AI permissions, proposals and execution EN

Follow a Kai request from available tools to a proposed change, approval and execution, with explicit limits for configuration and external clients.

An assistant's ability to describe an action is not permission to perform it. Kaduno separates tool availability, the caller's authority, store policy, proposed changes and execution. The operator needs to know which state a request reached, not just whether the conversation sounded successful.

This is a source-reviewed explanation dated 13 November 2026. It describes control paths in the reviewed implementation, not a production security test or a promise that every interface and installed release behaves identically. No live payment, order, publication or customer communication was executed for this guide.

A request is not an executed change

Request + authenticated caller + store context
                       |
             Available tools and policy
                       |
          Read / prepare / propose / refuse
                       |
             Applicable approval path
                       |
         Recheck authority and current policy
                       |
              Execute permitted operation
                       |
             Inspect result and evidence

This is a logical guide, not a complete call graph. Reads do not pass through a write-confirmation card. Some prepared changes are saved on the relevant page by the person, while eligible setup changes can use a restricted automatic path.

StateWhat it meansWhat it does not prove
ReadA permitted tool retrieved informationPermission to change the records or send their contents elsewhere
Prepared on a pageKai helped populate an editing workflowThat the person saved or published the change
ProposedA change is available for reviewThat the underlying business operation ran
RefusedA control rejected this attemptThat no other unrelated action occurred
ExecutedThe execution path returned a resultThat every downstream provider completed the intended business outcome

Tool selection is only the first check

Kai's store policy supports denied tool groups, blocked action prefixes and a maximum risk level. The implementation filters advertised tools and checks calls again, including action names inside domain routers. Composite operations and plans have additional checks rather than treating a convenient wrapper as new authority.

The caller's role, store context and applicable business-operation checks still matter. A model-generated store identifier is not the authority to choose another store. A reassuring prompt, hidden button or absent tool in a menu is not sufficient evidence of enforcement. See store boundaries.

Risk labels are classifications, not business consent. The reviewed defaults do not make Kai read-only: the risk ceiling permits high-risk proposals and there are no default blocked prefixes. Some invalid policy settings fall back to permissive defaults. Verify the effective policy and observed refusals instead of assuming that saving a restriction made it active.

Approval in Kai is tied to a stored action

For the reviewed Kai confirmation path, the action manager checks ownership and conversation scope, proposed status and expiry. It resolves current policy again; a card created before a restriction changed is not intended to preserve the old permission. Role and operation checks are also part of execution.

High-risk cards support typed confirmation derived from the stored target, such as a record number or affected-record count. The setting is on by default in the reviewed source, but it can be disabled and specific upstream-confirmed paths can satisfy confirmation differently. This is not a claim that every write always requires typing, or that a typed response can authorise an otherwise forbidden operation.

The implementation also flags some proposals whose deciding arguments derive from external text and requires an additional explicit acknowledgement. That control is not proof that all prompt injection is detected. Treat customer messages, documents and provider responses as information, not instructions granting access.

The setup-stage exception must be understood

Eligible settings changes during store setup can be applied without a separate confirmation card, through the action manager's execution path. Eligibility, caller authority and other checks still apply. This is not unattended permission for every write, and unattended watch turns require confirmation even during setup.

Store stage is derived from configuration and operational signals, with an explicit override. Those signals are heuristics, not proof that a store is safe for automatic changes. Before importing real data or opening a storefront, verify the effective stage and confirmation behaviour. Do not equate a preview hostname or an empty order history with a harmless environment.

Some CMS publication operations are classified as blocked for Kai, with publication remaining a human editing workflow. A prepared draft and a published page must be reported as different outcomes.

ChatGPT and other MCP clients are separate interfaces

The Kai card workflow does not automatically describe every external MCP client. For the selected client, establish the connected identity, store, allowed tools, server-side checks and how approval is obtained. A client's approval dialog does not replace the server's permission checks. Conversely, a server-side permission does not establish that the user approved a consequential action in that client.

Do not assume that a REST key, signed-in session and MCP connection share one credential contract. Start with a permitted read and verify its store before considering a write. Module and integration boundaries explains this distinction and the separate processing boundary of model providers.

Execution evidence and recovery

Inspect the action result and the affected records before reporting success. Audit and result records can help distinguish proposals, refusals, failures and execution, but their coverage and availability need verification for the selected operation and release. Do not promise that every downstream event is captured.

An available undo function is operation-specific, not a universal rollback of emails, payments or external-system effects. For multi-step work, inspect each step's outcome rather than assuming transaction-wide rollback. A timeout can occur after an external system acted; check that destination before retrying.

A controlled acceptance plan

The following are proposed checks, not tests performed for this article. Use an authorised isolated environment, synthetic records and a recorded release and configuration. Agree the exact permitted actions before any write.

ScenarioEvidence to collect
Permitted readCaller, intended store, tool and sanitised result
Proposed write with approval withheldProposal state and evidence that the target remains unchanged
Denied actionEffective policy, refusal and unchanged target
Policy changed after proposalThe old proposal cannot retain permission removed by the new policy
Wrong store or insufficient roleRefusal without revealing another store's records
Expired or already handled proposalRefusal and no duplicate business operation
Setup versus live settings changeEffective stage, applicable confirmation path and recorded outcome
Controlled execution and recoveryExact approval, operation result, downstream state and supported recovery limits

Passing website documentation tests does not establish any of these runtime results. A production acceptance statement needs observed evidence, release identity, operational review and the relevant owner's approval. No security certification, measured performance or universal isolation guarantee is supplied here.

Continue with the conceptual architecture map, API read exercise and security review information.