sample_run.yaml

# Sample run-spec for the cross-phase orchestrator.
#
#   python -m evolution.orchestrator --spec examples/orchestrator/sample_run.yaml --dry-run
#
# Phases run in the order listed (canonical: skills -> tools -> prompts -> code).
# Each phase's `args` keys are that phase's own CLI flag names in snake_case
# (e.g. `eval_source` -> --eval-source); the orchestrator supplies --output-dir
# itself for deterministic verdict capture, so don't set it here.
#
# Propose-only: nothing deploys, and no PRs open unless you both set `create_pr`
# on a phase AND pass --allow-pr. A create_pr phase must also carry its spend
# ceiling (skills/tools: max_total_cost_usd; prompts has no PR path).

# Optional. Merged into every phase (a phase's own args win on conflict).
# Caveat: a default key must be a valid flag for EVERY phase it reaches, so keep
# it empty unless you're running a subset that shares the flag (--only).
defaults: {}

phases:
  - phase: skills
    name: web_research
    args:
      eval_source: golden
      iterations: 8
      max_total_cost_usd: 50

  - phase: tools
    name: fetch_url
    args:
      manifest: manifests/fetch_url.json
      iterations: 5
      max_total_cost_usd: 50

  - phase: prompts
    name: tool_selection
    args:
      tasks: datasets/prompts/tool_selection.jsonl
      target: hermes

  - phase: code
    name: hermes/tools/fetch_url.py
    args:
      repo: /path/to/hermes-checkout
      visible_test: tests/test_fetch_url_visible.py
      holdout_test: tests/test_fetch_url_holdout.py
      repair_rounds: 5