Pillar D: Performance Optimization Prompts

Seven prompts exercising v1.5.0 Pillar D — performance optimization and budget discipline across FCC's event bus, workflow engine, and federation layer.

Pillar D gives teams a repeatable methodology for profiling, measuring, and regression-gating performance without letting the cure become worse than the disease.

Context for Claude

Pillar D uses OTO's per-lane TracerProviders + SMC's scorecards + the CLEAR+ benchmark framework (src/fcc/evaluation/) to measure 7-dimensional quality (Cost, Latency, Efficacy, Assurance, Reliability, Coverage, Explainability). Prompts focus on measurement discipline before optimization discipline.

Prompt 1: Baseline capture

Capture the v1.5.3 performance baseline for the 5-node workflow
graph. For 1,000 runs, report:
- p50 / p95 / p99 total latency
- p50 / p95 / p99 per-node latency
- Event count per run
- Memory high-water
- Producer/consumer backpressure incidents

Expected outcome: baseline report.

Prompt 2: CLEAR+ benchmark suite

Run the CLEAR+ benchmark suite on the extended_84 workflow.
Report the 7 dimensions:
- Cost (tokens, $)
- Latency (p95)
- Efficacy (task completion)
- Assurance (hallucination rate)
- Reliability (retry rate)
- Coverage (scenario coverage %)
- Explainability (trace usefulness)

Name 2 dimensions with room to improve and propose interventions.

Expected outcome: 7-dimension scorecard + interventions.

Prompt 3: Bottleneck discovery

Run a 1-hour load-test harness. Use OTO's flame graph to identify
the top 3 bottlenecks. For each:
- Bottleneck location
- Root cause
- Proposed fix
- Expected gain (quantified)
- Risk of regression in another dimension

Expected outcome: bottleneck triage.

Prompt 4: Back-pressure tuning

PAOMBus is dropping 2% of events under peak load. Propose 3
back-pressure policies:
- Current (drop-oldest)
- Alternative 1 (keep-latest, block producer)
- Alternative 2 (block producer, rate-limit at source)

Model each on latency, drop rate, producer stall time.

Expected outcome: policy comparison.

Prompt 5: Regression gate

Design a CI regression gate for Pillar D:
- Fails on p95 latency regression > 10%
- Fails on event drop rate > 0.5%
- Warns on CLEAR+ Explainability drop > 5 pts
- Observability required before failing (no flaky gates)

Expected outcome: regression gate spec.

Prompt 6: Cost dimension optimization

CLEAR+ Cost for a representative workflow is $0.18. Can we halve
it without regressing Efficacy? Explore:
- Provider switch (Anthropic → Ollama for some steps)
- Context window trimming
- Caching discoverable state
- Parallelism vs sequential

For each: expected cost savings + efficacy risk.

Expected outcome: cost optimization plan.

Prompt 7: Performance story for release notes

Write the v1.5.3 performance story (150 words) for release notes:
- Headline change
- Dimension-by-dimension delta
- The tradeoff accepted
- What the user should expect in the next release

Expected outcome: performance story.

Related

• Tutorial: docs/tutorials/advanced-capabilities/performance-optimization-guide.md (Pool B)
• Personas: OTO, LAR, SMC
• Notebook: notebooks/46_graphrag_advanced_patterns.ipynb (performance subsections)
• Benchmarks: src/fcc/data/evaluation/baseline_benchmarks.yaml