repeatable-agent-loops

name	repeatable-agent-loops
description	Use when designing, running, auditing, or reviewing repeatable agent loops for recurring audits, iterative fixes, maintenance sweeps, monitors, documentation checks, quality passes, test reliability sweeps, post-deploy checks, or any bounded workflow where an agent observes state, takes one action, records evidence, decides whether to continue, and stops safely under APIVR.
activation	Activate when the description trigger applies to the current task.
required_inputs	Task request, relevant repository context, constraints, and authority dependencies.
required_outputs	Skill-specific artifact, verification evidence, canonical verdict, and next action.
authority_dependencies	00_start_here/SOURCE_OF_TRUTH.md; 10_governance/APIVR_EXECUTION_LIFECYCLE.md; 10_governance/source_of_truth/Elite_Build_Goals_v3.md.
evidence_requirements	Executed checks or an honest Unknown, Not Run, or Blocked state for every material claim.

Repeatable Agent Loops

Use this skill when work should repeat until a measurable condition is met, safely exhausted, or blocked.

A loop is not permission for open-ended autonomy. Every loop must have a bounded scope, a single-step action rule, evidence checks, stop conditions, and APIVR ownership.

Required Files

Load these when designing or running a loop:

• 40_knowledge/REPEATABLE_AGENT_LOOP_PATTERNS.md
• 60_templates/LOOP_DESIGN_TEMPLATE.md
• 60_templates/LOOP_RUN_RECEIPT_TEMPLATE.md

APIVR Routing

• Phase 1 Audit: inspect current state, authority, target condition, risks, permissions, and evidence sources.
• Phase 2 Plan: define the loop design, iteration limit, stop conditions, rollback path, and receipt format.
• Phase 3 Implement: run one bounded action per iteration.
• Phase 4 Audit Implementation: check whether the action changed only the intended scope and whether evidence supports the result.
• Phase 5 Verify Implementation: record receipt evidence and decide continue, stop, escalate, or block.
• Phase 6 Re-Audit: review the full run for drift, missed risks, repeated failures, and future automation suitability.

Loop Decision Flow

flowchart TD
  A["Task appears repeatable or iterative"] --> B["Load APIVR tier and loop skill"]
  B --> C{"Is the objective measurable?"}
  C -- "No" --> D["Stop: write clearer objective or use normal APIVR plan"]
  C -- "Yes" --> E{"Can one safe action be defined?"}
  E -- "No" --> D
  E -- "Yes" --> F{"Are permissions and stop limits explicit?"}
  F -- "No" --> G["Stop: request authority or narrow scope"]
  F -- "Yes" --> H["Run one iteration"]
  H --> I["Record loop receipt"]
  I --> J{"Target met?"}
  J -- "Yes" --> K["Stop: PASS or CONDITIONAL PASS"]
  J -- "No" --> L{"Blocked, unsafe, or no progress?"}
  L -- "Yes" --> M["Stop: BLOCKED / PARTIAL / FAIL"]
  L -- "No" --> N{"Iteration budget remains?"}
  N -- "No" --> O["Stop: exhausted with evidence"]
  N -- "Yes" --> H

Loop Contract

Every loop must define:

• Objective: the exact condition the loop is trying to reach.
• Scope: files, systems, users, data, environments, and actions allowed.
• Non-scope: forbidden files, systems, changes, or irreversible actions.
• One-step action: the largest action allowed in a single iteration.
• Evidence check: command, observation, test, log, screenshot, report, or human approval used to judge the step.
• Continue condition: what justifies another iteration.
• Stop conditions: success, clean no-op, blocked, unsafe, approval-required, exhausted, or no measurable progress.

Run up to 5 iterations. Each iteration fixes one broken internal link, runs the link checker, records the changed file and command output, then stops when all links pass or a missing source cannot be resolved. Keep cleaning the docs until they look better.