0.138.0: A terminal study can run on Modal

humanish 0.138.0 can run a terminal study's shell on Modal as well as E2B. Set
execution.target: modal-terminal and execution.runtimeAuth: openai-env, sign in with
modal token new, and install the optional modal package beside humanish. The sandbox runs as a
non-root user with sudo, as on E2B, lives up to Modal's 24 hours, is journaled and reclaimed by id
like an E2B sandbox, and is priced at Modal's sandbox rate. Desktop studies still run on E2B or a
local VM.

Added

• A terminal study can run its shell on Modal. execution.target: modal-terminal creates the
  sandbox with Modal's JavaScript SDK (modal, a new optional peer, like @e2b/desktop) in an app
  named humanish, in the environment MODAL_ENVIRONMENT or the Modal profile names. It runs
  Modal's VM runtime with 2 cores (4 vCPU) and 4 GiB, from a Debian image Modal builds in your
  workspace on first use from the same pinned base and package snapshot as the local guest recipe,
  and lives up to Modal's 24 hours instead of the E2B plan's ceiling. Commands run as user with
  passwordless sudo, as on E2B. The run needs MODAL_TOKEN_ID and MODAL_TOKEN_SECRET, which key
  discovery reads from ~/.modal.toml after modal token new, and execution.runtimeAuth: openai-env: Modal can add a key to outbound requests only from a Secret stored in your
  workspace, so the default openai-egress is refused on Modal. An egressAllow list becomes
  Modal's domain and CIDR allowlists. Receipts and owner lines name modal, humanish reclaim
  kills and searches by tag through Modal, and the cost line prices the requested size at Modal's
  sandbox rate (resourceSource: modal.create). Desktop studies still run on E2B or a local VM
  (#1788).

Checked live

• Three live Modal terminal runs of the first-contact study (openai-env, 8-minute cap) graded share_ready with 16 checks. Each participant installed humanish, ran the dry-run flow and reported that it had no keys for a live run, as the E2B release dogfood does. Each sandbox's cleanup was confirmed by id, and its cost line priced at Modal's rate.
• A Modal run killed with SIGKILL after its receipt landed was reclaimed: reclaim --check reported the sandbox running, reclaim killed it by id and reported clean, and Modal listed no container afterwards. This ran twice.
• Earlier live runs found three differences from E2B that the release handles: Modal's default gVisor runtime blocks sudo (sandboxes use Modal's VM runtime), a Modal command's stdin stays open unless closed (Codex waited on it), and Modal reports a terminated sandbox finished a moment after terminate returns.

The release changes nothing a participant sees or does, and nothing the analysis reads, on E2B or local studies, so the efficacy benchmark was not rerun; the 0.133.0 result stands.