Continue ONE eligible task once in the same Luna Codex thread with an opaque single-use continuationReference and one bounded instruction. The original objective, allowedFiles, forbiddenFiles, changeIntent, acceptance, and verification contract remain immutable; no widening fields exist. Luna cannot delegate, and continuation never starts automatic repair. Verification, scope checks, and evidence reconciliation run again. handoff is default; compact/full are compatibility modes. While pending with no meaningful new state, remain silent; do not narrate waiting or polling. Report only a result, error, cancellation, timeout, or actionable state change.

Delegate ONE substantial, bounded executable seam to gpt-5.6-luna; no second seam is required. Keep small, simple, or tightly coupled work solo. Tasks may be implementation, tests, bug fixing, refactoring, investigation, or chores. The parent owns architecture, decomposition, unresolved design, sequencing, interfaces, scope, acceptance, and final judgement. Luna owns scoped exploration, implementation, verification, and bounded repair; it cannot see the conversation or delegate.

Provide a self-contained objective, effortReason, acceptanceCriteria, verificationCommands, changeIntent, and honest scopes; add a concise activityLabel when safe and only repository-unavailable context. automaticRepair permits at most one conservative same-thread repair. Results include one evidence-derived failureDecision; parent owns nonautomatic actions. resultDetail=handoff is the default.

The runtime reruns declared checks and reconciles observed edits. A clean PASS returns a text-only VERIFIED_COMPLETE handoff: finish without rereading worker-owned files or rerunning passed checks unless a listed risk changes architecture. FAILED/BLOCKED, untrustworthy, discrepant, scope-violating, refused/skipped, or runtime-error results expand with evidence. Worker claims are not authoritative.

Delegate only when ownership, isolation, context, verification, latency, coordination risk, quality, and current parent-conditional credit economics beat fixed overhead; raw tokens are not credit cost and no saving is guaranteed. While pending with no meaningful new state, remain silent; do not narrate waiting or polling. Report only a result, error, cancellation, timeout, or actionable state change.

Delegate a batch intended for two or more owned seams to gpt-5.6-luna; one task remains accepted for compatibility, but prefer delegate_task when no scheduling is needed. Use sequential for dependencies/shared workspace state and parallel only for genuinely independent disjoint declared scopes. Do not create artificial seams. At most 12 tasks are accepted and at most 3 run concurrently; the rest queue. Each task needs a self-contained contract and a concise activityLabel when a safe label exists. The parent owns architecture/interfaces and exceptional judgement; Luna owns exploration, implementation, verification, and repair. automaticRepair is one bounded task-local turn.

Parallel same-file edits prevent automatic integration. allowOverlappingScopes:true only accepts the declared overlap; it is not a write sandbox and does not permit same-file integration. integrate=false skips copying and retention follows operator policy. Partial outcomes remain visible. automaticRecovery defaults true: at most one evidence-eligible timeout continuation or exact process-exit retry; a counter alone never authorizes retry. Repair precedes recovery and neither nests. Successes, cancellation, scope/security/evidence failures, refused checks, discrepancies, and conflicts are never retried. Successful streams survive sibling failure.

After integration, deterministic code reruns the deduplicated union of declared checks in the final workspace. completionState=verified-complete means all seams, integration, and final checks passed; the default text-only handoff then tells the parent to finish without rereading files or rerunning checks. Any failure/refusal/conflict returns rich evidence for targeted diagnosis. resultDetail is one batch-level compatibility choice. More workers are not automatically cheaper; raw tokens are not credit cost and savings depend on the parent and task mix. While pending with no meaningful new state, remain silent; do not narrate waiting or polling. Report only a result, error, cancellation, timeout, or actionable state change.

Explicitly explore an admitted repository, API, or documentation scope with gpt-5.6-luna; fixed read-only disposable execution returns provenance-marked worker claims, runtime facts, inferences, and unknowns. Implements nothing, cannot delegate, and is never automatic.

Cheap deterministic check of whether delegating is worthwhile. First use cheap, bounded repository/test inspection to identify candidate delegated leaves; describe those leaves, not the whole objective, and defer broad exploration. The parent may retain shared contracts and final integration. "unknown" biases advice solo but never refuses. Creates no worker, batch, worktree, or thread, refuses nothing, and returns route (solo | either | delegation-plausible), matched rule, explicit/defaulted provenance, signals, and structural parallel eligibility. Advisory only and never required: the parent owns sequential vs parallel, worker count, effort, and final decision; choosing zero workers is normal. either means fixed delegation overhead needs explicit justification, otherwise stay solo.