rulvar v1.1.0

The 13 lockstep packages move to 1.1.0 (@rulvar/compat stays 0.1.0). Published through npm trusted publishing (OIDC) from release.yml, with provenance.

New in @rulvar/evals (minor): the M12 measured-value checkpoint harness. runValueCheckpoint executes the OQ-09 gate as two A/B experiments under one fixed pool (rung selection per ladder and taskClass cell; agentType selection with and without the knowledge store), with renderCheckpointReport producing the docs-ready record. The fixed mixed corpus and the budget-guarded live Anthropic runner ship as repo scripts. The gate itself ran live and FAILED (criterion 1 held, criterion 2 failed the cost bar), so ModelKnowledge phase 3 stays unshipped; the record lives in docs/checkpoints/2026-07-11-m12-value-checkpoint.md and docs/14 OQ-09.

Fixes found by the live checkpoint runs:

• @rulvar/anthropic: a Haiku 4.5 capability entry (enabled-budget thinking form, real pricing); the router scrubs the effort parameter off the wire for models that reject it; the adapter scrubs constrained-decoding-unsupported schema keywords (minimum, maximum, exclusiveMinimum, exclusiveMaximum, multipleOf, maxItems) from the wire copy of strict tool schemas and output format schemas, with the engine still validating against the full schema (docs/04 4.3 as amended).
• @rulvar/core: plain orchestrate treats ladder-declaring profiles as declaration-only: the spawn vocabulary advertises concrete profiles with declarers on a context line, and spawning a declarer is rejected with a typed ConfigError before admission (docs/07 section 10 as amended).
• @rulvar/evals: criterion 2 cannot pass vacuously (an all-zero arm fails); orchestrated arms take their own suite budget (orchestratedSuite) so the orchestrator cap math stays solvent under a small per-case ceiling.

Full per-package changelogs: https://github.com/o-stepper/rulvar/blob/main/packages/core/CHANGELOG.md and siblings.

rulvar v1.0.0

rulvar v1.0.0

The first published release: an embeddable TypeScript engine for durable, budget-bounded, testable multi-agent LLM workflows. No server, no database, no control plane.

What ships

• Durable execution as a library. An append-only journal with byte-deterministic replay and crash resume over plain JSONL files or SQLite; multi-process workers with lease fencing. A completed LLM call is never paid for twice.
• Testability as a contract. Hermetic VCR cassettes record real LLM exchanges once and replay them in CI with zero live calls; a frozen catalog of 60 defect cassettes gates every commit of the engine itself.
• Hard budget accounting. Per-run USD ceilings, orchestrator sub-budgets with finalize reserves, spawn/revision/escalation unit limits, and admission control: runs terminate cleanly under a hard ceiling, provably from the journal.
• Adaptive orchestration. Typed plan revisions with rebase, worker escalation protocols, model ladders (cheapest first, escalate on evidence), wake digests, lineage and lesson tracking, dedup and result reuse.
• ModelKnowledge (phases 1 and 2). A git-reviewed, per-project claim store about model suitability per task class: TTL decay, eval-measured claims from matrix sweeps through a dedicated committer identity, canary fingerprints for silent model swaps, and a one-rung-clamped verified layer. Advisory only: floors and caps stay hard.
• Provider-neutral. First-class adapters for Anthropic, OpenAI-compatible endpoints, and Google, plus a Vercel AI SDK bridge; model routing by role with quality floors.
• Tooling. The rulvar CLI (run, resume, runs ls, inspect, plan, kb list/sweep), an eval framework (golden, rubric, LLM-judge graders, matrix sweeps), store conformance kits, and a testing harness.

Versioning note

This release ships the scope of internal milestones M9 through M11 together (docs/12, section 2, amendment of 2026-07-11); the 1:1 milestone-to-version mapping resumes from the next milestone. Thirteen packages release in lockstep at 1.0.0; @rulvar/compat versions independently (0.1.0).

The seams

The six core SPI seams (ProviderAdapter; JournalStore/LeasableStore; TranscriptStore; ScriptRunner; ToolSource; IsolationProvider) are FROZEN as of this release, guarded by committed .d.ts baselines in CI; ModelKnowledgeStore froze alongside as the sanctioned seventh seam (docs/02, section 5.1).

Provenance

This first release was published by the maintainer directly while the repository was private: npm provenance attestations are NOT attached to 1.0.0. Provenance guarantees begin with the first release published from CI via OIDC trusted publishing after the repository goes public (docs/12, section 6).

License

Apache-2.0. Contributions are accepted under the DCO.

Full changelogs live per package (see packages//CHANGELOG.md); the complete 1.0.0 history spans the m9-, m10-, and m11- changesets plus the rename and licensing changes.