Skip to main content

Command activity tracking

Guard records every command it evaluates — the decision, the reason, and the proof level — as tamper-evident evidence. This evidence is viewable in the dashboard and via CLI inspection commands.

What gets recordedDirect link to What gets recorded​

Each command evaluation includes:

  • Command — the evaluated command text
  • Harness — which AI agent submitted it (Codex, Claude, etc.)
  • Decision — allowed, blocked, or queued for approval
  • Reason — the policy rule that triggered the decision
  • Timestamp — when the command was evaluated

Dashboard viewsDirect link to Dashboard views​

Evidence workbenchDirect link to Evidence workbench​

The dashboard's evidence views show command activity across harnesses. Filter by harness, status, or risk category.

AnalyticsDirect link to Analytics​

The evidence insights surface provides:

  • Day-by-day breakdown — commands evaluated per day
  • Top apps by activity count — most active harnesses
  • Activity heatmap — visual command volume over time

A health indicator states when counts may be incomplete (e.g., after a daemon restart or evidence recovery).

CLI inspection commandsDirect link to CLI inspection commands​

Test a command against current policyDirect link to Test a command against current policy​

hol-guard command test 'git push --force'

Shows what Guard would do with the command — allow, block, or queue for approval — without executing it.

Explain a command's classificationDirect link to Explain a command's classification​

hol-guard command explain 'grep "rm -rf" README.md'

Explains how Guard classifies the command and which rules apply.

List installed command extensionsDirect link to List installed command extensions​

hol-guard command extensions

Shows built-in command safety extensions that Guard uses to classify commands.

Detect command ecosystemsDirect link to Detect command ecosystems​

hol-guard command setup --detect --workspace .

Detects command ecosystems in the current workspace and previews recommended protection coverage.

Compound command pipelinesDirect link to Compound command pipelines​

Guard evaluates compound shell commands (chained with &&, |, or ;) by splitting them into segments and evaluating each independently:

  • Pipeline segments — each segment gets a pipeline_index so Guard can track producer-to-consumer data flow
  • PipelineMatcher — detects when a pipeline sends sensitive data (like secrets) from a producer to a network consumer
  • Per-segment decisions — safe segments (like git log) are allowed; risky segments (like git push --force) require approval

If any segment is unresolved (Guard can't statically prove safety), the entire compound command is queued for review.