Description

Command telemetry records whether --attach was used, but not how many times it was provided. An invocation with several attachments is therefore indistinguishable from one with a single attachment.

This adds an attachment_invocation event to issue and pull request create, edit, and comment commands. The event includes the command path and an attach_count measure. Invocations using --attach are sampled at 100%, which also captures the paired command_invocation event at the same rate.

The event is recorded before attachment validation. It captures the supplied flag count when a conflict, missing file, or other validation error stops the command. It does not include file paths, alt text, body content, repository information, or asset data.

How did you test this change?

I set telemetry to log in an isolated configuration and ran all six supported commands with two missing image paths. Each command stopped before upload and printed exactly one attachment_invocation event with the corresponding command, attach_count set to 2, and sample_rate set to 100.

I then ran gh version without --attach and saw no attachment event.

Key points

• The attachment flag owns its telemetry event. Generic command telemetry remains unaware of attachment behavior.
• The recorder is passed explicitly from the root command through the issue and pull request command sets, following the existing skills telemetry pattern.
• Sampling applies to the whole invocation, so an attachment invocation sends both the attachment event and standard command event at 100%.
• The count describes flags provided, not files validated or uploaded.

Notes for reviewers

Start with the attachment flag telemetry method. The remaining production changes pass the recorder to each command. Assertions are integrated into the existing command constructor tables.

This follows the dependency injection and domain-specific event approach established by skills-specific telemetry.

Authorship and follow-up

Who wrote this:

• A human wrote it.
• An agent wrote it under close human direction.
• An agent wrote it independently, and no human has guided the implementation beyond the initial prompt.

Who answers review comments:

• @BagToad will read and reply directly. Name the account.
• An agent will draft replies and @username will read them before they are posted.
• Nobody has explicitly committed to replying.