Skip to content

Latest commit

 

History

History
132 lines (100 loc) · 5.72 KB

platform-asks.md

File metadata and controls

132 lines (100 loc) · 5.72 KB

Rayrun platform asks

Status: 2026-11-11. What RayrunOS needs to add or confirm for Rayrun Bot, and what blocks which milestone. This is a handoff list for the platform team. The MVP is deliberately designed to ship without most of it; each ask lists the workaround we use in the meantime and the shape of the API we would want.

We are a BYO open-source client: we do not need SaaS, multi-tenant control planes, SSO/SCIM, marketplace or billing features. We need the substrate to be explicit and safe for a 24/7 bot product.

Blocks the v0.1 MVP

1. Session and long-lived runtime contract (P0, needs confirmation)

Today: one agent task per machine (409 sandbox_busy, no queue), and the docs do not say whether a session id can be created or continued across prompts. A machine's processes are not documented to survive stop/start, and there is no startup hook.

Need: a documented contract for (a) creating and continuing a session by id across prompts; (b) running a supervised long-lived process inside the machine — start command, restart policy, and what happens on stop/start/reboot; (c) a startup hook or platform supervisor so our bot host comes back after a machine restart.

Why: Bot↔session mapping and "laptop closed, the bot keeps working" are the core of the product.

Workaround without it: per-turn headless runs plus transcript replay (slower, more expensive) and a local watchdog that restarts the host through exec.

Requested shape: session_id accepted and returned by the prompt API; a services/entrypoint API (command, restart: on-failure, started on boot), or documented systemd support; a machine-started event.

2. Balance grace and spend guardrails (P0)

Today: balance exhaustion force-destroys every sandbox in the organization, including disks; only max_lifetime_seconds exists, and it stops without restarting; there is no spend cap.

Need: suspend instead of destroy on zero balance, a low-balance event/notification, and a documented way to cap spend per sandbox or account.

Why: a 24/7 bot losing its workspace is data loss and a support disaster, not merely a billing event.

Workaround: local balance polling, warnings, and idle stop. We cannot prevent a force-destroy.

Requested shape: balance.low / sandbox.suspended events; a grace period before destroy; a per-machine budget field.

3. File control-channel semantics (P1)

Today: we use workspace files as the approval/control mailbox and poll them every 3 s.

Need: atomic write/rename semantics, revision/ETag on download, list pagination and documented rate limits; or the message channel below.

Why: reading a half-written approval file is a correctness bug, and polling has a request budget.

Workaround: unique file names, write-then-rename, and a checksum inside the payload.

4. Outbound control channel (P1, recommended)

Need: a local-authenticated subscribe channel (SSE or WebSocket) that the cloud process can publish to, with no inbound networking, ordered delivery, at-least-once semantics and rate limits.

Why: it replaces file polling for approvals and status, cuts latency, and keeps the no-public-ingress rule intact.

Workaround: the file mailbox above.

Needed for v0.2 / beta

  • Safe inbound networking. Per-instance hostname and certificate, token/gateway auth, WebSocket and SSE passthrough. Today the token tier of the ingress API is unusable (403) so only public works; we never expose it as the product panel. This blocks remote UI, phone approvals and desktop takeover from another device. Workaround: local polling and IM.
  • Scoped credentials. API tokens or a device flow limited to sandbox CRUD and prompt, named, expirable, revocable, and without billing or org rights. Today we hold an account token in the Harness vault. Workaround: keep it in the vault only and recommend a dedicated account.
  • Generic secret injection beyond the model slots, so users can deliver their own GitHub, IM or SaaS tokens to the machine, unreadable by exec.
  • Machine lifecycle events. Webhook or SSE for started/stopped/failed and low balance, so local or IM clients react without polling.
  • Per-sandbox usage. Time-series usage (vm_hours, egress) per sandbox, so the cost bar can attribute spend per bot. Today we show balance and total hours only.
  • Prompt queue. Platform-side queueing instead of 409 sandbox_busy, so several bots on one machine serialize gracefully.
  • Golden image and DSH version. An official image with a current DSH (the image ships 0.2.0-rc.2; the local baseline is 0.2.2-rc.1) and a documented version matrix; optionally image build and pinning.

Later / enterprise — not needed for the OSS MVP

SSO/SAML/SCIM, org RBAC, audit export/SIEM, DPA and data residency, marketplace/agent APIs, multi-region and SLA. Revisit only if an enterprise edition appears.

API hygiene

  • Version the API, publish the OpenAPI document as the source of truth, and keep a changelog and deprecation policy. Today the endpoint count differs between pages (67 vs 93).
  • An error-code dictionary, documented rate limits, and Retry-After.
  • Confirm or mark beta: runtime PUT /egress, arbitrary-port ingress, PTY /terminals, and image builds were all documented but not usable in the 2026-11-01 test.

Priority summary

Ask Needed by Blocks ship today?
1. session + long-lived runtime contract v0.1 confirmation required
2. balance grace / guardrails v0.1 risk accepted with warnings
3. file channel semantics v0.1 no (workaround)
4. outbound control channel v0.1 quality no (polling)
safe inbound v0.2 no
scoped credentials v0.2 no
secrets, lifecycle events, usage, queue, image beta no
enterprise items later no