Molds

A Mold is the Foundry’s unit of action: an abstract, typed source artifact that casting turns into a skill an agent runs. Each is atomic at the harness-step tier, which is a statement about granularity, not size. mold-spec owns the authoring contract, casting owns what a cast does with one, and harness-pipelines owns which Molds each journey composes.

This record owns the shape of the Mold set rather than its membership — the axes a Mold buckets on, the boundary that decides whether a piece of knowledge becomes a Mold at all, and where the set is still uneven. It does not list the Molds that exist. content/molds/ is the membership; the generated content/Index.md and content/Dashboard.md project it with summaries; the site’s Mold browse renders it. A hand-kept catalog here would be a fourth copy of that, and the first Mold authored after it was written would make it wrong.

The axes a Mold buckets on

axis is a required frontmatter enum, not a mental model for grouping: source-specific, target-specific, tool-specific, or generic. The chosen value requires the field that names the specialization — source, target, or tool — and the schema rejects an axis whose companion field is missing. mold-spec owns that contract. What each value means is here.

• source-specific — the input format determines the content. A paper, a Nextflow tree, a CWL document, and an existing Galaxy workflow each demand different reading, and no shared summary shape survives forcing them together.
• target-specific — the output target determines the content. Galaxy and CWL disagree about what a step, a test, and a validation failure are.
• tool-specific — reserved for an action that genuinely depends on one external tool’s behavior. No Mold carries it. Whole-CLI reference surfaces are not what it is for.
• generic — none of the above.

The enum records one specialization, not a pair. Many Molds are genuinely source × target — nextflow-summary-to-galaxy-interface reads a Nextflow summary and designs a Galaxy interface — and declare source-specific, leaving the target to the Mold’s name and its input_artifacts[] / output_artifacts[] contracts, which is where the handoff is actually pinned. Whether one value plus the artifact contracts carries enough is open below.

Mold or reference content

A Mold acts; reference content explains. The test is whether the content terminates in a decision — emit this artifact, pick this wrapper, route back to that phase — or whether a reader consults it. Getting this wrong toward reference is cheap: a pattern page that should have been a Mold just sits there being read. The other direction puts explanation inside a cast bundle, where nothing cites it and nothing keeps it current.

Excluded by design, named so the boundary stays visible:

• Pure reference content. Pattern pages, CLI manual pages under content/cli/<tool>/<command>.md, schema notes, prompt fragments, examples, and research notes are referenced by Molds, not Molds themselves. reference_contract.yml registers the kinds and casting dispatches each one.
• Harnesses. Hand-authored orchestration that sequences Molds and is never cast from one.
• Approval gates, scope confirmation, plan presentation, and tool-discovery routing. Harness-level concerns; harness-pipelines holds the reasoning for each.
• Hand-authored prior-art skills. The gxwf-cli help-text dump and the find-shed-tool skill design are prior art. Their content feeds CLI manual pages and action Molds; their form does not.

Wrapping a CLI is not a disqualifier. discover-shed-tool, advance-galaxy-draft-step, validate-galaxy-workflow, and run-workflow-test all wrap CLIs and are Molds. The criterion is whether there is procedural content worth casting — when to run, how to read the result, when to loop back — not whether the mechanism underneath is deterministic.

Worked example: compare-against-iwc-exemplar

A case that could have gone the other way. “Compare the design against the IWC corpus” sounds like knowledge: the corpus-first principle is already reference content (corpus), and one plausible shape was a pattern page describing the idiom, cited by the template Mold.

It landed as an action Mold because the criterion above resolves the same way it does for the CLI wrappers — there is procedure worth casting. It runs after the design briefs and before the template Mold, so corpus divergence is caught before per-step effort is spent; it ranks candidates by a feature hierarchy rather than superficial similarity; and it hands off a structural-diff artifact the template Mold consumes. The corpus-first principle stays reference content; the act of locating the nearest exemplar, ranking it, and gating the template step is a Mold.

Where the Mold set is uneven

Direction, not inventory. Each of these is a property of the set that membership alone does not show.

• The Galaxy target is far ahead of the CWL target. Galaxy paths have a per-step orchestrator, a corpus-grounding Mold, and four test-plan producers; the CWL paths have leaf-shaped per-step work and one test-plan producer, which is Nextflow-sourced. The paper-sourced CWL journey has no test-plan producer at all — issue #448.
• The run/debug tier is thinner than the authoring tier, on purpose. run-workflow-test and the two debug-*-workflow-output Molds are sized for work an agent could often do ad-hoc. They are Mold-shaped for inventory completeness; whether they earn casting is a question a walk should answer, not this record.
• Walk status is not kept here. A Mold earns its eval.md and scenarios.md after a first walk, not before, so the validator’s missing-companion warnings are the closest thing to a walk-status surface. They are meant to stand until a walk closes them, and clearing one by authoring would remove the signal without changing the fact.

Open questions

• Does axis need a pair form for source × target Molds, or do one value plus the artifact contracts carry enough?
• tool-specific has no member. Does it survive the next ten Molds, or was the fourth value speculative?
• Is the CWL-target lag a deliberate scope decision or an unclosed gap? The pipelines disagree with each other today.
• Do the run/debug Molds survive their first walk as Molds?