Performance and Observability Budgets

Production-grade agent workflows need predictable runtime behavior. EasyEDA MCP Pro exposes latency budgets, in-memory metrics, vendor timing, cache counters, and retention expectations through a diagnostics tool.

Tool

easyeda_observability_report

The tool is read-only and returns:

• latency budgets by category;
• runtime metrics snapshot;
• cache hit/miss/write/delete counters;
• vendor API timing summary;
• storage and artifact retention policy;
• default bridge timeout policy.

Latency budgets

Category	P50	P95	Timeout	Scope
diagnostics	50 ms	250 ms	1 s	Local diagnostics and metadata-only tools
bridge-read	250 ms	1.5 s	15 s	Read-only EasyEDA bridge calls
bridge-write	500 ms	3 s	30 s	Confirmed bridge write/apply calls
vendor-api	800 ms	5 s	30 s	External vendor API requests including retry overhead
cache	10 ms	50 ms	250 ms	SQLite cache and metadata lookups
export	1 s	10 s	60 s	Manufacturing/export generation workflows
analysis	100 ms	1 s	5 s	Local DRC/ERC, BOM, power-tree, PCB and QA analysis
documentation	50 ms	500 ms	2 s	Generated docs, summaries and manifests

Budget evaluation uses three states:

ok   -> duration <= p95 budget
warn -> duration > p95 budget but <= timeout budget
fail -> duration > timeout budget

Runtime metrics

The in-memory metrics collector records:

{
  toolCalls: number;
  bridgeCalls: number;
  budgetWarnings: number;
  budgetFailures: number;
  byCategory: Record<
    string,
    {
      count: number;
      ok: number;
      errors: number;
      averageDurationMs: number;
      maxDurationMs: number;
    }
  >;
  cache: {
    hits: number;
    misses: number;
    writes: number;
    deletes: number;
    hitRate: number;
  }
  vendors: Record<
    string,
    {
      requestCount: number;
      errorCount: number;
      averageDurationMs: number;
      lastStatusCode?: number;
    }
  >;
}

The snapshot is intentionally memory-only. It is useful for local diagnostics, Inspector sessions, CI smoke checks, and support reports. It must not include secrets or raw vendor credentials.

Timeout behavior

Bridge calls use the configured bridge timeout unless a tool supplies a narrower timeout. Long-running live diagnostics and raw API execution accept explicit timeoutMs inputs with schema bounds.

Timeouts are surfaced as structured failures and recorded as failed budget events. The bridge manager also has a stale request sweep as a defensive cleanup path.

Cache and artifact retention

Default retention policy:

Field	Default
cache default TTL	24 h
vendor cache TTL	7 d
artifact retention	30 d
telemetry retention	14 d
cleanup cadence	manual
max artifact bytes	512 MiB

Cache entries with explicit TTL expire automatically on read. Artifact cleanup is conservative and should be invoked explicitly before deleting handoff files.

Regression guard

Representative unit tests cover:

• budget status evaluation;
• diagnostics output schema;
• registry-level tool duration instrumentation;
• cache hit/miss/write/delete counters;
• vendor request timing on success/failure;
• bounded retention defaults.

These tests prevent accidental removal of observability fields or unbounded retention defaults.