LRLoops RadarSubmit loop

Pipeline loops

automatic-loop

>

Source recipe

Copyable Markdown

# automatic-loop

## Purpose

>

## Source

- Source: Pi Pipelines
- Repository: Rybens92/pi-pipelines
- Path: pipelines/automatic-loop.pipeline.yaml

## Workflow

- analyze
- converge
- execute
- fix-gaps
- synthesize

## Verifier

Includes review gates or goal-loop checks in the source pipeline.

## Source Artifact

```yaml
name: automatic-loop
description: >
  Universal automatic-looping pipeline for arbitrary project tasks. Classifies the
  task, extracts a testable Definition of Done and a holistic success goal, then
  converges inside a goal-loop: each iteration executes the work under a scored
  review gate, fixes the reviewers' critiques under a second gate, and a critic
  decides whether the whole mission is achieved. Combines per-stage scored gates
  (quality bar) with a goal-loop (holistic mission convergence). Best-effort on
  exhaustion — never hard-fails on iteration/round limits.
preflight:
  topics:
    - type of task and expected output (research / planning / implementation / refactoring / bug-fix / review / other)
    - quality requirements (expected level of thoroughness, specific constraints or boundaries)
    - project context and affected files/modules
report: false


stages:
  - id: analyze
    agent: planner
    tier: big
    task: >
      Analyze the following query about this project: {task}

      Produce a structured analysis with:

      1. **Task type** — classify as one of:
         research, planning, implementation, refactoring, bug-fix, code review, or other.

      2. **Definition of Done** — explicit, testable criteria that verify
         successful task completion. Each criterion must be written so a
         reviewer can definitively say "yes, met" or "no, not met".

      3. **Affected files or modules** — what areas of the project will
         be touched or researched.

      4. **Edge cases and risks** — what could go wrong, what to watch for,
         what unusual scenarios the executor should handle.

      5. **Reviewer focus guidance** — what aspects the reviewers should pay
         special attention to when scoring the output.

      6. **Holistic success goal** — 1-2 sentences describing what "the whole
         mission is accomplished" means for this task. It must be a single,
         coherent success criterion that a critic can verify as met or not met
         (not a checklist). It is the high-level bar the goal-loop converges on,
         complementing the detailed DoD checklist above.

      Be specific and concrete. The DoD criteria and the holistic goal together
      define what success means. The DoD is used by the gate reviewers; the goal
      is used by the goal-loop critic.

      ### Output format (required)
      Focus on producing: (1) a numbered Definition-of-Done checklist, (2) the affected
      files/modules as a bullet list, (3) edge cases/risks as a bullet list, and
      (4) the one-sentence holistic success goal. Keep the whole output under 500 words — it is
      injected verbatim into the worker stages.
    report:
      mode: summary
      maxLength: 700
      instruction: >
        Distill to: the Definition-of-Done checklist, affected files, edge cases,
        and the one-sentence holistic goal. Drop any narrative or implementation plan.

  - id: converge
    goal-loop:
      maxIterations: 3
      goal:
        derive: true
        hint: >
          Extract the "Holistic success goal" from item 6 of the analysis.
          If the analysis does not have an explicit item 6 or the goal is
          absent, derive a single testable success criterion from the
          Definition of Done (item 2). The criterion must be verifiable by
          a critic reading the task output and/or the changed files.
        context:
          - task
          - outputs.analyze
      critic:
        agent: oracle
        tier: big
    stages:
      - id: execute
        agent: worker
        tier: small
        task: >
          Execute the following task for the project: {task}

          Goal-loop iteration {iteration}.
          Goal: {goal}
          Prior critic feedback: {goalFeedback}

          ### Analysis & DoD

          {outputs.analyze}

          Follow the Definition of Done criteria carefully — address each one
          completely. Keep changes focused and follow project conventions.
          If you discover something unexpected, handle it cleanly and note it.

          ### Editing is required
          You MUST create or edit files now — do not return a plan or a summary.
          Make the actual changes to the codebase.

        gate:
          type: review-loop
          maxRounds: 2
          targetScore: 8
          judgeModel: big
          reviewers:
            - focus: >
                Correctness against Definition of Done: does the output satisfy
                ALL DoD criteria from the analysis stage? Identify specific
                criteria that are not met or only partially met.
            - focus: >
                Completeness: are there gaps, missing pieces, overlooked edge
                cases, or parts of the task that were not addressed at all?
                Is the output complete relative to the original task?
            - focus: >
                Project conventions and quality: does the output follow the
                project's established patterns, naming conventions, code style,
                and structure? Is it maintainable, clean, and well-organized?

      - id: fix-gaps
        agent: worker
        tier: small
        task: >
          Review the previous attempt and fix ALL gaps and issues identified
          by the reviewers.

          Goal-loop iteration {iteration}.
          Goal: {goal}

          ### Original Task

          {task}

          ### Analysis & DoD

          {outputs.analyze}

          ### Previous Output

          {outputs.execute}

          ### Reviewer Critiques

          {lastFeedback}

          Address each reviewer critique point by point. If the output already
          meets all DoD criteria and no critiques remain, confirm it is
          production-ready and complete — do not invent work.

          Do not introduce new issues or unrelated changes. If reviewers
          identified missing or incomplete DoD criteria, complete them now.
          If reviewers found quality issues, fix them now.

          Do NOT skip any critique. Each one must be explicitly addressed.

          ### Editing is required
          Make the edits now — do not return a plan. Apply fixes directly to files.
          If every critique was already addressed, confirm explicitly and stop.

        gate:
          type: review-loop
          maxRounds: 2
          targetScore: 8
          judgeModel: big
          reviewers:
            - focus: >
                Precision of fixes: did the worker address EVERY specific
                critique from the execute review? No critique should remain
                unaddressed or only partially addressed.
            - focus: >
                Final polish: is the output thorough, well-structured, and
                production-ready? Are there any remaining rough edges,
                incomplete parts, or verbosity issues?
            - focus: >
                No regressions: did the fixes introduce new issues,
                inconsistencies, or break existing functionality? Is the
                overall result stable and coherent?

  - id: synthesize
    agent: planner
    tier: big
    task: >
      Synthesize a final summary of the completed work.

      ### Original Task

      {task}

      ### Analysis

      {outputs.analyze}

      ### Goal-Loop Result (execute → fix-gaps, possibly across iterations)

      {outputs.converge}

      ### User Answers

      {answers}

      Produce a structured report:

      1. **What was accomplished** — clear summary of the work done.
      2. **What changed** — files modified, outputs produced, decisions made.
      3. **Goal achievement** — whether the goal-loop critic confirmed the
         mission was achieved, and how many iterations were needed.
      4. **Quality assessment** — how well the Definition of Done criteria
         were met, based on the gate scores.
      5. **Residual risks** — any known limitations, edge cases not covered,
         or things to be aware of before merging or using the output.
      6. **Next steps** — recommended follow-up work or actions.
```