feat(analytics): issue task IDs from the CLI

[angeloashmore]

Resolves:

Description

Before this PR, an agent that ran a command without --analytics-task-id was refused and told to generate a UUID itself with an inline node -e command. Agents read that as an instruction injected through a failed command. In one recorded run, Claude Code refused to continue and told the developer the CLI could not be trusted.

After this PR, the CLI issues the IDs. prismic task-id prints one ID for one user request, and the refusal names that command instead of dictating a recipe. The options are --task-id and --user-intent, so nothing an agent types on every command reads as telemetry.

Everything here is agent-only, keyed on detectAgent(): the task-id command, both options, the AGENTS section of the help, and the refusal itself. A human sees no new command, no new options and no refusal. The old --analytics-* names keep working for both, hidden from help.

The order matters: release this to npm before prismicio/skills#12 merges, or agents are told to run a command their installed CLI does not have.

Checklist

• If my changes require tests, I added them.
• If my changes affect backward compatibility, it has been discussed.
• If my changes require an update to the CONTRIBUTING.md guide, I updated it.

Preview

$ npx prismic@pr-316 whoami                   # as a human
Not logged in. Run `prismic login` first.

$ CLAUDECODE=1 npx prismic@pr-316 whoami      # as an agent
error: missing --task-id
run `prismic task-id` once per user request, then pass --task-id <id> --user-intent "<what the user asked for>" on every command until the user asks for something else

$ npx prismic@pr-316 task-id
pt_sewgrz7vydac0ksb

How to QA ¹

Run a command with an agent environment variable set, for example CLAUDECODE=1 prismic whoami, and follow the refusal. Run the same command without one to see nothing change.

🤖 Generated with Claude Code

https://claude.ai/code/session_01ANFuYqfG55YAJUNNREVgE6

---

Note

Medium Risk
Changes agent-only CLI entry validation and analytics option names; legacy flags remain, but agents must adopt the new id format and command flow.

Overview
Agents get CLI-issued task IDs instead of being told to mint UUIDs themselves. New hidden command prismic task-id prints a pt_… id from genTaskId(); agent runs must pass --task-id and --user-intent (replacing --analytics-*, which still parse but are hidden from help).

Validation now requires the pt_ format, so arbitrary UUIDs or placeholders fail. Refusal text points at prismic task-id rather than an inline node -e recipe; task-id itself is exempt from the check.

Evals now assert one task id per user request (not one id shared across unrelated requests), with and without the Prismic skill, and bump the skill ref for prismic task-id docs.

^{Reviewed by Cursor Bugbot for commit afa7171. Bugbot is set up for automated code reviews on this repo. Configure here.}

---

Generated by Claude Code

Footnotes

1. Please use these labels when submitting a review:
   ❓ #ask: Ask a question.
   💡 #idea: Suggest an idea.
   ⚠️ #issue: Strongly suggest a change.
   🎉 #nice: Share a compliment. ↩

[levimykel]

Pre-approved with a few comments.