Return Bourne's canonical ExecutionRequest version-2 JSON Schema.

Validate and normalize an ExecutionRequest v2 without discovery, planning, persistence, or execution.

Run bounded local compute-site discovery and persist a new immutable inventory snapshot.

Read an existing inventory by latest, full ID, unique prefix, or @N; this never performs discovery.

List configured non-secret local and SSH site contexts.

Inspect one configured site, its policy claims, and inventory identities.

Discover one configured site and persist a new immutable inventory snapshot linked to it. reference is the site's exact name or canonical ID. Use this for a named local or SSH site; use bourne_discover for Bourne's current local context and bourne_site_inspect to read existing state. SSH discovery may require existing user authorization, uses only bounded typed probes, never accepts arbitrary commands, and never executes a scientific workload.

Append one durable structured policy claim and provenance record to an existing configured site. reference is the exact site name or canonical ID; claim contains the asserted fact, evidence classification, source identity, and applicability. Use this for reviewed site constraints or advice before candidate generation; use bourne_site_discover to observe infrastructure. This stores no source document, fetches no URL, runs no command, and does not modify previous claims.

Generate at most 64 candidate plans for one configured site from an ExecutionRequest, an existing inventory, and optional declarative provider constraints. Use after site discovery and before bourne_site_select; use bourne_plan when site-aware candidate comparison is unnecessary. This does not execute or durably persist a request or plan, but stores an ephemeral candidate session in this MCP process; a restart loses that session.

Choose one candidate returned by bourne_site_candidates, persist selection evidence, and create a new immutable execution plan without executing it. request_id and candidate_id must come from the same live candidate session; regenerate candidates after a server restart. This writes a new plan without editing existing plans. Review the returned plan before calling bourne_execute_plan; selection fails when required approvals, declarations, or provider trust are missing.

Persist and resolve an ExecutionRequest v2 against an existing inventory. Planning never executes the workload and preserves ambiguity.

Execute one immutable persisted Bourne plan without changing its command, resources, placement, backend, or inventory.

Read a Bourne execution, including request, plan, lifecycle, scheduler, allocation, experiment, telemetry, and verification state.

Reconnect and reconcile one exact Bourne-owned remote execution; never resubmit.

Wait on one existing Bourne-managed scheduled execution with an optional bounded caller timeout; this creates no execution.

Cancel only the exact scheduler job owned by an existing Bourne execution; arbitrary scheduler job IDs are not accepted.

Trace a recorded output artifact to its producer, inputs, and experiment ancestry without guessing across ambiguous matches.