[sandl99]

Failure

Section 4 of the duplicated-checks list on #14060 asked for a review of the helm_recovery and remote_service integration tests against the SDK and provider tests. The #14061 list also asked where the remote models live test's checks belong.

Part of #14060 and #14061.

Decision

Tests	Decision	Reason
helm_recovery (5) and helm_recovery::destroy_faults (4)	Keep all	The SDK's Kubernetes tests (kubernetes_interruption, kubernetes_lost_create_response, kubernetes_read_failures, the recovery_* unit tests) run a scripted OpenTofu. These tests run the bundled Helm provider, which treats a rejected release lookup as an absent release, against the in-memory Kubernetes API. Only they show the SDK restoring that binding.
remote_service::remote_vllm_preserve_refuses_image_change…	Dropped	The SDK's docker_model_updates_preserve_running_containers_unless_recreation_is_explicit owns the Preserve and Recreate decision for vLLM, Ollama, and a remote vLLM. The Ollama scenario keeps a real plan reaching it.
Repeated remote_service steps	Each runs once	Capacity-free planning, incompatible images, export and unchanged reapply, and the no-probe check ran in up to five scenarios. Each scenario costs a Docker-over-SSH runtime.
tests/e2e/rust/remote_models.rs and service_images/ (live)	Deleted	No CI step ran it: nothing sets its NEMOCLAW_TEST_AGENT_PROFILE. remote_service owns its observation failure, export, unchanged apply, and retained storage. The SDK's runtime_compile::remote_service_is_independent_of_the_external_sandbox_gateway owns where the remote service publishes its port. An agent answering through a remote model needs a real SSH host, which the #14061 checklist already tracks.

Rust integration tests: 17 to 16 in these two modules. Net −878 lines.

Validation

• The 16 remote_service and helm_recovery tests pass against a bundle built from v1.
• cargo ci test on darwin_arm64: 1755 tests passed. Clippy for nemoclaw-e2e passed.

[sandl99]

Authoring session — opening

Field	Status	Value	Evidence or limit
Harness	partial	Pi coding agent 1.2.0	The session record names the Pi session format; the version comes from the active install path, not the record
Skills	recorded	nemoclaw-contributor-create-pr, v1 copy (complete read)	Skill reads to cutoff; other skills were listed only
Instruction files	recorded	AGENTS.md of main (injected); AGENTS.md of v1 (complete read); WRITING.md (complete read); .github/PULL_REQUEST_TEMPLATE.md (complete read); .github/SESSION_REPORT_TEMPLATE.md (complete read)	Startup injection and reads to cutoff
Tool calls	partial	typescript: 585	Wrapper calls, including failed calls, to cutoff; nested shell, file, and HTTP calls inside the wrapper are not counted
Session window	recorded	2026-11-02T05:58Z to 2026-11-03T11:33Z; 29 h 35 min	First event to cutoff before publication, including idle time; the session also authored #14063, #14121, #14142, and #14145
Compactions	recorded	0	Complete record to cutoff; context elisions are not compactions
Models	recorded	claude-opus-5-5	Model change event and assistant turns; no changes
Reasoning levels	recorded	medium; high	Thinking-level changes, in order

[github-actions]

v1 documentation preview: https://nvidia-preview-nemoclaw-v1-pr-14146.docs.buildwithfern.com/nemoclaw