author-subagent

{!% require_skill 'instructions:subagent-standards' %!}

{!% require_skill 'instructions:agent-prompt-standards' %!}

{!% require_skill 'instructions:skill-standards' %!}

<objective> A finished {{! term('configured_agent') !}} definition and its calling-skill integration, ready for independent verification, with native configuration and a declared result contract. </objective>

<constraints>

• ALWAYS: apply every rule in the /subagent-standards and /agent-prompt-standards rule catalogs to the finished definition, and return it ready for independent verification once the product's author command and deterministic checks pass on it.
• NEVER: dispatch a {{! term('configured_agent') !}} auditor or wait on an audit verdict — a verdict on the definition comes from /audit-subagent run in an agent session separate from the one that authored it.
• ALWAYS: write to the invocation checkout. Reach a user-scope destination outside it — creating, editing, or deleting — only after the operator confirms the exact absolute destination; one approval covers one write.
• NEVER: widen an approved checkout-scope write to user scope on the judgment that the configured agent suits other projects.

</constraints>

<intake>

Select the route from the task message:

Task message carries	Route
A target path or role only	<workflow>
A target followed by a sealed run file and the instruction to attend to every rule the run names	<repair_from_run_file>
A target followed by the command line and complete output of a deterministic command that failed on it	<repair_from_command>

</intake>

<workflow>

1. Require $configuration_target. When it is empty, return a blocked result naming the missing configuration path or role. Read the requested target and its governing requirements. Identify the role's owning skill, exact configured name, destination, selected profile, and result. Return a blocked result naming the missing input when a product decision or a required input is absent. When the definition already holds the requested result, return an unchanged result naming the committed head and the path that holds it, and edit nothing.
2. Apply /subagent-standards to placement, capabilities, profiles, isolation, invocation, and evidence. Read the relevant references from the index below.
3. For a skill-backed role, author a thin wrapper that invokes the owning skill and relays its result. Keep task-specific behavior in that skill.
4. For a marketplace source, declare profile: standard, profile: strong, profile: executor, or profile: fast according to the governing selection. Let the product build emit the complete native configuration. For a product-owned native definition, use the selected complete configuration supplied by the loaded standards and examples.
5. Use skill instructions:author-skill. Update the calling skill through it so its explicit launch names the configured role and supplies only the target.
6. Run the product's author command, then apply /subagent-standards <configuration_subject> to identify the authored input and exact emitted definitions, or the directly authored native definition.
7. Check the finished definition against every row of the /subagent-standards and /agent-prompt-standards rule catalogs, each row at the surface its stating section names — template syntax and profile selection at the authored input, native fields, permissions, and complete profile values in each emitted definition. Repair each violation at its authored source and repeat steps 6 and 7 until no row is violated.
8. Run the focused deterministic checks, then preserve the exact definition through the product's normal checkpoint workflow. A later edit repeats steps 6 to 8, so the returned definition is the state those checks passed on.
9. Return the configuration path, the configured role, the emitted definitions, and the verification still outstanding under /subagent-standards: an /audit-subagent run in a separate agent session, and native loading with one minimal isolated execution of the emitted definition.

This skill launches no configuration it authors.

</workflow>

<repair_from_run_file>

Read every finding to repair, by its <unit>:<rule-id> key, from the run file alone, and read the definition to repair from the unit the run records for it. Ignore any summary, remedy, or class name in the task message. Attend to every rule the run names and sweep each named class across the whole target. Then follow the <repair> steps.

</repair_from_run_file>

<repair_from_command>

Repair the defect the command's output names at its authored source, never in an emitted definition. Then follow the <repair> steps.

</repair_from_command>

<repair>

1. Repair each finding across every instance the cited clause covers, and change nothing the clause does not require.
2. Before returning, list every same-class site and every dependency the repair changed — a renamed field, a changed profile, a changed result contract — and resolve each, in the authored input and in each emitted definition.
3. Run workflow steps 6 to 8 on the repaired definition.
4. Return the repaired definition with, for each repair, the clause applied and the check that shows the class closed, plus the outstanding verification the workflow names. Return a blocked result naming the missing input when the run file or the command output is absent.

</repair>

<reference>

Read only the references relevant to the role's actual work:

• subagents.md: native definitions, complete profile examples, capability selection, and skill-backed wrappers.
• write-subagent-prompts.md: converting a role requirement into a focused prompt and result contract.
• evaluation-and-testing.md: evidence design, native loading, and retained invocation results.
• error-handling-and-recovery.md: diagnosing failure boundaries and returning actionable evidence.
• context-management.md: independent target discovery and bounded context for long work.
• orchestration-patterns.md: explicit dependencies between skill-requested calls.
• debugging-agents.md: inspecting native observations without changing the attempted launch.

</reference>

<failure_modes>

Claude retained independent configuration fields.

Claude's examples selected a model separately from its reasoning controls and omitted controls the harness requires. A model chosen apart from its profile drifts from the configuration owner, and a hand-assembled block leaves out whatever that owner adds. Declare one central profile and use its complete native configuration as one unit.

Claude prescribed a retired tool contract in invocation advice.

Claude's handoff prescribed return fields and lifecycle tools an earlier tool version exposed. The advice was copied from earlier guidance instead of read from the current native tool schema. Use the invocation mechanics of the root harness instruction file /subagent-standards <invocation> names and the current native schema, and report the actual result.

</failure_modes>

<success_criteria>

• The definition implements the governing role and the complete selected native profile.
• The finished definition violates no rule in the /subagent-standards or /agent-prompt-standards rule catalog and is returned with no audit dispatched or awaited.
• The calling skill's launch instruction names the exact configured role and supplies only the target.
• The product's author command and focused deterministic checks exit zero on the returned definition, and every emitted definition the result names exists at its path.
• The result names the configuration path, the configured role, each emitted definition, the pending /audit-subagent run, and the pending native loading and minimal isolated execution.
• A repair returns, for each finding, the clause applied and the check that shows the class closed, with every same-class site and changed dependency resolved.
• A definition that already holds the requested result returns unchanged; an absent input or composed skill returns blocked.
• No configuration load or execution approval is claimed before its actual result exists.

</success_criteria>