# Sample deployment overlay — for trying the `config` eval run type.
#
# Modeled on a real Last Light instance overlay (~/work/lastlight-instance),
# trimmed to what the eval harness actually consumes. It's merged over core's
# `config/default.yaml` exactly as production does (maps deep-merge per key).
#
# Run it:
#   lastlight-evals run code-fix --mode config --overlay examples/overlay
#   lastlight-evals run triage   --mode config --overlay examples/overlay
#
# NOTE: the eval reads ONLY `models` and `variants` from here — it forces
# sandbox:none and disables approval gates (the two sanctioned deviations from
# prod), so any `sandbox` / `approval` / `managedRepos` keys would be ignored.
# That's why this sample omits them.

# Per-step model map. Keys are workflow task names; anything unset falls back to
# `default`. This deliberately VARIES per phase so the dashboard's "Per-step
# models" panel shows the assignment: cheap/fast model on the gate + review
# phases, a stronger model where the real work happens (architect/executor/fix).
#
# These ids must resolve in pi-ai's registry and have their provider key set
# (here: OPENAI_API_KEY). Swap to your available provider as needed — e.g.
# anthropic/claude-haiku-4-5 + anthropic/claude-opus-4-8.
models:
  default: openai/gpt-5.4-mini
  guardrails: openai/gpt-5.4-mini
  architect: openai/gpt-5.5
  executor: openai/gpt-5.5
  reviewer: openai/gpt-5.4-mini
  fix: openai/gpt-5.5

# Reasoning-effort per task (pi-ai "thinking": off|low|medium|high|xhigh).
# Empty/unset ⇒ the model's default effort.
variants:
  guardrails: low
  reviewer: low