Update Bub SDK and JSON result parsing

docs/bub-agent.md

@@ -1,8 +1,8 @@
 # Agent integration with Bub main
 
 Dohnuts provides one decision tool through Bub. The `agent` extra pins Bub main at
-`9bf70488e22a44eeb875a9d55d9a7d352b82e381`. Integration uses its
-[Python SDK](https://github.com/bubbuild/bub/blob/9bf70488e22a44eeb875a9d55d9a7d352b82e381/website/src/content/docs/docs/build/sdk.md)
+`6f23fee36a8aa55b71d5be015529a4b486b31d89`. Integration uses its
+[Python SDK](https://github.com/bubbuild/bub/blob/6f23fee36a8aa55b71d5be015529a4b486b31d89/website/src/content/docs/docs/build/sdk.md)
 to provide one batch decision tool and persistent session tapes.
 
 Follow the [runtime and checkpoint setup](inference.md) first.

docs/upstream-alignment.md

@@ -2,7 +2,7 @@
 
 References are pinned to Laya `d113dca2512fb3eaca313534bc54c7162d87c1d4`,
 Laya Vision `86ccec115ef3d72d1851168fc9b7194e8dbed35a`, and Bub main
-`9bf70488e22a44eeb875a9d55d9a7d352b82e381`. The training mixture includes upstream public tasks and soft teacher
+`6f23fee36a8aa55b71d5be015529a4b486b31d89`. The training mixture includes upstream public tasks and soft teacher
 labels. Existing Laya weights are benchmark references, not training targets.
 
 ## Publicly documented training coverage

pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
 
 [project.optional-dependencies]
 train = ["pyarrow>=25", "scikit-learn>=1.9", "pandas>=3.0", "datasets>=4"]
-agent = ["bub @ git+https://github.com/bubbuild/bub.git@9bf70488e22a44eeb875a9d55d9a7d352b82e381"]
+agent = ["bub @ git+https://github.com/bubbuild/bub.git@6f23fee36a8aa55b71d5be015529a4b486b31d89"]
 
 [dependency-groups]
 quality = ["ruff==0.16.8", "ty==0.0.82", "prek>=0.3.1,<1", "pytest>=9,<10"]

src/dohnuts/bub_agent.py

@@ -7,7 +7,7 @@
 
 from PIL import Image
 
-BUB_REVISION = "9bf70488e22a44eeb875a9d55d9a7d352b82e381"
+BUB_REVISION = "6f23fee36a8aa55b71d5be015529a4b486b31d89"
 
 
 def create_agent(predictor, workspace: Path, tape_directory: Path):

tests/acceptance.py

@@ -1,7 +1,6 @@
 """Accept an exported checkpoint through prediction, reload, and the real Bub SDK."""
 
 import argparse
-import ast
 import asyncio
 import gc
 import json
@@ -118,7 +117,7 @@ async def verify(args):
                 command += " image_path=" + shlex.quote(str(image_path.resolve()))
             stream = await agent.run_stream(session_id="acceptance-" + label, prompt=command)
             events = [event async for event in stream]
-            actual = ast.literal_eval(events[-1].data["text"])
+            actual = json.loads(events[-1].data["text"])
             assert actual == reference, "Bub changed the decision response"
             responses.append(actual)
     report = {