Author Flowglad Ows

Author Flowglad OWS

Create the canonical OWS source for a Flowglad automation.

Authoring process

1. Gather the workflow inputs, outputs, tasks, dataflow, artifacts, external systems, reusable capabilities, and side effects.
2. Call the connected Flowglad MCP tool automations.get_authoring_spec with {}. Record its specVersion, specDigest, compilerVersion, and section index. Treat this deployed specification as authoritative. If the tool is unavailable, stop and report that the current authoring specification could not be loaded; do not substitute remembered or bundled profile documentation.
3. Fetch getting-started, required-envelope, task-shape, dataflow, and validation-diagnostics by calling the same tool with { "section": "<id>" }. Use the index's summary and appliesWhen fields to fetch every additional section relevant to the workflow. Fetch artifacts whenever an input or output originates as a file or is a large value. Require every section response to match the index's specDigest before authoring.
4. For reusable code, read references/SKILL_RESOURCES.md and resolve an exact verified Skill revisionId, Python module path, and selected function. If the capability is missing and author-flowglad-skill is available, create and verify it before resuming. Use a space-file only for intentionally automation-specific code or when the user requests it.
5. Read references/CODE_ABI.md before writing Python selected by a code task.
6. Inspect the closest complete example as supporting material after loading the relevant live sections:
   • references/skill-backed-code.ows.yaml: verified Skill revision.
   • references/invoice-reconciliation.ows.yaml: authenticated reads, artifacts, and a terminal DAG.
   • references/yooz-bounded-document-number.ows.yaml: code, inference, artifacts, and terminal output.
   • references/mock-bank-browser-statement.ows.yaml: browser agent, PDF artifact, and code consumer.
   • references/page-todo-update.ows.yaml: exact Page reads and Page edit proposals.
   • references/mcp-ows-v1.ows.yaml: pinned MCP reads and command proposals.
7. Copy assets/starter.ows.yaml for a simple Skill-backed workflow and replace every placeholder.
8. Write YAML with the .ows.yaml extension. Create companion Python only for a selected space-file resource.
9. Validate with the Flowglad compiler or an advertised Flowglad MCP validation tool. Resolve every diagnostic before claiming compiler validity.
10. Publish or modify a remote Automation only when the user asks. An update publishes a new immutable revision; it never rewrites an existing revision.
11. Report the live specification identity, compiler validation, resource resolution, publication, activation, and execution as separate states.

Integrity guardrails

• Never invent Flowglad IDs, connection authorities, catalog pins, file paths, or browser-auth connections.
• A Skill resource contains revisionId and path; Flowglad derives the parent Skill. A Space resource contains path; Flowglad resolves it in the Automation's Space.
• Do not author runtime for code tasks. Flowglad seals Python 3.11 and the memory limit.
• Do not author inference provider, model, reasoning effort, or output limits. Flowglad seals them from organization policy.
• Browser profiles contain authored instructions, not prompt, vendor-skill, tool-set, or model-route identities.
• Resolve every pinned MCP catalog revision, schema hash, operation hash, and input schema from advertised capabilities. Never guess them.
• Do not infer OWS compatibility from Skill verification alone. Publication validates the selected function from the exact revision.
• Compiler acceptance of a placeholder revision proves document shape only; it does not prove that the Skill resource exists or is publishable.
• Leave an obvious placeholder when an exact value is unavailable and list it in the handoff.

Profile guardrails

• Author OWS 1.0.3 with Flowglad automation profile v1 and the versioned Flowglad catalog.
• Keep portable behavior in standard OWS fields and Flowglad-specific declarations under with.fg.
• Use inline JSON Schemas, a sequential top-level do list, and an explicit timeout on every task.
• Use only task shapes documented in the fetched live authoring-spec sections.
• Keep orchestration in OWS and reusable computation in governed Skills.
• Do not generate legacy v0 definitions, StepSpec[], runner-specific representations, or unsupported loops, retries, waits, forks, schedules, composite tasks, third-party catalogs, or generic run tasks.
• Successful compilation does not authorize publication, activation, or execution.

Handoff

Return the complete .ows.yaml, any companion space-file Python, every pinned revision/path/function or MCP capability identity, unresolved publication dependencies, exact validation and remaining diagnostics, and the distinct Skill-verification, Automation-publication, activation, and execution states.