chore(ci): add dependency audit and Dependabot

.github/dependabot.yml

@@ -0,0 +1,18 @@
+version: 2
+updates:
+  - package-ecosystem: 'uv'
+    # uv workspace root: one uv.lock covers every packages/* member.
+    directory: '/'
+    schedule:
+      interval: 'daily'
+    cooldown:
+      default-days: 7
+  - package-ecosystem: 'github-actions'
+    # '/' covers .github/workflows only; the composite actions pin their own actions.
+    directories:
+      - '/'
+      - '/.github/actions/*'
+    schedule:
+      interval: 'weekly'
+    cooldown:
+      default-days: 7

.github/workflows/ci.yml

@@ -4,6 +4,9 @@ on:
   push:
     branches: [main]
   pull_request:
+  # Weekly, so an advisory published against an unchanged uv.lock still gets reported.
+  schedule:
+    - cron: "0 13 * * 1"
 
 permissions:
   contents: read
@@ -23,6 +26,20 @@ jobs:
       - run: uv run ruff check .
       - run: uv run ruff format --check .
 
+  # ─── Dependency audit ───────────────────────────────────────────────────────
+  audit:
+    name: Dependency audit
+    runs-on: ubuntu-latest
+    # A new advisory must not block unrelated PRs; push and the weekly schedule still fail on it.
+    continue-on-error: ${{ github.event_name == 'pull_request' }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v5
+        with:
+          python-version: "3.12"
+      # Audits the committed uv.lock as-is (no `uv lock` first), so a vulnerable pin fails here.
+      - run: make audit
+
   # ─── Setup ──────────────────────────────────────────────────────────────────
   setup:
     name: Install & sync

Makefile

@@ -1,4 +1,11 @@
-.PHONY: start test typecheck lint format format-check lint-fix build clean api-report api-report-check
+.PHONY: start test typecheck lint format format-check lint-fix build clean api-report api-report-check audit
+
+# Report known vulnerabilities in the locked dependency set (pip-audit on uv.lock)
+# Exports to a file, not a pipe: make's /bin/sh has no pipefail, so a failed export would audit nothing and pass
+audit:
+	reqs=$$(mktemp) && trap 'rm -f "$$reqs"' EXIT && \
+	uv export --all-packages --no-emit-workspace --frozen -q -o "$$reqs" && \
+	uvx pip-audit@2.9.0 --disable-pip --no-deps -r "$$reqs"
 
 # Run main.py (uv run python main.py)
 start:

packages/claude-agents/src/launchdarkly_ai_claude_agents/handler.py

@@ -119,10 +119,14 @@
 #:   ``strict_mcp_config``, ``agents``: what the agent is allowed to do and which tools or
 #:   subagents it gets. The handler wires tools and permissions itself.
 #: * ``resume``, ``session_id``, ``fork_session``, ``continue_conversation``, ``session_store``,
-#:   ``session_store_flush``, ``enable_file_checkpointing``: session state on the host.
+#:   ``session_store_flush``, ``enable_file_checkpointing``, ``resume_session_at``,
+#:   ``resume_drops_turn``: session state on the host.
 #: * ``stderr``, ``debug_stderr``, ``include_partial_messages``, ``include_hook_events``,
-#:   ``max_buffer_size``, ``load_timeout_ms``: process I/O and transport plumbing, including the
-#:   streamed message shape the handler reads.
+#:   ``forward_subagent_text``, ``max_buffer_size``, ``load_timeout_ms``: process I/O and transport
+#:   plumbing, including the streamed message shape the handler reads.
+#: * ``verbatim_prompts``: whether Claude Code expands ``@path`` mentions and slash commands in
+#:   the prompt, which decides whether prompt text can make it read host files. The application
+#:   decides that, not a config.
 #: * ``output_format``: structured output, which the handler owns through the config's
 #:   ``outputFormat`` by adding its schema to the system prompt. Forwarding a second schema here
 #:   would have the reply follow one schema while the prompt asks for another.
@@ -145,6 +149,7 @@
         "env",
         "extra_args",
         "fork_session",
+        "forward_subagent_text",
         "include_hook_events",
         "include_partial_messages",
         "load_timeout_ms",
@@ -154,6 +159,8 @@
         "permission_prompt_tool_name",
         "plugins",
         "resume",
+        "resume_drops_turn",
+        "resume_session_at",
         "sandbox",
         "session_id",
         "session_store",
@@ -165,6 +172,7 @@
         "strict_mcp_config",
         "task_budget",
         "user",
+        "verbatim_prompts",
     }
 )
 

packages/client/tests/test_lifecycle.py

@@ -276,18 +276,25 @@ async def test_is_idempotent(self) -> None:
     async def test_a_failed_telemetry_setup_leaves_no_client_behind(
         self, restore_otel_globals: Any
     ) -> None:
-        # A malformed OTEL_EXPORTER_OTLP_TIMEOUT makes the real _setup_telemetry
-        # raise. _client used to be assigned first, so the next call returned the
-        # half-built client as a silent success with no telemetry, and its
-        # connection was never closed.
+        # _client used to be assigned before _setup_telemetry ran, so after a
+        # telemetry failure the next call returned the half-built client as a
+        # silent success with no telemetry, and its connection was never closed.
+        # The failure is injected: OpenTelemetry 1.45 logs a malformed
+        # OTEL_EXPORTER_OTLP_TIMEOUT and uses its default instead of raising.
         failed = _make_stub_client()
         healthy = _make_stub_client()
         mock_ld = MagicMock()
         mock_ld.Config = MagicMock(return_value=MagicMock())
         mock_ld.LDClient = MagicMock(side_effect=[failed, healthy])
         with patch("importlib.import_module", return_value=mock_ld):
-            bad_env = {"LD_SDK_KEY": "k", "OTEL_EXPORTER_OTLP_TIMEOUT": "soon"}
-            with patch.dict(os.environ, bad_env):
+            with (
+                patch.dict(os.environ, {"LD_SDK_KEY": "k"}),
+                patch.object(
+                    lifecycle_module,
+                    "_setup_telemetry",
+                    side_effect=ValueError("bad config"),
+                ),
+            ):
                 with pytest.raises(ValueError):
                     await init_client()
             assert lifecycle_module._client is None