> ## Documentation Index
> Fetch the complete documentation index at: https://docs.openclix.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Retention Automation

> Run a repeatable, agent-based mobile app retention and engagement automation loop with OpenClaw, Claude Code, and Codex.

Use this guide after the base workflow (`openclix-init`, `openclix-design-campaigns`, `openclix-analytics`, `openclix-update-campaigns`) is already producing artifacts.
Run `openclix-update` separately when template baseline updates are needed.
This is OpenClix's primary agent-based mobile app retention automation operating loop.

## Why this layer exists

OpenClix already gives deterministic config + metrics artifacts. This guide adds an operations layer so teams can run a weekly or biweekly decision loop with coding agents while preserving a strict human approval gate.

## Capabilities by skill/script

* `openclix-init`: integration automation
* `openclix-design-campaigns`: campaign config automation
* `openclix-analytics`: retention/engagement measurement automation
* `openclix-update-campaigns`: campaign action proposal automation
* `openclix-update`: source-template synchronization for baseline maintenance
* `retention_ops_automation.sh`: OpenClaw/Claude Code/Codex review prompt automation

## Operations loop

```text theme={null}
impact metrics + campaign metrics + active config
  -> evaluate_campaigns.sh recommendation draft
  -> agent-specific review prompt (OpenClaw / Claude Code / Codex)
  -> human approval
  -> apply by delivery mode
```

## Evidence confidence model

Use this grading model when citing examples in retention automation docs:

| Grade | Criteria                                                                          |
| ----- | --------------------------------------------------------------------------------- |
| A     | Public implementation artifacts and reproducible workflow evidence are available. |
| B     | Workflow and tooling are public, but KPI impact proof is limited.                 |
| C     | Anecdotal/operator report without reproducible KPI evidence.                      |

## Run automation helper script

```bash theme={null}
bash scripts/retention_ops_automation.sh \
  --root <target-project-root> \
  --agent all \
  --delivery-mode auto \
  --dry-run
```

Run from the OpenClix repository root, or invoke the script with an absolute path when your current working directory is different.

CLI options:

* `--agent openclaw|claude-code|codex|all`
* `--delivery-mode auto|bundle|hosted_http|dual`
* `--impact-file <path>`
* `--campaign-metrics-file <path>`
* `--config-file <path>`
* `--output-dir <path>` (default: `.openclix/automation`)

## Generated artifacts

* `.openclix/automation/run-summary.json`
* `.openclix/automation/prompts/openclaw.md`
* `.openclix/automation/prompts/claude-code.md`
* `.openclix/automation/prompts/codex.md`

The script also generates evaluator outputs in `.openclix/automation/evaluator/` so dry-run usage can stay isolated from active campaign files.

## Weekly / biweekly operating cadence

1. Refresh analytics artifacts with `openclix-analytics`.
2. Run `retention_ops_automation.sh` in dry-run mode.
3. Paste generated prompt(s) into the selected agent.
4. Compare agent recommendations with `update-recommendations.json`.
5. Approve only high-confidence, reversible changes.
6. Apply changes by delivery mode (`bundle`, `hosted_http`, `dual`).
7. Record rationale in your team changelog.

## Approval gate and rollback policy

Before apply:

* Confirm sampling thresholds were met (`delivered >= 200`, `opened >= 20`).
* Reject recommendations that rely on low-confidence data.
* Keep delete actions pause-first.

Rollback defaults:

* `bundle`: restore previous bundled config and redeploy app build.
* `hosted_http`: roll back to previous JSON version at stable HTTPS URL.
* `dual`: revert both hosted primary and local fallback together.

## Third-party OpenClaw plugin review

OpenClix's own skills (`openclix-init`, `openclix-design-campaigns`, `openclix-analytics`, `openclix-update-campaigns`, `openclix-update`) are first-party and source-auditable in this repository.

When the retention ops loop incorporates **additional third-party OpenClaw plugins** that are not part of this repository, apply these review steps to those external plugins:

* Review the external plugin's source before execution.
* Run unfamiliar external plugins in sandboxed or isolated environments.
* Prefer read-only analysis prompts for first-pass review of external plugins.
* Require human approval before applying config changes from any source.

## Failure codes from helper script

* `10`: required command/script missing
* `20`: no supported analytics provider detected
* `21`: OpenClix integration not detected
* `30`: required input artifact missing
* `31`: delivery mode unresolved (`unknown`)
* `40`: evaluator execution failed