Blender example on the router sandbox, with renders in chat

[tschellenbach]

How it works

flowchart LR
  user["You: render a lighthouse"] --> conv["Conversation model (llm-fast)"]
  conv -->|"ask skill render"| sub["Subagent (llm-thinking)"]
  sub -->|"run_code code + files"| box["Daytona sandbox built from sandbox_options (bpy 5.2)"]
  box -->|"output + render.png"| sub
  sub -->|"publish"| chat["Upload to the conversation's Chat channel"]
  chat -->|"URL"| settled["task_settled files"]
  settled --> reply["Agent reply with an image attachment, kept in Chat"]

Loading

Why

Thierry's decisions on the first version of this PR:

1. Extend the existing router sandbox (sandbox: daytona) instead of the example running its own Daytona sandbox behind an MCP server. That means long-running jobs, a custom image or setup with Blender, and returning files as well as text. Then switch the example to it.
2. Leave passing images back to the model alone.
3. Post renders to chat as image attachments on the agent's reply, persisted so they show on history reload.

Changes

• Router sandbox. sandbox_options on agent configs (API, sync, patch, store and migration) has image, setup, timeout_ms (at most 30 minutes), cpu, memory_gb and disk_gb. Any of them builds a Daytona image (Base/DebianSlim plus a RUN per setup command), with a 20-minute start timeout for the build. The harness's fixed 60 s code deadline is gone, so a run is bounded by the sandbox timeout and the skill's own deadline.
• Files from run_code. A new files argument lists paths, at most 4 of up to 20 MiB each. Daytona downloads them, and the harness publishes them and tells the subagent what was attached, missing or not shown. A task's result carries the latest file of each name.
• Chat attachments. Conversation.Publish uploads to the channel as the agent (UploadChannelImage or UploadChannelFile). On task_settled the files go onto the reply's Chat attachments (image with image_url, or file with asset_url), and history reads them back as files. The task_settled frame gains files.
• SDKs. Go reads sandbox_options: in agent.yaml (with timeout: 5m) and Event.Files; Python has SandboxSettings and RemoteEvent.files. Only the models this needs were regenerated in the Python client, to stay clear of Sentry and Google Calendar plugins, user logins in chat #736's connector regeneration. The SDK skill notes what the other SDKs still need.
• Example. blender_artist is now agent.yaml (sandbox_options installs Blender's libraries and bpy==5.2.2, 5 min runs, 2 CPU / 4 GiB), a render skill (20 min deadline, a starting program with camera framing, files=["/tmp/render.png"]), and instructions. Its MCP server and Daytona code are deleted.
• Kept from before: Agent.ask() waits on pending_work, and MCPServerLocal holds its session on one task.

Change	Test
Built image, sizes, timeout, ceiling	DaytonaSuite: TestASetupIsBuiltOnTheImageItNames, TestARunMayTakeAsLongAsTheConfigAllows, TestNoRunMayTakeLongerThanTheCeiling
Files downloaded, missing, capped	TestFilesTheCodeWroteComeBackWithItsOutput, TestAFileTheCodeDidNotWriteIsReportedMissing, TestOnlySoManyFilesComeBackFromOneRun
Harness publishes and reports files	HarnessSuite: TestFilesTheCodeHandsBackAreAttachedToTheAnswer and four more
Attachment on the reply and on reload	TestARenderIsAttachedToTheReplyAndComesBackWithItsHistory (fake Chat), TestARenderReachesStreamChatAndSurvivesAReload (real Stream Chat, downloads the CDN URL)
Config round trip and bounds	ConfigsSuite and SyncSuite (integration, Postgres)
SDK loaders and events	Go TestSyncSendsHowTheSandboxIsBuilt, TestASettledTaskCarriesTheFilesItsCodeMade; Python test_how_the_sandbox_is_built_is_read_from_a_nested_block, test_work_that_made_a_file_says_where_it_was_put

Evidence

• Router with a fake Daytona key. The sync stored sandbox_options, the agent delegated to render, and the subagent's run_code went to Daytona with the build request and got 401 Invalid credentials. The agent reported it.
• Router with the sandbox temporarily swapped for local bpy. This is uncommitted debug code, since removed. The subagent wrote the scene, Blender rendered it, and the router uploaded render.png to Stream's CDN, sent it in task_settled and attached it to the reply. GET /v1/agents/conversations/{cid}/messages returned it under files:

The agent

• Unverified: a real Daytona image build and render (no DAYTONA_API_KEY here).
• Go: go test ./... passes in acceleration/ and sdks/go/. Integration runs of api, conversation, store, session and harness pass, except TestPhoneSuite/TestSearchingEveryVendorSaysNoneCanBeAsked, which fails because this environment has real Twilio and Telnyx credentials, and one data-move test that passed on two reruns.
• dev.py check (with UV_FROZEN=1): ruff and mypy are clean, 1,660 passed. The 44 errors are the Docker-only Redis tests, as on accelerate.

_{To show artifacts inline, enable in settings.}