MazeGraph’s free version uses a deterministic policy for every reply and tool call. A request from the model is only a proposal: the policy engine checks it against the site’s controls and the signed-in person’s WordPress permissions before it can run. Asking again, changing the wording, or choosing a less cautious permission mode does not override a refusal.
The decision order
For a chat tool call, MazeGraph checks the request in three layers. A deny at any layer stops the call. An ask from a guardrail or policy rule remains in force while the later permission checks run; those checks cannot silently turn it into an allow.
| Layer | What it checks | Typical result |
|---|---|---|
| Guardrails | Whether replies are enabled, the person has accepted the data notice, usage limits allow the request, and a write is supported and undoable. | A failed check refuses the request. A write may instead wait for approval. |
| Additional policy rules | Rules from MazeGraph definitions, plugins, or the active theme that match an action and, optionally, a tool. | A rule can deny a match or ask for approval of a matching tool call. It cannot grant access. |
| Tool permissions | Whether tool use is on site-wide, the individual tool is enabled, the selected assistant includes it, and the person’s WordPress capability check passes. | The call proceeds only if the applicable checks allow it. No permission rule allowing the call means denial. |
The first layer includes checks such as Allow new replies, first-message data notice acceptance, per-person message pace, and the site’s daily token allowance. These checks apply at different points in a reply: for example, the daily allowance is checked before model calls, while the data notice is required before a person’s first message.
The site-wide Allow tool use switch is part of tool permission checks. Turning it on does not enable every individual tool. Likewise, enabling a tool or widening the built-in assistant’s tool list does not grant the current person a WordPress capability.
What the policy can decide
MazeGraph can allow, ask, or deny an action:
- Allow: the request has passed the applicable checks. The tool still runs as the signed-in WordPress person.
- Ask: the call pauses until a person approves or refuses the exact proposed tool call. Approval applies to those arguments, not to a later, changed request.
- Deny: the action stops. Approval modes, assistant instructions, and another model request cannot change a denial.
Policy rules supplied by the site or an extension can only make access stricter. A deny rule refuses a matching action. An ask rule can require a person to approve a matching tool call, even if Auto or Full would otherwise skip the prompt. If an added rule fails while running, MazeGraph refuses the action. Review available policy definitions in Tools → Extensions.
How this works in real requests
A user count tool is enabled, but the person cannot run it
An editor may be able to open MazeGraph and use other tools but still be unable to ask how many WordPress accounts the site has. The Count users ability checks the current person’s list_users capability. If WordPress does not grant it, the call is refused even when the tool is enabled and included in the assistant. Full mode cannot grant that capability; a WordPress administrator must change the person’s WordPress access if appropriate.
Site content keeps a later write in review
Suppose an assistant reads page content and then proposes updating a draft. MazeGraph treats site content that enters the conversation as untrusted input. A supported, undoable MazeGraph write based on that context waits for approval, even in Auto or Full mode. A writer without MazeGraph’s trusted undo support is refused instead. The approval card shows the specific change; if the person refuses, MazeGraph does not retry it unless asked to do something else.
The relevant write tool must also be enabled, and the current person must have permission for that exact action. MazeGraph’s model-initiated writes are limited to its own tools with trusted undo support. Draft tools create or update draft and pending content. Other tools can move content to the trash, edit media text or terms, and change WooCommerce data. A published-product update takes effect in the live shop; an order-status change may send an email. Content Studio hands drafts to the editor and does not publish posts. See Approval modes and Review changes and undo.
Permission modes affect approval, not access
Ask requires approval for each model-initiated write. Auto can skip the prompt only for two marked common operations: creating a draft and updating a draft. Full can skip the prompt for supported, undoable writes. In every mode, site content in the request can require approval, and the tool, assistant, WordPress capability, availability, usage, and policy checks still apply.
No mode can enable a disabled tool, expand a specialist assistant’s tool list, grant a WordPress capability, bypass a denying rule, or make an unsupported write safe. For the mode details, see Approval modes.
Find a recorded refusal
Administrators can open MazeGraph → Activity → Plugin logs and look for a policy.denied event. Its safe metadata can help identify the action and the rule code behind a refusal. The logs do not need conversation text or tool arguments to record the policy decision. For practical checks, see A tool is missing or refused and Policy Administration and Safeguards.