What problem does it solve? LLM spend routed through the Caveman gateway lands in one anonymous unlabeled-workflow bucket, making it impossible to attribute cost to the jobs the code actually performs. This Skill finds every LLM workflow in a repository, names each one with a valid slug, and wires the label into the callsite so spend is grouped by workflow. ## Core Features & Use Cases - Workflow inventory: Walks entry points (HTTP handlers, cron jobs, queue consumers, CLI scripts, eval harnesses, agent definitions) to find every job that calls an LLM. - Slug naming and labeling: Proposes gateway-valid slugs ([a-z0-9_-], 1-96 chars) and wires labels via SDK options, x-cave-workflow headers, caveman wrap --workflow flags, or raw HTTP headers. - Proposal-first workflow: Presents a labeling table for user approval before changing code, stays idempotent on re-runs, and verifies one labeled path after applying. - Use Case: A service with a support-reply bot, a nightly digest cron, and an eval script all calling OpenAI through the gateway gets three labeled workflows so the dashboard shows spend per job instead of one bucket. ## Quick Start Ask the assistant to discover the LLM workflows in this repository and propose a labeling table for Caveman Cloud.
