capx policy plugin: code-as-policy agents via CaP-X perception servers

.github/workflows/ci.yml

@@ -301,12 +301,48 @@
           uv run --no-sync python -m pytest plugins/inspect-robots-agent/tests -q
           --cov=inspect_robots_agent --cov-report=term-missing --cov-fail-under=0
 
+  # The CaP-X code-as-policy suite needs no GPU, sockets, or CaP-X checkout:
+  # httpx.MockTransport scripts the LLM plus all three model-server protocols,
+  # and an in-test joint-space embodiment exercises the complete rollout.
+  plugin-capx:
+    name: plugin · capx
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+          cache-dependency-glob: |
+            **/pyproject.toml
+            **/uv.lock
+      - name: Sync workspace (core + plugins, dev extras)
+        run: uv sync --locked --python 3.11 --all-packages --extra dev
+      - name: Ruff (plugin, lint)
+        run: uv run --no-sync ruff check plugins/inspect-robots-capx
+      - name: Ruff (plugin, format check)
+        run: uv run --no-sync ruff format --check plugins/inspect-robots-capx
+      - name: Mypy (plugin, strict)
+        run: >
+          uv run --no-sync mypy
+          --config-file plugins/inspect-robots-capx/pyproject.toml
+          plugins/inspect-robots-capx/src/inspect_robots_capx
+          plugins/inspect-robots-capx/tests
+      - name: Pytest (plugin, with coverage)
+        # Report-only, like the other plugin jobs. The root 100% gate remains
+        # scoped to the numpy-only inspect_robots core package.
+        run: >
+          uv run --no-sync python -m pytest plugins/inspect-robots-capx/tests -q
+          --cov=inspect_robots_capx --cov-report=term-missing --cov-fail-under=0
+
   # Single required status check for branch protection: green iff every job
   # above is green. Add new jobs to 'needs' here or they won't gate merges.
   ci-ok:
     name: ci-ok
     if: ${{ always() }}
-    needs: [quality, test, test-rerun, test-extra, core-only-import, plugin-isaacsim, plugin-xpolicylab, plugin-ros, plugin-agent]
+    needs: [quality, test, test-rerun, test-extra, core-only-import, plugin-isaacsim, plugin-xpolicylab, plugin-ros, plugin-agent, plugin-capx]
     runs-on: ubuntu-latest
     steps:
       - name: Fail unless every needed job succeeded
@@ -319,7 +355,7 @@
   alert-red-main:
     name: alert on red main
     if: ${{ failure() && github.event_name == 'push' }}
-    needs: [quality, test, test-rerun, test-extra, core-only-import, plugin-isaacsim, plugin-xpolicylab, plugin-ros, plugin-agent]
+    needs: [quality, test, test-rerun, test-extra, core-only-import, plugin-isaacsim, plugin-xpolicylab, plugin-ros, plugin-agent, plugin-capx]
     runs-on: ubuntu-latest
     permissions:
       issues: write

.github/workflows/release.yml

@@ -175,3 +175,26 @@ jobs:
           # The plugin is versioned independently; a core release that doesn't
           # bump it must not fail on the already-uploaded file.
           skip-existing: true
+
+  publish-capx:
+    name: publish inspect-robots-capx
+    needs: [cut]
+    if: ${{ !cancelled() && (github.event_name == 'release' || needs.cut.result == 'success') }}
+    runs-on: ubuntu-latest
+    environment: pypi-capx
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ needs.cut.outputs.tag || github.ref }}
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+      - name: Build sdist + wheel
+        run: uv build --package inspect-robots-capx
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1
+        with:
+          # The plugin is versioned independently; a core release that doesn't
+          # bump it must not fail on the already-uploaded file.
+          skip-existing: true

CLAUDE.md

@@ -35,7 +35,9 @@ rollout, scores it, and writes an immutable `EvalLog`. Mirrors Inspect AI's
   (policy adapter speaking the XPolicyLab websocket protocol — 40+ served VLAs,
   no xpolicylab dep), and
   `plugins/inspect-robots-agent/` (LLMs as policies via the OpenAI-compatible
-  wire format — httpx only, no provider SDKs; registered as `agent`).
+  wire format — httpx only, no provider SDKs; registered as `agent`), and
+  `plugins/inspect-robots-capx/` (CaP-X code-as-policy over perception and IK
+  HTTP servers; registered as `capx`).
 
 ## Working here
 

README.md

@@ -154,6 +154,27 @@ Read the recorded agent conversation with
 `inspect-robots view LOG.json`, including the camera frames the model saw (for
 `--store-frames` runs).
 
+### Generate robot policy code with CaP-X:
+
+The [inspect-robots-capx](plugins/inspect-robots-capx/) plugin evaluates a
+code-as-policy agent in the same policy slot. The LLM writes Python against
+SAM3 segmentation, Contact-GraspNet planning, Pyroki IK, and speed-limited
+joint-motion helpers. CaP-X model servers run separately, while the persistent
+code namespace and action queue run inside the evaluator.
+
+```bash
+uv pip install inspect-robots-capx
+
+inspect-robots "place the fork on the plate" --policy capx \
+    --embodiment <joint-space-embodiment> \
+    -P model=anthropic/claude-fable-5 -P sam3_url=http://gpu-box:8114
+```
+
+The v1 adapter requires one `joint_pos` arm with a declared gripper, full
+joint-state proprioception, a control rate, and a camera. See the
+[plugin README](plugins/inspect-robots-capx/) for model-server bringup, depth
+metadata, gripper polarity, and the model-code trust boundary.
+
 ### Run in simulation
 
 The same instruction runs on your configured simulator instead of the
@@ -264,6 +285,10 @@ adapter shipped from this repo as separate packages:
   embodiment through tool calls, as a first-class policy. The same
   `--policy agent` runs ad-hoc instructions and scores on registered tasks
   next to fine-tuned VLAs.
+- **[inspect-robots-capx](plugins/inspect-robots-capx/)**: evaluate CaP-X-style
+  code-as-policy agents against a joint-space embodiment. Model-generated
+  Python calls separately served SAM3, Contact-GraspNet, and Pyroki helpers,
+  then queues approver-checked joint targets behind `--policy capx`.
 
 ```bash
 # Isaac Lab world + a π0 checkpoint served by XPolicyLab, evaluated end to end: