all docs
/ reference

Onboarding a Skill (Shadow → Prod)

The three skill modes (off, shadow, prod), what each does to a request, where its numbers land, and the four-step rollout from shadow to prod.

Each skill runs in one of three modes per developer key, set on the console's Optimization Skills page (/dashboard/settings). Onboard in shadow first and promote once the numbers justify it.

Mode Effect on your request Where the numbers land
off Not in the data path. Absent from telemetry.
shadow None — decides, records, forwards untouched. No risk. shadow block — counterfactual only.
prod Acts: cache hit served, prompt pruned, request rerouted, redacted, or refused. Realized totals — saved, blocked, cache_hit.

Counterfactual metrics

A shadow skill computes its full decision — cache lookup, routing pick, compaction plan, classifier score — then discards it, and the gateway stores the delta: tokens that would have been pruned, requests that would have hit cache, spend avoided, turns refused. Those dollars were still spent, so the figures project what promotion would save, never what was banked — they never join the realized totals.

Rollout

  1. Set the skill to shadow on your key and save — nothing changes.
  2. Let traffic accumulate.
  3. Read the projection on the console's Telemetry page, or per request in trace.stages[].
  4. Flip to prod — the skill starts acting.