4.1 How the Platform Works

4.1.1 Execution lane: plan before running

The execution lane creates an isolated workspace and record for each BioLine Job. fkit discovers Flows for the objective, inspects their interfaces, and lists input files. Planning is read-only and records real IDs, bindings, defaults, and risks; no Job is created before user confirmation. Command-scoped help is read by fkit from the official BioLine AI documentation service without login; fkit does not parse the user-facing HTML site.

Only after confirmation does the lane submit the Job and record the ID returned by BioLine. Polling is non-blocking. Only successful results may be finalized into stage and passed downstream.

4.1.2 Downstream Tools and local isolation

Downstream capabilities are exact BioLine Tool versions. OpenClaw exports and checks their summary/config, then pulls the selected NAME:VERSION on demand. The summary declares analysis_lane: downstream and fixed_command: false; config command fields are empty or absent. A task-specific script is generated from the current inputs and objective, saved before execution, and run in a container with networking disabled. Stage and reproducibility material are read-only; analysis is writable. The historical image set does not need to be preinstalled.

4.1.3 Authoring lane

The authoring lane uses a separate workspace to create a Tool, Tool Version, or Flow. It records source material, specifications, static checks, and server read-back. The tested phase performs static validation; it does not run a real analysis or data smoke task. Build, publish, and register each require confirmation. After publication, read the BioLine object back and compare key fields; handoff follows successful verification.

4.1.4 Documentation gaps and capability discovery

fkit command documentation always uses https://www.bioline.cloud/ai/commandLineAndFkit/; it does not use a local bundle or source override. If documentation or an interface is missing, record document_gap and stop instead of inferring rules. First discover existing BioLine Tools and Flows; author one on demand only when no suitable object exists. Historical images and Skill workflows are migration material, not active runtime dependencies.