Skip to content

interop.md

Latest commit

 

History

History
55 lines (45 loc) · 2.8 KB

interop.md

File metadata and controls

55 lines (45 loc) · 2.8 KB
Raw
subsystem interop
owned_paths
dev/interop/
scripts/lint-interop.mjs

interop (offline development harness)

An offline contract simulator for #198 and #200, not a shipped adapter. Commands, semantics and limitations are in the harness README.

The separate opt-in isolation probe (#221) runs synthetic scripts in disposable Docker containers. It is outside the no-process simulator's command and required CI. It records selected OS observations and configured limits; it does not connect a runtime or certify a live deployment. Its ordinary unit tests use injected Docker calls only.

Owns

  • dev/interop/: bounded versioned codec, durable synthetic task coordinator, two-owner grant policy, fake runtime/clock, demo and deterministic tests.
  • scripts/lint-interop.mjs: lint using core's installed harness tooling.

Invariants

  • The contract simulator has no model, credential, real process, DSH data or network access. The command preload rejects accidental network/process calls; it is not a security sandbox.
  • Owner identities stay distinct. Same-owner social friendship is rejected; sending and executing require separate owner-bound mock decisions.
  • Only selected task data reaches a typed untrusted mock input; there is no real prompt builder or tool executor here.
  • Persist inbox/cursor, dedupe, work reservation and execution intent before advancing/dispatching. Ambiguous outcomes stay uncertain and are not retried.
  • Cancel tombstones survive reopen; runtime stop must be observed separately. A returned result needs explicit requester acceptance and can be cancelled.
  • Limits survive restart. The last allowed turn may finish; new turns are denied. The coordinator mirrors proposed host behavior without modifying tasks.js.
  • Only explicit RuntimeRecoveryError transport/evidence failures become uncertain; assertion, storage and programming errors propagate. A cancellation racing completion records stopped without reviving execution.
  • result-ready tasks do not occupy an execution slot while awaiting acceptance; they still count toward the journal's total task bound.

Tests and gates

npm run test:interop covers the collaboration path, isolation, approvals, revocation, malformed/out-of-order/duplicate messages, ambiguous recovery, cancellation, budgets and journal ownership. npm run lint:interop lints the harness. Root npm run verify includes both. Required CI runs the suite in the package job and lint in the core test composite.

No real OpenCode API, Bot Social enrollment, model output, OS sandbox or human acceptance is validated. These mock results satisfy none of the team-mode gates.