Rho-plugins develops concrete functional components for both headless callers and the Rho application. Architecture defines ownership; Status records implemented behavior and unresolved limits.
The plugin model separates reusable capability plugins, their optional entries and project workflows. The headless 0.4 plugins independently package the reusable foundation; its source trials use both an ordinary R project and the paired example. The following gates define the delivered headless boundary and the remaining UI acceptance.
- Remove the fixed paired recipe from base catalog/packaging. Keep the example, scientific checks and historical evidence as an explicit consumer.
- Extract declared script execution, requirement providers, artifact readers and checks behind reusable owners. Rho keeps request mapping and the Core bridge.
- Declare table types/keys, chart field mappings, media resources and bounded object snapshots. Remove gene/fold-change/model assumptions from generic readers and application routing. Live R inspection requires its own native adapter.
- Deliver versioned capability/view contracts, client and fixtures with explicit compatibility for old records. Restore missing entries without execution.
- Verify the base install without scientific example dependencies, an ordinary non-gene R project and the paired recipe through the same capabilities. Then implement and accept browser entries against those contracts.
The application extraction plan records owners, source locations, migration and completion conditions. Actual source and installation evidence lives in Rho's extraction acceptance. Browser entries remain pending.
The first application flow develops scientific workbench interaction, freely composed views and Agent-driven UI together. Its selected components cover an Agent bridge, document/code, runtime, requirements, table, media and chart capabilities, exercised by an explicitly selected project recipe.
- A user gives an analysis request. The Agent selects the component's action and shows its parameters through a registered UI component. Headless and UI entries use the same native implementation and public Core execution capability.
- Native execution produces actual script, table, plot and provenance resources. Execution completion, artifact completeness and scientific checks are visible as separate facts, including failures and partial results.
- Selecting a stable domain object in a table or plot updates participating views and the Agent's explicit context within the same resource version.
- Views dock, split and move within one browser window while retaining Unicode input, unsaved content and undo. Agent edits respect buffer ownership and conflicts.
- Closing an observer and reconnecting find the original task and its actual output. Missing acknowledgements preserve the original request identity and unknown outcome; they do not cause automatic replay.
- The Agent composes another view using the same typed catalog, resources and actions. Invalid, stale or out-of-scope references produce an observable result.
Rho first delivers an npm local service and React/TypeScript, Vite and Dockview WebUI. Electron and cross-window movement belong to a later client. Functional views select CodeMirror 6, AG Grid Community and ECharts; Agent-composed views use a json-render catalog. MCP Apps is optional for an isolated independent HTML consumer. Rho's technology direction is the application selection reference. The headless plugin packages and its Rho HTTP/Core/Agent composition have been checked with two explicit R projects. UI assembly remains pending.
Acceptance also covers input methods, keyboard access, missing components, stale resources, failed view restoration and version mismatch. Record the actual chosen source/artifact identities and overrides. A successful component check alone does not accept the combination. Set and measure limits for the actual protected processes, output retention, buffer sizes, subscriptions and rendered resources; name configuration and exceeded-limit behavior.
The paired airway recipe is an example, not a plugin. It supplies real batch execution, conditional simulation, ordinary artifacts and native/independent checks. Reuse its evidence within the recorded acceptance. Existing R tools, RDS, manifests and ordinary file operations remain valid paths. Scientific owners retain method choice, model assumptions and interpretation.
Reusable domain extensions may contribute parameterized native operations and object adapters when a demonstrated consumer needs them. Dataset/design/contrast choices and workflow sequencing stay in recipes and do not become base contracts. Extend a recipe or add native-session support only for a demonstrated user need. A request requiring interactive R memory must specify object identity, lifetime, input and recovery guarantees before implementation. Waiting continuation does not establish R-memory recovery after Host replacement.
Artifact work preserves input identities, native versions, completion declarations, partial outputs and provenance. A copied file is checked against its declared source. Missing or mutable files stay observable; a name or current path alone does not establish an earlier task's output.
The evaluation component provides isolated Core CLI trials, domain-owned result checks and offline regrading. An improvement experiment is a separate selected action; its current implementation is pending.
An experiment fixes cases, expected behavior, subject identity, model configuration and scorer version. It declares allowed candidate changes and compares attributable patches against a retained baseline. Domain checks stay with their owners; model judgments cannot override native failure or scientific limits. Exceptions, timeout, interruption and missing evidence remain in the denominator.
Development cases guide diagnostics, validation cases compare candidates, and a blind final set supplies held-out acceptance. Material used repeatedly in selection is no longer an unseen final set. Regrading writes new judgments against retained materials and scorer identities without replacing original reports. Report gains within trial variation as insufficient evidence.
Before running, configure trial count, process timeouts, model/token budget and evidence retention, and define stop behavior for exhaustion. The existing evaluator implements the documented trial/process/output limits; experiment-wide model budgets, candidate comparison, Git rollback and blind-set management remain work. Candidate results do not implicitly publish source, install software or restart user runtimes.
A second non-gene consumer is delivered as an ordinary base-R project. Subsequent domain extensions and interoperability adapters require a selected consumer. Deliver owner changes independently, then accept their actual application combination. Distribution and installation have separate acceptance and do not follow automatically from source publication.