Skip to main content

Release Notes — v0.7.10

AgentPlane 0.7.10 retires Blueprint runtime ownership and makes the Task Kernel plan, execution contract, and native task obligations the only active sources for decomposition, route authority, policy, evidence, mandatory lifecycle stages, and stop rules.

Summary​

  • Remove Blueprint selection, snapshots, plans, cursors, and execution state from new tasks, WorkOrders, prompts, verification inputs, evaluator evidence, and ACR records.
  • Preserve planning, approval, independent evaluation, deterministic verification, provenance, freshness, recovery, effect-in-doubt, and hosted-integration requirements on native owners.
  • Keep Recipe V1 guidance and runtime assets. Convert only exact legacy requirements; require manual conversion for custom graphs without a native equivalent.
  • Add explicit preview/apply/rollback migration for legacy tasks and a non-executable historical audit decoder.

Compatibility and migration​

The Blueprint registry, resolver, graph engine, mutation and catalog commands, project-local trust store, prompt projections, generated live schemas, and new-task fields are removed. Existing automation must stop calling agentplane blueprint ..., agentplane blueprints ..., or advanced init Blueprint installation options.

Inspect an existing legacy task with:

agentplane task kernel-migrate <task-id>

Apply only the exact reviewed source digest:

agentplane task kernel-migrate <task-id> \
  --apply \
  --source-digest <sha256:digest> \
  --yes

Migration preserves original bytes and emits a guarded rollback proof. Unknown schemas, ambiguous inputs, unreconciled effects, active leases, and unsupported backends stop or quarantine the task; AgentPlane does not infer a replacement graph.

Historical Blueprint data remains readable only for version-labelled audit and migration. The cold decoder cannot select a route, issue a WorkOrder, execute a graph, or mutate current task state.

Preserved lifecycle boundary​

  • PLANNER remains required for new task planning.
  • EVALUATOR remains required wherever current policy requires independent review.
  • direct and branch_pr retain their repository, provider, approval, verification, and closeout boundaries.
  • Scenario V2 and lifecycle-owner convergence remain deferred beyond 0.7.10.
  • Publication remains a separate exact-head operator action after local and hosted qualification.

M02 efficiency disposition​

M02 is NOT ESTABLISHED for 0.7.10. The benchmark corpus and deterministic replay are required release checks, but the repository contains no matched, quality-equivalent paid-provider 0.7.9 and 0.7.10 campaign. Therefore 0.7.10 makes no token-, cost-, latency-, or safety-efficiency claim from the available harness results. A future preregistered matched campaign must preserve every assigned attempt and may not selectively retry or replace failed observations.

Upgrade notes​

  • The CLI, @agentplaneorg/core, @agentplaneorg/recipes, and @agentplane/testkit move together to 0.7.10 with exact internal package versions.
  • Preview managed-file changes with agentplane upgrade --dry-run, then apply the reviewed upgrade with the repository's normal agentplane upgrade flow.
  • Migrate legacy tasks individually. Do not hand-edit task state or recreate retired Blueprint artifacts.

Verification boundary​

The release candidate requires the packed-install smoke, release-critical suite, architecture checks, Blueprint retirement guards, schema and documentation checks, benchmark corpus and replay, full local CI, independent semantic review, hosted CI, and protected exact-head integration. Publication and distribution readback occur only after the candidate is merged to the protected release base.