MarkusNeusinger · Nov 1, 2026 · Nov 1, 2026 · Nov 1, 2026 · Nov 1, 2026 · Nov 1, 2026
diff --git a/.github/workflows/ci-tests.yml b/.github/workflows/ci-tests.yml
@@ -97,8 +97,10 @@ jobs:
         # tests/unit/test_agent_instructions.py and tests/unit/skills read CLAUDE.md,
         # the Copilot instructions, .claude/guardrails.md, agentic/docs/project-guide.md,
         # the skills and the commands (#12493 changed only those, skipped this job, and
-        # left main red).
-        CONFIG_CHANGES=$(echo "$CHANGED_FILES" | grep -E '^(pyproject\.toml|pytest\.ini|conftest\.py|uv\.lock|package\.json|package-lock\.json|Project\.toml|Manifest\.toml|\.python-version|CLAUDE\.md|AGENTS\.md|\.github/copilot-instructions\.md|\.claude/guardrails\.md|agentic/docs/project-guide\.md)$|^(\.github/workflows/|\.github/actions/|prompts/|\.claude/skills/|agentic/commands/)' || true)
+        # left main red), and tests/unit/test_cloudbuild_configs.py and
+        # tests/unit/agents/test_image.py read the Cloud Build configs and the renderer's
+        # deploy smoke script.
+        CONFIG_CHANGES=$(echo "$CHANGED_FILES" | grep -E '^(pyproject\.toml|pytest\.ini|conftest\.py|uv\.lock|package\.json|package-lock\.json|Project\.toml|Manifest\.toml|\.python-version|CLAUDE\.md|AGENTS\.md|\.github/copilot-instructions\.md|\.claude/guardrails\.md|agentic/docs/project-guide\.md|([^/]+/)+cloudbuild\.yaml|agents/renderer/deploy_smoke\.sh)$|^(\.github/workflows/|\.github/actions/|prompts/|\.claude/skills/|agentic/commands/)' || true)
 
         # Force run on workflow_dispatch with force_run=true
         if [[ "${{ github.event_name }}" == "workflow_dispatch" && "${{ inputs.force_run }}" == "true" ]]; then

diff --git a/agents/README.md b/agents/README.md
@@ -335,7 +335,7 @@ The image's `fonts` stage installs Debian's `fonts-noto-core` and `fonts-noto-cj
 
 Since 2026-10-30 the toolchain image also carries the JavaScript row's toolchain (JS-1 of step 3 of `agentic/specs/261024-agent-all-libraries.md`; the agents service sends JavaScript jobs since JS-4a, 2026-10-30): Node 22 under `/opt/node`, which the row runs by its path and the sandbox's `PATH` never carries; the repository's `package-lock.json` installed with `npm ci --ignore-scripts` under `/opt/anyplot/node_modules`; Playwright's Chromium headless shell alone under `/opt/pw-browsers`; the catalogue's harness, `automation/js-render/render.mjs`, under `/opt/anyplot/js-render/` with the import allowlist of MUI X snippets that the build derives from the installed packages' exports maps (`import-allowlist.json`); and Liberation Sans, Serif and Mono under `/opt/anyplot/js-fonts`, which only the row's `FONTCONFIG_FILE` (`renderer/toolchain/js-fonts.conf`) names, so Python's renders keep their fonts. Its build renders a Chart.js chart and a MUI X component through the harness with the row's environment, the CI job renders a catalogue file of each, and the deploy's candidate smoke renders a Chart.js chart from a typed table and, since JS-4a, refuses the promotion when it fails twice. Since JS-2 (2026-10-30) the row's environment switches on the harness's sandbox mode (`ANYPLOT_SANDBOX=1`, `ANYPLOT_NO_HTML=1`, `ANYPLOT_IMPORT_ALLOWLIST`): every request and WebSocket of the page refused, no popup, no WebRTC, a content security policy, UTC, a 1 GiB V8 heap for the page, MUI X's imports closed to the allowlist, a bounded console and one final stderr line `render: error: <kind> <Class> <file>:<line>: <message>`; a JavaScript job may carry the table as typed JSON (`data_json`, from `anyplot/data/typed_json.py`), which the harness injects as `window.ANYPLOT_DATA`. The CI job checks the injection, the route and the import plugin in the built image. Since JS-4b (2026-10-31) `POST /page` runs the same row with `ANYPLOT_PAGE_ONLY=1`: the harness writes the page a render would screenshot and starts no browser, and the agents service wraps it into the downloadable `plot.html`. The CI job builds such a page for Chart.js and, since JS-4c (2026-10-31), for MUI X, whose page is a React bundle, wraps each and opens it from `file://` in the image's Chromium with `renderer/page_probe.mjs`: the chart draws and answers a hover, the frame's origin is `null`, and the plot code's attempts to navigate the window or its own frame, open a popup, fetch, or read storage and cookies all fail. Since JS-4b-2 (2026-10-31) the chat shows that page in the result card (`.interactive()`) through the API's relay page, `GET /preview/relay` (`api/preview_relay.py`): a static page of opaque origin that only the site's origins may frame, that tells its parent it is ready with a constant signal without content, and that writes the first page its parent then posts into a sandboxed `srcdoc` frame. The CI job's next step runs the probe's relay mode on the same wrapper: the relay with its real headers on an origin of its own, framed and scaled by a parent at an allowed origin that posts only on the relay's signal, shows the plot, which draws and answers a hover; the plot code cannot navigate the wrapper or the relay, a second page and a page posted by the plot, a sibling frame or another origin are ignored, the plot's own ready signal starts no post, another origin cannot frame the relay, and a listening 404 page at the relay's URL receives nothing.
 
-`renderer/cloudbuild.yaml` builds, deploys a `candidate` revision without traffic, smokes it (IAM refuses an anonymous caller; with `_SMOKE_RENDER=true`, `/status` reports the launcher and the volume and a real matplotlib render goes through the candidate, with a token minted for the service URL), and promotes it. The first build also creates the service; the deploy adds the service URL to `RENDERER_AUDIENCES` itself, and `_MIN_INSTANCES=1` keeps an instance warm. CI builds the same image before merge and renders a seaborn, a plotnine, an altair, a pygal and a lets-plot plot through the harness in it (`.github/workflows/ci-image.yml`). The deploy flags and the owner tasks are in [Agent network design](../docs/concepts/agent-network.md#serving-and-infrastructure).
+`renderer/cloudbuild.yaml` builds, deploys a `candidate` revision without traffic, smokes it (IAM refuses an anonymous caller; with `_SMOKE_RENDER=true`, `/status` reports the launcher and the volume, and a matplotlib, a Makie, a Chart.js and a plotly render and a data run go through the candidate with a token minted for the service URL, each one a gate; these render probes live in `renderer/deploy_smoke.sh`, because Cloud Build refuses a step argument over 10,000 characters), and promotes it. The first build also creates the service; the deploy adds the service URL to `RENDERER_AUDIENCES` itself, and `_MIN_INSTANCES=1` keeps an instance warm. CI builds the same image before merge and renders a seaborn, a plotnine, an altair, a pygal and a lets-plot plot through the harness in it (`.github/workflows/ci-image.yml`). The deploy flags and the owner tasks are in [Agent network design](../docs/concepts/agent-network.md#serving-and-infrastructure).
 
 ### Test
 

diff --git a/agents/renderer/cloudbuild.yaml b/agents/renderer/cloudbuild.yaml
@@ -272,133 +272,11 @@ steps:
           curl -fsS --retry 5 --retry-delay 5 --retry-all-errors -H "Authorization: Bearer $$TOKEN" -o status.json "$$URL/status" \
             || { echo "/status failed with this build's token: check _ALLOWED_CALLERS, roles/run.invoker and RENDERER_AUDIENCES ($$SERVICE_URL)"; exit 1; }
           python3 -c "import json; s=json.load(open('status.json')); assert s['name']=='anyplot-renderer' and s['sandbox'] and s['volume'] is True, s; print('OK: /status', s['version'], s['mem_available_mb'], 'MiB available, volume in place')"
-          python3 - <<'EOF'
-        import json
-        source = "\n".join([
-            "import matplotlib.pyplot as plt",
-            "fig, ax = plt.subplots(figsize=(16, 9), dpi=200)",
-            "ax.plot([1, 2, 3, 4], [3, 1, 4, 2], linewidth=6)",
-            "fig.savefig('plot-light.png')",
-        ])
-        job = {"job_id": "smoke", "language": "python", "library": "matplotlib", "source": source,
-               "data_csv": "x\n1\n", "themes": ["light"], "timeout_s": 60}
-        json.dump(job, open("job.json", "w"))
-        EOF
-          curl -fsS --retry 3 --retry-delay 10 --retry-all-errors --max-time 200 -H "Authorization: Bearer $$TOKEN" \
-            -H "Content-Type: application/json" --data @job.json -o render.json "$$URL/render" \
-            || { echo "the smoke render request failed"; exit 1; }
-          python3 -c "import json; r=json.load(open('render.json')); o=r['outputs']['light']; assert o['exit_code']==0 and o['png_base64'], {k: o[k] for k in ('exit_code','reason','attempts','wall_s')}; print('OK: smoke render', o['wall_s'], 's,', o['attempts'], 'attempt(s)')"
-          # 2b. A Makie render through the Julia row (step 5 of
-          #     agentic/specs/261024-agent-all-libraries.md): the depot's package images
-          #     load under gVisor with --compiled-modules=strict and the preloaded exec
-          #     filter, or the render fails within seconds and nothing is promoted. The
-          #     first Julia render of a fresh instance, so its wall time is the cold one.
-          python3 - <<'EOF'
-        import json
-        source = "\n".join([
-            "using CairoMakie",
-            "THEME = get(ENV, \"ANYPLOT_THEME\", \"light\")",
-            "fig = Figure(size = (1600, 900))",
-            "ax = Axis(fig[1, 1], title = \"Größe – smoke\")",
-            "lines!(ax, [1, 2, 3, 4], [3, 1, 4, 2], linewidth = 6)",
-            "save(\"plot-$$(THEME).png\", fig; px_per_unit = 2)",
-        ])
-        job = {"job_id": "smoke-makie", "language": "julia", "library": "makie", "source": source,
-               "data_csv": "x\n1\n", "themes": ["light"], "timeout_s": 90}
-        json.dump(job, open("makie_job.json", "w"))
-        EOF
-          curl -fsS --retry 3 --retry-delay 10 --retry-all-errors --max-time 240 -H "Authorization: Bearer $$TOKEN" \
-            -H "Content-Type: application/json" --data @makie_job.json -o makie.json "$$URL/render" \
-            || { echo "the Makie smoke render request failed"; exit 1; }
-          python3 -c "import json; r=json.load(open('makie.json')); o=r['outputs']['light']; assert o['exit_code']==0 and o['png_base64'], {k: o[k] for k in ('exit_code','reason','attempts','wall_s','stderr_tail')}; print('OK: Makie smoke render', o['wall_s'], 's wall,', o['cpu_s'], 's CPU,', o['max_rss_mb'], 'MiB,', o['attempts'], 'attempt(s)')"
-          # 2c. A Chart.js render through the JavaScript row (JS-1 of step 3): Node by its path
-          #     starts Chromium's headless shell under gVisor, which no local run can, and the
-          #     page reads the job's typed table (`data_json`, `window.ANYPLOT_DATA`), as every
-          #     agent job does. A gate like the Makie render since JS-4a served JavaScript: a
-          #     failure refuses the promotion, and the revision serving now keeps serving. curl's
-          #     retries cover the transport only, and a render that fails answers 200, so the
-          #     render gets a second try 30 s later; each failure prints the exit code, the
-          #     reason, the attempts and the stderr tail (JS renders failed on Cloud Run for about
-          #     40 minutes after a fresh toolchain build on 2026-10-30, cause unproven).
-          python3 - <<'EOF'
-        import json
-        source = "\n".join([
-            "const t = window.ANYPLOT_TOKENS;",
-            "const data = window.ANYPLOT_DATA;",
-            "const canvas = document.createElement('canvas');",
-            "document.getElementById('container').appendChild(canvas);",
-            "new Chart(canvas, {",
-            "  type: 'bar',",
-            "  data: {labels: data.map((d) => d.Gruppe), datasets: [{data: data.map((d) => d.Wert), backgroundColor: t.palette[0]}]},",
-            "  options: {responsive: true, maintainAspectRatio: false, animation: false,",
-            "    plugins: {title: {display: true, text: 'Größe – smoke', color: t.ink}}},",
-            "});",
-        ])
-        table = {"columns": ["Gruppe", "Wert"], "data": [["a", 3], ["b", 1], ["c", 2]]}
-        job = {"job_id": "smoke-chartjs", "language": "javascript", "library": "chartjs", "source": source,
-               "data_csv": "Gruppe,Wert\na,3\nb,1\nc,2\n", "data_json": json.dumps(table),
-               "themes": ["light"], "timeout_s": 60}
-        json.dump(job, open("chartjs_job.json", "w"))
-        EOF
-          chartjs_ok=""
-          for try in 1 2; do
-            if curl -fsS --retry 3 --retry-delay 10 --retry-all-errors --max-time 200 -H "Authorization: Bearer $$TOKEN" \
-              -H "Content-Type: application/json" --data @chartjs_job.json -o chartjs.json "$$URL/render" \
-              && python3 -c "import json; r=json.load(open('chartjs.json')); o=r['outputs']['light']; assert o['exit_code']==0 and o['png_base64'], {k: o[k] for k in ('exit_code','reason','attempts','wall_s','stderr_tail')}; print('OK: Chart.js smoke render', o['wall_s'], 's wall,', o['cpu_s'], 's CPU,', o['max_rss_mb'], 'MiB,', o['attempts'], 'attempt(s)')"; then
-              chartjs_ok=1
-              break
-            fi
-            echo "the Chart.js smoke render failed (try $$try of 2)"
-            if [ "$$try" = 1 ]; then sleep 30; fi
-          done
-          [ -n "$$chartjs_ok" ] || { echo "the Chart.js smoke render failed twice"; exit 1; }
-          # 2d. A plotly render through the two-step figure mode (PB-1, a gate since PB-2a served
-          #     plotly and bokeh): the harness forks the plot under its guard and the exec filter,
-          #     the plot reads data.csv and writes its figure as JSON through the pinned save (an
-          #     axis range included, which runs plotly's `build_validator` under the guard), and
-          #     the JavaScript row's harness draws it in Chromium under gVisor, which no local run
-          #     can. The same two tries as the Chart.js render; a failure refuses the promotion.
-          python3 - <<'EOF'
-        import json
-        source = "\n".join([
-            "import os",
-            "",
-            "import pandas as pd",
-            "import plotly.graph_objects as go",
-            "",
-            "THEME = os.getenv('ANYPLOT_THEME', 'light')",
-            "df = pd.read_csv('data.csv')",
-            "fig = go.Figure(go.Bar(x=df['Gruppe'], y=df['Wert'], marker_color='#009E73'))",
-            "fig.update_layout(title={'text': 'Größe – smoke'}, yaxis={'range': [0, 3.5]})",
-            "fig.update_layout(width=800, height=450).write_json(f'plot-{THEME}.json')",
-        ])
-        job = {"job_id": "smoke-plotly", "language": "python", "library": "plotly", "source": source,
-               "data_csv": "Gruppe,Wert\na,3\nb,1\nc,2\n", "themes": ["light"], "timeout_s": 60}
-        json.dump(job, open("plotly_job.json", "w"))
-        EOF
-          plotly_ok=""
-          for try in 1 2; do
-            if curl -fsS --retry 3 --retry-delay 10 --retry-all-errors --max-time 200 -H "Authorization: Bearer $$TOKEN" \
-              -H "Content-Type: application/json" --data @plotly_job.json -o plotly.json "$$URL/render" \
-              && python3 -c "import json; r=json.load(open('plotly.json')); o=r['outputs']['light']; assert o['exit_code']==0 and o['png_base64'], {k: o[k] for k in ('exit_code','reason','attempts','wall_s','stderr_tail')}; print('OK: plotly figure smoke render', o['wall_s'], 's wall,', o['cpu_s'], 's CPU,', o['max_rss_mb'], 'MiB,', o['attempts'], 'attempt(s)')"; then
-              plotly_ok=1
-              break
-            fi
-            echo "the plotly figure smoke render failed (try $$try of 2)"
-            if [ "$$try" = 1 ]; then sleep 30; fi
-          done
-          [ -n "$$plotly_ok" ] || { echo "the plotly figure smoke render failed twice"; exit 1; }
-          # 3. A real data run (POST /run-data, run_data_code): the harness in
-          #    data mode with its audit-hook guard armed returns a 2-row table.
-          python3 - <<'EOF'
-        import json
-        source = "import pandas as pd\ndf = pd.DataFrame({'x': [1, 2], 'y': [3, 4]})\n"
-        json.dump({"job_id": "smoke-data", "source": source, "timeout_s": 30}, open("data_job.json", "w"))
-        EOF
-          curl -fsS --retry 3 --retry-delay 10 --retry-all-errors --max-time 120 -H "Authorization: Bearer $$TOKEN" \
-            -H "Content-Type: application/json" --data @data_job.json -o data.json "$$URL/run-data" \
-            || { echo "the smoke data run request failed"; exit 1; }
-          python3 -c "import json; r=json.load(open('data.json')); assert r['exit_code']==0 and (r['csv_text'] or '').splitlines()==['x,y','1,3','2,4'], {k: r[k] for k in ('exit_code','reason','attempts','wall_s')}; print('OK: smoke data run', r['wall_s'], 's,', r['attempts'], 'attempt(s)')"
+          # The render probes (matplotlib, Makie, Chart.js, plotly and a data run, each a gate)
+          # live in their own script: Cloud Build refuses a step argument over 10,000
+          # characters, and this step passed it when the plotly probe joined (build dc1cf3ae,
+          # 2026-11-01). The token goes through the environment, not the command line.
+          SMOKE_TOKEN="$$TOKEN" bash agents/renderer/deploy_smoke.sh "$$URL"
         else
           echo "render smoke skipped (_SMOKE_RENDER=${_SMOKE_RENDER})"
         fi