[eric] runner: the image had node but no npm, so CreateApp scaffolded an app it could never build

openswarm-runner/Dockerfile

@@ -39,6 +39,21 @@ RUN printf '{"name":"router-stage","version":"0.0.0","private":true}\n' > packag
  && test -f node_modules/9router/app/server.js \
  && test -z "$(find node_modules/9router -name '*.node' -print -quit)"
 
+# The App Builder's template dependencies, installed once here and shipped ALREADY EXTRACTED at
+# the digest path backend/apps/outputs/view_builder_templates.py already probes. Without it the
+# first CreateApp in a run pays a cold npm install against the public registry, and a run with no
+# egress just fails. NOT $BUILDPLATFORM: vite pulls in a platform-specific esbuild, so this has to
+# resolve on the arch the container will actually run on.
+FROM node:${NODE_VERSION}-bookworm-slim AS webapp-template
+WORKDIR /stage
+COPY backend/apps/outputs/webapp_template/frontend/package.json ./package.json
+RUN set -eux; \
+    npm install --no-audit --no-fund --loglevel=error --ignore-scripts; \
+    test -x node_modules/.bin/vite; \
+    digest="$(sha256sum package.json | cut -c1-12)"; \
+    mkdir -p "/out/${digest}"; \
+    mv node_modules "/out/${digest}/node_modules"
+
 # Webpack output is architecture-independent, so this runs natively on the build host rather than under emulation.
 FROM --platform=$BUILDPLATFORM node:${NODE_VERSION}-bookworm-slim AS frontend
 WORKDIR /src
@@ -104,6 +119,10 @@ RUN set -eux; \
     rm -rf /var/lib/apt/lists/*
 
 COPY --from=node /usr/local/bin/node /usr/local/bin/node
+# npm and npx too, not just node. They are shims into lib/node_modules, so copying the tree and
+# re-linking is the only way to get them; a `node` with no `npm` is what left the App Builder
+# scaffolding an app it could never install, build or serve.
+COPY --from=node /usr/local/lib/node_modules /usr/local/lib/node_modules
 COPY --from=pydeps /opt/pydeps /usr/local
 
 # Numeric owner on every /app copy, because a `chown -R /app` afterwards rewrites the whole tree into a second layer and the image pays for it twice (that cost 493MB before this line existed). Numeric, not `runner`, because the user is created further down.
@@ -119,8 +138,14 @@ COPY --from=frontend --chown=10001:10001 /src/dist /app/frontend
 COPY --from=uv --chown=10001:10001 /stage/uv /app/backend/uv-bin/uv
 COPY --from=uv --chown=10001:10001 /stage/uvx /app/backend/uv-bin/uvx
 
+# Also after backend/, and at the exact path bundled_extracted_modules() looks for.
+COPY --from=webapp-template --chown=10001:10001 /out /app/backend/apps/outputs/webapp_template_cache
+
 RUN set -eux; \
     if ls /app/backend/.env* >/dev/null 2>&1; then echo "a dotenv reached the image; fix Dockerfile.dockerignore" >&2; exit 1; fi; \
+    ln -s ../lib/node_modules/npm/bin/npm-cli.js /usr/local/bin/npm; \
+    ln -s ../lib/node_modules/npm/bin/npx-cli.js /usr/local/bin/npx; \
+    npm --version >/dev/null; \
     mkdir -p /app/python-env/bin; \
     ln -s /usr/local/bin/python3 /app/python-env/bin/python3; \
     find /app/backend -name '__pycache__' -type d -prune -exec rm -rf {} +; \
@@ -129,7 +154,8 @@ RUN set -eux; \
     ln -s /data/openswarm /app/backend/data; \
     mkdir -p /tmp/.X11-unix; \
     chmod 1777 /tmp/.X11-unix; \
-    chown runner:runner /data
+    chown runner:runner /data; \
+    printf '[user]\n\tname = OpenSwarm Cloud Run\n\temail = cloud-run@openswarm.local\n[init]\n\tdefaultBranch = main\n[safe]\n\tdirectory = *\n' > /etc/gitconfig
 
 USER runner
 WORKDIR /app
@@ -144,6 +170,8 @@ ENV HOME=/home/runner \
     OPENSWARM_PORT=8324 \
     DATA_DIR=/data/9router \
     NODE_ENV=production \
-    ELECTRON_BIN=/app/electron-runtime/electron
+    ELECTRON_BIN=/app/electron-runtime/electron \
+    OPENSWARM_RUN_WORKSPACE=/data/workspace \
+    OPENSWARM_NODE_PATH=/usr/local/bin/node
 
 ENTRYPOINT ["python3", "-m", "runner.main"]

openswarm-runner/README.md

@@ -42,6 +42,50 @@ Exit codes: `0` ok, `1` runner crash, `2` bad spec, `3` credential expired on ar
 `4` backend never came up, `5` workflow failed, `6` wall-clock cap hit,
 `7` no Electron window ever registered.
 
+## Files the run makes
+
+`/data/workspace` is the agent's working directory and **the only path whose contents survive**.
+It is seeded as `default_folder` before the backend boots, and the agent is told in its system
+prompt that files saved there come back and everything else is destroyed.
+
+After the workflow reaches a terminal state (including a timeout, so partial work still lands) the
+runner walks that directory and POSTs each file to `callback.artifacts_url`, then sends the
+terminal report. That order is load-bearing: the per-run callback token is refused once the run is
+closed, so uploading afterwards would be rejected.
+
+Caps, applied in the runner AND again at the control plane, which is the one that counts:
+
+| Limit | Value |
+| --- | --- |
+| One file | 20 MB |
+| One run, all files | 50 MB |
+| Files per run | 40 |
+
+Nothing is ever truncated. A file past a cap is not sent and instead arrives as a row in the
+report's `files[]` carrying a written reason, so the user reads "your 512 MB render could not be
+sent" rather than finding a 20 MB fragment. `.git`, `node_modules`, `__pycache__`, `.venv`,
+`.claude` and the usual caches are skipped, and symlinks are never followed.
+
+## Skills and connected apps
+
+`skills[]` in the run spec is written to `~/.claude/skills/<id>/` before boot. This is not a
+nicety: the backend registers the Skill tool only when at least one non-built-in skill exists on
+disk, so a container without them has no Skill tool at all and answers from general knowledge in
+the same confident voice it would use with the real thing.
+
+`unavailable_mcp_servers[]` is **names only**. The user's MCP credentials (Slack session cookies,
+Notion and GitHub access tokens, Google refresh tokens) never leave their machine, so the names go
+up purely so the run's system prompt can tell the agent which apps exist and are out of reach.
+`McpServerNote` forbids extra fields, so there is no shape a secret could travel in.
+
+## What the image carries for the App Builder
+
+`node`, `npm` and `npx`, plus the App Builder template's `node_modules` pre-installed at the digest
+path `bundled_extracted_modules()` probes. Without npm, `CreateApp` scaffolded an app that could
+never install, build or serve; without the baked cache, the first `CreateApp` in a run would pay a
+cold registry install. `git` also carries a system identity (`/etc/gitconfig`), so a workflow that
+commits does not die on "Author identity unknown".
+
 ## The renderer
 
 OpenSwarm's browser tier is not an HTTP client. Element serialization and every click,