cli: retire generate in favor of skill-based eval authoring

[christso]

Objective

Retire the top-level agentv generate CLI surface and move grader/assertion authoring guidance into skill-based eval authoring. Keep the skill segment short and avoid over-specifying implementation choices so the AI can choose the right mix of LLM graders, deterministic graders, or simple assertions for the stated criteria.

Architecture Boundary

Primary boundary: external-first eval-authoring workflow and docs. Keep core/runtime changes minimal and focused on removing a redundant CLI entrypoint plus any compatibility/deprecation behavior needed for migration.

Design Latitude

• Preferred direction: expand agentv-eval-writer rather than introducing another narrow skill.
• Keep the skill guidance compact. It should frame the available authoring tools and tradeoffs, not prescribe one fixed grader recipe.
• The migration can use a staged deprecation if immediate removal would create too much churn.
• Keep the user-facing authoring story coherent across CLI help, docs, examples, and skill guidance.

Acceptance Signals

• agentv generate is removed or clearly deprecated with a migration path to the supported skill-based authoring workflow.
• agentv-eval-writer (or a justified replacement skill) covers eval authoring guidance well enough that users do not need the CLI subcommand for grader/assertion generation.
• Docs stop presenting agentv generate rubrics as the default path and instead point users to the supported authoring workflow.
• Error/help text that currently points users at agentv generate rubrics is updated to reference the new workflow.
• Tests/docs/examples are updated so there is one clear supported path for authoring evaluation criteria.

Non-Goals

• Reworking evaluator internals.
• Redesigning the full eval file schema.
• Forcing one evaluator style when multiple styles fit the test criteria.
• Expanding scope into broader skill marketplace or trigger-quality work.

Context

• apps/cli/src/commands/generate/index.ts currently exposes only a single namespace: generate rubrics.
• Docs currently duplicate this path in CLI/web docs, while eval authoring guidance already exists in agentv-eval-writer and the skill-improvement docs.
• This makes grader/assertion authoring feel like core product surface area even though it is closer to assisted eval authoring.

Related

• None identified.