cx-coding-agents

Coding Agents Skill

Use this skill as the entry point for any investigation or data question about AI Center Coding Agents. It identifies which data source each agent uses and tells you which reference files to load before querying.

Supported Agents

Loading References

Load the agent reference first, then the shared query-language files it depends on:

Safety

All query commands (cx metrics, cx logs, cx spans, cx dataprime) are read-only. They never modify data and can be run freely without --yes.

CLI Commands

Output format: append -o toon to any command when passing results to an agent or further processing.

Multi-profile: use -p <profile> (repeatable) to fan out across profiles simultaneously.

First Response

Identify the requested agent and analysis goal. If any required scope is missing, ask before querying:

• Agent: Claude Code, Codex, Cursor, Gemini CLI, or Copilot CLI.
• Time range.
• Optional filters: application, subsystem, user, model, session, repository.
• Output shape: quick answer, table, trend, top-N ranking, investigation notes, or query only.

If the user says "GitHub Copilot Codex", clarify whether they mean Codex or Copilot CLI. AI Center has separate Codex and Copilot CLI dashboards.

Use the same data source as the corresponding dashboard whenever possible:

• Claude Code metrics: PromQL over claude_code_* metrics.
• Claude Code session text: DataPrime over ai_sessions_claude logs.
• Codex: DataPrime over Codex logs for tokens, sessions, models, users, and tools; spans only for run_turn latency.
• Cursor: DataPrime over Cursor spans. Cursor reports prompt length and file edits, not provider token counts.
• Gemini CLI: PromQL over gemini_cli_* metrics.
• Copilot CLI direct OTel: DataPrime over Copilot spans.
• GitHub Copilot Collector: PromQL over github_copilot_* org/user/billing metrics when that collector data is present.

Time and Filters

PromQL (Claude Code, Gemini CLI, GitHub Copilot Collector)

Use increase(metric[<range>]) for counters over the selected window. Common label filter pattern:

{user_email="<user>",model="<model>",cx_application_name="<app>",cx_subsystem_name="<subsystem>"}

• Use sum by (...), count by (...), or topk(N, ...) for grouping.
• For chart trends, use an interval window based on the selected range rather than the full range.

DataPrime (Codex, Cursor, Copilot CLI direct, Claude Code sessions)

• Add | filter ... clauses for user, model, application, subsystem, session, and operation.
• Prefer groupby ... aggregate ... for dashboard-like tables.
• For spans: app/subsystem labels are $l.applicationName and $l.subsystemName (mixed case).
• For logs: app/subsystem labels are $l.applicationname and $l.subsystemname (lowercase).

App/subsystem filter pattern:

| filter $l.applicationName == '<application>' && $l.subsystemName == '<subsystem>'

Cross-Agent Questions

When comparing agents:

• Normalize units before comparing: cost in USD, tokens as input/output/total, runtime in milliseconds, sessions as distinct conversations.
• State which signals are not equivalent. Cursor prompt length is not a provider token count. Direct Copilot CLI users are pseudonymous IDs (enduser.pseudo.id), not email addresses. GitHub Copilot Collector users may be logins, names, or emails.
• Prefer per-agent sections over a single merged table when dimensions differ.
• For "which agent is most used": compare sessions and users first, then tokens or span counts with caveats.
• For "which agent costs the most": Claude Code has first-class USD metrics. Copilot direct spans expose github.copilot.cost. Collector billing metrics expose net/gross/discount amounts by organization/SKU. Do not merge Copilot span cost and collector billing without explaining the source and unit difference.

Answer Style

Always include:

• The agent and data source used.
• The query or a compact query excerpt.
• The time range and filters applied.
• A plain-language interpretation of the result.
• Caveats about empty data, approximate counts, pseudonymous users, missing labels, or non-comparable metrics.

Do not invent fields not listed in the agent reference files. If the user asks for data an agent does not emit, say what is available and propose the nearest supported query.