Native questions & evidence

Native questions & answers

Where passo.build needs a human decision, it returns native questions — structured prompts (text, keyed options, a default) — that the agent renders in its own interface. Most simple questions stay native to the agent chat. When the user needs to inspect state visually, passo.build returns an agentUi descriptor that the host can render as a compact widget or wizard view. Hosts that cannot render widgets must fall back to the structured JSON and ask the same question in chat. State advances server-side, so answers survive a resume.

The same native-question loop carries you through the full lifecycle, all agent-native:

1. Scope — discovered vs. missing components, sizing tier, database exposure (REST / GraphQL / both), account mode, spend cap.
2. Security — credential and supply-chain checks surface as pre-flight results the agent presents inline.
3. Architecture — the agent proposes the target topology (components, providers, sizes, extensions) it inferred locally and submits it for validation; passo.build returns validity, warnings, and a cost estimate.
4. Compare — provider/cost comparison is returned as data the agent renders, so you choose a target with the numbers in front of you.
5. Deploy — pre-flight checks (credentials secure, supply chain clean, comparison approved) are confirmed, then deployment runs step-by-step with explicit confirmation and a rollback path on failure.

Every step is a question/answer round-trip over MCP; none of it requires opening the console.

Chat, widget, and wizard escalation

Step	Chat-only	Widget / wizard
Scope	One missing answer or simple default confirmation.	Multiple required gaps or topology-affecting choices.
Security	Passing gate.	Vulnerabilities, secrets, warnings, or remediation tasks.
Implementation plan	Agent-only task execution.	Human approval of generated mesh tasks.
Architecture	Stable JSON handoff.	Topology inspection or post-SAST architecture change.
Compare/deploy	Evidence readback.	Provider, region, spend, account, billing, or deploy choice.

Evidence handoff

Each step returns artifacts the agent uses to drive the next one — and that you keep as a record:

• Scope → a finalized scope object (database replacements, gateways, sizing) keyed to your project.
• Security → pass/fail pre-flight results (exposed-credential count, supply-chain audit) gating deploy.
• Architecture → a validation verdict with warnings and an estimated monthly cost + breakdown.
• Compare → a provider/cost comparison the deploy step references for approval.
• Deploy → step-by-step deployment status with success/failure per step and a rollback action that tears down provisioned resources to avoid charges.

The handoff is sequential: scope feeds architecture, architecture + compare feed deploy, and deploy gates on the security and comparison evidence. Because state is persisted server-side and keyed to the project, the handoff survives across agent sessions and across the supported hosts — start in Claude Code, finish in Cursor, same project, same evidence chain. For the exact evidence tools and shapes, use tools/list; the catalog is authoritative.

See Security gates for the operations that always require explicit human approval.

For a user-facing explanation of evidence, see What counts as evidence.