harden agent builder render guidance

agent_builder/builder.py

@@ -61,6 +61,35 @@ runtime:
 Names must be plain Debian package slugs (lowercase, ``[a-z0-9.+-]``).
 Don't reach for this for Python deps — those go in ``requirements.txt``.
 
+If the agent needs more than the tiny default pod (long subprocesses,
+rendering, browser work, data processing, media conversion, Manim, etc.),
+declare resources in BOTH places:
+
+  - Python Card/runtime: ``resources = Resources(cpu="2", memory="2Gi",
+    max_runtime_seconds=900)``
+  - ``a2a.yaml`` for the Gitea/Argo scaffold, which cannot import user code
+    before the image is built:
+
+```yaml
+runtime:
+  resources:
+    cpu: "2"
+    memory: 2Gi
+    max_runtime_seconds: 900
+```
+
+For render/media agents, subprocesses must 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.
+
+For Manim presentation agents specifically, generate a stable PDF fallback
+that does not require Manim rendering. If HTML/video rendering is requested,
+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.
+
 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).
@@ -113,6 +142,10 @@ Discipline:
     scaffold.
   - When deploying, return the URL the platform gave back to the user so
     they can curl it / share it.
+  - After cp_deploy_tarball returns, compare ``live_skills[].input_schema``
+    with the skill signatures you wrote. If an argument is missing from the
+    live schema, treat the deploy as failed, fix the source/schema version,
+    redeploy, and refresh before declaring success.
   - Don't fabricate functionality the user didn't ask for. One or two
     well-scoped @skill methods beats a kitchen sink.
 """

tests/test_builder_prompt.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+import unittest
+
+from agent_builder.builder import SYSTEM_PROMPT
+
+
+class BuilderPromptTests(unittest.TestCase):
+    def test_prompt_requires_resource_mirror_for_heavy_agents(self) -> None:
+        self.assertIn("declare resources in BOTH places", SYSTEM_PROMPT)
+        self.assertIn("resources = Resources", SYSTEM_PROMPT)
+        self.assertIn("runtime:", SYSTEM_PROMPT)
+        self.assertIn("max_runtime_seconds", SYSTEM_PROMPT)
+
+    def test_prompt_requires_live_schema_check_and_manim_contract(self) -> None:
+        self.assertIn("live_skills[].input_schema", SYSTEM_PROMPT)
+        self.assertIn("class GeneratedDeck(Slide)", SYSTEM_PROMPT)
+        self.assertIn("manim-slides render --CE -ql", SYSTEM_PROMPT)
+        self.assertIn("render: bool = False", SYSTEM_PROMPT)