Scheduled maintenance

Scheduled agent sessions keep bounded maintenance work visible without granting an unattended system broad authority. The schedule starts a normal Spine task; the same skills, ownership rules, and human gates still apply.

Maintenance families

The current operating model covers three families:

• drift triage compares freshness budgets and proposes mechanical repairs;
• package radar reviews upstream releases against actual ecosystem usage;
• research radar reviews new methods and libraries for ecosystem relevance.

Exact task prompts, local machine paths, and dated run artifacts are operational state and are not published on this site. The owning skill is the durable procedure for each family.

Bounded execution

Each task starts from the Spine workspace, reads the relevant skill, gathers evidence, and writes a reviewable result. External publication, merges, releases, security changes, and methodological judgments remain human-gated. Mechanical proposals should stay focused on one repository and concern.

Durable state

The repository records machine-readable snapshots and ledgers so a later run can distinguish a new signal from an acknowledged one. Those histories are tracked for maintainers but excluded from the reader-facing site. Public docs explain the stable procedure rather than reproducing exact prompts or transient findings.

Failure and recovery

A failed schedule is a maintenance signal, not permission to broaden scope. Use the drift guide to render current budgets and the workspace health guide to verify local capabilities. When a schedule depends on repository visibility or credentials, restore that external prerequisite and rerun the owning task deliberately.

See capability surfaces for why the procedural knowledge lives in skills while scheduling remains an execution mechanism.