Release summaries
agent-host-extraction-lane-x8
Agent Host protocol, automation, model-route, lifecycle and Runtime Node V3 execution behavior moves to its platform owners
integrator
New explicit subpaths expose the governed automation run lifecycle, model qualification/canary head transitions, model route V2 admission, the application qualification port and agent export artifacts (ai-agent); the application installation lifecycle, application qualification contracts/validation and the markdown migration material schema (agent-protocol-spec); the PostgreSQL application lifecycle adapter and migration (lssm-os-state-postgres); and the Runtime Node V3 execution service (lssm-os-runtime).
agent-operations-harness-ownership
Give agent operations and Harness contracts their canonical runtime owners
integrator
Legacy imports preserve operation schemas and the same scenario and suite registry classes through canonical owners.
ai-canonical-runtime-owners
Preserve complete AI runtime bodies through canonical ownership imports.
maintainer
Preserve provider factory signatures, policy gates, nominal registry identity and complete runtime defaults.
ai-improvement-v2-human-promotion
Governed V2 AI improvement work orders reject failed evaluations and promote to production only with an explicit human approval; policy approvals and auto-apply policies stop after canary.
integrator
Breaking safety correction for V2 only. Previously a failed evaluation waited for approval and any approval (including approverType policy) or a qualified auto-apply policy could promote it. Now a failed evaluation ends as rejected with failure code evaluation_failed, a policy approval covers canary only, and a healthy canary without a human approval waits at awaiting_approval. Approving it resumes observation and then promotes. The V1 runtime and the receipt binding digest are unchanged; the additive AiImprovementApprovalStageV2 type and optional receipt stage field are new. Major bump rationale (semver for a breaking change, compatibility-versioning rule): the removed behaviour violated the documented no-autonomous-promotion policy and has no in-repository production consumer.
ai-protocol-v2-clock-boundary
Preserve historical optional-clock model validation at its runtime owner.
maintainer
Preserve guarded lazy clock reads, field-read order, issue order and fallback short-circuiting through complete structural phases.
cmp-agent-epochs-host-composition
Agent Host composes the bounded-epoch runner for ContractSpecAgent runs, and an open epoch's token and tool-call spend survives a restart
integrator
runAgentBoundedEpochsV1 stores the open epoch's token and tool-call spend, with the number of steps it counts and the number of steps that had reached a journaled effect (openEpochUsage), with its checkpoint. A restarted process continues those counters, so a crash-looping epoch exhausts its budget instead of getting fresh counters; it runs the admitted steps again without a budget check, so each replays its recorded effects or stops the run at its pending one, even after the wall-time budget ran out, and a journaled step is never sent again in a later epoch; and a restart past maxResumes closes the epoch with the persisted spend. An epoch that ends on its budget or a cancellation with a pending journal entry stops as effect_indeterminate instead of renewing, and the result's optional stopDetail names the effect. Steps keep reporting their spend as before: one whose effects all replay is counted once. Records stored before the change load unchanged and resume their open epoch from zero. Agent Host adds createAgentHostBoundedAgentRunnerV1, which bounds a ContractSpecAgent run by a host-owned epoch policy instead of maxSteps alone and journals every generate call, plus the optional boundedAgentRuns server binding and bounded_agent_runs readiness row.
maintainer
The durable record schema accepts an optional openEpochUsage (tokens, toolCalls, steps, admittedSteps) that must belong to the open epoch_started checkpoint. executeAgentEpochV1 takes the optional spent, onAdmit and onSpend(spent, durable) inputs: steps below spent.admittedSteps run without the budget check or the wall-time abort, onAdmit fires when a step first calls context.effect so its first journal write carries the admission, a step below spent.steps whose effects all replayed is not counted again, and durable is false when the step ran a journaled effect, so the next write carries its spend (the runner always stores it before the boundary decision). An execution that ends budget_exhausted or cancelled with a pending journal entry ends as effect_indeterminate, with a detail the runner returns as stopDetail. The Agent Host step runs each generate call through context.effect under the key agent-host:generate:<epoch>:<step>, reports the recorded spend of a replayed call, and turns any failure or abort of a call whose journal entry is written into effect_indeterminate, naming the abort reason or the error's class name. The bounded-epoch DocBlock and the architecture-invariants ledger row describe the persisted spend and the remaining gaps.
headless-chat-runtime
Move canonical headless chat implementation to the lower AI runtime while preserving module compatibility.
integrator
Headless consumers may use the additive AI runtime chat entry; existing module imports keep their canonical identities.
refactor-move-dev-health-repairs
Repair relocation regressions in locale imports, Core type imports, workspace ranges and repository metadata
integrator
Relocated packages resolve their locale helpers and Core types again and publish with accurate repository provenance metadata.