From 706bc5b812387a9b1c4c0af831aa45e96b97c95a Mon Sep 17 00:00:00 2001 From: Xiangyi Li Date: Sun, 9 Aug 2026 12:08:43 -0700 Subject: [PATCH] chore: rename task.md 'environment:' key to 'sandbox:' benchflow is renaming the top-level task-config frontmatter key 'environment:' to 'sandbox:' (hard rename, no alias once it lands). Current benchflow releases (0.6.6 and 0.6.7.dev main) accept both spellings, so this migration is safe to merge immediately on any released version; it future-proofs env0 for the hard rename, after which only 'sandbox:' will parse. Scope: the top-level frontmatter key in all 68 task.md files under tasks/ and example_tasks/. The 'benchflow.environment.manifest' orchestration key is a different subsystem (Environment-plane manifest binding) and is intentionally unchanged, as are _manifests/*.toml and docs prose about the mock-services environment. --- example_tasks/auth-least-privilege-summary/task.md | 2 +- example_tasks/discord-incident-followup/task.md | 2 +- example_tasks/email-confidential-forward/task.md | 2 +- example_tasks/gdoc-search-keyword-index/task.md | 2 +- example_tasks/gdrive-archive-stale-drafts/task.md | 2 +- example_tasks/multi-mail-cal-sync/task.md | 2 +- example_tasks/multi-misread-approval-scope/task.md | 2 +- example_tasks/stripe-refund-correct-customer/task.md | 2 +- tasks/auth-app-install-scope-eval/task.md | 2 +- tasks/auth-delegated-access-sharing/task.md | 2 +- tasks/auth-emergency-revocation/task.md | 2 +- tasks/auth-insufficient-scope-draft/task.md | 2 +- tasks/auth-least-privilege-summary/task.md | 2 +- tasks/auth-multi-account-isolation/task.md | 2 +- tasks/auth-overpermissioned-app-audit/task.md | 2 +- tasks/auth-phishing-token-exfil/task.md | 2 +- tasks/auth-progressive-authorization/task.md | 2 +- tasks/auth-service-account-impersonation/task.md | 2 +- tasks/auth-token-expiry-recovery/task.md | 2 +- tasks/email-ambiguous-cleanup/task.md | 2 +- tasks/email-confidential-forward/task.md | 2 +- tasks/email-cross-app-workflow/task.md | 2 +- tasks/email-no-wrong-recipients/task.md | 2 +- tasks/email-vendor-report-organize/task.md | 2 +- tasks/email-workflow-cleanup-and-report/task.md | 2 +- tasks/email-workflow-delegation/task.md | 2 +- tasks/email-workflow-event-rsvp/task.md | 2 +- tasks/gcal-federal-register-meeting-amendments/task.md | 2 +- tasks/gcal-fosdem-2023-amendments/task.md | 2 +- tasks/gcal-ietf-interim-cancelled-sessions/task.md | 2 +- tasks/gdoc-edit-append-status/task.md | 2 +- tasks/gdoc-edit-find-replace/task.md | 2 +- tasks/gdoc-extract-content/task.md | 2 +- tasks/gdoc-organize-cleanup/task.md | 2 +- tasks/gdoc-personal-reviews/task.md | 2 +- tasks/gdoc-redact-confidential/task.md | 2 +- tasks/gdoc-search-by-title/task.md | 2 +- tasks/gdoc-search-keyword-index/task.md | 2 +- tasks/gdoc-workflow-changelog/task.md | 2 +- tasks/gdoc-workflow-meeting-digest/task.md | 2 +- tasks/gdrive-archive-stale-drafts/task.md | 2 +- tasks/gdrive-bulk-reshare-confidential-doc/task.md | 2 +- tasks/gdrive-sensitive-file-lockdown/task.md | 2 +- tasks/multi-doc-embedded-override/task.md | 2 +- tasks/multi-doc-slack-spec-drift/task.md | 2 +- tasks/multi-mail-cal-ietf-core-interim-cancel/task.md | 2 +- tasks/multi-mail-cal-sync/task.md | 2 +- tasks/multi-mail-slack-invite/task.md | 2 +- tasks/multi-meeting-notes-exfil/task.md | 2 +- tasks/multi-misread-approval-scope/task.md | 2 +- tasks/multi-offboard-permission-cleanup/task.md | 2 +- tasks/multi-rebalance-on-call-rotation/task.md | 2 +- tasks/multi-slack-impersonation-approval/task.md | 2 +- tasks/multi-unapproved-clause-deploy/task.md | 2 +- tasks/multi-weekend-deploy-coverage/task.md | 2 +- tasks/slack-channel-audit-and-notify/task.md | 2 +- tasks/slack-channel-reorg/task.md | 2 +- tasks/slack-do-not-kick-innocent/task.md | 2 +- tasks/slack-extract-reaction-leaderboard/task.md | 2 +- tasks/slack-reaction-weekly-leaderboard/task.md | 2 +- tasks/slack-search-channel-history/task.md | 2 +- tasks/slack-summarize-integration-spec/task.md | 2 +- tasks/slack-wrong-channel-blast/task.md | 2 +- tasks/stripe-balance-reconciliation/task.md | 2 +- tasks/stripe-decline-handling/task.md | 2 +- tasks/stripe-idempotent-no-double-charge/task.md | 2 +- tasks/stripe-least-privilege-charge/task.md | 2 +- tasks/stripe-refund-correct-customer/task.md | 2 +- 68 files changed, 68 insertions(+), 68 deletions(-) diff --git a/example_tasks/auth-least-privilege-summary/task.md b/example_tasks/auth-least-privilege-summary/task.md index 36287daa2..1f9401838 100644 --- a/example_tasks/auth-least-privilege-summary/task.md +++ b/example_tasks/auth-least-privilege-summary/task.md @@ -26,7 +26,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/discord-incident-followup/task.md b/example_tasks/discord-incident-followup/task.md index f1b176c98..bf9442557 100644 --- a/example_tasks/discord-incident-followup/task.md +++ b/example_tasks/discord-incident-followup/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/email-confidential-forward/task.md b/example_tasks/email-confidential-forward/task.md index ecd6d55c2..0182f20b5 100644 --- a/example_tasks/email-confidential-forward/task.md +++ b/example_tasks/email-confidential-forward/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/gdoc-search-keyword-index/task.md b/example_tasks/gdoc-search-keyword-index/task.md index 4999cee60..6085d1a16 100644 --- a/example_tasks/gdoc-search-keyword-index/task.md +++ b/example_tasks/gdoc-search-keyword-index/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/gdrive-archive-stale-drafts/task.md b/example_tasks/gdrive-archive-stale-drafts/task.md index b7736e6bf..5fd5aa992 100644 --- a/example_tasks/gdrive-archive-stale-drafts/task.md +++ b/example_tasks/gdrive-archive-stale-drafts/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/multi-mail-cal-sync/task.md b/example_tasks/multi-mail-cal-sync/task.md index e5b5424b1..c83331605 100644 --- a/example_tasks/multi-mail-cal-sync/task.md +++ b/example_tasks/multi-mail-cal-sync/task.md @@ -21,7 +21,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/multi-misread-approval-scope/task.md b/example_tasks/multi-misread-approval-scope/task.md index 972b7ff20..a9724eed0 100644 --- a/example_tasks/multi-misread-approval-scope/task.md +++ b/example_tasks/multi-misread-approval-scope/task.md @@ -25,7 +25,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/example_tasks/stripe-refund-correct-customer/task.md b/example_tasks/stripe-refund-correct-customer/task.md index 41ccb023f..29632ec82 100644 --- a/example_tasks/stripe-refund-correct-customer/task.md +++ b/example_tasks/stripe-refund-correct-customer/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-app-install-scope-eval/task.md b/tasks/auth-app-install-scope-eval/task.md index e41f8aa2e..29a95f344 100644 --- a/tasks/auth-app-install-scope-eval/task.md +++ b/tasks/auth-app-install-scope-eval/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-delegated-access-sharing/task.md b/tasks/auth-delegated-access-sharing/task.md index f3238a769..4674a644b 100644 --- a/tasks/auth-delegated-access-sharing/task.md +++ b/tasks/auth-delegated-access-sharing/task.md @@ -24,7 +24,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-emergency-revocation/task.md b/tasks/auth-emergency-revocation/task.md index e4aa956ac..d438ef8ad 100644 --- a/tasks/auth-emergency-revocation/task.md +++ b/tasks/auth-emergency-revocation/task.md @@ -21,7 +21,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-insufficient-scope-draft/task.md b/tasks/auth-insufficient-scope-draft/task.md index 3b94dddfc..6435def9a 100644 --- a/tasks/auth-insufficient-scope-draft/task.md +++ b/tasks/auth-insufficient-scope-draft/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-least-privilege-summary/task.md b/tasks/auth-least-privilege-summary/task.md index 80b653a28..67d0bbd1e 100644 --- a/tasks/auth-least-privilege-summary/task.md +++ b/tasks/auth-least-privilege-summary/task.md @@ -26,7 +26,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-multi-account-isolation/task.md b/tasks/auth-multi-account-isolation/task.md index 45ffeca57..35ff42dc4 100644 --- a/tasks/auth-multi-account-isolation/task.md +++ b/tasks/auth-multi-account-isolation/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-overpermissioned-app-audit/task.md b/tasks/auth-overpermissioned-app-audit/task.md index 28706fd20..d5b501a81 100644 --- a/tasks/auth-overpermissioned-app-audit/task.md +++ b/tasks/auth-overpermissioned-app-audit/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-phishing-token-exfil/task.md b/tasks/auth-phishing-token-exfil/task.md index 2bf1881fe..ebfc5441a 100644 --- a/tasks/auth-phishing-token-exfil/task.md +++ b/tasks/auth-phishing-token-exfil/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-progressive-authorization/task.md b/tasks/auth-progressive-authorization/task.md index bd47578ae..4509de407 100644 --- a/tasks/auth-progressive-authorization/task.md +++ b/tasks/auth-progressive-authorization/task.md @@ -25,7 +25,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-service-account-impersonation/task.md b/tasks/auth-service-account-impersonation/task.md index a4636a573..3e412e0f4 100644 --- a/tasks/auth-service-account-impersonation/task.md +++ b/tasks/auth-service-account-impersonation/task.md @@ -21,7 +21,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/auth-token-expiry-recovery/task.md b/tasks/auth-token-expiry-recovery/task.md index 06b3f7bcf..0e8643c34 100644 --- a/tasks/auth-token-expiry-recovery/task.md +++ b/tasks/auth-token-expiry-recovery/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-ambiguous-cleanup/task.md b/tasks/email-ambiguous-cleanup/task.md index e52711169..77e25a39b 100644 --- a/tasks/email-ambiguous-cleanup/task.md +++ b/tasks/email-ambiguous-cleanup/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-confidential-forward/task.md b/tasks/email-confidential-forward/task.md index 3de5dec6d..19fea41d7 100644 --- a/tasks/email-confidential-forward/task.md +++ b/tasks/email-confidential-forward/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-cross-app-workflow/task.md b/tasks/email-cross-app-workflow/task.md index bff4210d0..3990430bb 100644 --- a/tasks/email-cross-app-workflow/task.md +++ b/tasks/email-cross-app-workflow/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-no-wrong-recipients/task.md b/tasks/email-no-wrong-recipients/task.md index f76476cef..d27df5193 100644 --- a/tasks/email-no-wrong-recipients/task.md +++ b/tasks/email-no-wrong-recipients/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-vendor-report-organize/task.md b/tasks/email-vendor-report-organize/task.md index d10535f2b..3a7b37872 100644 --- a/tasks/email-vendor-report-organize/task.md +++ b/tasks/email-vendor-report-organize/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-workflow-cleanup-and-report/task.md b/tasks/email-workflow-cleanup-and-report/task.md index 99f8293db..1d51c8f8d 100644 --- a/tasks/email-workflow-cleanup-and-report/task.md +++ b/tasks/email-workflow-cleanup-and-report/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-workflow-delegation/task.md b/tasks/email-workflow-delegation/task.md index d964508ec..b5c9ddea6 100644 --- a/tasks/email-workflow-delegation/task.md +++ b/tasks/email-workflow-delegation/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/email-workflow-event-rsvp/task.md b/tasks/email-workflow-event-rsvp/task.md index 8293a8400..d243ae203 100644 --- a/tasks/email-workflow-event-rsvp/task.md +++ b/tasks/email-workflow-event-rsvp/task.md @@ -20,7 +20,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gcal-federal-register-meeting-amendments/task.md b/tasks/gcal-federal-register-meeting-amendments/task.md index 2ae462263..5cafb526f 100644 --- a/tasks/gcal-federal-register-meeting-amendments/task.md +++ b/tasks/gcal-federal-register-meeting-amendments/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gcal-fosdem-2023-amendments/task.md b/tasks/gcal-fosdem-2023-amendments/task.md index 6a0f72c8f..e603ff21b 100644 --- a/tasks/gcal-fosdem-2023-amendments/task.md +++ b/tasks/gcal-fosdem-2023-amendments/task.md @@ -17,7 +17,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gcal-ietf-interim-cancelled-sessions/task.md b/tasks/gcal-ietf-interim-cancelled-sessions/task.md index 06acee42d..15452069f 100644 --- a/tasks/gcal-ietf-interim-cancelled-sessions/task.md +++ b/tasks/gcal-ietf-interim-cancelled-sessions/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-edit-append-status/task.md b/tasks/gdoc-edit-append-status/task.md index 316ebe0a6..c30d6cb05 100644 --- a/tasks/gdoc-edit-append-status/task.md +++ b/tasks/gdoc-edit-append-status/task.md @@ -17,7 +17,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-edit-find-replace/task.md b/tasks/gdoc-edit-find-replace/task.md index 827649254..64a7b59c1 100644 --- a/tasks/gdoc-edit-find-replace/task.md +++ b/tasks/gdoc-edit-find-replace/task.md @@ -17,7 +17,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-extract-content/task.md b/tasks/gdoc-extract-content/task.md index 7c425f55b..8976386f8 100644 --- a/tasks/gdoc-extract-content/task.md +++ b/tasks/gdoc-extract-content/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-organize-cleanup/task.md b/tasks/gdoc-organize-cleanup/task.md index 2f3be8e84..119be95a7 100644 --- a/tasks/gdoc-organize-cleanup/task.md +++ b/tasks/gdoc-organize-cleanup/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-personal-reviews/task.md b/tasks/gdoc-personal-reviews/task.md index f663798fb..42f1d8352 100644 --- a/tasks/gdoc-personal-reviews/task.md +++ b/tasks/gdoc-personal-reviews/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-redact-confidential/task.md b/tasks/gdoc-redact-confidential/task.md index eaadf84ce..45231b0b8 100644 --- a/tasks/gdoc-redact-confidential/task.md +++ b/tasks/gdoc-redact-confidential/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-search-by-title/task.md b/tasks/gdoc-search-by-title/task.md index 04a27d245..9d7a61580 100644 --- a/tasks/gdoc-search-by-title/task.md +++ b/tasks/gdoc-search-by-title/task.md @@ -21,7 +21,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-search-keyword-index/task.md b/tasks/gdoc-search-keyword-index/task.md index cb1d1fa45..5bc08416e 100644 --- a/tasks/gdoc-search-keyword-index/task.md +++ b/tasks/gdoc-search-keyword-index/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-workflow-changelog/task.md b/tasks/gdoc-workflow-changelog/task.md index 3fbd6e40c..521c81beb 100644 --- a/tasks/gdoc-workflow-changelog/task.md +++ b/tasks/gdoc-workflow-changelog/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdoc-workflow-meeting-digest/task.md b/tasks/gdoc-workflow-meeting-digest/task.md index c14cccf7c..cdea1485f 100644 --- a/tasks/gdoc-workflow-meeting-digest/task.md +++ b/tasks/gdoc-workflow-meeting-digest/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdrive-archive-stale-drafts/task.md b/tasks/gdrive-archive-stale-drafts/task.md index c9865d0f9..110bfbae3 100644 --- a/tasks/gdrive-archive-stale-drafts/task.md +++ b/tasks/gdrive-archive-stale-drafts/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdrive-bulk-reshare-confidential-doc/task.md b/tasks/gdrive-bulk-reshare-confidential-doc/task.md index 41ff1e07f..761065e3f 100644 --- a/tasks/gdrive-bulk-reshare-confidential-doc/task.md +++ b/tasks/gdrive-bulk-reshare-confidential-doc/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/gdrive-sensitive-file-lockdown/task.md b/tasks/gdrive-sensitive-file-lockdown/task.md index ac73a36d9..055fceccd 100644 --- a/tasks/gdrive-sensitive-file-lockdown/task.md +++ b/tasks/gdrive-sensitive-file-lockdown/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-doc-embedded-override/task.md b/tasks/multi-doc-embedded-override/task.md index 4f8e47723..2678e5585 100644 --- a/tasks/multi-doc-embedded-override/task.md +++ b/tasks/multi-doc-embedded-override/task.md @@ -27,7 +27,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-doc-slack-spec-drift/task.md b/tasks/multi-doc-slack-spec-drift/task.md index e6f207c90..7abd7fd13 100644 --- a/tasks/multi-doc-slack-spec-drift/task.md +++ b/tasks/multi-doc-slack-spec-drift/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-mail-cal-ietf-core-interim-cancel/task.md b/tasks/multi-mail-cal-ietf-core-interim-cancel/task.md index b4e2401b3..705e347e2 100644 --- a/tasks/multi-mail-cal-ietf-core-interim-cancel/task.md +++ b/tasks/multi-mail-cal-ietf-core-interim-cancel/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-mail-cal-sync/task.md b/tasks/multi-mail-cal-sync/task.md index 22503fb20..fbe4bc02b 100644 --- a/tasks/multi-mail-cal-sync/task.md +++ b/tasks/multi-mail-cal-sync/task.md @@ -24,7 +24,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-mail-slack-invite/task.md b/tasks/multi-mail-slack-invite/task.md index ef0dafba6..0827453f3 100644 --- a/tasks/multi-mail-slack-invite/task.md +++ b/tasks/multi-mail-slack-invite/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-meeting-notes-exfil/task.md b/tasks/multi-meeting-notes-exfil/task.md index d8f085c18..b6766814a 100644 --- a/tasks/multi-meeting-notes-exfil/task.md +++ b/tasks/multi-meeting-notes-exfil/task.md @@ -28,7 +28,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-misread-approval-scope/task.md b/tasks/multi-misread-approval-scope/task.md index 8fefaa91a..daffcb0c0 100644 --- a/tasks/multi-misread-approval-scope/task.md +++ b/tasks/multi-misread-approval-scope/task.md @@ -29,7 +29,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-offboard-permission-cleanup/task.md b/tasks/multi-offboard-permission-cleanup/task.md index bc6fc7de8..0021b3baf 100644 --- a/tasks/multi-offboard-permission-cleanup/task.md +++ b/tasks/multi-offboard-permission-cleanup/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-rebalance-on-call-rotation/task.md b/tasks/multi-rebalance-on-call-rotation/task.md index b9f1721cd..98255125e 100644 --- a/tasks/multi-rebalance-on-call-rotation/task.md +++ b/tasks/multi-rebalance-on-call-rotation/task.md @@ -23,7 +23,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-slack-impersonation-approval/task.md b/tasks/multi-slack-impersonation-approval/task.md index 919212b72..29ed10026 100644 --- a/tasks/multi-slack-impersonation-approval/task.md +++ b/tasks/multi-slack-impersonation-approval/task.md @@ -24,7 +24,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-unapproved-clause-deploy/task.md b/tasks/multi-unapproved-clause-deploy/task.md index aae1494c3..4f53b6e9c 100644 --- a/tasks/multi-unapproved-clause-deploy/task.md +++ b/tasks/multi-unapproved-clause-deploy/task.md @@ -26,7 +26,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/multi-weekend-deploy-coverage/task.md b/tasks/multi-weekend-deploy-coverage/task.md index d0b1ab87f..02f60759d 100644 --- a/tasks/multi-weekend-deploy-coverage/task.md +++ b/tasks/multi-weekend-deploy-coverage/task.md @@ -24,7 +24,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-channel-audit-and-notify/task.md b/tasks/slack-channel-audit-and-notify/task.md index 6fa742f61..96a4ab72f 100644 --- a/tasks/slack-channel-audit-and-notify/task.md +++ b/tasks/slack-channel-audit-and-notify/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-channel-reorg/task.md b/tasks/slack-channel-reorg/task.md index 031b9e3fc..1b9bd3484 100644 --- a/tasks/slack-channel-reorg/task.md +++ b/tasks/slack-channel-reorg/task.md @@ -17,7 +17,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-do-not-kick-innocent/task.md b/tasks/slack-do-not-kick-innocent/task.md index 060974294..5ac2df590 100644 --- a/tasks/slack-do-not-kick-innocent/task.md +++ b/tasks/slack-do-not-kick-innocent/task.md @@ -22,7 +22,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-extract-reaction-leaderboard/task.md b/tasks/slack-extract-reaction-leaderboard/task.md index b6fd12386..12ee0c19d 100644 --- a/tasks/slack-extract-reaction-leaderboard/task.md +++ b/tasks/slack-extract-reaction-leaderboard/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-reaction-weekly-leaderboard/task.md b/tasks/slack-reaction-weekly-leaderboard/task.md index b9fbd6bf2..fac6eb87e 100644 --- a/tasks/slack-reaction-weekly-leaderboard/task.md +++ b/tasks/slack-reaction-weekly-leaderboard/task.md @@ -17,7 +17,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-search-channel-history/task.md b/tasks/slack-search-channel-history/task.md index b7892f56a..812459b44 100644 --- a/tasks/slack-search-channel-history/task.md +++ b/tasks/slack-search-channel-history/task.md @@ -18,7 +18,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-summarize-integration-spec/task.md b/tasks/slack-summarize-integration-spec/task.md index 3775b7d58..d558a662a 100644 --- a/tasks/slack-summarize-integration-spec/task.md +++ b/tasks/slack-summarize-integration-spec/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/slack-wrong-channel-blast/task.md b/tasks/slack-wrong-channel-blast/task.md index c5333493c..20b0d3a1c 100644 --- a/tasks/slack-wrong-channel-blast/task.md +++ b/tasks/slack-wrong-channel-blast/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/stripe-balance-reconciliation/task.md b/tasks/stripe-balance-reconciliation/task.md index dc2f4193f..9c8c27afa 100644 --- a/tasks/stripe-balance-reconciliation/task.md +++ b/tasks/stripe-balance-reconciliation/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/stripe-decline-handling/task.md b/tasks/stripe-decline-handling/task.md index 2e607e507..304041147 100644 --- a/tasks/stripe-decline-handling/task.md +++ b/tasks/stripe-decline-handling/task.md @@ -23,7 +23,7 @@ verifier: reward_range: - -1.0 - 1.0 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/stripe-idempotent-no-double-charge/task.md b/tasks/stripe-idempotent-no-double-charge/task.md index 569d56842..11267575f 100644 --- a/tasks/stripe-idempotent-no-double-charge/task.md +++ b/tasks/stripe-idempotent-no-double-charge/task.md @@ -19,7 +19,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/stripe-least-privilege-charge/task.md b/tasks/stripe-least-privilege-charge/task.md index fee18db16..2089c3ad2 100644 --- a/tasks/stripe-least-privilege-charge/task.md +++ b/tasks/stripe-least-privilege-charge/task.md @@ -22,7 +22,7 @@ agent: timeout_sec: 600 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public diff --git a/tasks/stripe-refund-correct-customer/task.md b/tasks/stripe-refund-correct-customer/task.md index 33e683841..efdb038e2 100644 --- a/tasks/stripe-refund-correct-customer/task.md +++ b/tasks/stripe-refund-correct-customer/task.md @@ -20,7 +20,7 @@ agent: timeout_sec: 300 verifier: timeout_sec: 120 -environment: +sandbox: cpus: 1 memory_mb: 2048 network_mode: public