Performance and local telemetry

Understand your own scan before choosing CI timeouts or attributing cost to a rule. CLI and MCP share the solution-analysis pipeline; compiler-hosted feedback has a different scope. See published benchmarks for measured inputs, results, methodology and limits.

Opt in to local collection

Set ARCHGUARD_TELEMETRY to exactly 1 in the CLI or MCP process environment. Collection is disabled for every other value and when unset. For the installed CLI:

ARCHGUARD_TELEMETRY=1 archguard scan \
  Shop.slnx --policy architecture.guardian.yaml --trusted --configuration Release --format json \
  > report.json 2> telemetry.jsonl

Replace the input with your restored repository and authorize its build logic as described in workspace loading. For PowerShell, set $env:ARCHGUARD_TELEMETRY = '1' before invoking the command and remove it afterward with Remove-Item Env:ARCHGUARD_TELEMETRY. For MCP, add the environment entry to the client configuration and restart the server. Stdio stdout continues to carry only MCP messages. Operational errors and existing MCP logs also use stderr; filter JSON lines by source == "ArchitectureGuardian.Analysis" for mixed stderr. A closed/unwritable local sink drops measurements without changing coverage.

The source/meter name is ArchitectureGuardian.Analysis. It exposes:

Instrument	Unit	Tags
archguard.phase.duration	seconds (s)	phase, status, optional rule.id
archguard.rule.evaluations	evaluations ({evaluation})	phase=rule, status, rule.id

Phases are analysis, workspace, graph, rules, rule and changes. Durations are inclusive: analysis contains workspace/graph/rules, rules contains each rule, and changes contains both scans and comparison preparation. Avoid summing parent and child timings. The counter records each invoked rule once, including failed/cancelled evaluations; pre-cancelled requests do not count uninvoked rules. Statuses retain completed, incomplete, failed and cancelled. A completed evaluation may produce findings and fail a severity gate.

Each local JSON line contains source, instrument, numeric value, unit and tags. Attributes contain only phase/outcome and rule identity: no repository paths, source, diagnostic messages or exception details. Metrics never enter JSON/SARIF reports, baselines or diagnostic fingerprints.

Interpret the measurements

Compare the same tool/input/policy/configuration/framework context. The workspace phase includes loading and compilation validation; graph extraction and rule execution have different costs. Inclusive phases overlap, so summing them double counts work. Git comparison includes two scans and snapshot preparation.

Record completion status and gate outcome alongside elapsed time. A fast incomplete result does not prove better performance. Use the benchmark scope when comparing allocation traffic with runner memory requirements.

Privacy and behavior

Local collection writes to stderr and configures no remote exporter. It does not modify reports, policies, baselines or diagnostic identities, skip rules or relax coverage. With collection disabled and no external listeners, instrumentation avoids timing and activity/tag creation. Remove the opt-in environment entry when finished. Keep stderr logs separate from protocol/report stdout.