Skip to documentation
Code Voucherby Pyratz
Open
Browse documentation
How-to guideSetup assistant12 min readDogfood alpha

Prepare the repository with the setup skill

Use a local coding agent to map runtime services, external integrations, existing test mocks, initialization needs, and browser capability before proposing repository-owned QA files.

Local assistant, not remote execution

This page does not ask Code Voucher to fetch or run the repository. The skill analyzes the selected local checkout, proposes changes for review, and never publishes them without explicit approval.

Start with analysis

Invoke the installed repository-local skill from the repository you want to prepare.

Agent prompt
Use $agentic-qa-setup to analyze this repository and propose a safe, deterministic Agentic QA setup.
  1. 1

    Map the runtime

    Statically inventory entrypoints, infrastructure, external integrations, initialization, and existing test assets without executing repository code.
  2. 2

    Classify the target

    Report browser_ready, composition_required, or api_only without manufacturing a browser surface.
  3. 3

    Review dependency strategies

    Inspect the evidence, proposed substitutes, negative cases, confidence, and unresolved gaps.
  4. 4

    Accept the file plan

    Implementation begins only after an evidence-backed proposal exists and local repository edits are in scope.

What the proposal contains

  • An evidence-backed inventory of every material runtime dependency
  • An attempt-local service topology and acyclic readiness chain
  • A per-file create, no-op, additive, or conflict decision
  • Synthetic data and explicitly authorized, classified fixture handling
  • Deterministic positive and negative browser scenario candidates
  • Separate static, isolated execution, and cleanup validation stages
  • Fidelity limits, unsupported workflows, confidence, and unresolved items

Choose runtime-compatible substitutes

Monkeypatches are evidence of expected calls and outcomes, not mocks for an application running in another process or container. Unknown stub operations fail closed.

existing_fake
Select an existing runtime-compatible fake through non-secret QA configuration.
local_http_stub
Serve only evidenced request and response shapes from an attempt-local service.
stateful_stub
Model polling, callbacks, retries, replay, and idempotency with bounded local state.
local_infrastructure
Run disposable databases, caches, queues, mail sinks, or compatible emulators.
not_mocked_with_reason
Keep an integration outside the proposed workflows and record why.
unresolved
Block affected workflows when evidence or a safe configuration seam is missing.

See how to replace external services.

Understand readiness

The setup report uses analysis_complete, proposal_generated, and static_valid as classifications; these are not dashboard or backend states. environment_ready and ready require isolated execution with executor attestation. Without an attested executor, local setup stops at static_valid. mocks_incomplete, drifted, and failed require action.

api_only and composition_required cannot run Agentic QA today because the current runner requires a real browser surface. Do not manufacture a browser UI to change that result.

Safety boundaries

Host enforcement is required

Plain local Docker Compose cannot prove no-egress isolation. Execution requires user authorization plus an attested disposable, credential-free executor with host-enforced runtime egress limited to attempt-local services.
  • Require a deterministic broker or sandbox outside the model to enforce the readable source corpus
  • Do not read credential files or environment-example values
  • Do not contact external providers or provider sandbox endpoints
  • Do not copy production data, sessions, personal information, or real credentials
  • Inspect fixture contents only after explicit authorization, then classify or quarantine them
  • Do not execute repository code during analysis or file generation
  • Do not push, open a merge request, or change SaaS settings without explicit approval

A model-authored exclusion list is not an isolation boundary. If the local runtime cannot enforce a readable-file allowlist before repository content reaches the assistant, analysis stops with analysis_blocked: unenforced_read_boundary.

Continue to the environment

Define the QA environment, replace External services, then add Browser scenarios.