# ralph.yml — goal-acceptance as the loop's completion gate.
#
# Ralph iterates an agent until it emits LOOP_COMPLETE. This config makes
# goal-acceptance the "definition of done": the agent manages criteria via
# the goal-acceptance MCP server while it works, and the pre.loop.complete
# hook hard-blocks any LOOP_COMPLETE claim the criteria don't back.
#
# Setup (once per workspace):
#
#   1. Share ONE data dir between the agent's MCP server and the hooks.
#      Set it on the `ralph run` process so hooks inherit it:
#        PowerShell:  $env:PLUGIN_DATA = "$PWD\.ralph\goal-acceptance"
#        bash:        export PLUGIN_DATA="$PWD/.ralph/goal-acceptance"
#
#   2. Register the MCP server for the agent's backend with the SAME dir.
#      Claude backend → .mcp.json in this workspace:
#        {
#          "mcpServers": {
#            "goal-acceptance": {
#              "command": "goal-acceptance",
#              "args": ["mcp"],
#              "env": { "PLUGIN_DATA": "<same absolute dir as step 1>" }
#            }
#          }
#        }
#      (`goal-acceptance install --target claude` writes the entry; add the
#      env block so both sides read the same store.)
#
#   3. ralph run -c ralph.yml -p "Implement <feature>; use goal-acceptance
#      to track the acceptance criteria."

event_loop:
  completion_promise: "LOOP_COMPLETE"
  max_iterations: 50

cli:
  backend: "claude"

core:
  guardrails:
    - "Track the work with the goal-acceptance MCP tools: set criteria first, then validate each criterion with real evidence."
    - "Do not emit LOOP_COMPLETE until every required criterion is formally passed; self-claimed passes need confirm_criterion by a reviewer."
    - "The goal-acceptance gate hook rejects premature LOOP_COMPLETE claims and the loop continues."

hooks:
  enabled: true
  events:
    # Optional visibility: log the current goal state when the loop starts.
    post.loop.start:
      - name: goal-acceptance-status
        command: ["goal-acceptance", "status"]
        on_error: warn
    # The hard gate: `goal-acceptance check` exits 1 while required criteria
    # are unmet, which blocks the completion and keeps the loop going.
    pre.loop.complete:
      - name: goal-acceptance-gate
        command: ["goal-acceptance", "check"]
        on_error: block