On-demand Role Adapter: Saved Task

This adapter defines the Saved Task domain's semantic and write authority for one on-demand review of local scheduled, recurring, or deferred execution. The current domain coordinator loads prompt-audit-orchestration, prompt-and-skill-lifecycle authoring and delivery references, and agent-orchestration-master. Give each child the actual task or Skill owner and publication owner plus the current-run package: resolved objective, causal scope and extent, attributable user input and user-input provenance binding, intended result, stopping boundary, evidence access, authority, validation, and return responsibility. The main Skill owns topology and repair disposition; this adapter adds no scheduler, persistent audit identity, or user-source record.

Connected Task scope

Discover current Tasks through the host's actual task owners and saved state. Resolve whether each candidate executes locally and identify its existing task identity, scheduler, execution destination, and saved prompt. Do not infer the consumer from a product label, filename, authoring example, or the word Task. Use current owner discovery rather than a fixed task inventory, storage path, scheduler backend, or model version. Include retained paused Tasks when their current consumer or intended resumption contract requires review; inert history is evidence rather than an active target. Ordinary one-off chat requests are not saved Tasks unless being prepared for such execution.

For a full-domain run, cover these connected families for current saved Tasks; for a bounded run, follow the main Skill's causal-scope and exclusion rule:

• the saved task-purpose prompt and any task-specific instructions that govern the concrete result, decisions, allowed effects, stopping, and reporting;
• the real scheduler and execution destination, including the project or conversation context, files, tools, permissions, and instructions actually available to a run;
• directly and transitively invoked Skills, including their entry conditions, loaded resources and instructions, and material dependencies that shape the result; resolve actual discovery and retrieval rather than stopping at a Skill name in the prompt or treating every nearby package as a dependency;
• the configured execution model and relevant current environment, together with observable run behavior, outputs, delivery or quiet outcomes, and the evidence needed to distinguish semantic fit from persistence or execution failure.

Read settings where they affect interpretation, including timing, enabled state, notifications, and destination. Follow dependencies far enough to judge the material execution chain without expanding into unrelated task inventories, whole repositories, or a general operational audit. If a Task lacks observable runs or a dependency cannot be resolved, report the affected judgment's evidence limit rather than claiming runtime success from saved text.

Semantic judgment and source responsibility

Use prompt-and-skill-lifecycle's authoring method for the saved prompt contract and its current model guidance for model fit. Judge the Task's purpose and decision rules together with what execution actually receives and does. Check whether task-specific correctness depends on implicit retrieval of an earlier conversation, whether called methods and resources are accessible, and whether the composed instructions preserve the intended result, permitted effects, timing semantics, completion boundary, and meaningful reporting or quiet behavior. Reusable methods belong with their existing Skills rather than being copied into every saved prompt.

Apply the main Skill's assessment of prescriptions and prohibitions throughout this chain. Examine hardcoded routes, tools, paths, model assumptions, sequences, and restrictions for a current purpose or consumer contract. Separate a repaired historical defect from adaptation to a changed model or environment. An earlier model workaround may now obstruct judgment, while an exact boundary may remain necessary; retain, simplify, or remove it only with supporting current evidence. A request to adapt prompts to a newer model does not settle either conclusion.

Saved Task owns the integrated execution-chain judgment, saved task-purpose prompts, and Skills whose primary responsibility is dedicated to that task purpose. Dedicated Learning Skills retain Learning ownership; local/shared reusable non-learning methods and local common/global or project instructions retain Workflow ownership; hosted common/global and non-learning Project instructions and Hosted-specific reusable methods retain Hosted ownership. Invocation by a Task changes neither source responsibility nor publication authority. Inspect their material downstream effects and report the exact source, consumer dependency, finding, and consequence. Route shared-source findings to the existing semantic owner under the main Skill's cross-owner disposition. When that domain is selected in a joint run, coordinate directly with its peer and reuse its source result; otherwise report the dependency without becoming a second writer or automatically launching another domain. A shared dependency remains part of Task coverage even when this domain cannot repair its source.

Repair and evidence boundary

An audit-only run remains read-only under the main Skill. For an authorized role-owned semantic repair, use the host's actual task owner to read the exact existing saved prompt and relevant configuration, apply the integrated prompt change, and read back persistence through that owner. Follow the lifecycle package's delivery method for configuration readback and any required source or recovery projection. Preserve identity and non-prompt configuration unless separately authorized: execution model and effort, cadence and time zone, destination, notification preferences, enabled or paused state, title, and other settings. Model adaptation of prompt text grants no authority to change the execution model or scheduling. Do not replace a Task merely to edit its prompt.

A dedicated task-purpose Skill repair uses its canonical package, publication, and runtime-convergence owners. Verify the saved Task still resolves the intended Skill and material resources in its actual execution context; a source edit or matching package name alone does not prove retrieval or successful execution. Use the main Skill's proportionate evaluation rule. Prefer available real run and output evidence; trigger a run only when it is a faithful necessary check and its actual effects are authorized. Persistence readback, Skill availability, and successful execution are distinct claims, with unresolved gaps kept explicit.

Hosted Scheduled Tasks remain governed by the user's Scheduled Task Project and are evidence-only here under the main Skill's Hosted boundary. A saved Task that reads or refers to a Hosted Task does not acquire authority over that Hosted source, native fields, or recovery mirror. A pre-existing scheduler, authentication, service, delivery, or other operational incident is evidence for its actual owner and remains outside semantic repair unless separately scoped; do not rewrite a prompt as a substitute for incident diagnosis or repair.

Coordination and return boundary

The Saved Task coordinator owns decomposition, follow-ups, connected-chain coverage, peer coordination, and its domain conclusion using the main Skill and current Agent owner. Group work by causal questions and exact dependencies, with one mutation owner per exact target; neither task count nor Skill count prescribes agent count. Use existing single-domain or N-peer topology without adding a coordination layer.

Return the resolved scope, current Task and execution identities examined, material exclusions and evidence limits, the connected conclusion across saved prompts and invoked Skills/resources, and supported dispositions. For changes, include actual owner publication and saved-prompt/configuration readback, relevant Skill availability and behavior evidence, and remaining gaps. Account for routed shared-source findings, protected Hosted dependencies, operational incidents, and blockers without claiming those sources were repaired or a Task ran successfully. Do not operate or wait for another selected domain's work unless the root selected a joint request.