validator.yaml

# Validator Workflow
#
# Demonstrates the `validator:` block (issue #220): a second LLM call that
# grades a provider-backed agent's output against a rubric, and re-runs the
# agent once with feedback when it fails.
#
# This is distinct from the two correctness mechanisms Conductor already has:
#   - `retry:`  handles *transient* failures (provider errors, timeouts) by
#               re-running the same prompt unchanged.
#   - `output:` schema handles *shape* failures (missing field, wrong type).
#
# The validator catches the case neither does: a structurally valid output
# that is semantically wrong, incomplete, or off-rubric — here, a code review
# that misses a real bug or fabricates a function name.
#
# Mechanics:
#   1. `code_reviewer` produces a review of the diff.
#   2. The validator runs a second LLM call checking the review against
#      `criteria`. It returns {passed, issues}.
#   3. If it fails, `code_reviewer` re-runs ONCE with the validator's issues
#      appended under a "## Validation feedback" section. The second output is
#      taken as final (no second validation loop).
#
# Validator token cost shows up as a separate `code_reviewer (validator)` row
# in the usage summary, and validator activity is surfaced in the web
# dashboard and `--verbose` console output.
#
# Usage:
#   conductor run examples/validator.yaml
#   conductor run examples/validator.yaml --input diff="$(git diff HEAD~1)"

workflow:
  name: validator
  description: Code review with a semantic validator that re-runs once on failure
  version: "1.0.0"
  entry_point: code_reviewer

  runtime:
    provider: copilot

  input:
    diff:
      type: string
      required: false
      description: A unified diff to review
      default: |
        --- a/auth.py
        +++ b/auth.py
        @@ -10,6 +10,9 @@ class TokenStore:
        -    def get(self, key):
        -        return self._cache[key]
        +    def get(self, key):
        +        entry = self._cache.get(key)
        +        return entry.value

agents:
  - name: code_reviewer
    description: Reviews a diff for bugs and suggests concrete fixes
    model: claude-sonnet-4-5
    prompt: |
      Review the following diff for bugs, with a focus on null-safety and
      correctness. For each problem, give a concrete, actionable suggestion.

      Diff:
      {{ workflow.input.diff }}
    output:
      summary:
        type: string
        description: One-line summary of the review
      issues:
        type: array
        description: List of problems found, each with an actionable suggestion

    # The validator runs a SECOND LLM call after code_reviewer completes,
    # grading its output against the criteria below. On failure the agent is
    # re-run once with the validator's feedback appended.
    validator:
      # Optional: validate with a cheaper/different model than the primary.
      # Defaults to the primary agent's model when omitted.
      model: claude-sonnet-4-5
      criteria: |
        Verify the review:
        1. Identifies the null-safety issue in the diff (`entry` may be None
           when the key is missing, so `entry.value` can raise AttributeError).
        2. Each entry in `issues` includes a concrete, actionable suggestion
           (not just "fix it").
        3. Does not reference any function or attribute names that do not
           appear in the diff (no fabrication).
      # Hard-capped at 1. Use 0 to validate-and-report without re-running.
      max_retries: 1
    routes:
      - to: $end

output:
  summary: "{{ code_reviewer.output.summary }}"
  issues: "{{ code_reviewer.output.issues }}"