Execution Files, Flows, and Jobs

The execution lane locates BioLine objects only through real fkit responses. A request keyword is a query, not a local identifier or a static catalogue key.

Files: list inputs first

After docs_loaded, call fkit ls <input-path>. Planning records the input path, file listing, and each binding in execution; paths must belong to the execution boundary or be a BioLine-readable input. Missing files, ports, or defaults are not guessed: stop with document_gap or plan_failed and preserve the evidence.

Flows: discover, then inspect

Call fkit flow list with the requested keyword. Use the response's id and version_id only when there is one exact match. Then call fkit flow inspect <id> and check input and output ports, file groups, and parameters. If several approximate matches remain, show the candidates and stop instead of selecting by list order.

Jobs: submit after confirmation

The sequence is fixed:

1. Write Flow, inspect, file-list, binding, and default evidence to the plan record.
2. Enter awaiting_plan_confirmation and show the readable plan and impact.
3. After confirmation, create the reproducible spec and call fkit job create.
4. Record the returned Job ID, enter submitted → running, and poll non-blockingly with fkit job get.
5. Finalize only a successful status and use fkit download to organize stage/.

Submission is not publication, and polling is not success. Preserve response, timestamp, and phase for every failure and use its corresponding *_failed value.

Query command documentation with fkit

OpenClaw requests the page for the exact command it is about to use. The request does not require a BioLine login:

fkit flow list --docs --docs-json --docs-lang en
fkit tool summary export --docs --docs-json --docs-lang zh

The response contains the selected command path, language, document ID, canonical page URL, fixed source, and Markdown. The source must be https://www.bioline.cloud/ai/commandLineAndFkit/, and Markdown assets must share the index origin. Missing or invalid command pages fail with evidence and become document_gap. OpenClaw does not scrape the documentation site's HTML, read a local documentation bundle, or use a Skill fallback or pipeline map.

Handoff to downstream

After finalization writes stage/manifest.json and the reproducibility record, prepare an exact BioLine NAME:VERSION. OpenClaw checks its downstream summary and empty config command, then pulls that version. The historical image list is migration inventory only. stage/ and task reproducibility are read-only, the network is disabled, analysis is writable, and the generated script, outputs, and logs stay inside the Job workspace.