[2547 by Claude Opus 5] Bundle Agent Tank into the unified agent image and add a three-state (disabled / bundled / external) integration mode

.github/workflows/docker-images.yml

@@ -278,6 +278,15 @@ jobs:
           AGENT_TAG: ${{ env.DOCKERHUB_NS }}/agent:${{ steps.version.outputs.version }}
         run: ./scripts/verify-antigravity-image.sh
 
+      # Bundled Agent Tank mode runs the CLI inside this image, so the release
+      # has to prove it returns real usage - not just that the binary is there.
+      # Antigravity is the provider this runner is authenticated for.
+      - name: Verify authenticated bundled Agent Tank usage from packaged agent image
+        env:
+          AGENT_TAG: ${{ env.DOCKERHUB_NS }}/agent:${{ steps.version.outputs.version }}
+          AGENT_TANK_PROVIDERS: agy
+        run: ./scripts/verify-agent-tank-image.sh
+
       - name: Validate serialized SQLite startup with packaged images
         env:
           APP_TAG: ${{ env.DOCKERHUB_NS }}/app:${{ steps.version.outputs.version }}

.github/workflows/pr-build-check.yml

@@ -553,6 +553,7 @@ jobs:
               scripts/smoke-test-images.sh \
               scripts/smoke-test-preview-runtime-images.sh \
               scripts/smoke-test-sqlite-startup.sh \
+              scripts/verify-agent-tank-image.sh \
               scripts/verify-antigravity-image.sh
           } > >(tee -a build_log.txt) 2>&1
 

Dockerfile.agent

@@ -154,6 +154,21 @@ RUN case "${VIBE_CLI_VERSION}" in \
     && if [ ! -e /usr/local/bin/vibe-acp ]; then ln -s /usr/local/bin/vibe /usr/local/bin/vibe-acp; fi
 
 
+FROM agent-base AS agent-tank-cli
+# Pinned here, not in CI: this literal is part of the Dockerfile content that
+# feeds the agent bundle content hash, so bumping it produces a new image tag.
+# An env-only override in CI would NOT change the tag and would silently ship a
+# different binary under an existing tag - do not do that.
+ARG AGENT_TANK_CLI_VERSION=0.9.11
+USER root
+# node-pty compiles against the toolchain already present in agent-base
+# (build-essential + python3), which is why this needs no extra apt packages.
+RUN npm install -g "agent-tank@${AGENT_TANK_CLI_VERSION}" \
+    && npm cache clean --force \
+    && rm -rf /root/.npm \
+    && agent-tank --version
+
+
 FROM agent-base AS antigravity-cli
 ARG ANTIGRAVITY_CLI_VERSION=1.2.4
 ARG ANTIGRAVITY_CLI_RELEASE_ID=6085322963025920
@@ -196,18 +211,21 @@ ARG CODEX_CLI_VERSION=0.154.0
 ARG ANTIGRAVITY_CLI_VERSION=1.2.4
 ARG OPENCODE_CLI_VERSION=1.18.31
 ARG VIBE_CLI_VERSION=2.25.4
+ARG AGENT_TANK_CLI_VERSION=0.9.11
 
 LABEL dev.propr.agent-bundle="true" \
       dev.propr.agent.claude.version="${CLAUDE_CLI_VERSION}" \
       dev.propr.agent.codex.version="${CODEX_CLI_VERSION}" \
       dev.propr.agent.antigravity.version="${ANTIGRAVITY_CLI_VERSION}" \
       dev.propr.agent.opencode.version="${OPENCODE_CLI_VERSION}" \
-      dev.propr.agent.vibe.version="${VIBE_CLI_VERSION}"
+      dev.propr.agent.vibe.version="${VIBE_CLI_VERSION}" \
+      dev.propr.agent-tank.version="${AGENT_TANK_CLI_VERSION}"
 
 USER root
 COPY --from=claude-cli /usr/local/lib/node_modules/@anthropic-ai /usr/local/lib/node_modules/@anthropic-ai
 COPY --from=codex-cli /usr/local/lib/node_modules/@openai /usr/local/lib/node_modules/@openai
 COPY --from=opencode-cli /usr/local/lib/node_modules/opencode-ai /usr/local/lib/node_modules/opencode-ai
+COPY --from=agent-tank-cli /usr/local/lib/node_modules/agent-tank /usr/local/lib/node_modules/agent-tank
 COPY --from=vibe-cli /usr/local/bin/uv /usr/local/bin/uv
 COPY --from=vibe-cli /usr/local/bin/uvx /usr/local/bin/uvx
 COPY --from=vibe-cli /usr/local/bin/vibe /usr/local/bin/vibe
@@ -223,6 +241,7 @@ ENV UV_TOOL_DIR=/opt/uv/tools \
 
 COPY --chown=node:node \
     scripts/agent-entrypoint.sh \
+    scripts/agent-tank-runtime.mjs \
     scripts/claude-entrypoint.sh \
     scripts/codex-entrypoint.sh \
     scripts/antigravity-entrypoint.sh \
@@ -241,6 +260,7 @@ RUN set -eu; \
     link_npm_bin @anthropic-ai/claude-code claude \
     && link_npm_bin @openai/codex codex \
     && link_npm_bin opencode-ai opencode \
+    && link_npm_bin agent-tank agent-tank \
     && ln -sf /home/node/.local/bin/agy /usr/local/bin/agy \
     && chmod +x \
         /home/node/agent-entrypoint.sh \
@@ -255,7 +275,8 @@ RUN set -eu; \
     && codex --version \
     && opencode --version \
     && vibe --version \
-    && agy --version
+    && agy --version \
+    && agent-tank --version
 
 USER node
 RUN git config --global user.name "ProPR Agent Bot" \

apps/desktop/scripts/packaged-connect-journey-fixture.mjs

@@ -213,8 +213,12 @@ export const createPackagedJourneyFixture = async ({ approvalReadinessDelayMs =
         return;
       }
       if (request.method === 'GET' && record.authorization === `Bearer ${token}`) {
-        response.writeHead(200, cors);
-        response.end('{}');
+        // Optional dashboard reads can finish before the authenticated socket.
+        // Returning 200 with {} supplies invalid data that crashes the route and
+        // removes the selector before it can report REACT_CONNECTED. Let these
+        // unimplemented reads use the renderer's normal request-error handling.
+        response.writeHead(404, cors);
+        response.end('{"code":"UNIMPLEMENTED_SMOKE_ENDPOINT"}');
         return;
       }
     } catch {

apps/desktop/scripts/packaged-connect-journey.test.mjs

@@ -17,7 +17,7 @@ const encryption = {
 const confirmationPath = '/api/auth/user?desktop_account_confirmation=1';
 
 for (const delay of [0, 300]) {
-  test(`real credential service pairs and reprobes against Connect fixture (approval delay ${delay}ms)`, async () => {
+  test(`real credential service pairs, rejects unimplemented API reads, and reprobes against Connect fixture (approval delay ${delay}ms)`, async () => {
     const fixture = await createPackagedJourneyFixture({ approvalReadinessDelayMs: delay });
     const directory = await mkdtemp(join(tmpdir(), 'propr-connect-account-test-'));
     const profiles = new ProfileStore(directory, encryption);
@@ -43,6 +43,20 @@ for (const delay of [0, 300]) {
       confirmation.assertComplete();
       assert.deepEqual((await profiles.list()).profiles[0].account, PACKAGED_CONNECT_ACCOUNT);
       assert.equal((await service.probe(profile)).status, 'ready');
+      // The connected renderer reads these before its socket is necessarily ready.
+      // A successful {} is invalid dashboard data and can unmount the entire route,
+      // preventing REACT_CONNECTED even though authentication succeeded.
+      for (const path of [
+        '/api/dashboard/active?repository=all',
+        '/api/dashboard/stats?repository=all&period=7d',
+        '/api/unimplemented-connect-smoke-endpoint',
+      ]) {
+        const response = await fetch(`${fixture.endpoint}${path}`, {
+          headers: { Authorization: `Bearer ${fixture.secrets[2]}` },
+        });
+        assert.equal(response.status, 404, path);
+        assert.deepEqual(await response.json(), { code: 'UNIMPLEMENTED_SMOKE_ENDPOINT' });
+      }
       await service.dispose();
       await profiles.close();
       const reloaded = new ProfileStore(directory, encryption);

docker/Dockerfile.app.prod

@@ -92,6 +92,7 @@ COPY --from=builder /build/packages/api/package.json ./packages/api/
 # through the mounted Docker socket, using /usr/src/app as the Docker context.
 COPY Dockerfile.agent ./Dockerfile.agent
 COPY scripts/agent-entrypoint.sh ./scripts/agent-entrypoint.sh
+COPY scripts/agent-tank-runtime.mjs ./scripts/agent-tank-runtime.mjs
 COPY scripts/claude-entrypoint.sh ./scripts/claude-entrypoint.sh
 COPY scripts/codex-entrypoint.sh ./scripts/codex-entrypoint.sh
 COPY scripts/antigravity-entrypoint.sh ./scripts/antigravity-entrypoint.sh

dockerhub/agent.md

@@ -12,6 +12,11 @@ agent's credentials and task worktree. Version-specific bundle tags contain a
 complete CLI version matrix, so every agent instance can switch to the same
 image without another pull.
 
+The image also bundles the [Agent Tank](https://github.com/integry/agent-tank)
+CLI (`agent-tank`). ProPR's optional bundled usage-tracking mode runs it here
+on demand, so operators do not have to install it — or a second copy of the
+agent CLIs — on the host. It is inert unless that mode is enabled.
+
 The common Debian runtime is an internal Dockerfile stage, not a separately
 published image. Custom installation-level packages create one derivative of
 the selected bundle. The base includes `build-essential` for native extension

docs/ci-runners.md

@@ -106,7 +106,8 @@ expected contract, not a claim that the workers are configured or validated:
   `DOCKER_CONTEXT`, `DOCKER_TLS_VERIFY` and `DOCKER_CERT_PATH`. Job setup replaces
   HOME and Docker client config, so a saved HOME-based Docker context is not a
   reliable endpoint. `ci-rootless-preflight.sh` rejects default/remote/production
-  endpoints, checks the daemon reports rootless, and requires cgroup v2/systemd.
+  endpoints, checks the daemon reports rootless, requires cgroup v2/systemd, and
+  requires an init binary (the Redis helper starts `--init` containers).
   It does not prove socket ownership, host mount isolation or effective limits.
 - CI paths used as Docker bind sources must contain the same files at the same
   absolute path inside the runner and the daemon's host mount namespace. Map
@@ -323,10 +324,25 @@ older attempts matching the exact owner; it preserves newer attempts and all
 other owners. An unexpected owner fails closed. Existing callers with no
 instance, including nightly, retain one container per job and attempt.
 
+Each container runs with `--init`. The container's PID namespace reparents every
+health-check process to PID 1 once its runc parent exits, and `redis-server`
+does not reap them; one check every two seconds for the length of a shard
+therefore filled `--pids-limit` with zombies and left a container the rootless
+daemon could not kill, which failed the teardown step of a shard whose tests had
+all passed. tini as PID 1 reaps them instead.
+
+Teardown (`stop`) is the only caller that tolerates a failed removal. It runs
+after the tests have decided the job's result, and no step in the job can reap a
+zombie PID, so a container whose ownership fully verifies but which the daemon
+still refuses to remove is reported as a run warning and left for host cleanup.
+Its state file is kept, so a later teardown of the same owner retries. Ownership
+violations, and failed removals during `start`, still fail.
+
 Regression tests prove `job=shard, instance=default` and
 `job=shard-default, instance=<omitted>` coexist and either stop order preserves
 the other. They also cover foreign labels, tampered state, field-boundary
-collisions, retries and resource limits using a Docker CLI double.
+collisions, retries, resource limits, `--init` and the teardown tolerance using
+a Docker CLI double.
 
 ## Coverage, required check and partial reruns
 

docs/docs/features/observability.md

@@ -29,7 +29,7 @@ The task detail view exposes progress during execution, including streamed outpu
 
 ## Provider Capacity
 
-With the optional [Agent Tank](../operations/agent-tank.md) integration enabled, provider capacity becomes a visible signal too: the sidebar shows live usage bars per subscription provider, and each LLM log entry records the usage delta its call consumed. Turn it on from the dashboard banner ProPR shows when it detects a running instance, from **Settings → LLM Usage Tracking**, or with `propr tank on`.
+With the optional [Agent Tank](../operations/agent-tank.md) integration enabled, provider capacity becomes a visible signal too: the sidebar shows live usage bars per subscription provider, and each LLM log entry records the usage delta its call consumed. Turn it on from the dashboard banner, from **Settings → LLM Usage Tracking**, or with `propr tank bundled` — bundled mode runs Agent Tank inside the ProPR agent image, so there is nothing to install.
 
 ## Recovery
 

docs/docs/features/propr-cli.md

@@ -56,7 +56,7 @@ The full-screen wizard requires an interactive terminal. Over SSH or in shells w
 - `propr init stack [--root <dir>]` creates `data/`, `logs/`, `repos/`, writes `.env` from the bundled template, and auto-detects agent credential directories on the host (`~/.claude`, `~/.codex`, `~/.gemini`, `~/.config/opencode`, `~/.vibe`).
 - `propr check` reports the detected [GitHub auth mode](../operations/github-auth.md) (own App, relay, or demo) and flags missing or placeholder configuration before anything starts. `--verify` additionally runs an image/CLI smoke test per agent.
 - `propr start --no-tui` starts without the interactive dashboard (for scripts/CI); `--no-pull` skips image pulls; `--restart` recreates running services.
-- `propr tank [on|off] [--url <url>]` toggles [Agent Tank](../operations/agent-tank.md) LLM usage tracking on a running stack (omit the state to print the current setting).
+- `propr tank [bundled|external|off] [--url <url>]` configures [Agent Tank](../operations/agent-tank.md) LLM usage tracking on a running stack (omit the mode to print the current one). `bundled` runs Agent Tank inside the agent image with nothing to install; `external` needs `--url` pointing at a daemon you run. `on` remains a deprecated alias for `external`.
 
 ### Agent Skill