Skip to content

Declare checkpoint sandbox paths on samples, tell agent about checkpointed paths - #77

Merged
tadamcz merged 6 commits into
mainfrom
bloom68-checkpointing
Oct 20, 2026
Merged

tadamcz merged 6 commits into
mainfrom
bloom68-checkpointing

Conversation

@tadamcz

@tadamcz tadamcz commented Oct 6, 2026

Copy link
Copy Markdown
Collaborator

Every sample built by build_dataset now carries a CheckpointSampleConfig declaring what Inspect checkpointing should capture: all of /workspace in the agent sandbox, and nothing in the comparator (it is reset per check).

This does not enable checkpointing. The sample layer is customize-only in Inspect; enabling is left to the eval layer, e.g. in a Hawk eval-set config:

checkpoint:
  enabled: true

Without explicit paths, Inspect 0.3.273 would capture each sandbox's home directory (/root, including the Lean toolchain under /root/.elan) in both sandboxes, contrary to Hawk's docs, which say only host state is captured.

Samples capture all of /workspace in the agent sandbox and nothing in the
comparator. This only configures checkpointing; enabling it is left to the eval
layer (e.g. checkpoint.enabled in a Hawk eval-set config).
Adds a WORKSPACE constant to apn.layout, used by PROJECT, the prompt line and
the sample checkpoint sandbox_paths.
@tadamcz

tadamcz commented Oct 20, 2026

Copy link
Copy Markdown
Collaborator Author

Checkpointing was confirmed to be working by manually killing the runner in erdos3-sol61-9600usd-14d-g5ji9eqc7bgre38b. Checkpointing was every 6 hours. After ~68 hours of running, Giles manually killed the runner. It took ~8 hours to restore the state and begin making model calls again. 8 hours is pathologically long but still better than restarting from scratch.

@tadamcz
tadamcz marked this pull request as ready for review October 20, 2026 10:30
@tadamcz tadamcz changed the title Declare checkpoint sandbox paths on samples Declare checkpoint sandbox paths on samples, tell agent about checkpointed paths Oct 20, 2026
@tadamcz
tadamcz merged commit 509ee49 into main Oct 20, 2026
12 of 13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant