agent-invocation-tracking

name	agent-invocation-tracking
description	Emit OTEL spans when an agent starts and completes its task. Use when instrumenting an agent to measure duration, success rate, and invocation patterns.
license	MIT
compatibility	Claude Code, GitHub Copilot, OpenCode, Cursor, Codex, Gemini CLI
metadata	{"author":"paruff","suite":"uFawkesAI"}

Skill: Agent Invocation Tracking

Load trigger: "load agent-invocation-tracking skill" > DORA: Cap 2 (Observability) Token cost: Low

Purpose

Track when agents start, complete, and how long they take. Enables answering: "Which agents are used most? Which take longest? Which fail most often?"

Span Specification

agent.invocation.started

Emit when the agent begins processing.

Attributes:

Attribute	Type	Description
agent.name	string	Agent identifier (matches filename without .md)
session_id	string	Unique invocation identifier
mode	string	Operational mode (if agent supports multiple)
trigger	string	What triggered this invocation (human request, CI, direct invocation)
input_files	string[]	Files read as context for this invocation

agent.invocation.completed

Emit when the agent finishes.

Attributes (all from started, plus):

Attribute	Type	Description
duration_ms	integer	Wall-clock duration in milliseconds
total_skills_loaded	integer	Count of skills loaded during invocation
total_findings	integer	Count of findings produced
decision	string	Final decision (PASS/FAIL/BLOCKED/APPROVE/REQUEST_CHANGES/ESCALATE)

agent.invocation.failed

Emit when the agent encounters an error it cannot recover from.

Attributes (all from started, plus):

Attribute	Type	Description
error	string	Error message or summary
stage	string	Stage where failure occurred (validate, execute, report)

Implementation

After completing the task, write the Phase 0 log entry (.agents/logs/YYYY-MM-DD.jsonl). The telemetry spans can be derived from the log data by the aggregation script.