[Epic] R0: prove a minimal VoiceClaw connection to one NemoClaw v1 agent

## Outcome

Prove one independent client can reach one v1-deployed OpenClaw agent through VoiceClaw. Use a private NemoClaw API endpoint and one short-lived, agent-scoped credential. Brev CLI is sufficient to operate the experiment.

R0 starts with hand-authored configuration consumed directly by v1. Main onboarding and configuration export are deferred. During apply, v1 verifies agent readiness and invokes VoiceClaw-owned onboarding. VoiceClaw owns installation, prompts, configuration, credentials and external processes.

## First experiment

1. Apply hand-authored configuration directly in v1 on Brev, with one VoiceClaw integration referencing one agent. Import/validation alone has no runtime effects.
2. v1 deploys or retains the agent and verifies readiness, then invokes trusted VoiceClaw-owned onboarding. Starting with VoiceClaw absent, its tooling installs its package, collects its settings and credentials, configures its stack and starts its external processes.
3. Give VoiceClaw the private semantic API endpoint and a short-lived credential scoped to that verified agent, through protected storage or a pipe.
4. Connect an independent client through VoiceClaw's real interface and prove it reaches the agent.
5. Stop the run. A disconnect, credential expiry or restart ends the experiment; another run requires a fresh operator-authorized credential.

This is direct authentication with one credential. It does not require an offer document, single-use redemption, a second attachment credential, lease recovery or automated reconnection.

## Ownership boundary

| Responsibility | Owner |
| --- | --- |
| Hand-authored integration configuration, validation, apply and onboarding invocation | NemoClaw v1 |
| Agent/sandbox deployment, target authorization, scoped credential and semantic API | NemoClaw v1 / OpenShell |
| Acquisition, installation, onboarding, internal settings/credentials and external processes | VoiceClaw |
| Client API, media/WebRTC, ASR/STT, TTS, TURN, interruption and playback | VoiceClaw |
| Native/ACP adapters, native IDs, protocol frames and agent credentials | Private to NemoClaw/OpenShell |

VoiceClaw calls NemoClaw's semantic API. Only finalized text, semantic controls and semantic events cross that boundary. NemoClaw does not import VoiceClaw's internal configuration or manage its processes; their lifecycle is `not-managed`.

## Contract status

The user selected this temporary R0 profile to reduce the first experiment. It preserves the ownership boundary in `VOICECLAW_NEMOCLAW_API_CONTRACT.md` but deliberately simplifies its proposed offer/redemption/attachment-lease admission rules. This is not full `v1alpha1` conformance or release qualification. The proposed full contract remains the later reference; #11747 records this bounded profile explicitly.

## R0 completion

- [ ] v1 consumes hand-authored integration intent without VoiceClaw's internal settings. Import/validation alone has no runtime effects.
- [ ] v1 provides one verified real agent and correct integration ownership.
- [ ] After agent readiness, v1 invokes VoiceClaw-owned onboarding. VoiceClaw starts absent and owns installation, prompts, credentials, configuration and startup.
- [ ] One protected, short-lived credential authorizes access only to that actual agent.
- [ ] An independent client reaches the agent through VoiceClaw's actual interface and NemoClaw's semantic API.
- [ ] Invalid/expired credentials, wrong/replaced targets and unavailable agents fail correctly.
- [ ] Native protocols/credentials remain private, and VoiceClaw failure preserves the healthy agent.
- [ ] Pinned versions, Brev commands and redacted evidence make the experiment repeatable.

## R0 tasks

- [x] [#11747: [R0] Specify the v1-to-VoiceClaw onboarding handoff and scoped connection fixtures](https://github.com/NVIDIA/NemoClaw/issues/11747)
- [ ] [#11750: [R0] Verify one plain OpenClaw deployment on NemoClaw v1 through Brev CLI](https://github.com/NVIDIA/NemoClaw/issues/11750)
- [ ] [#11751: [R0] Apply voice integration intent on v1 and hand off to VoiceClaw onboarding](https://github.com/NVIDIA/NemoClaw/issues/11751)
- [ ] [#11749: [R0] Connect VoiceClaw to one v1 agent with a scoped credential](https://github.com/NVIDIA/NemoClaw/issues/11749)
- [ ] [#11752: [R0] Demonstrate a v1-deployed agent reached through onboarded VoiceClaw using Brev CLI](https://github.com/NVIDIA/NemoClaw/issues/11752)

The contract and plain v1 baseline can proceed in parallel. #11751 owns the v1 configuration fixture, apply and onboarding trigger. #11749 supplies scoped access. #11752 proves the complete path. #11748 does not block R0.

VoiceClaw counterpart: [VoiceClaw #7](https://gitlab-master.nvidia.com/jarvis/voice-claw/-/issues/7).

## Deferred from this experiment

Main onboarding/configuration export (#11748), separate offer resources/redemption, credential exchange, attachment leases, idempotent attach reconciliation, transport recovery, automatic reconnection/renewal, reusable broker/OIDC identity, public ingress and durable sessions/replay. #8403 remains outside R0. These are not hidden acceptance gates.

A real spoken lookup, correlated work results and request-loss reconciliation remain R1a:

- [#11753: [R1a] Define and implement correlated VoiceClaw submission receipts and speech/detail results](https://github.com/NVIDIA/NemoClaw/issues/11753)
- [#11754: [R1a] Reconcile duplicate submissions and lost VoiceClaw responses without blind redispatch](https://github.com/NVIDIA/NemoClaw/issues/11754)
- [#11755: [R1a] Configure and verify real weather lookup access for the selected OpenClaw agent](https://github.com/NVIDIA/NemoClaw/issues/11755)
- [#11756: [R1a] Demonstrate one real voice question with correlated speech and source detail](https://github.com/NVIDIA/NemoClaw/issues/11756)

Request idempotency in R1a is separate from deferred attachment idempotency. The full contract's remaining release requirements belong to later agreed phases.

## Historical work

#8275 and #9646 retain prior planning/evidence. #11204 remains a separate proposal; this epic does not accept its entire implementation design. No implementation or deployment is performed by this planning update.

## Category

Feature

Start with VOICECLAW_NEMOCLAW_API_CONTRACT.md and the R0 child issues #11750, #11751, #11749, and #11752 to understand the planned handoff, deployment, credential, and end-to-end proof. The epic is complete when the checklist is demonstrated with pinned Brev commands, protected scoped credentials, failure checks, and redacted repeatable evidence.

Written by the indexing model from the issue text.