Activity Ledger

Overview

Dashboard stats

• Unique Agents - Count of distinct agent_ids
• Unique Action Types - Count of distinct action names

Gateway decisions timeline

• Time and Action
• Decision badge: allow, deny or require approval
• Reason code explaining the decision
• Approval request opened (or redeemed) by the decision, linked to the Approvals page
• Reserved, Actual and Remaining budget in USD. A — means no value was recorded (for example, no reservation, or headroom not captured), not $0.
• Chain status for the row:
  • verified: the row hash matches and the signature verifies
  • unsigned: the row has no signature (whatever its hash)
  • failed: a signed row whose hash or signature doesn’t check out

Advanced filtering

Filter by outcome

• Success - Completed successfully
• Error - Failed with exception
• Blocked - Prevented by policy
• Replayed - Executed via replay mode

Filter by action name

• Exact match - Type action name
• Auto-complete - Suggests from existing actions
• Case-sensitive

• call_llm - All LLM calls
• search_web - All web searches
• transfer_funds - All fund transfers

Filter by agent

• Select from dropdown of known agents
• Type to search by agent_id
• Shows display name if set

Filter by run

• Enter run_id
• Shows only actions from that specific run
• Useful for drilling down from Runs page

Filter by cost range

• Min cost - Show actions costing at least this much
• Max cost - Show actions costing at most this much

• Min: $0.10 - Expensive actions only
• Max: $0.01 - Cheap actions only
• Min: $1.00, Max: $5.00 - Mid-range actions

Filter by date range

• Start date - Show actions after this date/time
• End date - Show actions before this date/time
• Date picker with time selection
• Timezone-aware

• Last 24 hours
• Last 7 days
• Last 30 days
• Custom range

Full-text search

• Action name
• Input parameters (JSON content)
• Output results (JSON content)
• Tags
• Agent ID

• Use quotes for exact phrases: "customer email"
• Searches are case-insensitive
• Matches partial strings

• Search error - Finds actions with “error” in outputs
• Search user@example.com - Finds actions with that email
• Search "payment processed" - Exact phrase match

Sort options

• Timestamp (asc/desc) - Chronological order
• Cost (asc/desc) - By expense
• Duration (asc/desc) - By execution time
• Action name (asc/desc) - Alphabetical

Action table

• 🟢 Success - Green
• 🔴 Error - Red
• 🟠 Blocked - Orange
• 🔵 Replayed - Blue

• Click row to open Action Inspector
• Hover for quick preview

Action inspector modal

Overview section

• Action ID (UUID)
• Action name
• Timestamp (full precision)
• Outcome with description
• Cost (USD)
• Duration (nanosecond precision, displayed in ms)
• Agent ID (link)
• Run ID (link)
• Tags (if any)

Inputs section

• Syntax-highlighted JSON
• Expandable/collapsible nested objects
• Copy to clipboard button
• Type information (string, number, object, array, etc.)

{
  "prompt": "What is the capital of France?",
  "model": "gpt-4o",
  "max_tokens": 100,
  "temperature": 0.7
}

Outputs section

• Syntax-highlighted JSON
• Expandable/collapsible nested objects
• Copy to clipboard button
• Size indicator (bytes)

{
  "response": "The capital of France is Paris.",
  "tokens_used": 15,
  "finish_reason": "stop"
}

Metadata section

• Error details (if outcome=error)
  • Exception type
  • Error message
  • Stack trace (if available)
• Compliance data (if applicable)
  • Approval status
  • Approver information
  • Decision rationale
  • Data lineage
  • Model card
• Original vs Modified (if downgraded by policy)

Related actions

• Other actions in this run - View siblings
• Similar actions - Same action name, different runs
• Agent’s recent actions - Last 10 from this agent

Pagination

• Page size: 50, 100, 500, 1000 per page
• Page controls: First, Previous, Next, Last
• Jump to page: Enter page number
• Total count: Shows total matching actions

Export functionality

Export formats

1. JSONL (JSON Lines) - One JSON object per line
   • Best for programmatic processing
   • Stream-friendly
   • Newline-delimited
2. JSON - Standard JSON array
   • Best for data analysis tools
   • Human-readable
   • Single file
3. CSV - Comma-separated values
   • Best for Excel/spreadsheets
   • Flat structure (nested objects flattened)
   • Universal compatibility

Export options

• Respects current filters - Only exports what’s visible
• Includes all fields - Inputs, outputs, metadata, compliance data
• Limit: 100,000 records per export
• Streaming: Large exports stream to prevent timeouts

Export use cases

• Compliance reporting - Generate audit reports
• Cost analysis - Import into BI tools
• Data science - Train models on action patterns
• Debugging - Analyze errors in detail
• Backup - Archive historical data

Verify chain integrity

• Chain valid: every row’s hash and signature checks out.
• Chain invalid: shows the first row (seq) where the chain breaks or a signature fails, which is where tampering or corruption starts.

Real-time updates

1. Appears at top of table (if matches filters)
2. Row highlights briefly (fade-in animation)
3. Stats refresh
4. Notification (optional)

• 🟢 REALTIME - Live updates
• 🟡 POLLING - Fallback mode (30s refresh)

Common workflows

Audit trail for compliance

1. Filter by date range (e.g., last quarter)
2. Export as JSONL
3. Store in compliance archive
4. Generate report showing:
   • All actions logged
   • Human oversight events (approvals)
   • Policy violations (blocked actions)
   • Data lineage

Find expensive LLM calls

1. Filter by action name: call_llm
2. Sort by cost (descending)
3. Review top 10 most expensive
4. Check inputs for optimization opportunities:
   • Reduce max_tokens?
   • Use cheaper model?
   • Add caching?

Debug errors for specific agent

1. Filter by agent_id: my-agent
2. Filter by outcome: Error
3. Sort by timestamp (newest first)
4. Open first error in inspector
5. Review error message and stack trace
6. Check inputs that caused error
7. Reproduce and fix

Track approval decisions

1. In the gateway decisions timeline, filter to Require approval to see which actions opened approval requests
2. Full-text search: approved or rejected
3. Filter to actions with compliance metadata
4. Review approver decisions
5. Analyze patterns:
   • Who approves most?
   • What actions get rejected?
   • How long do decisions take?

Monitor policy enforcement

1. In the gateway decisions timeline, filter to Deny and read the reason codes
2. Filter the ledger by outcome: Blocked
3. Review what’s being blocked
4. If many legitimate actions blocked:
   • Policies may be too strict
   • Adjust in Settings → Policies
5. If many violations:
   • Agent may be misbehaving
   • Review agent logic

Saved filters (future)

• “Expensive LLM calls” (action=call_llm, cost>$1)
• “Recent errors” (outcome=error, last 24h)
• “High-risk blocks” (outcome=blocked, risk=critical)

Best practices

Keyboard shortcuts

• / - Focus search bar
• f - Open filter panel
• e - Export
• Esc - Close inspector modal
• Arrow keys - Navigate table rows

Troubleshooting

”No actions showing”

• Check if filters are too restrictive
• Try “Clear all filters” button
• Verify date range includes recent data
• Check if any runs have been created

”Search not finding results”

• Ensure spelling is correct (case-insensitive but must match)
• Try partial search without quotes
• Check if searching in right field (inputs vs outputs)
• Clear other filters that may exclude results

”Export timing out”

• Reduce result set with filters
• Try JSONL instead of JSON (streams better)
• Break into multiple exports by date range
• Contact support if exporting > 100k records needed

”Real-time updates not working”

• Check connection indicator (should be green “REALTIME”)
• Refresh page to reconnect WebSocket
• Verify no network/firewall issues
• Fallback polling should activate automatically

See also

• Platform Ledger API - API documentation
• Runs - View actions within specific runs
• Agents - View actions by agent
• Compliance - Export compliance reports