DR. Agent Backup

Use when setting up, auditing, repairing, or restoring backup continuity for a Daniel-owned agent.

Outcome

One deterministic, non-LLM systemd user timer is the only process allowed to commit and push workspace backups. Every six hours it stages safe restore-relevant files, commits when needed, and pushes to the approved Azure DevOps remote. A failed run invokes one rate-limited failure notifier. Conversational agents and health checks never run an independent git push.

Required design

1. Use Pattern B1: a manifest-backed runner, not an agent/LLM or OpenClaw cron.
2. Use one systemd --user timer with OnBootSec=5min, OnUnitActiveSec=6h, AccuracySec=5min, and Persistent=true.
3. Make the runner load its manifest and credential contract at run time, set GIT_TERMINAL_PROMPT=0, and use the approved askpass helper.
4. Stage only allowlisted restore-relevant paths. Exclude credentials, env files, auth profiles, databases, logs, caches, media, raw transcripts, dreaming/reflection output, temporary/download/archive files, generated artifacts, and state migrations.
5. Scan staged names and text for high-risk credential patterns before committing.
6. Commit only intentional staged changes, push the approved existing remote and branch, write a local ledger outside Git, and exit non-zero on any credential, staging, commit, remote, or push failure.
7. Wire OnFailure= to one deterministic notifier. It sends an immediate concise failure message, suppresses identical repeats for 24 hours, and never sends success messages.
8. Do not run a scheduled health monitor by default. Add one only when its distinct detection value is documented and it cannot push or commit.

Ownership rule

The backup runner exclusively owns routine Git commit/push for the workspace. Agents may edit approved durable files but must leave them for the next runner. Do not use plain git push from an agent runtime; it does not inherit the runner's credential contract and creates false backup failures.

Preflight

Before enabling or manually running a new or repaired runner:

1. Confirm the existing approved Azure DevOps remote, branch, credential method, and valid provider target for failure notifications.
2. Inspect git status --short and classify each restore-relevant dirty path as intentional durable content, intentional deletion, generated/runtime artifact, or unresolved.
3. Keep raw session exports and state-migration artifacts untracked; do not broadly exclude curated daily memory files.
4. Refuse to arm a runner over unresolved or sensitive content. Correct the manifest, exclusions, or worktree first.
5. Verify the credential file exists with mode 600, required key names are present without exposing values, and the askpass helper is executable.

Setup and repair

1. Audit the existing manifest, runner, installed service/timer, failure handler, Git remote/branch, and worktree.
2. Create or repair the manifest, runner, local ledger, service, timer, failure notifier, and focused tests.
3. Validate JSON, Python/shell syntax, unit rendering, and systemd-analyze --user verify.
4. Dry-run notification rendering without live delivery.
5. Run one real backup after preflight is clean; verify its commit/push, ledger, and remote alignment.
6. Install and enable the timer; verify it is enabled/active and has a next trigger.
7. Retire duplicate backup timers, health checks, or agent-driven push paths after confirming the new owner is live.

Default tracked scope

Include: AGENTS.md, SOUL.md, USER.md, TOOLS.md when present, MEMORY.md, comments-only HEARTBEAT.md, curated memory/**/*.md and intentional durable JSON, context_pipeline/**, automation/** excluding runtime output, workspace-local skills/**, and active doc_reviews/**/*.md.

Approval boundaries

Ask Daniel before creating a repo or remote, changing credentials/PATs or access controls, using a new notification target or changing its cooldown, committing sensitive/customer data, deleting history, or rewriting commits.

Shared release

This is a shared fleet skill. Follow the shared-skill source-of-truth rule in dr-agent-baseline: after a behavior change, publish a new ClawHub version, verify the registry resolves it, then install/apply that exact release on each in-scope agent. Do not present a local Workshop or workspace revision as canonical.

Completion evidence

Report the runner and timer state, next trigger, real commit/push and remote alignment, notification policy, tracked/excluded scope, validation results, and any limitation. For a skill revision, also report the ClawHub version and verified fleet-application status. This skill does not replace VM snapshots or system backups.