[kaihackney-scale]

Summary

Lets a deployment choose the default cpu and memory of sandboxes while every size a step sets still wins.

• [sandbox.sizer], a new impl / config table under [sandbox] (the same shape as [runner]), names a SandboxSizer (providers/sandbox_providers/sizing.py). The in-tree default is the base class, which chooses nothing, so without the table nothing changes.

• deploy_env, deploy_agent and deploy_sandbox resolve each of cpu and memory separately, in this order:

  1. the step's own value;
  2. else the sizer's;
  3. else the step's default (deploy_sandbox's 1 vCPU / 8 GiB), or, with none, the provider's.

  The sizer is asked only when the step leaves something unset.

• Each sandbox's attribution gains cpu_source and memory_source (step, sizer or default), so a deployment can see how many sandboxes each source sizes. Providers carry attribution as they do today; on Modal these are two more sandbox tags.

• deploy_sandbox keeps an unsized cpu and memory as None on the step and in its stored dict, instead of writing 1.0 and 8192. It still reserves 1 vCPU and 8 GiB, but a size the step chose can now be told apart from one it left to the default. Steps already stored keep the values they were written with.

Notes for review

• Not an entry-point group. A sizer is one policy per deployment, like the runner, so it's a config seam with an in-tree default rather than a registry of named types.
• No pinning across retries. The size is resolved when the step deploys, so a retried deploy asks the sizer again. A sizer whose answer can change mid-run is responsible for keeping one answer per run (attribution carries run_id).
• A failing sizer fails the deploy, as a failing provider would. No caller produces a sizer that should be ignored.
• A linked sandbox: deploy_agent into an existing sandbox still resolves a size (and its sources) though the sandbox already exists; the value is unused there, as cpu / memory_mb already were.

Tests

• tst/unit/task_step/sandbox_sizer_test.py:
  • the precedence for each field and its source;
  • a step that sets both never asks the sizer;
  • an impl that isn't a SandboxSizer is refused;
  • deploy_env hands the sized values and sources to the env;
  • deploy_sandbox falls back to its own default when the sizer chooses nothing.
• tst/unit/task_step/deploy_sandbox_default_cpu_test.py: an unsized step stays None through to_dict / from_dict and still reserves 1 vCPU and 8 GiB; stored sizes are kept; a memory-only step keeps the default cpu.
• tst/unit/task_step/attribution_serialization_test.py: the dict deploy_env hands the env is the step's attribution plus the two source keys.
• make lint is clean.
• tst/unit plus the protocol tests, locally on macOS (-n 4): 12,992 passed, 6 failed before the attribution fix, 4 after.
  • The 4 left are test_disk_recipe.py's two launcher tests (QEMU/KVM) and test_mcp_cli_builder.py's two codegen tests (click missing from the generated venv). They fail the same way on an untouched main checkout on this machine.
  • tst/unit/artifact/test_docker_image_github_clone.py was left out locally. Its git credential fill waits on the macOS Keychain helper that Apple's Command Line Tools git configures outside GIT_CONFIG_NOSYSTEM. CI runs it.

Docs: the site's [sandbox] configuration page should list [sandbox.sizer] and the two attribution keys.

🤖 Generated with Claude Code

Confidence Score: 5/5

The final change appears safe to merge.

Summary

The final change keeps unset sizes separate from explicit sizes in DeploySandboxTaskStep, while preserving the resources it reserves.

• Sandbox steps keep omitted sizes unset while still using the same defaults.

The latest commit deliberately removes the runtime sizer described in the PR text. The earlier saved-size finding is addressed; removing the sizer and generated size-source keys also removes the conditions behind the other two previous threads. Those threads were unnumbered, so they have no previousFindings entries.

Review checks were limited to code, tests, and commit history; tests were not run.

_{Reviews (6) · Last reviewed commit: "refactor(sandbox): drop the runtime size..." · Reviewed by Greptile}