ci(e2e): establish trusted Jetson execution through Colossus-backed dispatch

.agents/skills/nemoclaw-maintainer-cut-release-tag/SKILL.md

@@ -136,7 +136,7 @@ npm run release:e2e-evidence -- \
 
 The preflight derives every required execution from one empty-selector dispatch.
 The full run includes every default-selected workflow E2E plus `Exact staging Brev Launchable`.
-Accepted release evidence requires `allow_jetson_runner_queue=false` and `allow_dgx_spark_runner_queue=false`.
+Accepted release evidence requires `allowJetsonDispatch: false`, `allowJetsonRunnerQueue: false`, and `allowDgxSparkRunnerQueue: false`.
 The required denominator excludes `jetson-nvmap-gpu`, `llama-cpp-dgx-spark-plan`, and `llama-cpp-dgx-spark-qualification`.
 Each job that declares `RELEASE_E2E_ACTIVATION_PATH` requires that path at the candidate SHA.
 A missing activation path is a preflight failure.
@@ -148,7 +148,7 @@ Monitor the dispatched correlation ID with one bounded status query.
 Before accepting full-mode exact Brev evidence, require:
 
 - the workflow `head_sha` to equal the plan candidate SHA;
-- the trusted dispatch receipt to prove empty selectors, `include_staging_brev_launchable=true`, `allowJetsonRunnerQueue: false`, and `allowDgxSparkRunnerQueue: false`;
+- the trusted dispatch receipt to prove empty selectors, `include_staging_brev_launchable=true`, `allowJetsonRunnerQueue: false`, and `allowDgxSparkRunnerQueue: false`; a v1 receipt may omit `allowJetsonDispatch` but must set it to `false` when present, while a v2 receipt must include `allowJetsonDispatch: false`;
 - the workflow conclusion to be `success`;
 - the `Exact staging Brev Launchable` job conclusion to be `success`;
 - the job URL and selected successful Launchable job attempt;

.agents/skills/nemoclaw-maintainer-cut-release-tag/scripts/release-e2e-evidence.mts

@@ -484,6 +484,16 @@ export function buildReleaseE2eLedger(
       false,
       `${label}.dispatch.allowJetsonRunnerQueue`,
     );
+    if (
+      dispatch.kind === "nemoclaw-e2e-dispatch-v2" ||
+      Object.hasOwn(dispatch, "allowJetsonDispatch")
+    ) {
+      requireEqual(
+        booleanField(dispatch, "allowJetsonDispatch", `${label}.dispatch`),
+        false,
+        `${label}.dispatch.allowJetsonDispatch`,
+      );
+    }
     requireEqual(
       booleanField(dispatch, "allowDgxSparkRunnerQueue", `${label}.dispatch`),
       false,

.agents/skills/nemoclaw-maintainer-day/MERGE-GATE.md

@@ -85,7 +85,7 @@ After a failure, inspect the artifacts and remove resources that target cleanup
 
 Dispatch the trusted `main` workflow with the current PR number, lowercase 40-character head SHA, head repository, lowercase 40-character base SHA, trusted workflow SHA, and a review reason containing 10 to 500 printable characters.
 Leave job and target selectors empty and keep Launchable disabled.
-Keep `allow_jetson_runner_queue=false` and `allow_dgx_spark_runner_queue=false` for the default PR revision selection.
+Keep `allow_jetson_dispatch=false` and `allow_dgx_spark_runner_queue=false` for the default PR revision selection.
 If the DGX Spark flag is `true`, GitHub can pause `llama-cpp-dgx-spark-qualification` for the `approve-dgx-spark-image-qualification` environment.
 An authorized environment reviewer must approve it before qualification starts.
 The trusted pre-checkout step requires current `maintain` or `admin` access and validates the exact open PR before candidate code runs.

.agents/skills/nemoclaw-maintainer-e2e/SKILL.md

@@ -67,7 +67,8 @@ Choose exactly one mode:
   - every default-selected free-standing workflow E2E except `Exact staging Brev Launchable`;
   - every shared credential-free test; and
   - these controller-selected registry targets: `ubuntu-policy-custom-missing-presets-negative`, `ubuntu-repo-cloud-langchain-deepagents-code`, `ubuntu-repo-cloud-openclaw`, and `ubuntu-repo-docker-post-reboot-recovery`.
-  The run skips `jetson-nvmap-gpu`, `llama-cpp-dgx-spark-plan`, and `llama-cpp-dgx-spark-qualification` unless their separate runner-queue flags are `true`.
+  The run skips `jetson-nvmap-gpu` unless its Colossus dispatch flag is `true`.
+  It skips `llama-cpp-dgx-spark-plan` and `llama-cpp-dgx-spark-qualification` unless their runner-queue flag is `true`.
 - For protected managed-image runtime qualification, set `E2E_JOBS=managed-image-protected-runtime`. The exact candidate must contain `ci/protected-managed-image-multiarch-activation-v1.json` and `ci/protected-managed-image-runtime-activation-v1.json`.
 
 Leave `targets` empty and keep Launchable disabled:
@@ -81,7 +82,7 @@ esac
 REVIEW_REASON='Reviewed the PR head commit for credentialed E2E.'
 CORRELATION_ID="$(python3 -c 'import uuid; print(uuid.uuid4())')"
 INFERENCE_MODE=mock
-ALLOW_JETSON_RUNNER_QUEUE=false
+ALLOW_JETSON_DISPATCH=false
 ALLOW_DGX_SPARK_RUNNER_QUEUE=false
 gh workflow run .github/workflows/e2e.yaml \
   --repo NVIDIA/NemoClaw \
@@ -90,7 +91,7 @@ gh workflow run .github/workflows/e2e.yaml \
   -f "jobs=${E2E_JOBS}" \
   -f "inference_mode=${INFERENCE_MODE}" \
   -f include_staging_brev_launchable=false \
-  -f "allow_jetson_runner_queue=${ALLOW_JETSON_RUNNER_QUEUE}" \
+  -f "allow_jetson_dispatch=${ALLOW_JETSON_DISPATCH}" \
   -f "allow_dgx_spark_runner_queue=${ALLOW_DGX_SPARK_RUNNER_QUEUE}" \
   -f "pr_number=${PR_NUMBER}" \
   -f "checkout_sha=${HEAD_SHA}" \
@@ -162,7 +163,7 @@ Ask for clarification only when the request contains conflicting mode phrases.
 Ordinary mode selects every default-selected workflow E2E except `Exact staging Brev Launchable`.
 Launchable mode runs only `Exact staging Brev Launchable`.
 Full mode adds `Exact staging Brev Launchable` to the default E2E selection in the same workflow run.
-The documented invocations for all three modes keep both hardware runner-queue flags set to `false`.
+The documented invocations for all three modes keep the Jetson dispatch and DGX Spark runner-queue flags set to `false`.
 
 ## Resolve the Candidate
 
@@ -199,7 +200,7 @@ gh workflow run .github/workflows/e2e.yaml \
   -f jobs= \
   -f inference_mode=mock \
   -f include_staging_brev_launchable=false \
-  -f allow_jetson_runner_queue=false \
+  -f allow_jetson_dispatch=false \
   -f allow_dgx_spark_runner_queue=false \
   -f "correlation_id=${CORRELATION_ID}"
 ```
@@ -214,7 +215,7 @@ gh workflow run .github/workflows/e2e.yaml \
   -f jobs=staging-brev-launchable \
   -f inference_mode=mock \
   -f include_staging_brev_launchable=false \
-  -f allow_jetson_runner_queue=false \
+  -f allow_jetson_dispatch=false \
   -f allow_dgx_spark_runner_queue=false \
   -f "correlation_id=${CORRELATION_ID}"
 ```
@@ -229,7 +230,7 @@ gh workflow run .github/workflows/e2e.yaml \
   -f jobs= \
   -f inference_mode=mock \
   -f include_staging_brev_launchable=true \
-  -f allow_jetson_runner_queue=false \
+  -f allow_jetson_dispatch=false \
   -f allow_dgx_spark_runner_queue=false \
   -f "correlation_id=${CORRELATION_ID}"
 ```
@@ -240,11 +241,14 @@ The `include_staging_brev_launchable` input adds the Launchable E2E job to that
 The trusted `main` workflow verifies that the dispatching and rerunning actors have
 repository `maintain` or `admin` permission before the Launchable path's source
 checkout. That role check is the authorization.
-A user permitted to dispatch this workflow may set
-`allow_jetson_runner_queue=true` to add `jetson-nvmap-gpu` to an empty-selector
-manual run or enable its explicit selection. Set it only after a repository
-administrator confirms an online Jetson runner in the authoritative runner
-inventory.
+A user permitted to dispatch this workflow may set `allow_jetson_dispatch=true`
+to add `jetson-nvmap-gpu` to an empty-selector manual run or enable its explicit
+selection. Set it only after every deployment check in
+[Jetson Dispatch Through Colossus](../../../test/e2e/docs/jetson-colossus-dispatch.md)
+passes.
+The Jetson lifecycle removes and verifies only its fixed job-owned cleanup
+allowlist. Do not treat `cleanup: "succeeded"` as evidence that every possible
+candidate host change was reversed.
 A permitted dispatcher may set `allow_dgx_spark_runner_queue=true` to add
 `llama-cpp-dgx-spark-plan` and `llama-cpp-dgx-spark-qualification` to an
 empty-selector manual run or enable explicit qualification selection. Set it
@@ -350,6 +354,8 @@ The validator requires:
 
 - the workflow run to succeed for the selected SHA;
 - `dispatch.json` to bind the same run, empty selectors, `include_staging_brev_launchable=true`, `allowJetsonRunnerQueue: false`, `allowDgxSparkRunnerQueue: false`, and the selected successful Launchable job attempt;
+- `allowJetsonDispatch: false` in every v2 `dispatch.json` receipt;
+- `allowJetsonDispatch` to be absent or `false` in every v1 `dispatch.json` receipt;
 - `Exact staging Brev Launchable` to conclude `success` in the selected current or earlier attempt of the same workflow run;
 - `launchable-e2e.json` to identify the selected SHA in the repository and provision records;
 - the booted repository to be unmodified;
@@ -376,7 +382,10 @@ If the release candidate SHA changes, discard the earlier full run and dispatch
 No release-note-only delta exception is currently defined.
 
 When `nemoclaw-maintainer-cut-release-tag` invokes this skill, return the validated fields for its pre-tag E2E evidence ledger.
-The trusted `dispatch.json` receipt proves that full mode used empty selectors, included `Exact staging Brev Launchable`, and disabled both optional hardware paths.
+The trusted `dispatch.json` receipt proves that full mode used empty selectors and included `Exact staging Brev Launchable`.
+For Jetson dispatch, a v2 receipt requires `allowJetsonDispatch: false`.
+A v1 receipt may omit `allowJetsonDispatch`, but it must be `false` when present.
+Both receipt versions require the optional DGX Spark runner path to be disabled.
 The release evidence ledger proves the result of each workflow E2E.
 Do not ask for the release confirmation phrase in this skill.
 

.agents/skills/nemoclaw-maintainer-e2e/scripts/validate-full-e2e-evidence.mts

@@ -30,6 +30,7 @@ export interface FullE2eEvidenceSummary {
   };
   dispatch: {
     allowDgxSparkRunnerQueue: false;
+    allowJetsonDispatch: false;
     allowJetsonRunnerQueue: false;
     emptySelectors: true;
     includeStagingBrevLaunchable: true;
@@ -166,6 +167,9 @@ export function validateFullE2eEvidence(input: FullE2eEvidenceInput): FullE2eEvi
     "dispatch.includeStagingBrevLaunchable",
   );
   requireEqual(dispatch.allowJetsonRunnerQueue, false, "dispatch.allowJetsonRunnerQueue");
+  if (dispatch.kind === "nemoclaw-e2e-dispatch-v2" || Object.hasOwn(dispatch, "allowJetsonDispatch")) {
+    requireEqual(dispatch.allowJetsonDispatch, false, "dispatch.allowJetsonDispatch");
+  }
   requireEqual(dispatch.allowDgxSparkRunnerQueue, false, "dispatch.allowDgxSparkRunnerQueue");
   requireEqual(dispatch.emptySelectors, true, "dispatch.emptySelectors");
 
@@ -233,6 +237,7 @@ export function validateFullE2eEvidence(input: FullE2eEvidenceInput): FullE2eEvi
     },
     dispatch: {
       allowDgxSparkRunnerQueue: false,
+      allowJetsonDispatch: false,
       allowJetsonRunnerQueue: false,
       emptySelectors: true,
       includeStagingBrevLaunchable: true,

.agents/skills/nemoclaw-maintainer-policies/references/release-train.md

@@ -57,7 +57,7 @@ Before asking for the release confirmation phrase, build and show an evidence le
 - Require the workflow-produced trusted dispatch receipt to bind the accepted run candidate SHA, run ID, attempt, and selector inputs.
 - Run `nemoclaw-maintainer-e2e` in full mode when the ledger lacks complete evidence for the candidate SHA.
 - Require one completed, successful full workflow run for all default-selected workflow E2E jobs and the full-mode additions, including `Exact staging Brev Launchable`.
-- Require the trusted dispatch receipt to record `allowJetsonRunnerQueue: false` and `allowDgxSparkRunnerQueue: false`.
+- Require the trusted dispatch receipt to record `allowJetsonRunnerQueue: false` and `allowDgxSparkRunnerQueue: false`. A v1 receipt may omit `allowJetsonDispatch` but must record `false` when the field is present. A v2 receipt must record `allowJetsonDispatch: false`.
 - Exclude `jetson-nvmap-gpu`, `llama-cpp-dgx-spark-plan`, and `llama-cpp-dgx-spark-qualification` from the required denominator.
 - Require the trusted dispatch receipt to bind the workflow run and an attempt no later than the run's latest attempt. The receipt must record empty selectors and `include_staging_brev_launchable=true`.
 - Require the Launchable E2E receipt to identify the candidate SHA in the repository and provision records.

.github/workflows/e2e.yaml

@@ -15,7 +15,7 @@ on:
         default: ""
         type: string
       jobs:
-        description: "Optional comma-separated E2E test IDs. Empty selectors choose the default suite. Set include_staging_brev_launchable for Launchable. Jetson and DGX Spark require their queue flags. PR revisions use the trusted controller matrix."
+        description: "Optional comma-separated E2E test IDs. Empty selectors choose the default suite. Set include_staging_brev_launchable for Launchable. Jetson dispatch and DGX Spark require their opt-in flags. PR revisions use the trusted controller matrix."
         required: false
         default: ""
         type: string
@@ -33,8 +33,8 @@ on:
           - mock
           - internal-nvidia
           - public-nvidia
-      allow_jetson_runner_queue:
-        description: "Before setting true for jetson-nvmap-gpu, obtain repository administrator confirmation of an online Jetson runner in the authoritative NVIDIA/NemoClaw Settings -> Actions -> Runners inventory; queued jobs do not honor timeout-minutes before assignment."
+      allow_jetson_dispatch:
+        description: "Set true for jetson-nvmap-gpu only after the Colossus dispatcher, Cloudflare Tunnel, cleanup helper, and JETSON_DISPATCH_URL repository variable have passed the deployment checks in test/e2e/docs/jetson-colossus-dispatch.md."
         required: false
         default: false
         type: boolean
@@ -304,7 +304,8 @@ jobs:
         if: ${{ github.event_name == 'workflow_dispatch' }}
         env:
           ALLOW_DGX_SPARK_RUNNER_QUEUE: ${{ inputs.allow_dgx_spark_runner_queue && 'true' || 'false' }}
-          ALLOW_JETSON_RUNNER_QUEUE: ${{ inputs.allow_jetson_runner_queue && 'true' || 'false' }}
+          ALLOW_JETSON_DISPATCH: ${{ inputs.allow_jetson_dispatch && 'true' || 'false' }}
+          ALLOW_JETSON_RUNNER_QUEUE: "false"
           BASE_SHA: ${{ inputs.checkout_sha != '' && inputs.base_sha || github.sha }}
           CANDIDATE_REPOSITORY: ${{ inputs.checkout_repository || github.repository }}
           CANDIDATE_SHA: ${{ inputs.checkout_sha || github.sha }}
@@ -334,6 +335,7 @@ jobs:
             --arg workflowRunId "$RUN_ID" \
             --arg workflowSha "$WORKFLOW_SHA" \
             --argjson allowDgxSparkRunnerQueue "$ALLOW_DGX_SPARK_RUNNER_QUEUE" \
+            --argjson allowJetsonDispatch "$ALLOW_JETSON_DISPATCH" \
             --argjson allowJetsonRunnerQueue "$ALLOW_JETSON_RUNNER_QUEUE" \
             --argjson includeStagingBrevLaunchable "$INCLUDE_STAGING_BREV_LAUNCHABLE" \
             --argjson workflowRunAttempt "$RUN_ATTEMPT" \
@@ -351,6 +353,7 @@ jobs:
               jobs: $jobs,
               targets: $targets,
               allowDgxSparkRunnerQueue: $allowDgxSparkRunnerQueue,
+              allowJetsonDispatch: $allowJetsonDispatch,
               allowJetsonRunnerQueue: $allowJetsonRunnerQueue,
               includeStagingBrevLaunchable: $includeStagingBrevLaunchable,
               emptySelectors: ($jobs == "" and $targets == "")
@@ -4918,70 +4921,38 @@ jobs:
 
   jetson-nvmap-gpu:
     needs: generate-matrix
-    # The opt-in flag prevents assignment to an unavailable Jetson runner.
-    if: ${{ inputs.allow_jetson_runner_queue && github.repository == 'NVIDIA/NemoClaw' && github.ref == 'refs/heads/main' && ((github.event_name != 'workflow_dispatch' || (inputs.jobs == '' && inputs.targets == '')) || contains(format(',{0},', inputs.jobs), ',jetson-nvmap-gpu,') || contains(format(',{0},', inputs.targets), ',jetson-nvmap-gpu,')) }}
-    runs-on: ${{ vars.JETSON_E2E_RUNNER_LABEL || 'linux-arm64-gpu-jetson-orin-latest-1' }}
+    if: ${{ inputs.allow_jetson_dispatch && github.repository == 'NVIDIA/NemoClaw' && github.ref == 'refs/heads/main' && github.event_name == 'workflow_dispatch' && (inputs.checkout_repository == '' || inputs.checkout_repository == github.repository) && ((inputs.jobs == '' && inputs.targets == '') || contains(format(',{0},', inputs.jobs), ',jetson-nvmap-gpu,') || contains(format(',{0},', inputs.targets), ',jetson-nvmap-gpu,')) }}
+    runs-on: ubuntu-latest
     timeout-minutes: 60
-    env:
-      E2E_JOB: "1"
-      E2E_TARGET_ID: "jetson-nvmap-gpu"
-      E2E_ARTIFACT_DIR: ${{ github.workspace }}/e2e-artifacts/live/jetson-nvmap-gpu
-      NEMOCLAW_CLI_BIN: ${{ github.workspace }}/bin/nemoclaw.js
-      NEMOCLAW_RUN_LIVE_E2E: "1"
-      NEMOCLAW_NON_INTERACTIVE: "1"
-      NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE: "1"
-      NEMOCLAW_SANDBOX_NAME: "e2e-jetson-nvmap"
-      NEMOCLAW_RECREATE_SANDBOX: "1"
-      NEMOCLAW_PROVIDER: "ollama"
-      OPENSHELL_GATEWAY: "nemoclaw"
+    permissions:
+      contents: read
+      id-token: write
     steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - name: Check out trusted Jetson controller
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
-          repository: ${{ inputs.checkout_repository || github.repository }}
-          ref: ${{ inputs.checkout_sha || github.sha }}
+          repository: NVIDIA/NemoClaw
+          ref: ${{ github.workflow_sha }}
           persist-credentials: false
 
-      - *dockerhub-auth
-
-      - name: Prepare E2E workspace
-        uses: NVIDIA/NemoClaw/.github/actions/prepare-e2e@f6304bc25fc35bfaa441c8c2fbfee38f72805a75
-        with:
-          build-cli: "false"
-
-      - name: Restore exact-commit CLI artifact
-        uses: NVIDIA/NemoClaw/.github/actions/restore-e2e-cli-artifact@c246409193a31133cab10c8a3589001cc0d59eb3
+      - name: Set up Node for Jetson controller
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
-          provenance-json: ${{ needs.generate-matrix.outputs.cli_artifact_provenance }}
-
-      - name: Verify Jetson GPU availability
-        run: |
-          set -euo pipefail
-          echo "=== Tegra release ==="
-          cat /etc/nv_tegra_release 2>/dev/null || echo "(no /etc/nv_tegra_release)"
-          echo ""
-          echo "=== /dev/nvmap ==="
-          ls -l /dev/nvmap 2>/dev/null || echo "(no /dev/nvmap)"
-          echo ""
-          echo "=== Docker ==="
-          docker info --format '{{.ServerVersion}}'
-          docker info --format '{{json .Runtimes}}'
+          node-version: 22
 
-      - name: Run Jetson nvmap GPU live Vitest test
-        # Keeps the
-        # Jetson/Tegra host, NVIDIA Docker runtime, install.sh, OpenShell
-        # sandbox exec, /dev/nvmap, CUDA cuInit(0), and status-proof boundary.
-        run: |
-          set -euo pipefail
-          npx tsx tools/e2e/live-vitest-invocation.mts run --test-path test/e2e/live/jetson-nvmap-gpu.test.ts
+      - name: Dispatch exact commit to Jetson through Colossus
+        env:
+          E2E_ARTIFACT_DIR: ${{ runner.temp }}/e2e-artifacts/live/jetson-nvmap-gpu
+          JETSON_DISPATCH_CANDIDATE_SHA: ${{ inputs.checkout_sha || github.sha }}
+          JETSON_DISPATCH_URL: ${{ vars.JETSON_DISPATCH_URL }}
+        run: node --experimental-strip-types --no-warnings tools/e2e/jetson-dispatch-client.mts
 
       - name: Upload Jetson nvmap GPU artifacts
         if: always()
         uses: NVIDIA/NemoClaw/.github/actions/upload-e2e-artifacts@7768e15eb90d3ee2d33432f481dfe8747e4f6d57
-
-      - name: Clean up Docker auth
-        if: always()
-        shell: bash
-        run: bash .github/scripts/docker-auth-cleanup.sh
+        with:
+          name: e2e-jetson-nvmap-gpu
+          path: ${{ runner.temp }}/e2e-artifacts/live/jetson-nvmap-gpu/
 
   concurrent-gateway-ports:
     needs: generate-matrix