Status: 2026-11-11. Pre-alpha bootstrap. Two separate projects: the
cloud-computer plugin RayrunOS/dsh-rayrun-computer stays its own project and is
consumed as a dependency. The milestones below come from the product research
that shaped this repository; they are direction, not promises. Nothing on the
v0.1 list is built in this repository yet — the bot half does not exist. Dates
are omitted on purpose.
The repository exists, its rules and docs are written, and the product page describes both the target and the current honest status. No publishable code lives here yet.
Exit criteria:
- README, docs, CLAUDE.md and the community files exist and every link resolves;
- the honest status section says what works, what does not, and what is blocked on the Rayrun platform;
- the dependency boundary is explicit:
dsh-rayrun-computeris a separate project that stays untouched, not migrated, renamed or vendored.
Goal: one person, one cloud computer, one or more named bots, a fail-closed
approval card, and a visible bill. Everything shipped on the DSH plugin +
dsh web path; no desktop app.
Task-level plan, spikes and weekly sequencing: mvp-plan.md.
Scope:
- Rayrun Bot package lands here.
dsh-rayrun-botsis published from this repository;dsh-rayrun-computerstays a separate project and is consumed as a dependency (version contract + documented install path in v0.1). - Bot roster and bot↔session mapping. A bot has a name, instructions, skills, a memory directory and a persistent session; the mapping survives a DSH restart instead of living in host memory.
- Persistent DSH host on the machine. Replace "one
headlessrun per prompt" with a supervised host, so a bot can keep working after the laptop closes. The session-resume contract with Rayrun is the first spike. - Fail-closed approvals. DSH approval seam → host half → browser card → decision written back. Timeout, unknown or closed panel means deny.
- Routines, minimal. A scheduled prompt that wakes a bot; no marketplace of triggers.
- Cost bar. Machine hours and balance visible; idle stop remains a plugin option (stop only, never destroy).
- Packaging and CI. npm package(s), version compatibility with DSH, tests, and a release path.
Exit criteria (all of the following):
- A new user can go from README to a running bot in about three minutes, without reading source code.
- After the user closes their laptop, a bot still completes one background task.
- A gated action stops for approval and resumes only after the user approves it in the browser.
- The same bot's second conversation continues from where the first stopped.
- Balance and machine hours for the session are visible.
- Missing token or a failing cloud call produces a clear error, never a fake success.
- Both plugins install through the documented path — Rayrun Bot plus the
separate
dsh-rayrun-computer.
The earlier research target also mentioned 20 seed users, a first-bot success rate above 80% and weekly retention above 40%; treat that as internal validation, not a public promise. A focused estimate for v0.1 from a working cloud-computer base is 4–6 weeks of work — optimistic and not a commitment.
Goal: the user can answer a bot from a phone without opening the DSH page.
Scope:
- an optional IM channel using the user's own bot token (Telegram or Feishu) — no relay server run by this project;
- approvals and "needs attention" messages delivered to that channel;
- the same fail-closed semantics as the browser path;
- candidate, not committed: bot-to-bot handoff via the shared workspace, a primary bot that checks in, and a memory review panel — see inspiration.md.
Exit criteria:
- a gated action can be approved or denied from the user's own IM bot;
- the message path carries no project-hosted component and no user token to a project server;
- the browser path keeps working if IM is not configured.
Goal: only if evidence says a standalone app is needed, wrap dsh web in a thin
desktop shell; never rewrite the chat UI. The research view is that the plugin
plus dsh web is the v0.1/v0.2 form and a desktop app is a later bet, worth
1–2 weeks of wrapping rather than a new client.
Also in scope for a stable v1.0, as capacity allows:
- a documented version compatibility matrix (DSH × cloud image × packages);
- release and install polish (package manager installs, provenance, an update story);
- public, dated status docs that match the shipped code.
Exit criteria:
- the shell adds only packaging and OS integration; the UI is DSH's;
- a user can uninstall cleanly and revoke access at the provider;
- the docs' status claims are checked against the release, not aspirational.
These are explicitly out of scope. Some are permanent, some need a new decision before they could ever enter a milestone:
- multi-tenant SaaS, hosted accounts, or any project-run control plane;
- forking, vendoring or merging
dsh-rayrun-computer; it stays its own project, consumed as a dependency; - team roles/RBAC, audit export, SSO/SCIM, enterprise compliance packages;
- a bot or skill marketplace;
- a native mobile app (IM answers the "from my phone" need for now);
- a credential vault that signs into third-party sites or fills forms for the model;
- multi-region hosting, high availability, or an SLA from this project;
- billing, metering or reselling of Rayrun compute or model tokens;
- a provider abstraction that lets the backend be swapped for E2B-style services, unless that becomes an explicit decision. Today the plugin is Rayrun-specific, and pretending otherwise would mislead users.
Some items depend on the platform, not on this repository. v0.1 needs the first two confirmed; the rest are v0.2/beta:
- the session and long-lived-runtime contract (whether a session id can be created or continued, and how a supervised process survives a restart);
- balance grace and guardrails (today the only lifetime cap stops; exhaustion force-destroys machines and disks);
- safe inbound networking with auth (today the token ingress is unusable and
only
publicworks) — v0.2; - persistent browser profile and sign-in state across sessions and restarts (needed for the legacy-system use case) — v0.2;
- idle auto-stop or wake-on-request, spend caps, runtime egress updates and generic secret injection beyond model slots — beta.
The full handoff list, with the API shape we would want and our workarounds, is
in platform-asks.md. architecture.md
explains each one; faq.md has the user-facing version. One packaging
risk is not the platform's: neither package is on npm yet, so both install from
checkouts until dsh-rayrun-computer is published and the one-install bundle
dependency is verified (W0 spike S6, TBD).
The guardrail is the three-minute first bot: if a feature makes the first-run path longer or less honest, it waits. Safety invariants (fail-closed approvals, no secrets leaving the user's side, no auto-destroy, no telemetry by default) outrank milestones — a milestone does not justify weakening them.