Private AI

Create a workflow, or change one (only send what changes; every change raises config_version). Owner only. A new workflow needs label, provider, model, instructions and output_schema. The key of the provider must be in the vault (see get-workflows: provider_status). output_schema is a JSON Schema subset (type, properties, required, additionalProperties, items, enum, minLength, maxLength, minimum, maximum, minItems, maxItems, title, description) with "type": "object" or "array" at the top; other keywords are refused, so everything you write is enforced. allowed_refs lists the private files a run may read: {"type":"gated_file","file_id":990}, {"type":"private_asset","slug":"library/a.json"} or {"type":"private_asset","prefix":"library/"}; an empty list allows none. Admin sessions can run workflows (allow_admin) and a determined admin could get the model to repeat its instructions through a text field, so keep secrets out of instructions. Next: run once to test.

All workflows of this project (or one, with workflow_key) with today's AI calls against daily_cap, plus defaults, platform bounds and the supported schema keywords. The owner also gets instructions, output_schema, allowed_refs and provider_status (is the provider key in the vault); an admin session does not.

Delete a workflow and its usage counter permanently. Owner only. The run log (metadata only, see list-runs) stays until the project is deleted. To pause instead, use set-workflow with is_active false.

Run a workflow on the server. Send case_packet (the data of this case, a JSON object) and optionally knowledge_refs (private files from the workflow's allowed_refs; the server reads them itself, UTF-8 text only). Returns result (JSON that matches output_schema), run_id, attempts, token_usage and metadata of the files used (type, id, sha256, bytes) - never their content or the instructions. Errors: PRIVATE_AI_OUTPUT_INVALID (no valid answer after the repairs; context.errors names the problems), PRIVATE_AI_PROVIDER_ERROR, PRIVATE_AI_DAILY_CAP, REF_NOT_ALLOWED, CASE_TOO_LARGE, KNOWLEDGE_TOO_LARGE. Every AI call counts for daily_cap. Owner, admin session (when allow_admin) and server flows; never site visitors.

The run log of this project, newest first: status (ok, output_invalid, provider_error, limited), error_code, attempts, sizes, files used (metadata), token usage, caller type. Never case data, file content or answers.