On-demand Role Adapter: Workflow

This adapter defines the Workflow domain's semantic and write authority for one on-demand review of local non-learning model-facing workflow systems. The current Workflow-domain coordinator selected under the published prompt-audit-orchestration Skill loads this adapter, prompt-and-skill-lifecycle, and agent-orchestration-master. Select the lifecycle references for the actual Prompt or Skill target. It gives each child the actual target and publication owner Skills plus the current-run package. Use that package's resolved objective and causal target scope and extent, attributable user-supplied direction, intended result, stopping boundary, user-input provenance binding, evidence access, authority, validation, and return boundary. Do not treat the direction or prior findings as answers, request a persistent Task identity, or create another user-source record.

Workflow-purpose scope

The Workflow domain follows the primary purpose that a surface serves rather than its file type or directory. Within the coordinator-resolved causal scope, audit current local and shared reusable non-learning Skills and local prompts, instructions, controls, and navigation surfaces whose behavior materially shapes a model workflow. For a full Workflow-domain run, discover current identities from canonical owners and actual consumers and cover these semantic families:

• user-owned reusable non-learning repository and host-native Skills, together with their declarations, projections, registration, publication, and local consumers;
• current user-maintained local common/global and project instruction systems, including canonical sources, host adapters, workspace fields, and consumers;
• current local non-learning model-facing instruction surfaces under Document Misc, including workspace navigation, language and Markdown presentation policies, user-manual authoring instructions, native prompt publication and readback rules, AGENTS, injected shared-static workflow instructions, policy prose, and navigation; and
• another current or future non-learning document workflow's model-facing instructions and controls when its actual consumer and semantic, canonical, publication, authority, and safety owners place it in the resolved scope.

These categories define full-domain coverage, not a frozen inventory or one explorer or worker per category. Inclusion depends on current consumer behavior and primary purpose, not a filename, extension, directory, repository name, or residence under general document collection. A future document flow is not included merely because it contains prose. For a causally bounded run, apply the main Skill's scope rule to the resolved Workflow-purpose slice and name each excluded category and its boundary. Group or split evidence by current causal questions and dependencies.

The purpose is semantic fit with the actual consumer, including clear purpose and ownership, duplication, contradiction, hierarchy, dependency, discovery, retrieval, obsolescence, environment drift, and unsupported fixed detail that constrains current model judgment without serving a real contract.

Evidence and target-specific repair boundaries

Use current Agents and tools for broad reading of any current repository, general document collection or other document workflow, history, runtime, native, or Hosted surface that is material evidence. A repair authorized under the main Skill's completion policy may change a user-owned Workflow target only when its current canonical semantic owner and publication owner are known. The assigned target worker edits through the canonical owner under the main Skill's proportionate repair route. An independent review, Plan, or behavioral evaluation appears only when that route or another current owner requires it.

For a local common/global or project instruction repair, update the exact canonical saved source and publish through its actual local owner with persisted consumer readback. hosted common/global and non-learning Project Instructions belong to Hosted, including their saved sources wherever stored; dedicated learning-purpose surfaces remain Learning-owned. Workflow may inspect their local consequences and route findings, but does not repair their Hosted prompt fields or create a competing source writer.

Hosted owns hosted common/global and non-learning Project Instructions and the independent Hosted package and tool-chain audit, including Hosted-specific reusable methods and tool instructions. Workflow covers shared canonical non-learning methods used both locally and on Hosted; Hosted may carry a necessary shared-method repair through the existing owners under its adapter. Coordinate one source worker and reuse its judgment when both domains are selected. Same-named independent sources are not assumed to be shared. For an affected hosted consumer, coordinate with Hosted when selected and follow prompt-and-skill-lifecycle for source mapping, package closure, cloud publication, and hosted assistant readback. connected execution carrier supplies local execution without installed Skills; test that path only when affected. Treat source, package, deployment, discovery, resource access, and successful use as distinct claims. Workflow's local/shared audit does not duplicate Hosted's full package inventory; it checks material consequences of its own source changes and reports any remaining Hosted convergence or evidence gap under the main Skill.

Classify model-facing documents under general document collection or document and memory store by their actual consumer, not their location or filename. Local general instructions remain Workflow targets; hosted common/global and non-learning Project prompts and Hosted-specific connector/operator instructions belong to Hosted. For one canonical source feeding both consumers, retain its actual owner, coordinate one writer, and check both consumers' material consequences under the main Skill. Repair each source only through its actual owner. Generated Prompt Map or local-prompt projections are evidence rather than editable sources; return their repair to the repository source and current Skill or generator owner. Retained Artifacts, Plans, user-input provenances, historical design records, available activity history observations, personal memory records, and ordinary document data are evidence rather than prompt-normalization targets. An active producing workflow may still own an exact current mutable record.

Imported upstream Skill bodies follow their upstream or explicit overlay owner. Bundled, provider-owned, system, externally owned Project, or unresolved Hosted-only Skills are cross-owner or evidence-only targets rather than direct repairs. A pre-existing Prompt Map, sync, UI, or configuration incident is evidence-only unless the user separately requested incident repair.

Dedicated learning-method Skills and prompts, Course controls, and connected local workspace instructions belong to the Learning adapter. They are excluded from Workflow's target inventory. Workflow may inspect them as material evidence and report the exact dependency, but it does not audit or repair them on Learning's behalf. Likewise, Learning may inspect Workflow targets but cannot repair them. A learning-relevant clause does not transfer a local common/global target out of Workflow; transfer requires a dedicated Skill or prompt whose current primary purpose is learning. Dedicated learning-purpose Hosted Project Instructions, including learning, remain Learning-owned; a learning-related clause alone does not transfer a common/global instruction source.

Saved local scheduled, recurring, or deferred Task prompts and dedicated task-purpose Skills belong to Saved Task. Local and shared reusable non-learning Skills and local common/global instructions remain Workflow targets when a Task invokes them; hosted common/global and non-learning Project instructions and Hosted-specific methods retain Hosted ownership. Workflow may inspect the Task's material effect and exact dependency; Saved Task must likewise inspect material downstream effects of Workflow sources and route shared-source findings to Workflow. Follow the main Skill's overlap and single-writer contract; a single-domain request does not automatically launch a peer or transfer source responsibility.

Hosted Scheduled Task prompt content, native fields, and recovery mirrors remain maintained through the user's Scheduled Task Project under its current Project Instructions and explicit Task request. An on-demand Workflow run may read them as evidence and report an exact dependency, but it does not mutate them. Current local/shared general task-authoring methods remain Workflow audit targets; hosted common/global and non-learning Project prompts and independently owned Hosted-specific reusable authoring methods follow the Hosted adapter. Auditing those methods does not authorize changes to any Task configuration or recovery mirror.

Working and return boundary

The current Workflow-domain coordinator owns decomposition, child follow-ups, evidence synthesis, coverage, peer exchange in joint mode, and its domain handoff. Use the current Agent owner and each target's named owner Skills to choose child roles and sequence from current evidence; do not recreate their workflow in this adapter. A material owned change uses its actual canonical, publication, and validation owners and follows the main Skill's proportionate review, Plan, and evaluation route. Repository writes use an isolated assigned lane or the workspace owner's convergence route. Serialize only concrete dependencies, operations on the same exact mutable target, or a carrier-proven inability to support concurrent ownership; a shared account, login, profile, general document collection root, or authenticated context is not by itself a collision.

Return one result for the resolved Workflow-domain scope: scope examined, material exclusions and their boundaries for a bounded run, semantic conclusion, owned changes with publication readback and claim-appropriate acceptance evidence, Hosted, Saved Task, Hosted Scheduled Task, or Learning dependencies, and any other cross-owner finding, evidence-only incident, or blocker. A full-domain result accounts for all current target families above without treating every document or historical record as an active prompt. Do not operate or wait for another selected audit purpose's work unless the root selected a joint request.