[SaladDay]

Problem

Every sandbox node got the fixed limits Sandboxes at once 2 and Retained sandboxes 8, whatever the machine, so new conversations queued on large hosts. The literals lived in the enrollment table defaults, the Core API, Web's Add node and Edit node, and Web had a separate browser-only hint.

Design

• One rule in Core, deployment.RecommendCapacity (formula owned by sandbox-deployment.md#recommended-capacity): Sandboxes at once = min(⌊2 × CPUs / sandbox CPUs⌋, ⌊(memory − 2 GiB) / sandbox memory⌋), at least 1; CPUs are the cgroup-effective cores when reported. Retained sandboxes = max(active, ⌊free disk / 2 / root disk⌋) for a provider that checkpoints, else equal to active (existing capability rule, no provider names). CPU counts twice because agent sandboxes mostly wait on the model; memory is never overcommitted.
• Core–sandbox node protocol: the enrollment request carries the host measurements the node already takes for health (cpu_count, effective_cpu_cores, total_memory_bytes, available_disk_bytes), validated like health. Wire file, protocol doc and the node implementation change together.
• Enrollment tokens: an omitted max_active or max_retained is stored unset (migration 000003 drops the 2/8 defaults); at enrollment Core fills each unset limit from the recommendation, then retained = max(retained, active) and the provider rule apply. Explicit limits and later edits bind as before. An unset limit without host measurements fails with 409 sandbox_node_host_unmeasured and leaves the token unused.
• /core/v1 nodes carry recommended_capacity for the latest health and the current specification (null without either); Web uses it and the browser-side hint is deleted.
• Web: Add node offers Recommended (default, sends no limits) or Custom; Edit node shows the recommended values with Use recommended.
• The node configuration endpoint's unused max_active/max_retained fields (only validated by the Python installer, never used) are removed.

Measured host

mx2 (microsandbox, 4 vCPU, 14.4 GiB memory, 239 GiB free, Standard size 2 CPU / 4 GiB / 8 GiB root disk): 3 at once and 14 retained instead of 2 and 8. Its 8 retained checkpoints took 140–190 MiB each on disk; a suspended sandbox uses no memory.

mx e2e

Fresh install on mx of e2e-805-806 (main + #805 + this PR) from a release-shaped offline archive; mx2 enrolled through Add node with the default Recommended. README ~/.oac/tests/e2e-acceptance-20261020, logs logs/run-pr806/.

• mx2 got 3 at once and 14 retained. Nodes shows 0 / 3 and 0 / 14; Edit node shows "Recommended: 3 at once, 14 retained." with Use recommended.
• codex and claude_sdk, hosted (both cells at the same time) and self_hosted: 36/36; slowest case 30 s.

Tests run

• go test for services/core/internal/deployment/..., internal/sandbox/node, internal/api, the changed deploymentpg tests and the integration test TestRuntimeEnrollmentCapacityPrecedesResetConflict (Postgres 16 container); go vet of the touched packages; make check-openapi.
• New: table test for RecommendCapacity (mx2 → 3/14, a host fitting one sandbox → 1/1, cgroup-limited CPUs, provider without checkpoints); TestEnrollmentDerivesUnsetLimitsFromHost (omitted, partial, explicit, unmeasured).
• tsc for packages/agents-client and apps/web; Vitest for the touched Web folders; Python node tests; pnpm check-translations.
• Not run locally: the Playwright spec apps/web/e2e/nodes.spec.ts (no browser here); it runs in this PR's web-acceptance job.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}