Outcome
A git_lost_work plugin on the PDK: deterministically enumerate dangling commits and stashes (subject, date, author, touched files, reachable-from) as typed output, in one receipted call. Then measure it the Gym way: the same suite with and without the plugin loaded, the delta attributable to the plugin digest via provenance-stamped tool calls (#32).
Evidence
docs/terminalbench/2026-08-24-fix-git-run-analysis.md (OpenAgentsInc/openagents.com): every fix-git run spent 2-3 tool rounds on the same recon (status/branches/reflog/stash), and the local lane paid for those rounds in minutes. This capability is already possible with shell — which is what makes it the right first plugin delta: the plugin's claim is determinism and economy, and the Gym can check both.
Notes
Rust guest on plugins/pdk with a read-only mount of the workspace .git (the mounts and openagents.list_dir capability landed with the skeleton and the foreign-sessions pilot). Parse refs/reflog/stash files directly or via bounded reads; no shelling out inside the sandbox. Acceptance: recon rounds drop on the suite with no pass regression, per-lane, recorded in the Gym with the plugin digest in the report.
Outcome
A
git_lost_workplugin on the PDK: deterministically enumerate dangling commits and stashes (subject, date, author, touched files, reachable-from) as typed output, in one receipted call. Then measure it the Gym way: the same suite with and without the plugin loaded, the delta attributable to the plugin digest via provenance-stamped tool calls (#32).Evidence
docs/terminalbench/2026-08-24-fix-git-run-analysis.md (OpenAgentsInc/openagents.com): every fix-git run spent 2-3 tool rounds on the same recon (status/branches/reflog/stash), and the local lane paid for those rounds in minutes. This capability is already possible with shell — which is what makes it the right first plugin delta: the plugin's claim is determinism and economy, and the Gym can check both.
Notes
Rust guest on plugins/pdk with a read-only mount of the workspace
.git(the mounts andopenagents.list_dircapability landed with the skeleton and the foreign-sessions pilot). Parse refs/reflog/stash files directly or via bounded reads; no shelling out inside the sandbox. Acceptance: recon rounds drop on the suite with no pass regression, per-lane, recorded in the Gym with the plugin digest in the report.