Skip to main content

​
Overview

CelerFlow records every tool call an agent makes as a trace span. Traces are the foundation of the platform — they power the dashboard, agent scores, risk alerts, and permission auto-discovery.

​
Trace classification

Each trace is automatically classified into one of four categories:
CategoryExamplesDefault displayHighlight
Outbound writegmail_send, notion_create, slack_post✅ Shown🔴 Red
Outbound readgmail_read, web_fetch, web_search✅ Shown—
Local writewrite, edit, exec✅ Shown—
Local readread, memory_search❌ Hidden by default—
Users can toggle “Show all” to include local reads.

​
Service detection

Traces are automatically mapped to a service based on tool name patterns:
ServiceTool patternsCategory
Gmailgmail_*External
Notionnotion_*External
Slackslack_*External
GitHubgh_*, github_*External
Google Docsgdocs_*, google_docs_*External
Webweb_*External
System Toolsread, write, edit, exec, browser, memory_*System
Custom service definitions can be added from the dashboard.

​
Span structure

Each tool call produces two trace spans:
  1. Before span — recorded when the tool call starts, captures intent (tool name, parameters, policy result).
  2. After span — recorded when the tool call completes, captures result (status, duration, summary).
Both spans share a span_id to link them together.

​
Trace fields

FieldDescription
tool_nameThe tool that was called.
tool_args_summarySanitized parameters (truncated to 200 chars, no secrets).
categoryoutbound_write, outbound_read, local_write, or local_read.
service_nameDetected service (gmail, notion, system…).
statusallowed, blocked, confirmed, denied, or error.
result_summaryOne-line description of what happened.
duration_msExecution time in milliseconds.
task_nameSession or thread name for grouping.
task_group_idGroups related traces into a task.

​
Task grouping

Traces are grouped by task for easier browsing:
  • OpenClaw agents — grouped by session or conversation thread.
  • MCP Gateway agents — grouped by time window (tool calls within 5 minutes of each other).
Traces — Agent: product-manager
┌─────────────────────────────────────────────────────────┐
│ ▶ Task: "Write Phase 4 PRD"                    2h ago  │
│   12 operations · 3 outbound · 0 blocked               │
│                                                         │
│ ▼ Task: "Research MCP gateways"                4h ago  │
│   ├── 🔍 web_search "MCP gateway comparison"           │
│   ├── 🔍 web_fetch  composio.dev/...                   │
│   ├── 🔴 gmail_send → team@celerflow.ai                │
│   └── 📝 write shared/phase4/research.md                │
└─────────────────────────────────────────────────────────┘

​
Filtering

DimensionValues
CategoryOutbound write / Outbound read / Local write / Local read
ServiceGmail / Notion / Slack / GitHub / System / …
StatusAllowed / Blocked / Confirmed / Denied / Error
Time rangeCustom date picker
Task nameText search

​
Retention

PlanRetention
Free3 days
Pro30 days
Team90 days
EnterpriseUp to 365 days