Start with Gemini CLI native telemetry for semantic questions
Current Gemini CLI telemetry includes gemini_cli.tool_call with function name, arguments, duration, success, decision, and MCP metadata; file-operation events for tool-performed creates, reads, and updates; API request/response signals; token metrics; and detailed GenAI spans when tracing is enabled. That makes native telemetry the direct source for tool identity and approval semantics.
File operations emitted by tools are not a full descendant-process inventory
Gemini's documented file events describe operations performed by tools. A shell tool can still launch compilers, package managers, or helpers that open many more files below the tool implementation. AgentSight is useful when you need to compare the semantic tool event with the observed process and path activity generated by that tool.
Node.js changes where the TLS plaintext hook lives
AgentSight documents Gemini CLI as a Node-based workflow and treats Node as an embedded-OpenSSL case for supported builds. The record path resolves the Node executable behind the CLI and can attach there instead of assuming a system libssl.so mapping. If multiple NVM or system Node installations exist, an explicit binary path can pin the runtime actually used.
Use system views for non-model network and resource questions
Gemini native telemetry is strong for Gemini API activity and the CLI's own metrics. A coding task can also contact package registries, Git hosting, local services, or remote tool endpoints. AgentSight's network and process views can help attribute those destinations to child commands and line them up with CPU/memory phases in the same run.
Correlate by prompt or tool phase rather than duplicating every field
When a native tool_call is slow, inspect the matching child process and resource interval. When a system path or destination is surprising, return to the corresponding Gemini tool call and decision. The combined workflow is strongest when the semantic event answers “what did Gemini choose?” and the system profile answers “what did that execution do?”
Review content logging settings before central export
Gemini telemetry can include prompts and, with detailed traces, large attributes such as tool outputs and file reads depending on configuration. AgentSight local capture can also contain sensitive prompts, responses, headers, paths, and network targets. Keep those data paths independently controlled and avoid exporting duplicate raw content unless the investigation requires it.
$ sudo agentsight record -- gemini
$ agentsight report serve
$ sudo agentsight record -c node