Read the bounded, customer-scoped hosted-agent editor aggregate

Returns one selected agent, its resolved stack including saved turn_taking, voice_polish, answer_detection, navigation and audio_environment settings, which components were actually configured (for a voice the agent never stored, voice_stack.language_profiles.default_voices names the default voice model of each supported language the shown voice cannot run, so a language added in the editor gets one too), the same inbound-admission verdict the Agents list carries, the runnable-adapter catalog (each entry carrying component_metered, so the launch-supported view is one filter over it rather than a second copy), a wallet projection, a bounded metadata-only BYOK credential list, and the agent's measured reply time: the median of a bounded sample of its recorded worker-side reply turns, reported as unavailable rather than zero until real calls exist. That figure ends when the first audio frame leaves the worker, so carrier transit and remote playout are not in it. It is bound to the dashboard session organization and selected project; it never decrypts credentials, calls a provider, validates a key, or writes state. It also carries call_variables: the AI agent's normalized call-variables config (declared fields, SMS templates, SMS mode, caller lookup), the revision a write names as base_revision, the schema its calls are admitted with (a published workflow's variables when one is active, with workflow_version and the agent's own earlier fields as superseded_fields), whether its voice engine is live, its languages, and at most 50 of the project's contact lists for caller lookup, newest first, with lookup_lists_has_more. It is written only through PUT /v1/agents/{agent_id}/variables. The editor also carries the AI agent's failure_breakdown over the calls it handled in the last 30 days (window_days): failure_breakdown counts the labelled calls (classified_calls), those with a failure (failed_calls), and the failures by layer (telephony, media, agent; only layers with a failure appear) and within each layer by category, most frequent first; a breakdown with no failure has an empty layers list. One grouped read; labels are never decided on this read. It also carries capabilities, what the AI agent's configured tools, attached knowledge bases and calendar tools do on each conversation engine (engines, the saved one in engine) and call type (call_types: inbound, outbound, browser_test for the agent page's browser test, and homepage_demo in the homepage demo workspace only). calls[engine][call_type] says whether such a call starts (starts), whether it runs the published workflow (workflow: outbound calls do; an inbound call only through a workflow phone test pin) and, when it does not start, refused_reason. items lists each configured tool (at most 40, with items_has_more), an implied built-in such as request_callback (implied), the knowledge search when a knowledge base is attached, and phone menu navigation, each with id, kind (builtin, calendar, webhook, connection, legacy, knowledge, navigation, unknown), name and status[engine][call_type]: null when the item is offered on such a call, else the one reason it is not. knowledge counts the attached active knowledge bases and whether one has a ready document. It describes what is offered, not what a call does: suspension, the wallet, outbound authorization and a tool's own endpoint are still decided per call. The reasons are a closed set: engine_cannot_run_workflows: this call type runs the AI agent's published workflow, and its conversation engine (GPT-Live and Gemini Live today) cannot run workflow calls, so such calls are refused before they start; homepage_demo_refuses_workflows: the homepage demo does not run an AI agent that has a published workflow; tool_budget: the tools do not fit the call's 32-name tool contract (on a billed call it counts end_call and a ready knowledge base too): the call is refused, or it runs without the implied built-in (request_callback) or without phone menu navigation's three tools; tool_contract_unsafe: a tool name cannot be part of a call's tool contract (at most 32 lowercase letters, digits and underscores, starting with a letter), so calls are refused; workflow_call: this call type runs the AI agent's published workflow, and workflow calls do not offer the AI agent's own tools, knowledge bases or calendar tools yet; not_buildable: the voice worker cannot declare this entry (an unknown type or built-in, a repeated or reserved name, or a schema that is not an object or nests deeper than 32 levels), so it is left out of the call and named in a configured_tool_omitted event; sms_templates_missing: the SMS mode is Templates only and the AI agent has no SMS template, so there is nothing send_sms could send; integration_unavailable: the integration action, or its integration, is not active in this project; calendar_connection_inactive: the calendar connection the calendar tools name is not active in this project; calendar_fleet_not_ready: not every running voice worker can run the calendar tools yet, so they are not offered; knowledge_not_ready: no attached knowledge base has a ready document yet; needs_phone_caller: this built-in acts on a phone call (such as a text, a transfer or a callback), and this call type has no phone number on the line; calendar_test_writes_off: a workspace browser test books only on a dedicated test calendar connection with test writes allowed; outbound_only: phone menu navigation runs only on calls the AI agent places; worker_tool_name: the voice worker adds a tool of this name itself on these calls (switch_language while Follow the caller is on), so the configured tool is not built and a configured_tool_shadowed_by_worker_tool event names it; and, on the navigation item only, the reasons navigation.ineligible_reason of the voice-stack response documents: turned_off, not_released, engine_not_supported, languages_unreleased, languages_uncovered, too_many_sockets, no_language, tool_name_conflict.