builder: require durable workspace execution

agent_builder/builder.py

@@ -93,13 +93,18 @@ runtime:
     max_runtime_seconds: 900
 ```
 
-For render/media agents, subprocesses must be bounded with timeouts and broad
+For render/media/data agents, commands that create user-visible files MUST run
+through ``await ctx.workspace_shell(...)`` or ``await ctx.workspace_python(...)``
+and write under ``/workspace/outputs/...``. Do not use in-process
+``asyncio.create_subprocess_exec`` for durable outputs; the agent container is
+not workspace-mounted, so files created in ``/tmp`` or the image filesystem can
+vanish after the request. Commands must still be bounded with timeouts and broad
 exception handling, and failures must return structured JSON instead of
 crashing the worker. Use ``render: bool = False`` by default and expose every
 public skill argument with concrete type hints so the live ``input_schema``
 contains the full call surface. If a render command returns nonzero after
-emitting a usable file, persist or wrap that artifact and include the command
-failure as a warning instead of dropping the result.
+emitting a usable file, preserve the file under ``/workspace/outputs/...`` and
+include the command failure as a warning instead of dropping the result.
 
 For Manim presentation agents specifically, generate a stable PDF fallback
 that does not require Manim rendering. If HTML/video rendering is requested,
@@ -107,6 +112,12 @@ generate ``class GeneratedDeck(Slide)`` and call
 ``manim-slides render --CE -ql <source.py> GeneratedDeck``; do not assume raw
 ``manim`` is enough for a Manim Slides deck.
 
+Generated agents that invoke an inner DeepAgents graph MUST pass
+``config={"recursion_limit": 500}`` to ``graph.ainvoke(...)`` or
+``graph.astream_events(...)``. This matches agent-builder's own runtime budget
+and prevents complex skill/subagent workflows from failing at the default
+LangGraph recursion cap.
+
 And a ``requirements.txt`` listing any extra deps beyond a2a-pack
 itself (pandas, httpx, etc. — the base image ships a2a-pack already
 when you deploy through the control plane).
@@ -177,6 +188,9 @@ Discipline:
     Do not replace LLM reasoning with fake deterministic tools that return
     canned answers. Tools should do exact work; skills and subagents should
     carry reusable workflow knowledge and judgement-heavy behavior.
+  - When invoking the generated DeepAgents graph, use
+    ``config={"recursion_limit": 500}`` so child agents have the same graph
+    budget as agent-builder.
   - If generated agent.py calls ``ctx.workspace_backend()``, declare
     ``workspace_access = WorkspaceAccess.dynamic(...)`` on the A2A agent class
     so local/dev/platform invocations actually receive a workspace grant.