Legacy Maintenance And Recovery

This reference describes orchestration: legacy. New objective runs use the objective lifecycle and migration contract.

Explicit legacy runs use one root-maintainer automation. It creates a root-scoped maintainer task for the next decision: choose work, review a result, arrange a repair or close the phase. Workers and child maintainers own narrower outcomes. There is no new task store, scheduling profile or agent hierarchy.

Admission And Changed Evidence

The scheduler keeps at most one unfinished occurrence of the root automation. Its initial condition admits planning or actionable Forge work under the normal capacity, budget, lease and workspace checks. After its first pass, spare capacity alone cannot launch repeated maintenance sessions.

The existing run_coordination record stores the substantive evidence fingerprint seen at maintenance admission. Published commit contents, current PR heads and statuses, review decisions and nonrecurring worker outcomes can change it. Repeated root-session completion, queue churn and observation timestamps do not. Changed evidence is a reason to reconsider the next action, not mathematical acceptance.

An integration owner can checkpoint on a specific external event, including failure/cancellation, and release its execution slot. Its retained context resumes when that condition is satisfied. A deliberate event wait remains authoritative; an idle slot does not bypass it. Native children still need their parent to collect them within the same live execution.

The default root path does not also enter idle-planner refill or legacy coordination-recovery admission. Those mechanisms remain for previously configured separate planner/supervisor runs. Fresh orchestrated launches are rejected. Saved orchestrated: true runs retain that compatibility path; it is not the new default.

Failure And Missing Ownership

Ordinary execution recovery first preserves the same assignment, provider context, journal and source work under the configured retry policy. A failed worker or changed review result gives the root maintainer new evidence to act on.

When the root owner terminates unsuccessfully, or finishes an open mission with neither a follow-up owner nor an explicit event wait, the existing automation can own a recovery decision. The decision identifies the observed failure and asks for a concrete repair or justified wait. The agent may invoke a diagnostic native helper if the cause is unclear; an audit is not a mandatory extra stage.

Recovery is deduplicated against the substantive evidence. Repeating the same failure does not recursively launch more coordinators. An unchanged failed recovery remains visible as blocked, with its preserved decision and work. New substantive evidence or explicit operator recovery can make another useful attempt possible. This avoids an unbounded retry storm; it does not claim that every infrastructure failure can be repaired autonomously.

Journal reconciliation remains separate from mathematical progress. A lost API reply must be reconciled under its original idempotency key. Known no-effect validation failures are diagnostics. Unknown transport outcomes and state conflicts remain unresolved until inspected. Clearing a wrapper obligation alone cannot settle an uncertain local request.

On-Demand Diagnosis

horizon-pipeline agent context --view operations is available to live workers and maintainers. It exposes bounded owner/activity data, queue reasons, review coverage and sanitized host health, with scoped drilldowns. The orchestration-auditor native helper investigates an identified anomaly and returns evidence, cause, proposed action, responsible owner and confirmation condition. Its parent owns mutations and incident communication.

The helper cannot initiate itself when no parent is running. Scheduler recovery therefore retains responsibility for making the existing maintenance entrypoint available. Agents retain responsibility for mathematical decomposition, review and integration decisions. A diagnostic message alone is not a repaired run.

Completion And Deployment

Worker delivery, specialist assessment, maintainer acceptance and phase completion are different facts. The current API still requires explicit obligation and mission settlement; it does not infer them from final prose. The root maintainer records semantic completion, drains the run and returns. Infrastructure waits for physical execution, publication and external delivery settlement.

No schema migration is added for this workflow; it reuses existing assignments, automations, obligations and coordination state. Normal installations still need the repository's current migrations. Source updates do not rewrite existing automations, pinned instruction bundles or project reviewer descriptors.

Regression tests cover startup, unchanged-evidence suppression, failure recovery, review repair, explicit waits and root closure. Prompt/skill checks cover scoped reviewer startup and phase acceptance boundaries. These establish specific runtime invariants; they do not prove model judgment or a preprocessing duration target. A fresh representative Luna run is still needed to measure the revised workflow.