AgentPlat 1.1.0 — release preparation

Status: published and verified on 2026-09-29. All 66 coordinated packages are available at 1.1.0 under latest. See the distribution record and brief adoption guide. The preparation procedure below records the gates used before publication.

Why 1.1.0

This is an additive release under the 1.x stability contract. The importance of governed agent autonomy does not itself require a major version. A 2.0.0 release would be appropriate for removal/incompatible changes to supported APIs or incompatible persistence/protocol migration. Existing instruction behavior and omitted-field definition digests are retained; purpose execution is opt-in.

The stable compatibility gate uses the recorded 1.0.0 source commit d59e59e452711c834728470f93953c93908da1a3, preserves its 216 public entrypoints and compiles its 83 unchanged public TypeScript fixtures against the candidate. The candidate adds one public entrypoint for Jev. These checks complement behavioral and migration tests; they are not an exhaustive proof of every consumer's behavior.

Included changes

• Governed instruction/purpose modes, authenticated owner/delegate configuration, versioned history, inceptions, signals/references and narrowed execution limits.
• Qualified Room Planner mission/outcome cycles. Completed tasks and proxy changes cannot alone establish purpose fulfillment; suspended/stale work is fenced.
• Cumulative budget reservations, explicit reconciliation of unknown effects, bounded ancestry/delegation, progressive autonomy evidence and evolution receipts.
• PostgreSQL migrations 013–018 and a persistent support demonstration covering separate-process restart, owner correction, evidence-based recovery or escalation.
• Optional Jev assessor/client, content-free evidence, explicit request limits and conservative spend reservations; existing Rooms proposal-review wiring.
• Coordinated adapter/development dependency updates, including Temporal 1.24.0 and the locked Next.js 16.3.6 example. Existing security overrides remain enforced.

Harbor #181 is not part of this release. Neither the deterministic support demo nor Jev's fixture tests establish model quality, production-scale performance or AGI.

Optional Jev installation and setup

Applications that choose the optional Jev adapter can install:

pnpm add @agentplat/assessor-typesafe@1.1.0 @agentplat/inference-control@1.1.0

Other applications keep their current assessor. Core/runtime/framework/Rooms/Mesh and Inference Control do not depend on Jev or its SDK. The library does not read API credentials automatically, initiate calls on import or replace the host's policy/approval authority. Supply an API key only on the server and only for data the application is authorized to share. Pin a model identifier when reproducibility matters; provider confidence is not an accuracy guarantee.

The adapter README (opens in a new tab) provides a typed TypeSafeAssessorV1 configuration with buildRequest, mapResult and optional evidence sink. For the existing source example, select PROPOSAL_REVIEWER=jev, configure TYPESAFE_API_KEY outside the repository and optionally TYPESAFE_MODEL. The default PROPOSAL_REVIEWER=rules remains available. See the Rooms example (opens in a new tab) for execution and scope. Jev does not automatically implement every purpose-mission assessor or install a scheduler; those remain explicitly configured host integrations.

Upgrade from 1.0.0

1. Retain the existing deployment, lockfile and a verified database backup. Upgrade the coordinated AgentPlat packages together after publication is verified.
2. Compile the application and run its existing tests. Definitions that omit interaction remain instruction-driven and are not automatically enrolled.
3. For PostgreSQL, stop older workers and apply the ordered migrations through 018 with the same explicit schema used by all stores. The 012→018 test preserves legacy Room/definition records and runs instruction work before/after upgrade.
4. Opt into owner governance and qualified execution separately. Register identity, assessor, scheduling and effect-control ports before owner activation. Selecting a purpose revision alone grants no execution authority.
5. Exercise suspension, owner correction, restart, stale-task rejection and uncertain external-effect reconciliation in the application's staging environment.
6. If switching a governed agent back to instruction mode, use the owner mode command with a compatible published revision and qualified reactivation. Retain limits, budgets and origin; do not bypass governance with an older worker.

Prefer a forward fix to schema downgrade. Migrations 016–018 intentionally refuse rollback when enforcement history would be discarded. See the migration guide and adoption guide.

Artifact preparation and publication boundary

The existing protected release workflow must run from an approved, clean main commit with dry_run=true, scope all, and intended distribution tag latest. Its prepare job reads the repository's protected terminology denylist, runs strict public/dependency audits and complete verification, packs all 66 packages, checks the exact tarballs and uploads an immutable artifact manifest with source commit, SHA-512 integrity values and manifest digest. The publish job is disabled for this run. Never substitute a fabricated denylist, disable a gate or invoke staging/ publishing scripts to finish preparation.

Local checks use pnpm run check, pnpm run verify:stable1-compatibility, pnpm run verify:jev-consumer and the PostgreSQL-enabled purpose gate. The Jev consumer uses an injected transport and makes no provider API calls. Consumers may be verified from the exact prepared tarballs with AGENTPLAT_PREPACKED_TARBALL_DIRECTORY after download.

Actual publication is a separate action: confirm the final commit and artifacts, configure/verify the first-publication trusted publisher for the Jev package, and retain the existing protected environment, identity, staged-byte/OIDC and npm 2FA controls. No availability claim or npm tag change belongs to this preparation.