• overnight-checkpoint.md - Latest overnight checkpoint
• next-actions.md - What to do next

Overnight Rules

1. NO USER INTERACTION

• DO NOT use AskUserQuestion tool
• DO NOT wait for approval
• Make autonomous decisions
• Document decisions in .claude/overnight-decisions.md

2. STRICT TDD (Still Required!)

Even overnight, you MUST follow TDD:

RED    → Write ONE failing test FIRST
GREEN  → Implement MINIMAL code to pass
REFACTOR → Clean up while tests pass
MUTATE → Verify test quality (if tools available)

3. CHECKPOINT FREQUENTLY

After every significant action:

# Update checkpoint timestamp
TIMESTAMP=$(date -u +%Y-%m-%dT%H:%M:%SZ)
sed -i '' "s/last_checkpoint: .*/last_checkpoint: \"$TIMESTAMP\"/" .claude/auto-overnight.local.md

4. EVIDENCE-BASED VERIFICATION

Never claim "tests pass" without running them:

# WRONG
echo "Tests should pass"

# RIGHT
npm test 2>&1 | tee .claude/overnight-test-output.log
echo "Test output saved to .claude/overnight-test-output.log"

5. AUTO-RESTART ON CONTEXT LIMIT

If context is filling up:

1. Write checkpoint to .claude/auto-memory/overnight-checkpoint.md
2. Include: current task, next steps, important context
3. Stop hook will restart with /auto-continue

Error Handling

When Tests Fail

1. Max 3 fix attempts per issue
2. Use systematic-debugging (4-phase protocol)
3. If still failing: Document in .claude/overnight-issues.md and continue

When Unknown Error Occurs

1. Document everything in .claude/overnight-issues.md
2. Try to continue with remaining tasks
3. Flag for human review in overnight report

Completion Criteria

The stop hook will use prompt-based LLM evaluation to check:

1. All plan tasks completed?

   • Read .claude/auto-progress.yaml
   • Every task status should be done

2. All verification passing?

   • Tests pass (actual output required)
   • Lint clean (actual output required)
   • Build succeeds (if applicable)

3. No critical issues?

   • .claude/overnight-issues.md has no unresolved CRITICAL items

Only when ALL three are true, overnight mode completes.

Report Generation

When complete, generate .claude/overnight-report-{timestamp}.md:

# Overnight Development Report

## Execution Summary
- **Started:** {started_at}
- **Completed:** {timestamp}
- **Duration:** {duration}
- **Iterations:** {iteration}
- **Tasks:** {completed}/{total}

## Changes Made
| File | Action | Lines |
|------|--------|-------|
| src/foo.ts | Created | +150 |
| src/bar.ts | Modified | +30/-10 |

## Verification Results
### Tests
\`\`\`
{actual npm test output}
\`\`\`

### Lint
\`\`\`
{actual npm run lint output}
\`\`\`

### Build
\`\`\`
{actual npm run build output}
\`\`\`

## Issues Encountered
{contents of overnight-issues.md or "None"}

## Git History
\`\`\`
{git log --oneline for overnight commits}
\`\`\`

## Recommendations
{any follow-up work or concerns}

Safety Reminders

1. All changes are in git - can be reverted
2. Never modify outside project dir
3. Never push without explicit flag
4. When in doubt, checkpoint and document

Handoff to Human

If overnight mode ends with issues:

1. Report is at .claude/overnight-report-*.md
2. Issues are at .claude/overnight-issues.md
3. Human can review and decide next steps
4. Use /auto-continue to resume if needed