---

name: long-run-ops category: development-workflow/deployment-ops version: 1.0.0 description: Enable long-running agent operations without looping or drift author: obra/superpowers verified: true tags: [long-running, checkpoints, value-function, anti-looping, agent-ops]

Skill: Long-Run Operations (Anti-Looping + Adaptation)

<checkpoint_cadence> Every 30–60 minutes:

• Update the queue
• Record what changed
• Make one prioritization decision

</checkpoint_cadence>

<value_function> <before_work> Before a step, write: </before_work>

• "After this step, we will have <artifact/change>"
• "This improves the result because "
• "If this fails, fallback is "

</value_function>

<stop_conditions> <hard_stops> Stop and write blockers.md if:

• You cannot name a new artifact you can produce
• You keep scanning sources without updating your hypotheses
• You are rephrasing the same conclusions </hard_stops> </stop_conditions>

<value_generation> Every step must produce a tangible artifact or insight Prefer depth over breadth - one deep analysis > ten shallow scans Document why each step matters in the step file </value_generation>

<context_management> Compaction is mandatory, not optional Delete or archive low-value step files during review Keep compactions ≤ 1MB each </context_management>

<best_practices> Write the value function BEFORE starting work, not after Treat blockers.md as a success - it means you avoided infinite loops Use specific artifact names ("analyze-X.py") not generic ones ("script.py") Document patterns that work well in reviews for reuse </best_practices>

<error_handling> Stop immediately. Write blockers.md documenting:

• What action you're repeating
• Why it's looping
• What new artifact would break the loop

Value function BEFORE starting

echo "After this step, we will have:

• A comparison matrix of 5 React frameworks
• Performance benchmarks for each
• Use case recommendations

This improves the result because: Concrete data beats opinions Fallback if fails: Use existing documentation summaries"

Execute work (analyze frameworks)

...

Create step file

./.blackbox/scripts/new-step.sh --plan .blackbox/agents/.plans/oss-research-001 "Complete: React framework comparison"

After 10 steps, compact

./.blackbox/scripts/compact-context.sh --plan .blackbox/agents/.plans/oss-research-001

STOP immediately

cat > .blackbox/agents/.plans/oss-research-001/blockers.md <<'EOF'

Blocker: No New Insights

Current State

• Have analyzed 20 repositories
• Keep finding same patterns
• Converged on same conclusions 3 times

Why I Can't Continue

• Cannot name a new artifact that would add value
• Scanning more repos yields diminishing returns
• Reaching synthesis point, not discovery point

Next Steps (Human Decision Needed)

• Proceed to synthesis/report generation
• Pivot to different research angle
• Expand scope to different ecosystem EOF