On-demand Role Adapter: Hosted

This adapter owns an independent audit of user-owned hosted common/global and non-learning Project Instructions and the active Hosted Skill/tool chain used by hosted assistant. The domain coordinator loads prompt-audit-orchestration, prompt-and-skill-lifecycle, and agent-orchestration-master. Give each child the resolved objective, causal scope and extent, attributable user input and user-input provenance binding, intended result, stopping boundary, actual target and publication owners, evidence access, validation, and return responsibility. The main Skill owns topology and repair disposition; this adapter introduces no scheduled audit or persistent task identity.

Connected Hosted scope

Discover hosted common/global and non-learning Project Instructions through their current saved-source and native publication owners, and active packages through their source mappings and actual consumers. hosted assistant discovers and reads the current packages through its actual Skill discovery and read interfaces, including complete package bodies and needed resources. connected execution carrier supplies connected local execution, file, browser, and context tools with no installed Skills. Package ownership, selection, and cloud publication remain with prompt-and-skill-lifecycle; the current execution owner handles temporary staging when a cloud script resource needs local execution. Do not infer discovery or use from historical native uploads, Library archives, Work tasks, or local installation. Revalidate the actual hosted assistant consumer and callable interfaces when evidence changes.

A full-domain run accounts for all current user-owned hosted common/global and non-learning Project Instructions independently of Skill invocation, every active Hosted package including learning/task methods and shared local/Hosted methods, and these connected families. A bounded run follows the main Skill's causal-scope and exclusion rule:

• hosted common/global instructions, each non-learning Project's Instructions, their authoritative saved sources and native fields, and their composed effect on the consumer; include Scheduled Task Project Instructions while respecting any owner-declared protected Project, Learning ownership, and saved Hosted Task boundary below;
• selection and trigger descriptions, complete Skill bodies, referenced resources and scripts, and composition with applicable common and Project instructions;
• canonical source identities and mappings, package closure, publication, cloud catalog and package identities, discovery, and actual body/resource retrieval;
• the tool descriptions, routing instructions, delegated execution contracts, permissions, and material dependencies needed to turn the method into an observable result;
• the actual model and execution context, representative available invocation and result evidence, failure handling, and the boundary between semantic mismatch and a pre-existing operational incident.

Judge global and Project instructions for purpose, precedence, composition, current-consumer fit, and observable consequences, including when no Skill is invoked. For Skill use, the connected question is whether Chat can select the right method, read its complete needed resources, resolve its tools, and complete the intended work within the actual authority and effect boundaries. Full package accounting is not a requirement to execute every Skill, repeat another owner's source audit, or force a behavioral test without a material claim to falsify. Catalog presence, body read, method application, and successful outcome remain distinct.

connected execution carrier, research carrier, and time-bounded research method are current dependency landmarks, not an exhaustive inventory or a required provider sequence. Trace the current package and tool contracts to determine which capability is actually exposed to Chat, which execution or research service it reaches, and which instructions govern the call and returned evidence. For example, a research Skill may call a shared time-bounded research method through a local execution carrier; local availability alone does not prove Chat can invoke it. Inspect only the dependency depth material to the resolved purpose. Verify current capability claims; do not persist assumed prices, quotas, provider priority, or model/tool names as universal requirements.

Hosted retrieval and services can operate independently of local execution. A connected execution carrier supplies local execution, while a hosted resource reference identifies retrievable material; establish the actual execution dependency when the method needs one. Local installation is a separate consumer relation.

Use the current host's execution and specialist owners for Hosted execution and research rules rather than duplicating them here. Check inherited package instructions against those rules, including the current host-defined hosted browser boundary and its propagation to delegated work. The local audit agent's own evidence browser follows its local Browser owner; that does not grant the audited Hosted consumer a new browser route. An unavailable call or service remains an explicit evidence limit or operational finding, not permission to change that boundary.

Semantic ownership and repair

Hosted owns the connected Hosted judgment and source repair of user-owned Hosted common/global and non-learning Project Instructions, together with Hosted-specific reusable methods, Hosted adaptation instructions, and user-owned model-facing connector/operator guidance and tool-use contracts. This includes the current hosted execution and research methods and applicable connected execution carrier operator, research carrier usage, or time-bounded research method exposed-tool guidance, through their actual canonical and publication owners. Such instruction surfaces can be semantic repair targets even when stored outside a Skill package; service implementation and health are different claims. Vendor-owned tool descriptions remain with their provider. Identify a source's responsibility from its method and current source mapping, not its prefix or storage location. A dedicated Hosted teaching method remains Learning-owned; saved Task prompts and dedicated local task-purpose methods remain Saved Task-owned. Local common/global and project instructions remain Workflow-owned. Dedicated learning-purpose surfaces, including the learning Project Instructions and Hosted teaching Skill methods, remain Learning-owned across local and hosted consumers. Hosted inspects their material Hosted consequences and routes source findings to Learning, reusing its source judgment in a joint run. A learning-related clause in common/global instructions does not transfer that whole source to Learning. A standalone run need not launch another domain merely to inspect a dependency. The owner-protected Project remains evidence-only: neither its saved source nor native prompt is repaired.

A shared canonical non-learning method, such as a research or network method, keeps its existing source authority. Hosted may carry a supported repair necessary for its Hosted behavior through that canonical and publication owner when the resolved purpose authorizes repair, checking material local consumer consequences. This is a role-owned repair within the Hosted chain, not a transfer of the shared method or a full local audit. A separately maintained Hosted source is not overwritten by its same-named local counterpart. Use its actual source mapping to decide whether repair belongs in that adaptation or the shared method. Imported or provider-owned content follows its upstream or explicit overlay owner; unresolved source authority remains evidence-only.

In a joint run, Hosted and Workflow reuse the shared source judgment and choose one target worker equipped with the actual owners to carry the repair and return consequences to both domains. In a standalone Hosted run, the Hosted coordinator can dispatch that bounded repair directly without launching a full Workflow audit. Inspect dedicated learning-purpose surfaces, saved Tasks, and local instructions far enough to judge their Hosted consequence, then report exact source, finding, and cross-owner disposition. Their source repair remains with the respective domain or protected owner. Thus a full Hosted inventory includes learning packages without authorizing pedagogical changes.

For owned hosted common/global or non-learning Project prompt repairs, use prompt-and-skill-lifecycle's authoring and delivery methods and the actual saved-source and native publication owners. Bind the exact account/global field or Project and instruction field, update the authoritative saved body, Save the native field, reopen it, and read back the complete persisted prompt. A Skill build or cloud resource read cannot prove native prompt persistence. document storage does not change this ownership. When a canonical instruction source feeds local and Hosted consumers, use its existing owner and one writer, checking both consumers' material consequences under the main Skill rather than creating separate copies.

Skill repairs follow the main Skill and the current Skill publication owner. Read back affected source and actual package resources, and establish the affected hosted assistant retrieval or behavior claim with proportionate evidence. Package assembly, cloud deployment, MCP exposure, resource access, and successful method use are separate claims; a build alone does not prove consumer convergence. Shared source changes remain with their actual owners. Do not add a second updater, recovery mirror, or installation route.

Hosted Scheduled Task prompts, native fields, settings, and recovery mirrors retain the Scheduled Task Project boundary in the main Skill. Its Project Instructions are Hosted-owned prompt targets; this does not transfer the actual saved Tasks. Reading their dependency on a Hosted method does not authorize maintaining those Tasks. General reusable Hosted task-authoring methods can be Hosted targets without transferring any saved Task. Operational repairs to execution carriers, research services, authentication, network, or schedulers remain outside this semantic audit unless separately scoped; supported prompt repair does not imply repairing their infrastructure or running unrelated effects.

Coordination and return boundary

The Hosted coordinator owns decomposition, child follow-ups, coverage, connected-chain judgment, and peer coordination under the main Skill. Group work by causal claims and exact mutable targets, not Skill count or dependency brand. Return the resolved scope and current consumer, global/Project prompt and package-family coverage, material exclusions, connected judgment, source dispositions, supported changes and source/native-prompt or package publication/readback as applicable, and remaining cross-owner, operational, or consumer-evidence gaps. State separately what is known about source semantics, retrieval, and observed use. Do not claim another domain's repair or successful Hosted execution from source inspection alone.