Schedule Package

The internal/application/schedule/ package owns the full lifecycle from task declaration to execution.

Pipeline

Domain specs (TaskSpec, GlobalTaskSpec)
    ↓
Registry DSL — NewTasks()
    ↓  repos.Pack()     → multiply × partitions, wrap closures
    ↓  Preflight()      → mark phase
    ↓  globalTasks()    → wrap without multiplication
    ↓
[]spec.Task (flat list, ready to schedule)
    ↓
Planner — split by Phase
    ↓
Scheduler
    ├── runPreflights() → longrun.NewOneShotTask
    └── runWorkers()    → longrun.NewIntervalTask

Key Types

Type	File	Role
RepositoryTasks	repository_tasks.go	Multiplies specs × configured repos
Preflight()	repository_tasks.go	Marks a spec as preflight phase
globalTasks()	global.go	Wraps global specs without multiplication
join()	global.go	Concatenates task slices
Planner	planner.go	Splits tasks by Phase, validates duplicates
Scheduler	schedule.go	Executes preflights → workers
infraClassifier()	schedule.go	Classifies infra errors for retry

Registry DSL

The registry reads as a manifest:

func NewTasks(...) []spec.Task {
    return join(
        repos.Pack(
            Preflight(repoValidator.TaskSpec()),
            issuePoller.TaskSpec(),
            outboxRelay.TaskSpec(),
        ),
        globalTasks(
            issueExplainer.TaskSpec(),
        ),
    )
}

Adding a task = one line in the appropriate section. Adding a partition dimension = one new provider + one new section.

Pack Internals

Pack(entries ...any) accepts two types via type switch:

• repository.TaskSpec → default PhaseWork, lazy repoID resolution
• repoTask (from Preflight()) → PhasePreflight, repoID = 0

Panics on zero Interval (must use spec.OneShot) or unknown entry type. This is a conscious trade-off: any loses compile-time safety but enables clean DSL syntax. See REVIEW.md "DSL internals may use pragmatic shortcuts."

Partition Multiplication

For each spec × each configured repository, Pack creates a closure that:

1. Lazily resolves RepositoryID on first call (work-phase only)
2. Constructs repository.Partition{Owner, Name, RepositoryID}
3. Calls the domain's Work(ctx, partition) with honest arguments

Domain code never sees closures, config, or lazy resolution. It receives a typed Partition and works.

Phase Model

Phase	When	Degraded policy	repoID
PhasePreflight	Before all workers	nil — unknown errors crash	0 (row may not exist)
PhaseWork	After preflights complete	Set — survive and retry	Lazy-resolved

Planner splits []spec.Task by Phase. Scheduler runs each phase with its own longrun.Runner and Baseline config.

Extension Points

What	Where	Planner/Scheduler touched?
New per-repo task	repos.Pack(...) +1 line	No
New global task	globalTasks(...) +1 line	No
New partition	New XxxTasks provider + registry section	No
New phase	+1 Phase const, +1 Scheduler runner	Yes

Error Classification

infraClassifier() checks apierr interfaces on errors from infrastructure clients:

• WaitHinted with positive duration → Service + explicit wait
• ServicePressure → Service
• Retryable → Service
• Not classified → nil (baseline handles as Unknown/Degraded)