Connect Hosted

Inspect

Use known project endpoint, exact agent name/version, KB MCP identity, runtime principal and actual connection/toolbox/binding readbacks through authenticated, observed read operations. Missing read tooling is a tooling blocker, not a source request. Run these azd reads only with already available project/environment:

Never use azd ai project set as a read; context changes require approval. Inspect compatible connections/toolboxes before choosing a new name. Never overwrite a Prompt connection using ProjectManagedIdentity to make it Hosted.

Compare agent identity/version, model, telemetry, ACR, network, environment, connection target/auth/audience and toolbox/default tools. For omitted audience or UnknownConnectionPropertiesV2, use the exact nonsecret connection readback. Projection gaps are warnings only after authoritative equality is proven; denial, conflicting properties or still-missing required evidence blocks.

Resolve the actual published agent principal, not its blueprint or project MI. Check effective runtime access and plan only missing grants:

Reuse compatible effective grants. Disclose that Search service scope covers its indexes, while Foundry User grants broader project data operations, not just toolbox reads. No automatic grant, scope widening or caller-role change. A newly observed principal or access requirement needs a separate approved access plan.

Select binding before approval

Unchanged consumer: toolbox-only branch; no deploy.

Source-free changes require observed runtime support and an authoritative surface. Otherwise explain the exact code/config/deploy delta and handoff/request source only then. Hosted code orchestrates tools: existing agents may need source too; no arbitrary Hosted tool changes through Prompt APIs.

For the source-backed recipe, inspect its supported mode and emitted environment, not just local variables. For FoundryToolbox, an explicitly empty TOOLBOX_ENDPOINT raises instead of falling back to TOOLBOX_NAME.

The consumer endpoint is <project-endpoint>/toolboxes/<toolbox-name>/mcp?api-version=v1. Do not substitute the version-specific developer endpoint returned by some toolbox readers. Preserve an already compatible exact binding without standardizing its mode. For new/repair bindings, select one supported mode: if the manifest injects an empty endpoint, prefer endpoint mode when supported. Unsetting a local variable does not remove a manifest’s runtime environment key. If no valid binding is possible without source edits, stop for separate review.

If installed, validate resolution offline with the SDK; do not install it for this check. No token/model call or standalone agentic toolbox tools/list: runtime identity requires published-agent context.

Approved changes only

This source-backed recipe requires the approved existing azd ai agent source project; unsupported/image-only source blocks this recipe, not remote inspection. Retain source/config hashes for source/config changes or source-backed redeployment. Plan grants/resource/binding deltas, source digest, costs and verification. Exact state skips every mutation below. After approved access changes, create only absent exact resources:

Verify stored AgenticIdentityToken, audience, target and project before toolbox creation. YAML contains only description and the exact connection name; no credentials or extra tools. Creation auto-publishes the first toolbox version and writes TOOLBOX_<NORMALIZED_NAME>_MCP_ENDPOINT locally; disclose that write.

Apply only the selected approved binding delta. Endpoint mode:

If the source exposes an unused name variable, clear it only as an approved delta:

Name mode, only with the endpoint absent from the emitted runtime environment:

Deploy only when remote configuration differs, under the same agent name:

Independently read back the returned version, actual principal, binding and protected state. active metadata is not runtime-health proof. Unknown/changed identity, source drift or failure stops; no blind retry, --force, handcrafted mutation API, replacement agent or automatic cleanup.

Verify and reconcile

Pin the observed version at session creation. A new session does not guarantee a new Responses conversation. Start both fresh:

Capture the first response’s session ID as <first-response-session-id> and its conversation ID. Verify both are distinct from prior trial IDs. Missing or reused first-response IDs block the second call. Do not rely on automatic session selection; explicitly pin the captured session while starting a fresh conversation:

Sessions bind their version at creation; --version and --session-id are mutually exclusive. If the installed CLI lacks --session-id, block verification rather than falling back to a saved session. Require the second response’s session ID to equal the captured ID and its conversation ID to differ from the first and all prior trials. Missing or mismatched IDs make isolation unverified: stop before monitoring or further invocations. Do not silently add invocations to repair a bounded test.

Correlate actual knowledge_base_retrieve success with each request and response, using returned events or supported logs:

Require supported answers with original citations and unsupported answers exactly I don't know. without citations. Errors are not abstention. Tool success proves execution, not the contents of an unavailable payload: record missing arguments, results or usage and leave payload-level faithfulness unverified. Never infer empty retrieval or change source/telemetry to manufacture proof.

Compare identities, versions, tools, roles, protected source and binding. Repeat reconciliation through reads only: no setters, create, publish or deploy calls on exact state. Zero configuration writes excludes authorized invocation/conversation/session activity and local test metadata; report both. Retain resources. Cleanup remains separately approved and hosted-cleanup-unsupported in this skill.