# Changelog

## 0.1.0-beta.1 — unreleased

`npm run demo` now opens the research quickstart. The smaller ticket-routing example remains available as `npm run demo:triage`; `demo:typed` remains an alias for research.

The quickstart now demonstrates both a completed report and stopping on missing evidence. Research commands have readable summaries; saved reports retain full JSON. Evaluation reports derive their threshold from the workflow, including after user edits.

The standalone candidate now targets `Parcha-ai/agentrun` with clean DSL-only history. Research can continue with explicitly configured Jev and Pi adapters, and the same evidence-selection component has independently labeled evaluation cases. Offline results remain labeled as scripted.

Generated documentation now includes typed authoring, configured research, host integration and compatibility. CI checks local links, generated-file drift, minimum Node/TypeScript support and clean package consumers. Manual beta publication is gated by approved licensing, an exact reviewed tag and verified archive bytes; it does not run on push. Packages remain unpublished.

Typed authoring now accepts Standard JSON Schema contracts, with a fail-closed Zod subset and inferred input/output types. Definitions serialize to the existing v2 document and use the same interpreter. Imported JSON remains supported with runtime validation.

Child workflows now compose maps, routes, loops, parallel branches and nested children with isolated state and schemas. Execution paths distinguish nested occurrences; existing flat effect keys remain compatible. Recovery of new compositions requires an explicitly path-aware host. Terminal reports/artifacts remain parent-owned.

Recorded events and the regenerated website replay fixtures now include `executionPath` alongside display labels.

The new `inspect` CLI command and structural inspection API display a workflow without running authored code. A typed research example demonstrates component reuse and isolated Jev system one evaluation. Example tests now run in CI, and package verification checks typed authoring from a clean consumer install. Existing-harness integration is checked separately from standalone features.

Initial standalone AgentRun DSL candidate: neutral workflow interpreter, optional Jev adapter, Pi runner and candidate author, offline examples, generated schema, and inspectable website demo.

Verification rejects exhausted or unchecked candidates. Input schemas and concurrency bounds are enforced. Effect deadlines retain uncertain outcomes. Report failures propagate. Application-specific policies, brokers, budgets, and artifact delivery are outside the core.

Candidate corrections: the documented dry-run accepts declared or supplied input; validate always rejects unknown node fields; missing SOPs fail before effects. LLM instructions are literal, optional state maps scope their JSON input, and maps support explicit resultPath selection. Host question-count guards, typed concurrent escalations, and loop exit events make limits and stopping conditions observable.

The Pi example now demonstrates retained authoring, fixed host acceptance fixtures, and explicit execution without a key. Public exports are intentional; private input conventions and reference utilities are no longer exported from the default entrypoint. Unsupported enum_equals is removed; use field_equals with an explicit path.