Skip to content

Latest commit

 

History

History
238 lines (191 loc) · 33 KB

File metadata and controls

238 lines (191 loc) · 33 KB

Sources — annotated bibliography

Working note. The external references that ground Shinken's design decisions (D1–D15) and the five headline outcomes (stateful, branchable runtime — checkpoint/fork/resume — first; then the replay ledger that supports it, the sandbox capability manager, bandwidth optimization, and real-time streaming). Entries are deduplicated and grouped by topic; each has a one-line annotation explaining why it matters to Shinken. Sibling design docs are linked by relative path (e.g. 05-tech-decisions.md).

Conventions. All performance / density / cost figures cited from these sources are vendor-published and unverified unless they are first-party measurements; see open-questions.md for the measurement plan. URLs are external public pages only. Cloned prior-art repositories are listed separately under Cloned reference repos. Today's date for currency claims is 2026-05-30.


1. Agent frameworks, computer-use models & benchmarks

The competitive landscape Shinken slots into (see ../docs/design/landscape.md and eval-benchmarks.md) and the canonical action/observation interfaces it must natively speak (D2).

Models & operator products

Agent architectures & grounding research

Benchmarks (the conformance suite Shinken hosts — D7)


2. Sandbox, microVM, container isolation & fast fork

Substrate options (D1) and the fast-fork primitive that makes both instant reset and replay-branching the same operation (D5). See sandbox-infra.md.

microVM / VMM substrate

Container / K8s sandbox pattern

Fast snapshot / fork / branch platforms (build-vs-buy)

Cross-platform guest constraints (Windows / macOS — D1/D10)


3. Streaming, WebRTC & NVENC hardware video encode

The dual-channel transport (D4) and the optional NVENC-accelerated media tier (D11). See streaming-bandwidth.md.

WebRTC architecture & protocols

NVENC & GPU encode (optional GPU tier — D11)

OSS streaming stacks & GStreamer (Linux streaming reference)

Build-vs-buy pixel channel (NICE / Amazon DCV)


4. Replay, recording & on-disk trajectory formats

The event-sourced .skn bundle and branchable checkpoint DAG (D5). See replay.md.


5. Permissions & capability model

The three-layer resource-scoping system: Cedar + ocap + OS enforcement, plus egress/secret brokering (D6) — how the runtime scopes net egress, filesystem, and credentials for a Sandbox. See permissions.md and the isolation & capability note.

Declarative policy engine (Cedar — D6 layer 1)

OS-level sandboxing (D6 layer 3)

Egress scoping & secret brokering


6. ACI, observation model & MCP

The screenshot-first plus structured-upgrade observation rungs (D3), the typed action schema and version-pinned adapters (D2), and the native-SDK-core-plus-MCP-facade interface posture (D8). See ai-native-interface.md.

Accessibility tree / structured observation (D3 Rung 0)

MCP facade (D8)


7. GPU virtualization & confidential compute (optional GPU tier)

Public NVIDIA product facts framed as options for the opt-in GPU tier (D11). See sandbox-infra.md and ../docs/design/economics-and-build-vs-buy.md.


Cloned reference repos (references/)

These projects are cloned/vendored under references/ (git-ignored; only references/README.md is tracked) so we can study implementation details first-hand. Each is prior art we MATCH, BEAT, or DIFFERENTIATE against (see ../docs/design/landscape.md). Re-clone steps live in ../references/README.md.

Repo (path) Upstream Why we keep it
OSWorld/ https://github.com/xlang-ai/OSWorld Primary prior-art runtime+benchmark Shinken supersedes: in-VM action server, gym-style client env, multi-cloud providers, getter/metric evaluators. We document where it is too primitive — single-platform, screenshot polling, no streaming/replay/permissions (../docs/design/osworld-analysis.md, osworld-teardown.md).
cua/ https://github.com/trycua/cua Closest cross-platform competitor. Source of the Image→Runtime→Transport→Interfaces→Sandbox layering, three-mode lifecycle, per-OS handler-factory, dual-altitude MCP, and trajectory recorder we adopt — and the pull-a-PNG-per-step + coarse-auth gaps we BEAT.
codex/ https://github.com/openai/codex OpenAI Codex Rust CLI (codex-rs) sandboxing: the three-axis permission model (admissibility × confinement × approval), strictest-wins merge, ordered (path, Read/Write/Deny) FS rules, fail-closed egress — the most reusable permission design (D6).
anthropic-quickstarts/ https://github.com/anthropics/anthropic-quickstarts computer-use-demo + best-practices: the canonical hosted tool schema (computer/bash/str_replace), sampling loop, image-resize/coordinate math, and prompt-caching discipline Shinken must MATCH — plus the stock Xvfb+x11vnc+noVNC stack it aims to BEAT (D2/D4).
neko/ https://github.com/m1k1o/neko Self-hosted virtual desktop over WebRTC (GStreamer → VP8/VP9/AV1/H264 track + binary DataChannel control). Shared-encoder fan-out, lazy pipelines, runtime source-swap, keyframe-lobby, and GCC/TWCC+hysteresis ABR are the media-plane streaming reference (D4).
OpenAdapt/ https://github.com/OpenAdaptAI/OpenAdapt Desktop record-and-replay RPA: timestamp-correlated multi-stream capture (input/frame/window-a11y/DOM), H.264 video sidecar, nestable action schema, Strategy/get_next_action_event replay abstraction — validates the replay thesis (D5). Its host-coupled, no-sandbox execution is the anti-pattern Shinken avoids.
e2b-desktop/ https://github.com/e2b-dev/desktop ~600-LOC desktop domain layer over the generic E2B Firecracker sandbox. The "thin domain SDK over an isolated runtime" pattern + clean lifecycle API (create/connect/auto-pause/kill) — and the raw-VNC-push + one-process-per-action gaps Shinken BEATs (D1/D4).
UI-TARS-desktop/ https://github.com/bytedance/UI-TARS-desktop Electron app + @ui-tars/@tarko SDK: closest public analog to Shinken's Operator layer — operator pluggability (local nutjs/CDP, remote VM/browser behind one interface), a typed BaseAction union carrying normalized+pixel coordinates, typed-event stream. Source for the Operator contract + coordinate-normalization-at-the-boundary lesson (D2).
OmniParser/ https://github.com/microsoft/OmniParser Pure-vision screen parser: screenshot → labeled, ID-addressable, normalized-bbox interactable elements with Set-of-Marks overlay. Reference for Rung-1 observation (server-side, on-demand) and "never emit raw pixel coordinates" grounding (D3); per-frame re-parse with no caching is the inefficiency Shinken improves on.

Note: references/README.md currently documents only OSWorld/ in its table; the other repos above are present in the working tree and should be added to that table (with upstream URL + one-line rationale) when the references manifest is next updated.


Notes on currency & verification

Schema/version strings (Anthropic computer_2025xxxx, OpenAI computer-use-preview) move fast — OpenAI has announced a computer-use-preview shutdown and Operator sunset — which is exactly why Shinken pins dated adapters (D2). Every latency / density / bitrate / cost number drawn from a vendor blog or product page above is vendor-published and unverified; the first-party measurement plan (a11y coverage, fork P99, glass-to-glass latency, egress $/concurrent) is tracked in open-questions.md.