Optimization Pattern — Context

Pattern: Optimization
Component: context.md
Version: 1.1 | Updated: 2026-07-16

Preconditions (must be true before activation)

Target metric named (e.g. checkout p99_latency_ms, cost_per_1k_requests)
Minimum useful effect size declared (e.g. ≥15% p99 reduction or ≥20% unit-cost cut)
Guardrails listed (error rate, p99 ceiling, CPU/mem saturation, neighbor fairness)
Workload fixture frozen (RPS mix, concurrency, payload distribution, region)
Correctness oracle defined and runnable in CI or bench harness
Feature flag / config key exists to disable the optimization in <5 minutes
Human owner available for canary go/abort

Trigger conditions (apply the pattern)

Apply when any of the following hold:

SLO or error-budget burn attributed to latency/throughput/cost
Profiling shows a dominant resource (≥30% of critical path or spend)
Growth model shows capacity cliff within planning horizon
Provider/token unit cost exceeds budget with a hypothesized mechanism

Do not apply when

No metric owner and no success threshold (“make checkout faster”)
Active customer-harming incident — contain via hotfix-pattern
Change is purely correctness (wrong results) — fix, then re-measure
Theoretical maximum gain of the hot path cannot meet the minimum useful effect
Environment cannot run a representative workload (no staging parity and no safe canary)

Required inputs

Input	Source	Minimum quality
SLO / budget	Service catalog / FinOps	Numeric threshold + window
Baseline report	Bench harness	Warm-up + n≥30 + percentiles
Profiles / traces	Continuous profiler / APM	End-to-end critical path
Workload definition	Load-test repo	Versioned, checked in
Oracle	Test suite / golden set	Fail-closed on mismatch
Flag name	Flag service	Kill-switch documented

Authority boundary

The optimization agent proposes one change and evaluates evidence. It does not promote to 100% traffic. Named human owns canary abort and promotion. Coding agents implement only the hypothesized change behind the flag.