Operating model and artifact lifecycle

Day Shift turns intent into a repository-visible Human-Agent-Contract (HAC). The contract is a chain of artifacts, not hidden session state: each artifact has a producer, required relationships, a next consumer, and evidence that determines whether work may advance.

Choose the planning lane

Standard planning is the default for work with multiple stages, meaningful risk, or coordination needs:

specification → slice overview → phase overview → milestone overview → task definition → implementation summary → reconciliation

Lightweight planning is an explicit exception for narrow, low-risk work. It skips phase and milestone decomposition but still requires a task contract, implementation evidence, and reconciliation. If scope or uncertainty grows, promote the work into the standard hierarchy rather than silently stretching the lightweight lane.

Emergent work describes priority-interrupting intake, not a planning depth. Capture an operator report, production regression, or environment blocker as an emergent specification first, then choose standard or lightweight planning from explicit evidence.

Artifact responsibilities

Artifact	Producer and purpose	Required relationships	Evidence and next consumer
Specification	A human or agent captures source intent, constraints, and success conditions.	Registered source identity and source reference.	Recertification establishes current intent; slice planning consumes it.
Slice overview	Planning groups one coherent delivery outcome.	Upstream specification links and slice identity.	Slice review and validation inform phase decomposition.
Phase overview	Planning orders a major stage of the slice.	Slice and specification traceability.	Phase review informs declared milestone creation.
Milestone overview	Planning declares an ordered, reviewable outcome and task set.	Phase, slice, and specification traceability.	Milestone review and declared-child validation govern task scaffolding.
Task definition	A planner freezes implementation scope, targets, acceptance criteria, dependencies, and validation.	Milestone ancestry in standard planning, or the lightweight parent context.	Readiness review gates the implementation handoff.
Implementation summary	The implementing agent records changed paths, validation, acceptance results, deviations, and discovered work.	Its task definition and full planning ancestry.	Summary review supplies task-closeout evidence; it does not replace the task contract.
Reconciliation	A reviewer rolls completed task summaries into acceptance, residual-work, and verdict evidence.	The applicable milestone or lightweight parent plus completed summaries.	Review and explicit status synchronization support the next parent or release handoff.

Planning coordinates (slice, phase, milestone, and task) identify location in the hierarchy. planning_depth records standard or lightweight handling. Shared status is independent of family-specific lifecycle fields: task definitions own readiness, implementation summaries own implementation_status, and reconciliations own reconciliation_status. Upstream links accumulate by level so a later reader can recover both source intent and parent context.

Roles and ownership

• Humans own intent and authorization. They choose scope, risk tolerance, planning depth, acceptance tradeoffs, and any meaningful write or release decision.
• Agents own scoped execution and honest evidence. They follow the selected artifact, preserve unrelated work, validate proportionately, record deviations, and stop when evidence conflicts.
• The CLI owns deterministic mechanics. It scaffolds declared artifacts, validates structure and trace links, applies explicitly selected metadata or materialization writes, and reports findings.
• Reviewers own judgment. They assess whether evidence satisfies intent, classify residual work, and approve or reject lifecycle handoffs.

Ownership does not transfer implicitly. A recommendation from an agent or CLI is evidence for a human or workflow decision, not authorization to broaden scope.

Evidence gates and transitions

Creation and mutation commands are write-capable only when explicitly selected. By contrast, planning next-action, inventories, gap reports, readiness reviews, artifact reviews, validation links-check, and validation smoke are read-only. They can report findings or recommend a next command; they do not create children, repair links, update metadata, or advance lifecycle state.

The main gates are:

1. Recertify source intent before relying on stale or conflicting specifications.
2. Declare children at the parent layer before scaffolding them; never invent undeclared downstream work during expansion.
3. Require a task to be ready to implement before changing its runtime targets.
4. Close implementation with non-placeholder changed-path, validation, acceptance, deviation, and follow-up evidence.
5. Reconcile only after every in-scope summary is completed; review the reconciliation before any explicit parent-status synchronization.

Warnings require review or explicit acceptance. Errors block advancement. A review result may justify a later write-capable update, sync, build, or materialize command, but the write remains a separate action.

Resumability and lifecycle state

Repository artifacts make interrupted work recoverable. Re-open the selected task or parent overview, inspect scoped changes, identify the last passed gate, and rerun the current read-only review before continuing. Record structural pressure, scope growth, or newly discovered work in evidence instead of silently changing the declared contract. Archive or release only after reconciled evidence has reached the applicable parent and unresolved work has an explicit disposition.

Continue with Getting Started for workspace setup, workflow guides for procedural journeys, and reference for precise command contracts.