Data Collected

What we do NOT collect

No prompt or response content

We do not collect the content of your conversations — no prompts, no model responses, no raw API bodies, and no tool results/output. Content capture is disabled in every supported tool: log_user_prompt = false for Codex, captureContent: false for Copilot in VS Code, OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=false for Copilot CLI and the macOS desktop app, and for Claude Code OTEL_LOG_USER_PROMPTS=0, OTEL_LOG_ASSISTANT_RESPONSES=0, OTEL_LOG_RAW_API_BODIES=0, and OTEL_LOG_TOOL_CONTENT=0.

One exception applies to Claude Code tool inputs — see the caution below.

What we do collect

Metadata about each AI call — enough to measure usage, not to reconstruct the conversation: model, token counts, cost, latency, your enduser.id, and which tools and skills were invoked.

Tool inputs are collected for Claude Code

Claude Code runs with OTEL_LOG_TOOL_DETAILS=1, so in addition to tool names and outcomes it records tool input parameters — bash command strings, file paths, search patterns, URLs, skill / MCP / workflow / subagent names, and JSON-serialized tool arguments (individual values truncated at ~512 characters, full payload bounded to ~4 KB). It does not record tool results/output (OTEL_LOG_TOOL_CONTENT=0).

Because those arguments are captured, they can themselves contain snippets of file content or prompt-like text — for example the content written by a Write, the old_string/new_string of an Edit, or the prompt passed to a delegated agent. The truncation limits above bound how much is captured, but short files or snippets can be recorded in full. Codex and Copilot do not emit this tool-parameter detail — they report call-level metadata only.

Field	Meaning	Example
service.name	Which assistant emitted the data	claude-code, codex, copilot-chat
gen_ai.system	Model vendor	anthropic, openai, github
gen_ai.request.model	Model used	claude-sonnet-4-6, gpt-5-codex, gpt-4o
gen_ai.operation.name	Kind of call	chat
enduser.id	Who (your Aliz email)	jane.doe@aliz.ai
gen_ai.usage.input_tokens	Prompt/context tokens	1843
gen_ai.usage.output_tokens	Generated tokens	412
cost	Spend for the call (Claude/Codex); Copilot reports a premium-request multiplier, not USD	0.0091
duration	Call latency (ms)	2140
tool / skill usage (Claude Code)	Tool/skill name, outcome, and input parameters (bash commands, file paths, patterns, URLs, args) — but not results/output	tool Bash ran git status; skill code-review invoked
timestamp	When the call happened	2026-07-29T09:14:22Z

The collector normalizes these into shared gen_ai.* metrics so the three providers can be compared side by side.

Sample records

These are representative of what lands in BigQuery. Note there is no prompt or response text — but for Claude Code, tool events carry the tool input (see the second record below).

Claude Code

Claude Code emits log events (e.g. api_request) plus tool-usage events. With OTEL_LOG_TOOL_DETAILS=1, tool events include the input parameters:

{
  "service.name": "claude-code",
  "gen_ai.system": "anthropic",
  "event.name": "api_request",
  "gen_ai.request.model": "claude-sonnet-4-6",
  "enduser.id": "jane.doe@aliz.ai",
  "gen_ai.usage.input_tokens": 1843,
  "gen_ai.usage.output_tokens": 412,
  "cost_usd": 0.0091,
  "timestamp": "2026-07-29T09:14:22Z"
}

{
  "service.name": "claude-code",
  "event.name": "tool_result",
  "tool_name": "Bash",
  "tool_input": "{\"command\":\"git status\",\"description\":\"Show working tree status\"}",
  "success": true,
  "enduser.id": "jane.doe@aliz.ai",
  "timestamp": "2026-07-29T09:14:25Z"
}

OpenAI Codex

Codex emits spans; the LLM details live on the sampling-request span:

{
  "service.name": "codex",
  "gen_ai.system": "openai",
  "span.name": "run_sampling_request",
  "gen_ai.operation.name": "chat",
  "gen_ai.request.model": "gpt-5-codex",
  "enduser.id": "jane.doe@aliz.ai",
  "gen_ai.usage.input_tokens": 2210,
  "gen_ai.usage.output_tokens": 587,
  "duration_ms": 3120,
  "timestamp": "2026-07-29T09:20:03Z"
}

GitHub Copilot

Copilot (VS Code, CLI, and the macOS desktop app) emits chat spans, from which the collector derives gen_ai.calls and gen_ai.duration:

{
  "service.name": "copilot-chat",
  "gen_ai.system": "github",
  "gen_ai.operation.name": "chat",
  "gen_ai.request.model": "gpt-4o",
  "enduser.id": "jane.doe@aliz.ai",
  "gen_ai.usage.input_tokens": 1502,
  "gen_ai.usage.output_tokens": 240,
  "duration_ms": 1875,
  "timestamp": "2026-07-29T09:26:41Z"
}

Copilot CLI and the tested macOS desktop runtime both emit service.name=github-copilot; the dashboard's copilot-cli label covers both. See the attribution limitation for the verification scope.

note

Exact field names vary a little by provider and by the version of each tool. The collector maps provider-specific fields onto the shared gen_ai.* convention so they can be queried uniformly in BigQuery. The samples above are illustrative of shape and content, not a byte-for-byte schema.