Agent recipes, GAPE and AI connections

01 · Choose the right path: architecture, assistance, or AI execution

GAPE observes and reconciles a bounded set of architecture facts. In the implemented R1 path it reads declared committed-source inputs and creates deterministic projections without calling an AI provider. You do not need a Vertex AI connection simply to perform that GAPE operation.

An AI assistant can help you read an approved public-safe projection or draft an integration document. You provide the allowed material and review the proposed interpretation. The assistant's answer does not become a product contract or permission to execute.

VibePackr's Vertex AI integration is a separate product execution path. It needs a selected project, location, model and authorized service identity, together with the product's runtime binding and admission checks. A GAPE diagram or an agent prompt does not configure that path.

02 · Prepare the same small case for your agent

Download the service-status reading exercise. It describes one fictional request and response so you can practice identifying fields. It is not a Contract, Document, Helper request, observed API response or working integration.

Supply that file and the public Integration Helper guide to your organization's approved agent. Choose the recipe below that matches your task. These exercises request text review only and do not require a new connector, MCP server, appliance/Vertex AI call or product installation. Your chosen assistant may use its own hosted model, with its own data and billing settings. The prompts restrict additional task or tool calls; they do not disable the assistant’s model.

The expected deliverable is a grounded review table or draft correction. Check every cited field against the supplied file. If the agent invents a field, endpoint, credential, successful run or approval, reject that claim and ask it to identify the missing information.

A prompt is a task instruction, not an enforcement mechanism. Use the client's approved settings and inspect what actually happened. These recipes have not been run against a customer agent installation; no client version is certified by this guide.

{
  "case": "service-status-reading-exercise",
  "operation": "status.read",
  "input": {
    "service": "demo-service"
  },
  "illustrative_response": {
    "service": "demo-service",
    "healthy": true
  },
  "instruction": "Fictional reading exercise only. This is not a Contract, Document, Helper request, agent configuration, observed response or product success record."
}

For machine-format review, also download the actual CUSTOM_API document. Keep the fictional request/response exercise and the machine document as separate files. Their field names now describe the same status.read case.

03 · Agent recipe and verification matrix

These are ways to use an approved assistant to review supplied documents. They are not a list of certified VibePackr connectors. Record the client version before your own trial; official configuration syntax was checked on 2026-10-03 where available.

04 · Codex: a focused review recipe

1. Public guide and fictional service-status reading exercise
2. Supply the two files in the task context and paste the Codex review prompt. Compare every cited field with the input.
3. Observed fields, missing public schema facts, next local check and what was not executed

Review the attached service-status-reading-exercise.json and the approved public Integration Helper guide.
This is a documentation review task; do not run commands, install tools, open network connections, or edit files.
1. Separate the example's operation, request fields, and response fields.
2. Explain the role of Contract, Document, and Integration Helper using only the supplied guide.
3. List which exact input fields require a supplier-approved schema before a Helper request can be built.
4. For every conclusion, cite the input filename and JSON key or guide section.
5. Label any missing fact UNKNOWN. Do not invent contract IDs, endpoints, registry entries, credentials, or success output.
Return four headings: Observed input; Missing input; Proposed next local check; Not executed.
This prompt does not register, submit, execute, or approve any product operation.

Accept the review when: it identifies service as the input and service/healthy as the illustrative response, cites the supplied keys, and leaves target Contract values unknown unless actually supplied. A plausible-looking request it invents is not the expected result.

05 · Claude Code: a focused review recipe

1. An actual redacted Helper error and its matching public troubleshooting section
2. Supply only the allowed error/context, paste the Claude Code error-review prompt and inspect its quoted stage/reason before considering a proposed correction.
3. Evidence-based explanation and a smallest correction proposal; no applied change

Review only the supplied redacted Helper error text and the approved public troubleshooting guide.
Do not execute commands, read other project files, install packages, contact a provider, or edit configuration.
Quote the exact error stage and reason if present; write UNKNOWN for missing fields.
Explain what the error establishes and what it does not establish.
Propose the smallest correction to the supplied input as a draft for human review; do not invent schema fields.
If the required public schema is missing, list the missing schema/version rather than constructing a replacement.
Finish with a table: Observation | Evidence | Suggested correction | Verification still needed.
Do not claim that a proposed correction has been executed or that registration/execution succeeded.

Start with the recorded SCHEMA_INCOMPATIBLE example and its described missing uptime property. A useful suggestion either defines the genuinely needed primitive property or removes it from required. Check that the agent does not claim it reran the Helper.

06 · Cursor: a focused review recipe

1. Permitted source/target declaration copies and a public field reference
2. Provide those files as review context, paste the mapping prompt and inspect the field/type table. Review any textual diff before editing.
3. A field-mapping review with missing facts visible

Review the two supplied copies of the service-status declaration/example and the approved public field reference.
Stay in review mode: do not edit files, run a terminal command, enable MCP tools, or contact any service.
Build a mapping table: source field | target field | type | mapping supported by the supplied reference | open question.
For example, determine whether service is actually declared on both sides; do not assume similar names are compatible.
Keep missing required fields visible. Do not supply guessed credentials, contract identifiers, or endpoint paths.
Show a proposed textual diff only if the reference establishes the correct field; otherwise explain what information is needed.
End with Not executed: no Helper command, registration, provider call, or product operation.

For a first exercise, use the illustrative mapping table. A real target document must come from the supported package. When no target is supplied, the correct review says that compatibility is unconfirmed; it does not fill the target column with guessed names.

07 · Generic MCP client: a focused review recipe

1. Previously captured, redacted initialize/tools-list observations from an approved server
2. Review the saved observations with the prompt. Record actual protocol/tool schema details and compare them with the supported Helper input revision. Do not call tools/call to fill gaps.
3. An observation inventory and a compatibility question list

First label the supplied transcript SYNTHETIC_EXAMPLE or CAPTURED_OBSERVATION based on its stated provenance; use UNKNOWN if not supplied.
Review the supplied MCP initialize and tools/list transcript. If captured, identify the separately approved server from the supplied record; if synthetic, preserve that label.
Do not connect to a server or call a tool. Do not infer an endpoint or authentication method from these files.
Record the observed protocol version, server name/version, offered tool names, and inputSchema fields.
Mark missing, incomplete, paginated, or redacted observations explicitly.
Check that each required field and its type is described; do not call tools/call to discover behavior.
Explain which facts are observations and which compatibility checks remain pending for the supplied Helper release.
An advertised tool does not prove that it is authorized, callable, safe, or supported by VibePackr.
Return a review table, not a fabricated successful tool response.

The downloadable synthetic transcript gives you a safe reading exercise without connecting a server. Expected observations are revision 2025-03-26, tool status.read and the declared primitive field shapes. Label them synthetic. For real observations, preserve the negotiated revision and record client/server versions; do not call tools/call as part of this reading exercise.

08 · GAPE: observe structure, then explain the supplied projection

Use only the approved public-safe export supplied for your review. The public-safe viewer and JSON contain anonymous components, relationships and reconciliation states. Engineering snapshots, manifests and source references are different artifacts and are not included in this exercise.

Ask the agent to list the relationships shown and cite their component or edge IDs. Keep observations and interpretations in separate columns. Preserve states such as UNRESOLVED, STALE and CONSISTENT exactly as written.

UNRESOLVED means the comparison is still unresolved; it does not mean the feature is absent. CONSISTENT does not prove a running connection, execution permission or release readiness. Do not ask the agent to recover redacted names or fill an unknown state with an assumption.

The source-confirmed GAPE CLI commands are an operator reference. They require their matching input package and do not accept the public-safe reading export as a replacement for a full request or snapshot. A ready-to-run customer GAPE package is not established by this manual.

Explain the supplied, approved GAPE public-safe projection for a first-time reader.
Use only the supplied public-safe JSON or viewer content. Do not search for engineering exports or resolve redacted names.
List anonymous components and the relationships explicitly shown. Preserve each reconciliation state exactly.
Separate what the source projection says from your interpretation. For each interpretation cite the component or edge ID.
An UNRESOLVED state is an unanswered comparison; CONSISTENT is not runtime or authorization proof.
Do not infer hidden owners, credentials, endpoint addresses, live connectivity, execution permission, or release readiness.
Return: Observed relationships; Unanswered questions; Suggested documentation checks; Not executed.

Expected review: component/edge ID → exactly stated relationship → interpretation → unanswered question. If no approved export has been supplied, keep this prompt as a template; the manual does not fabricate an architecture snapshot.

09 · GAPE operator reference: input types matter

Source-confirmed syntax; not executed here. The operator must already have the matching tool and approved input package. This is not a Marketplace installation recipe.

packr-gape validate request.json
packr-gape project request.json new-output-folder
packr-gape snapshot manifest.json new-output-folder
packr-gape validate-snapshot gape-architecture-snapshot.json
packr-gape compare before-snapshot.json after-snapshot.json

An existing output directory is rejected by R1 export. Choose a new directory rather than deleting prior results. A failed command writes JSON with decision: FAIL to stderr and exits 1. Preserve that error; do not substitute a different input type.

10 · Vertex AI: identities, binding and the first controlled call

1. Complete the non-secret AI setup worksheet: project ID, supported location, exact model ID, service identity owner, allowed prompt data and test budget. The worksheet is not a product configuration file.
2. Have the cloud administrator confirm API/billing availability, the actual attached VM identity, inference permission, VM access scopes, model/location policy and the outbound network path. A broad administrator role is not a substitute for checking the actual inference identity.
3. Locate the supplier-supported binding/setup procedure for the exact release. The complete first-customer Linux setup path remains pending. Do not invent a configuration file or apply a macOS authentication command to the VM.
4. The product resolves an approved runtime binding and checks eligibility/admission. Persisting a selection is not a provider call. A missing eligible runtime should remain a configuration/readiness issue instead of silently using another provider.
5. On the implemented Linux path, the adapter obtains a token for the VM's attached service account through the metadata service and sends an outbound generateContent request. The macOS path uses existing application-default credentials. The token is not an item to paste into the guide or worksheet.
6. After the supported setup and authority are established, a controlled first call can use a harmless prompt such as Reply with exactly: onboarding-check. Record the actual result, exact model/location and product validation outcome. This guide does not report that call as executed.

AI connection worksheet — planning only; not a VibePackr configuration file.
Fill in non-secret values with your cloud administrator before a connected test.

Cloud project ID: <YOUR_PROJECT_ID>
Location supported by your chosen model: <YOUR_LOCATION>
Exact model ID: <YOUR_MODEL_ID>
Runtime service identity: <SERVICE_ACCOUNT_EMAIL>
Identity owner: <TEAM_OR_ROLE>
Vertex AI API enabled and billing available: <CONFIRMED / PENDING>
Inference permission checked: aiplatform.endpoints.predict <CONFIRMED / PENDING>
VM access scope and IAM both checked: <CONFIRMED / PENDING>
Chosen model/location allowed by organization policy: <CONFIRMED / PENDING>
Metadata and outbound Google API connectivity checked: <CONFIRMED / PENDING>
Approved prompt data class: <DATA_CLASS>
Allowed test prompt: Reply with exactly: onboarding-check
Approved test budget and retry limit: <LIMITS>
Supplier-supported binding/setup procedure and version: <REFERENCE / PENDING>

Do not paste tokens, API keys, private keys, credential files, or customer data here.
Completing this worksheet does not configure the product or authorize a provider call.

11 · Vertex request example and result interpretation

This is the public provider request shape for understanding the separate AI runtime. It is not a VibePackr setup file or a complete authenticated request. Replace project, location and model only after the supported deployment procedure supplies them; this guide selects no default model.

https://<LOCATION>-aiplatform.googleapis.com/v1/projects/<PROJECT_ID>/locations/<LOCATION>/publishers/google/models/<MODEL_ID>:generateContent

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "Reply with exactly: onboarding-check"
        }
      ]
    }
  ]
}

The product adapter obtains its own runtime credentials and adds its own request controls. The example does not expose those controls as user settings. Customer prompt content sent through this path leaves the VM for the selected Google API; use only the data approved for that service. The agent exercises above may have their own separate data-sharing policy.

For a future authorized first call, record: selected model/location, whether a call was actually sent, the returned text or sanitized failure, and the product validation result. The requested word onboarding-check is not a recorded response or a guaranteed model answer. HTTP success alone does not finish product validation.

12 · AI connection failures: the next bounded check

These categories explain the inspected adapter/selection behavior. They are not captured customer responses or a promise of public response-field spelling. Diagnose from the actual sanitized error. A PIN selection does not silently switch providers.

13 · Optional client syntax reference: keep outside active settings

Codex and Cursor publish configuration formats for connecting to MCP servers. The supplied templates illustrate those vendor formats only. They use a reserved .invalid address and contain no credentials. They are not VibePackr server registrations and should stay outside active configuration during this exercise.

The Codex template is explicitly disabled. The Cursor template is a reference JSON file with no assertion about disabled-server behavior. Before operational use, an administrator must supply the real approved server, transport, authentication procedure and exact client version.

Integration Helper's ability to read specified MCP initialize/tools-list observations does not mean that it runs a persistent MCP server or that VibePackr can be added at an invented URL. The bounded Helper CLI should not be entered as a STDIO server command without a separately documented server implementation.

# Vendor syntax reference only. No VibePackr MCP endpoint is supplied or verified.
# Keep disabled until your approved server address and access procedure are known.
[mcp_servers.approved_example]
url = "https://approved-server.invalid/mcp"
enabled = false

{
  "mcpServers": {
    "approved_example": {
      "url": "https://approved-server.invalid/mcp"
    }
  }
}

Codex documents mcp_servers in its TOML configuration; Cursor documents project .cursor/mcp.json and user ~/.cursor/mcp.json. File syntax was checked against the official references below. Reading or downloading these templates does not install them. Before any future connection, record the real approved endpoint, transport, identity procedure, client version and expected discovery result.

14 · Version records and self-service acceptance

For a repeatable connection example, retain client/tool version, product version, input revision, transport, exact approved setup procedure, positive and negative checks, expected outputs and cleanup. The recipes here do not certify client versions. Codex/Cursor configuration documentation was read on 2026-10-03; Claude configuration retrieval was unavailable, so no unverified Claude settings or CLI flags are supplied.

• Codex MCP reference
• Cursor MCP reference
• Claude Code MCP documentation
• Vertex AI inference reference

Next standard deliverables: complete customer request examples; a supported installation path; an approved GAPE reading export or operator package; the Linux AI binding/setup path; and version-specific connected agent tests. These are the remaining product/documentation standardization items, not instructions for customer-specific custom scripts.

Return to the worked-example checkpoints or customer AI preparation.