Host-owned Workbench executions

finos/morphir-engine/execution/executions owns bounded in-memory execution state for a composing host. It uses the exact morphir/workbench/executions 0.2.0-draft.1 profile and the shipped receipt cursor. It adds no wire methods or capability advertisement. Connected protocol v1 remains unchanged.

Create one Session for each host/session/provider authority, using the offer that the host will authenticate. Session identities must be fresh across owner replacement. The composing host serializes access, authenticates callers, enforces provider grants and dispatches work. This synchronous component performs no I/O and does not authenticate an offer or establish provider trust.

Ownership and execution

  1. submit(unit, suite) captures the full rich unit as an Ion binary checkpoint and the suite through the bounded invocation codec. Planning validates public entries and arguments before the host accepts the execution. The initial receipt binds host/session/provider plus new execution and snapshot IDs. These snapshot handles denote retained input, not a path or content hash.
  2. claim(identity) reconstructs an independent execution plan and publishes a running revision. Only one claim succeeds. The worker can use the existing execution provider or evaluator against this plan. Losing the worker connection does not authorize another execution claim.
  3. query(identity) returns a fresh copy of the authoritative receipt. Every operation checks the entire identity, including the snapshot.
  4. cancel(identity) acknowledges a request if the composing host offered cancellation. Workers must query this acknowledgement at their cancellation points. Repeated acknowledgement and queries do not advance the revision. Acknowledgement does not terminate work or replace returned data. Query-only offers reject cancellation while work remains pending.
  5. After the worker and cleanup finish, complete(identity, outcomes) reconstructs the submitted public declarations and validates all terminal outputs before publication. A worker cannot widen output types by editing its plan. Ordinary Result.Err remains returned data; SDK errors, failures, cancellation and cleanup retain their separate variants. Completion may occur before a claim for a host policy rejection or cancellation. An identical terminal is an idempotent replay; a different terminal rejects.
  6. release(identity) explicitly frees terminal storage. Pending work cannot be released, and capacity pressure never evicts another execution. Released identities become unknown, and subsequent submissions do not reuse their IDs.

Cancellation is cooperative. The host must only offer it when its worker can observe requests. complete is a trusted host publication API, not an endpoint for an untrusted worker to assert that cleanup finished. The component validates the evidence shape and declared types, not whether an external provider told the truth. Disconnect and transport timeout have no receipt-producing API.

Bounds

Defaults allow eight retained executions and 16 MiB of serialized retained evidence. Configuration allows 1 to 64 executions and 1 to 64 MiB. Each submission reserves the checkpoint bytes, encoded suite bytes and a full 1 MiB receipt allowance. This keeps terminal publication independent of later submissions. The existing data boundary limits each checkpoint to 4 MiB, and invocation and receipt codecs each limit their encoded payloads to 1 MiB. Quotas describe retained serialized data, not allocator overhead or temporary planning allocations. A quota too small for one validated snapshot simply refuses submissions.

Validation, capacity and transition failures leave retained work unchanged. IDs advance only on successful submissions and never wrap. State survives a client disconnect only while the composing host retains this owner. There is no durable storage, restart recovery, retry lease, deduplicated submission or exactly-once execution guarantee.

Scheme runtime and capability checking policy

The parallel submit_scheme(unit, suite, policy=...) route captures the bounded Library Scheme runtime and its shared SDK. Omitted policy resolves to Strict once, at acceptance. Explicit Permissive uses the qualified reached-only capability checks. The legacy submit and claim methods keep their existing compiled-plan validation; claims cannot cross between these routes.

describe_scheme(identity) returns a fresh domain projection containing the full execution identity, opaque plan_id, runtime, effective policy and selected boundary. It adds no fields to the frozen receipt or offer. A composing adapter supplies its own exact negotiated projection and authenticates the provider.

claim_scheme(identity, plan_id, provider_id, runtime, policy) requires every selection component to match the accepted execution before advancing its cursor. It returns the immutable SchemePlan once. SchemePlan.unit() and .suite() decode fresh copies for an isolated worker; edits to these copies, the original input or a projection cannot change accepted source, arguments, policy or boundary. Completion reconstructs from retained bytes and policy, compares the selected boundary, then validates every terminal against that context. Retained boundary bytes count toward the session quota alongside the source, suite and reserved terminal space.

replan_scheme(identity, plan_id, provider_id, runtime, policy) reconstructs only the original accepted inputs. It preserves policy and allocates fresh execution, snapshot and plan identities. It never claims or schedules either execution and leaves the original history and once-only eligibility unchanged. Editing source or arguments requires a new submission. An incompatible replacement selection rejects explicitly; it cannot downgrade policy or silently switch a provider.

Query, cancellation, completion replay and release use the same receipt cursor for both routes. These core methods do not authenticate a worker, establish cleanup truth, advertise native or HTTP support, or grant execution authority.

Validation and remaining integration

Tests run on JS, native, Wasm and Wasm GC. They exercise actual bounded Scheme evaluation, exact arithmetic, returned errors, SDK errors, mutation isolation, all identity dimensions, cancellation races, replay/conflicts, terminal cleanup, validation failures and retention quotas. The unchanged receipt reducer has shared MCK draft.6 qualification. These producer tests do not independently qualify an authenticated host adapter.

The existing Workbench execution bridge and native owner compose this core with their own transport and scheduling. Policy-aware UI, native and HTTP capability offers still require separate integration and qualification. Generated browser execution and durable recovery remain separate work. This package does not automatically use metadata host grants or advertise trusted naming providers.

View source on GitHub