portfolio/hats.yml

cli:
  backend: "codex"

event_loop:
  prompt_file: "Ralph/PROMPT.md"
  starting_event: "work.start"
  completion_promise: "LOOP_COMPLETE"
  max_iterations: 60

backpressure:
  gates:
    - name: "lint"
      command: "npm run lint"
      on_fail: "retry"
    - name: "typecheck"
      command: "npm run typecheck"
      on_fail: "retry"
    - name: "build"
      command: "npm run build"
      on_fail: "retry"

hats:
  planner:
    name: "Content Refactor Planner"
    description: "Plans one rollout stage at a time for centralizing all descriptive text into a single editable source."
    triggers: ["work.start", "review.changes_requested"]
    publishes: ["plan.ready"]
    memory:
      path: ".ralph/agent/memories.md"
      scope: "global"
    instructions: |
      You are the Planner. Read PROMPT.md first.

      Terminal rule (run this first):
      - If PROMPT.md already shows all rollout stages complete and contains LOOP_COMPLETE,
        print LOOP_COMPLETE immediately.
      - Do NOT emit plan.ready for verification-only or closure-only passes.
      - Do NOT create additional backpressure-recovery tasks when no unchecked stage exists.

      If triggered by review.changes_requested, read .ralph/review.md and incorporate feedback.

      Your job:
      1. Identify the NEXT unchecked rollout stage in PROMPT.md.
      2. Inspect the codebase and map only the files needed for that stage.
      3. Write/update .ralph/plan.md with:
         - stage name and objective
         - explicit file list with planned edits
         - migration approach that minimizes breakage
         - compatibility strategy (temporary adapters/re-exports if needed)
         - rollback-safe checkpoints
      4. Keep scope to one stage per iteration.
      5. Emit plan.ready.

      Planning only. Do not modify source files.

  builder:
    name: "Content Refactor Builder"
    description: "Implements the current stage, centralizes text content, and preserves behavior."
    triggers: ["plan.ready"]
    publishes: ["build.done"]
    memory:
      path: ".ralph/agent/memories.md"
      scope: "global"
    instructions: |
      You are the Builder. Read PROMPT.md and .ralph/plan.md.

      Terminal rule:
      - If planner signaled completion or PROMPT.md is already fully complete, print LOOP_COMPLETE.
      - Do not emit build.done for verification-only closure cycles.

      Your job:
      1. Implement ONLY the currently planned stage.
      2. Centralize descriptive/profile text into the single source defined by PROMPT.md.
      3. Update consumers for that stage to read from centralized content.
      4. Preserve runtime behavior and existing interactions.
      5. Run quality checks after meaningful changes:
         - npm run lint
         - npm run typecheck
         - npm run build
      6. Mark completed stage checkboxes in PROMPT.md.
      7. Emit build.done when the stage is complete and checks pass.

      Backpressure payload format requirement for build.done:
      - Include these exact evidence fields in plain text:
        - tests: pass
        - lint: pass
        - typecheck: pass
        - audit: pass
        - coverage: pass
        - complexity: <score>
        - duplication: pass
      - For unconfigured checks, still use `pass` and append `(not-configured)`.
      - Keep field names lowercase and exact.

      Constraints:
      - Keep TypeScript strictness intact.
      - Do not rewrite unrelated logic.
      - Prefer incremental migration with compatibility exports where useful.
      - Avoid duplicate text sources after each stage is completed.

  reviewer:
    name: "Content Refactor Reviewer"
    description: "Validates each stage against requirements and requests focused rework when needed."
    triggers: ["build.done"]
    publishes: ["review.changes_requested"]
    memory:
      path: ".ralph/agent/memories.md"
      scope: "global"
    instructions: |
      You are the Reviewer. Read PROMPT.md, .ralph/plan.md, and current source changes.

      Terminal rule:
      - If all stage checkboxes in PROMPT.md are complete and success criteria are complete,
        print LOOP_COMPLETE immediately and do not emit review.changes_requested.
      - Do not request verification-only recovery work after completion.

      Validate in this order:
      1. Run gates:
         - npm run lint
         - npm run typecheck
         - npm run build
         All must pass.
      2. Confirm the stage objective was fully delivered.
      3. Confirm migrated files now read from centralized content instead of hardcoded/duplicated text.
      4. Confirm no behavior regressions for navigation, detail panels, search/chat context, and timeline/constellation wiring.
      5. Confirm PROMPT.md status reflects reality.

      Decision rules:
      - If current stage is incomplete or quality fails: write actionable fixes to .ralph/review.md and emit review.changes_requested.
      - If current stage is complete but more stages remain: note approval in .ralph/review.md and emit review.changes_requested for the next stage.
      - If all stages and success criteria are complete: write final approval to .ralph/review.md and print LOOP_COMPLETE.

      Circuit breaker:
      - If the same blocker class repeats across 3 consecutive review cycles with materially identical evidence,
        stop retrying. Record blocker/evidence in .ralph/review.md with status "needs-human",
        assign owner + target date, and request human clarification before further loop progress.