Migrating to 0.6

Migrating To 0.6

Elio 0.6 changes coroutine ownership, cancellation, structured concurrency, worker-local I/O enforcement, and several runtime contracts. Review the items below when upgrading from 0.5.x.

Task Ownership And Virtual Threads

• coro::task<T> is move-only. Move unstarted tasks into containers, return values, scheduler handoffs, and other owners; do not copy them or pass an lvalue where ownership transfer is required.
• Normal task frames use ordinary coroutine allocation. The public vthread_stack bump allocator and its deferred cross-thread deletion protocol were removed. Direct allocator users must provide an allocator with suitable object lifetime and cross-thread free behavior.
• The logical vthread model remains. Virtual-stack ancestry, task chains, debugger inspection, affinity, and runtime execution context still describe coroutine execution independently of frame allocation.
• Lazy task ownership no longer installs creator-thread virtual-stack state. Ancestry is bound when a task is actually awaited; independent scheduler spawn detaches construction-time ancestry.

Cancellation And Structured Concurrency

• Every Elio task has a shared execution context with cooperative cancellation authority. Direct lazy-task awaits propagate runtime cancellation between Elio promises. Foreign coroutine promises and explicit token parameters stay separate unless an adapter deliberately bridges them.
• join_handle::request_cancel() is a best-effort request, not forced frame destruction. Pass this_coro::cancel_token() into every wait that should react and handle cancellation callback exceptions where callbacks may throw.
• mutex, shared_mutex, semaphore, event, condition_variable, and channel have explicit cancellable wait overloads. Existing no-token overloads retain their prior result shapes.
• Use coro::task_group or coro::task_scope() when children must be cancelled and joined before captured state is released. Select fail-fast, collect-all, and bounded-concurrency policy explicitly.

Combinator Behavior

• when_all(), when_any(), and with_timeout() now return move-only lazy tasks and structurally own every accepted branch.
• when_any() cancels and joins losers before returning. A token-ignoring loser can therefore delay the result.
• with_timeout() treats the duration as the point at which cancellation is requested, not as a hard return-time bound. It waits for the wrapped work to reach a terminal state.
• Parent cancellation that prevents a required result throws combinator_cancelled. Review code that previously assumed every failed with_timeout() was a deadline expiry.
• Later loser/secondary exceptions are routed through the scheduler unhandled exception handler after the primary result is fixed.

Scheduler And I/O Integration

• A scheduler is one-shot. Once shutdown begins, start() does not restart it; construct a new scheduler for another run.
• Pending scheduler-owned I/O pins a task to the exact worker/backend owner. Caller affinity changes take effect after the operation reaches a terminal state and cannot clear the internal I/O pin.
• Custom awaitables derived from io_awaitable_base must use setup_op_state() and prepare_op_state(). The old bind_to_worker()/restore_affinity() pattern and mutable raw-backend accessor are removed. Custom backend notify() implementations must be non-throwing.
• Elio does not serialize conflicting operations on the same stream or fd. Callers remain responsible for preventing overlapping reads, overlapping writes, close-versus-I/O races, and protocol ordering errors.

RPC And Process Lifecycle

• RPC server sessions structurally own accepted request, pong, and overload response tasks. Session teardown requests cancellation and joins accepted work; a handler that ignores both runtime and RPC tokens can delay teardown.
• RPC admission remains explicit per-session policy. Configure bounded concurrency and overload handling instead of relying on clients to self-limit.
• Follow Fork Safety. A single-threaded process that forks before Elio runtime use may create independent parent/child runtimes. With an active runtime or any other thread, the child must immediately call an explicitly async-signal-safe exec function such as execve(), or _exit(), and must not use or destroy inherited Elio state. Do not assume every exec-family wrapper is async-signal-safe.

Upgrade Checklist

1. Replace task copies and implicit lvalue handoffs with explicit moves.
2. Remove direct vthread_stack allocator use while preserving any logical virtual-stack/debugging integrations you need.
3. Pass runtime or explicit cancellation tokens into every operation that must stop during task-group or shutdown cancellation.
4. Re-evaluate when_any() and with_timeout() latency assumptions now that losers are joined.
5. Update custom I/O awaitables and backend notify() implementations to the 0.6 ownership contract.
6. Define RPC concurrency/overload policy and ensure handlers cooperate with session cancellation.
7. Audit process creation paths against the documented fork boundary.
8. Run normal, ASAN, and TSAN validation for application-specific adapters and cancellation callbacks.

See API Contracts for the authoritative guarantee/responsibility inventory and the Unreleased section of CHANGELOG.md for the complete change list.