Telemetry Sanitized Skill Invocation

SKILL: Telemetry Sanitized Skill Invocation

Domain: tool-orchestration
Trigger: Use when SkillTool emits tengu_skill_tool_invocation so built-in commands, remote MCP skills, and forked executions are logged without exposing command strings in open analytics tables. Source Pattern: Distilled from reviewed tool-loop and result-shaping patterns.

Core Method

Before executing any skill, compute whether the command is built-in, bundled, official marketplace, or custom and pass the sanitized value as command_name while storing the actual skill name under _PROTO_skill_name to satisfy both privacy and auditability. Add contextual metadata such as execution_context (inline, forked, remote), query_depth, parent_agent_id, and plugin fields when applicable, and wrap remote canonical invocations or discovered skills with additional flags so downstream dashboards know the skill source without revealing user-generated text. Emit separate telemetry for inline/forked/remote flows so metrics can split by execution path while normalizing the skill name for aggregated reporting.

Key Rules

• Derive sanitizedCommandName as commandName for built-in/bundled/official skills and fall back to 'custom' for user-authored commands; log the actual name only in _PROTO_skill_name to keep BigQuery columns safe.
• Always include execution_context ('inline', 'forked', 'remote') plus invocation_trigger (nested vs proactive) and query_depth so telemetry consumers understand the runtime scenario.
• When a plugin skill is involved, add both _PROTO_plugin_name and the sanitized plugin_name/plugin_repository fields along with marketplace metadata to enable plugin-level analytics without leaking repository URLs.
• For remote MCP skills, emit is_remote, was_discovered, and cached load info (latency, cache hit) so telemetry can diagnose discovery performance without exposing the raw prompt body.

Example Application

Use this skill to log any slash command invocation, ensuring that a cloned forked review or remote_skill call still records an execution_context flag plus proper naming sanitization for consumption by privacy-sensitive dashboards.

Anti-Patterns (What NOT to do)

• Do not log user-entered command text directly into public telemetry columns; that breaks PII rules and makes analytics unusable.
• Do not merge inline/forked/remote flows in a single metric stream without flags, because it becomes impossible to trace delays or diagnostics per execution context.