Update to Forward Deployed

{
  "description": "Connect AI work to enterprise systems through meaningful business objects, usable tools, identity, and reliable actions. Use for integration design, MCP and API tool contracts, asynchronous operations, or diagnosis of a workflow that cannot complete its work.",
  "included_files": [
    {
      "relative_path": "references/uncertain-write.md",
      "size_in_bytes": 4975
    }
  ],
  "name": "enterprise-integration",
  "skill_md_contents": "---\nname: enterprise-integration\ndescription: Connect AI work to enterprise systems through meaningful business objects, usable tools, identity, and reliable actions. Use for integration design, MCP and API tool contracts, asynchronous operations, or diagnosis of a workflow that cannot complete its work.\n---\n\n# Enterprise integration\n\nMake the agent's work count in the systems people use. Follow one assignment from its initiating request through context gathering, business operations, and verified completion. Build the integration around those effects.\n\n## Establish the business meaning\n\nInspect a representative case, its records, and its history. Identify the objects and relationships the agent needs: an employee, project membership, application role, and access request describe different things even when they share an identifier.\n\nResolve identifiers at the granularity of the action. Name the system that owns each consequential fact and command. Preserve distinctions such as requested, queued, provisioned, and usable. A ticket marked complete does not establish that an employee can use the application.\n\nCarry the time and version semantics that change the decision. Current eligibility can differ from eligibility when a request was submitted. An old policy can explain a historical action without authorizing a new one. Reconcile the decisive relationships before proposing a broad enterprise data model.\n\n## Discover the supported path\n\nInspect the customer's available product, administration, and developer surfaces. Check current primary documentation for the deployment in use, then exercise its interfaces under the intended identity. Entitlement, documented support, and observed behavior answer different questions.\n\nChoose among native extensions, APIs, events, connectors, extracts, and supported browser operation according to the work. An overnight extract may be enough for research; provisioning requires a command and a way to establish its result. Preserve an incumbent application's useful workflow instead of duplicating it by default.\n\nSeparate context reads from commitment checks. Cache stable descriptive material where useful; refresh the facts that determine whether the action remains valid. For recurring work, retain a position per source and account for pagination, late updates, and repeated events. A failed read leaves an unread interval; it cannot establish that nothing changed.\n\n## Give the agent effective tools\n\nExpose operations that preserve business meaning. A query returning an employee's project membership, role eligibility, and source revision can be more useful than unrelated field lookups. A supported provisioning command can carry validations and downstream events that direct table writes would bypass.\n\nMake the interface intelligible through its operations, parameters, result types, and descriptions. Expose the query or action choices the task needs. Return useful status, source references, versions, and errors that distinguish correction, denial, pending work, and uncertain outcome. Test tool use on complete assignments; repair ambiguity in the interface before adding repeated instructions about how to call it.\n\nLarge results should offer summaries or selectable records plus handles for deeper inspection. Preserve the detail the next decision needs. A successful tool call can still be an unusable interface if its output hides the relevant condition or overwhelms the context.\n\n## Enforce scope at execution\n\nDetermine whose authority each read and action uses, including unattended work. Enforce permissions, allowed targets, and action limits outside the model. Keep credentials out of model-visible material and use the platform's supported credential mechanisms.\n\nAlign caches, derived summaries, and shared task state with the source access boundary. Source content can inform the work; it cannot expand authority. Existing authorization should permit eligible actions without repeated approval. Seek a new decision when the requested effect changes materially or exceeds that scope.\n\n## Make completion recoverable\n\nKeep proposed intent, attempted command, and confirmed external result distinguishable. Persist an operation reference when the receiving interface supports it. After a timeout, reconcile the authoritative state before repeating a potentially completed action. Reuse the same idempotency reference for the same request; changing it can create a second effect.\n\nFor asynchronous operations, retain the returned job or resource identifier and continue through its documented status mechanism. \"Accepted\" can mean queued. Verify the result at the level promised to the user.\n\nIn a read-then-act workflow, mark an item handled only after its required effect is confirmed. If intake advances ahead of execution, preserve the item as durable pending work. After a lost response, reconcile that effect before moving the completion position or sending again. Existing queues or workflow state can supply this behavior.\n\nUse native version and concurrency checks where available. Across systems, describe partial completion and the next valid action. Reversing a completed operation can itself be a new business action. Preserve confirmed effects during retry, cancellation, or manual takeover.\n\nRead [the provisioning recovery example](references/uncertain-write.md) when an agent must complete a business operation despite a lost response.\n\n## Deliver an exercised integration\n\nBuild or specify a vertical slice that uses the difficult semantic relationship and completes a meaningful receiving-system action. Inspect the resulting state and the failure that could duplicate, misapply, or strand the work. Keep documented promises separate from observed behavior.\n\nLeave the receiving team with a usable contract, an example transaction, and ownership for unresolved outcomes. The format can be a working adapter, sequence diagram, replay, or concise engineering brief.\n"
}