v3.13.0

Resource/credential leasing ships in this release as a beta feature. The interfaces below are
stable enough for evaluation and feedback, but the capability is deliberately v1-scoped —
multi-holder admission (maxHolders > 1), budget counting, and phase-scoped leases are designed-for
but deferred, and details may evolve based on real-world usage. Deployments that declare no
resources: see zero behavior change.

Added

• Resource leasing (beta). Work items can now declare exclusive or advisory access to shared resources
  (a staging slot, a test database, a deploy credential, a fleet-wide deployment mutex) via a new
  resources: dimension on trait definitions, enforced as a gate inside advance_item at WORK
  entry (start and resume) — no new MCP tool, no explicit acquire/release verb. exclusive
  resources take a TTL-bounded lease (default 3600s, max 86400s); contention rejects with a
  transient resource_unavailable error (MCP) or 409 + Retry-After (REST), disclosing the
  contended keys only, never the current holder. advisory resources are audit-only. Lease
  acquisition retries up to 5 times (40ms delay) on SQLITE_BUSY/SQLITE_LOCKED contention before
  falling through to a transient DB-error response, so a losing writer in a genuine cross-connection
  race gets a clean re-evaluation instead of a spurious failure.
• credentialRefs audit field (beta). advance_item transitions accept an optional credentialRefs
  array of opaque credential/resource labels (never secret values), persisted to
  role_transitions.consumed_credentials and surfaced read-only via RoleTransitionDto.consumedCredentials.
  Once an item declares resources:, the field becomes a closed set validated against the
  declared/registered keys, and declared keys are auto-recorded on work entry.
• Resource-lease read/recovery surfaces. get_context(itemId=...) item mode gains a
  resourceLeases block (mirroring claimDetail). New REST routes GET /api/v1/resources/leases
  (list active leases; holder actor identity ADMIN-only) and
  DELETE /api/v1/resources/leases/{key} (ADMIN-only force-release — the operator recovery path for
  a crashed holder, without waiting out the TTL or disabling enforcement server-wide).
• Enforcement matrix and kill switch. Both MCP and REST enforce the resource-lease gate by
  default; REST additionally accepts an ADMIN-only overrideResourceLeases request flag (403 for
  non-admin callers who set it, WARN-logged when used). New RESOURCE_LEASES_ENFORCED env var
  (default true) is a deployment-wide kill switch, read per advance call rather than once at
  startup. CORS_EXPOSE_HEADERS now defaults to include Retry-After.
• Lease-interval history (beta, audit log). New append-only resource_lease_history table records one
  row per lease hold interval — answering "who held resource R at time T" after a lease has already
  been released, stolen, or force-released. Every acquire/refresh/release/steal/force-release write
  is mirrored into history in the same transaction as the live-row write; ResourceLeaseRepository
  gains findHoldersAt / findRecentIntervals, and forceReleaseByKey gains an optional actorId
  parameter recorded on the closed interval. New REST route
  GET /api/v1/resources/leases/history?key=&at=&limit= (holder/closer actor identity ADMIN-only,
  at optional ISO-8601 instant, 400 on an unparseable value, limit default 100 max 500). No
  pruning/retention in v1 — append-only, cardinality tracks lease events.

Changed

• Every exit from the work phase — primary transitions, cascades, and complete_tree — now releases
  the item's resource leases automatically; TTL expiry remains the crash backstop.
• Bumped plugin version to 3.5.2 — skills and hooks distinguish transient lease contention from
  note-gate blocks (schema-workflow, dependency-manager, status-progression, the dispatched-agent
  context, both orchestrator output styles), /manage-schemas recognizes and preserves the
  resources: config section, and /configure-server documents the RESOURCE_LEASES_ENFORCED
  kill switch.
• Documented the one item = one concurrent actor modelling assumption (from beta field
  feedback): item-keyed exclusivity arbitrates between items, not between actors sharing one item —
  cut items at actor granularity; claim_item is the opt-in enforcement. Stated in the
  workflow-guide guarantees section and fleet capacity planning.

Fixed

• Lease-interval history could report two simultaneous holders for one key (beta field report): a
  re-take of a previously-held key after an intervening expired holder left that holder's interval
  open, and a late release then stamped it closed at "now". Stale holder intervals now close on
  every acquire path, and history closes are clamped to the interval's own expiry with a truthful
  expired reason. Live-lease exclusivity was never affected — audit-view accuracy only.

---