Connect a NeMo workflow to permitted work

Use the matching Cohesix NeMo kit guide and agent protocol guide, pinned to the selected source revision or read from the installed kit. The 1.2.0 selection is Linux AArch64 with NeMo Agent Toolkit 1.9.0, MCP SDK 1.29.1 and A2A SDK 0.3.26. Preserve its locked environment separately from CUDA/PEFT. A newer upstream feature or a Mac NeMo installation does not qualify this kit. For setup, load cohesix-ai-host-setup; for another agent client, load cohesix-agent-delegation.

Establish one subject and a working client

Verify the kit wheel/lock identities and pip check. Resolve COH_NEMO to the absolute installed cohesix-nemo executable, including in non-login SSH or service contexts. Use the existing HTTP gateway as the sole Queen TCP owner. Give one client process a private 0600 settings file with gateway_url, subject, request_auth_ref and delegated_ticket_ref. The subject label does not grant authority; the verified delegated ticket does. Use HTTPS or protected loopback HTTP. Keep credentials out of argv, model input and executor payloads.

: "${COH_NEMO:?set the absolute installed kit executable}"
: "${NEMO_SETTINGS:?set the private subject settings JSON path}"
"$COH_NEMO" doctor --settings "$NEMO_SETTINGS"
"$COH_NEMO" mcp --settings "$NEMO_SETTINGS"
"$COH_NEMO" a2a --settings "$NEMO_SETTINGS"

Use discovery to select the requested protocol and advertised action; do not silently switch protocols. New admission needs live available jobs/skills, fresh scope and cumulative standing budget. A client cannot enable a disabled protocol or renew a spent scope. For two users, use separate processes, settings and credentials; another user’s task must remain inaccessible.

Run once, then recover the same job

For authorized execution, prepare the kit’s request JSON with scope_id and the complete immutable host-ticket/v2 from the selected CUDA/PEFT recipe. Use a stable original ticket/idempotency identity and a fresh private output path. Follow the guide’s mcp --request or a2a --request command. First prove the deterministic native client path; an optional model planner adds no scope.

After a timeout or disconnect, inspect the original ID using mcp --recover-id ORIGINAL_TICKET_ID or a2a --get-task-id ORIGINAL_TICKET_ID, retaining the same settings and a new output filename. Output files are create-only. Recovery reads consume the caller’s finite read quota even when no new admission is needed. Some kit paths require discovery before recovery: if a spent scope hides new job tools/skills and recovery reports them unavailable, retain that error and use the gateway’s same-ID cohesix.inspect_job/cohesix.recover_job or native A2A tasks/get under existing read authority. Load the delegation guide for that client. Do not mint scope, submit again or use a new job ID to get a reply. Missing recovery access leaves the result unresolved for the operator.

Run the independent CUDA byte check or exact coh peft release verifier and serving canary before claiming the requested effect. A2A’s provider_verified=false, a cancel request and recovered_failure must keep their actual meanings. Cancellation is native termination only after its owner observes it.

Add a planner or measure value when requested

Use the kit’s agent-mcp or agent-a2a mode with separate protected model settings, one fixed input file and an unused trace log. Pin the loaded model ID and endpoint. The kit verifies the selected name through /v1/models; only actual tool calls followed by independent job evidence establish work.

For an evaluation request, freeze 1–16 dataset rows, expected answers, model, task IDs, quality criterion and budget. Use eval-direct, then the selected eval-mcp or eval-a2a in unused private directories. Read the exact benchmark contract. Compare actual row completion, tool invocation, output quality, latency and manual recovery. Check Toolkit’s standardized CSV and native traces: zero exit with no completed model rows is insufficient, and a completed row does not independently verify an answer or effect. Return measured results, original job identities, refusals, recovery limitations and native evidence; do not infer savings or whole-client compatibility from discovery.