From 7bbf15a555d71dfaf3a2683fb6554ea9e5678124 Mon Sep 17 00:00:00 2001 From: SaibaWaipu Date: Thu, 15 Oct 2026 14:06:47 +0900 Subject: [PATCH 01/14] docs: define account-free Duga invitation direction --- AGENTS.md | 112 ++++++++++++++++++++++++++++++++++++++++++++ README.md | 4 ++ docs/design/duga.md | 30 ++++++++++++ 3 files changed, 146 insertions(+) create mode 100644 AGENTS.md create mode 100644 docs/design/duga.md diff --git a/AGENTS.md b/AGENTS.md new file mode 100644 index 00000000..f251590b --- /dev/null +++ b/AGENTS.md @@ -0,0 +1,112 @@ +# Duga: Product Direction and Agent Notes + +## Purpose and status + +Duga is a proposed website-first personal AI assistant at **duga.app**. A visitor should be able to open the site and start using an assistant without installing software or setting up a room. Over time, they can connect their own devices, contribute idle compute, and use agents for coding, everyday tasks, memory, voice, and data integrations. + +**This is a product plan, not a description of shipped Duga functionality.** The current repository is Pooled: a peer-to-peer inference engine, room runtime, coding-agent harness, CLI/API bridge, and OpenClaw integration. Reuse those components where appropriate; do not represent accounts, the Global Net, credits, long-term personal memory, general connectors, or end-to-end voice conversation as implemented today. + +The existing Pooled project expressly commits to *no accounts* and does not plan to open rooms to strangers by default. Duga's initial product also has **no accounts**: use invite links and room codes, preserving the current browser-room experience. Accounts are deferred, not a prerequisite for continuity or device participation. A public compute network still requires a distinct trust boundary and an explicit governance decision; it must not silently change the security model. See [GOVERNANCE.md](GOVERNANCE.md), [roadmap/README.md](roadmap/README.md), and [SECURITY.md](SECURITY.md). + +## Target experience + +1. Visit **duga.app** and get a useful answer immediately. Guest access should not require an account, installing a worker, pairing a device, or donating compute. +2. Use invite links or room codes to connect another device without registration or login. Preserve the existing distinction: invite links grant access to their holders; code-based requests require host approval by default. Cross-device continuity, durable memory and persistent device management remain proposed and must not imply an account service exists. +3. Let users connect their computers and other capable devices through invite links or host-approved room codes. Show availability, capacity, privacy mode, current jobs, and a one-click disconnect/revoke action. +4. Let users ask for coding help, everyday planning and actions, and voice interaction from one assistant interface. Show what an agent plans to do, what it did, and what requires approval. +5. Provide clear, per-task compute choices: **Local**, **Private Net**, or **Global Net**. Do not make network placement an invisible privacy decision. +6. Let users inspect, correct, export, and delete their memories and connected-service permissions. + +An immediate first answer needs a dependable inference route. An active Global Net alone cannot guarantee availability. Plan a clearly disclosed bootstrap route (such as reserved/hosted capacity or a suitable on-device model); measure cost, privacy, and quality before selecting a default. Never silently send a private task to Global Net because a user's devices are offline. + +## Two networks, plus local execution + +| Mode | Compute membership | Intended use | Data rule | +| --- | --- | --- | --- | +| **Local** | The device currently in use | Tasks that fit locally; offline-capable work where supported | No other compute node receives task data. | +| **Private Net** | The user's own devices explicitly admitted through invite links or host-approved room codes | Personal conversations, memories, coding work, and connected-service data | Only the owner's approved devices may act as inference workers. | +| **Global Net** | Opt-in idle devices contributed by participants; shared pool for participants | Tasks whose inputs the user explicitly agrees to expose to untrusted workers | Treat assigned workers as untrusted. No default access to personal memories, connector credentials, or private files. | + +Private Net membership is **the user's own devices only** in the initial product; invite-based teams are out of scope for now. A link or code does not prove ownership: the user must check which devices they admit. An existing Pooled room can include other people and must not automatically be labeled an owned-device Private Net. The current runtime does not enforce device ownership. Global Net is a participant-shared pool, not a free source of secretly enrolled devices. Contributing compute is optional and separate from using the website. The user can choose a mode per task or conversation; the system must visibly explain when a choice changes where data is processed. + +A Global Net worker may see task text, intermediate activations, outputs, model weights, and operational metadata, depending on the job. **Activations are not encryption**: they may reveal prompt content. Do not claim that public nodes cannot learn users' prompts, or that transport encryption makes a public worker trustworthy. If a task includes potentially sensitive content, recommend Local or Private Net. Require explicit, informed consent before sending it to Global Net. Provenance-based blocking of memory/connector data is defense in depth, not a promise that arbitrary pasted text is automatically classified correctly. See [SECURITY.md](SECURITY.md). + +## Execution architecture + +```text +Browser at duga.app + - account-free sessions, invite links/codes, chat, voice, agent activity, approvals + - per-task compute-mode choice, memories and device controls + | +Control plane (new product services, not the Pooled room broker) + - room-scoped admission and proposed device revocation (no account identity) + - presence, capability discovery, job scheduling and quotas + - contribution/usage ledger and abuse controls + - optional encrypted sync and connector authorization + | +Trusted agent orchestrator (user device or explicitly disclosed service) + - plans and routes tasks, enforces tool permissions and budgets + - owns memory retrieval and service credentials + - submits only permitted inference payloads to the selected compute mode + | +Compute/data plane + - Local: one device + - Private Net: owned and paired devices + - Global Net: opt-in, untrusted volunteer workers +``` + +Keep the control plane and inference data plane separate. An admission or scheduling service should not receive conversation plaintext merely because it handles membership or scheduling; whether it does receive plaintext through a specific fallback or hosted orchestrator must be stated truthfully. Never give Global Net workers connector tokens or unrestricted tool execution. Prefer running agent orchestration and sensitive tools in a trusted environment and using remote nodes only for explicitly approved inference jobs. + +The existing Pooled runtime splits **one model's layers** across GPUs over WebRTC; this is different from running several independent agents. Support both patterns deliberately: model-level distribution for a large model, and agent-level delegation only when tasks can be partitioned, measured, and safely combined. Cross-device latency, the slowest worker, and worker churn can make distribution slower than a single device. Implement scheduling and fallbacks based on measured performance rather than assuming more nodes are faster. See [docs/architecture.md](docs/architecture.md) and [README.md](README.md). + +## Agent capabilities and permissions + +- **Coding:** reuse the existing coding harness and isolated preview where it fits. Present diffs and require approval before touching real projects or running consequential code. Keep generated code sandboxed. +- **Everyday tasks:** start with read-only lookup and drafts, then add calendar/email/file changes behind review and explicit confirmation. Maintain an audit trail, cancellation, retry limits, and clear failure states. +- **Memory:** separate transient conversation context from durable, user-visible facts. Request or clearly present proposed saves; support correction, deletion, and scoped retrieval. Never assume a fact inferred by the model is true. +- **Voice:** add microphone permission, speech-to-text, spoken responses, interruption, and visible recording state. Begin with push-to-talk rather than always-on listening. Browser speech synthesis exists in the room, but voice conversation and audio API support are not already complete. +- **Data integrations:** add connectors one at a time with narrowly scoped OAuth or equivalent grants. Store credentials in a trusted boundary, never in prompts sent to volunteer nodes. Show what each connector may read or change, and allow revocation. +- **Multiple agents:** begin with one orchestrator and bounded tool loops. Add specialist/parallel agents only where evaluation shows a benefit, and keep a single authority for permissions, budgets, and final actions. + +Tool results and model output are untrusted. Validate arguments and schemas, isolate execution, bound steps/time/cost, and distinguish a proposed action from an action already performed. Do not automatically execute a volunteer worker's code or recommendations on a user's machine. + +## Contributed-node experience and credits + +A device owner must explicitly opt in to Global Net contribution and retain control over resource limits, allowed hours, idle/battery/thermal behavior, downloads, and immediate stop. A browser tab may contribute while active and supported; a closed or suspended tab is **not** an always-on worker. A future desktop or background node is an optional upgrade, not a prerequisite for using duga.app. + +Use **non-cash contribution credits** initially: earn credits by providing verified work and spend them on Global Net jobs; show earnings, charges, balances, quotas, and job history. Provide a bounded way for guests and newcomers to try the service without first donating compute. Define exact accounting rates after benchmarking and abuse analysis. Do not mint credits merely for claiming idle capacity or self-reported work; account for verifiable completed work, failed jobs, duplicate identities, manipulated results, and disputes. Correct execution by an arbitrary public worker is not guaranteed by the existing Pooled system; redundancy/spot checks and reputation can reduce risk, not prove universal correctness. Do not promise monetary redemption or a trustless economy. + +Public contribution also entails risks for providers: untrusted inputs, model and dependency downloads, disk use, power, heat, bandwidth, and potentially abusive workloads. Constrain jobs and isolate any code execution; do not turn a volunteer inference worker into a general remote shell. + +## Delivery sequence + +1. **Decision record and evaluation baseline:** define account-free session and invitation boundaries, guest inference bootstrap, explicit Global Net consent, data-retention policy, model licenses, target devices, task examples, cost and latency targets. Resolve Duga-versus-Pooled governance before changing core product commitments. +2. **Website-first assistant:** usable guest chat at duga.app with an honest, dependable inference route; responsive UI, streaming, visible privacy mode, and basic session handling. +3. **Private Net:** invite-link/code admission, key and device revocation, an explicit owned-device trust policy, offline handling, local/private routing, and quality/performance comparison against one device. +4. **Useful agents:** coding flow, limited everyday tools, user-controlled memory, push-to-talk voice, and at least one read-only connector. Add approvals and observable execution before write-capable integrations. +5. **Global Net pilot:** explicitly enrolled nodes and opted-in jobs, limited model/job types, admission and scheduling, rate limits, disconnect recovery, privacy warnings, abuse controls, and measured correctness/latency. Start with a restricted cohort, not an open public network. +6. **Credits and scale:** transparent ledger, verified-work policy, allocation fairness, audits, and multi-agent or background-task expansion once the pilot's reliability and security are demonstrated. + +This ordering is a planning recommendation, not a statement that later capabilities already exist. Parallel research is fine, but do not expose high-risk public workloads before the trust model and safeguards exist. + +## Working on this repository + +- Existing building blocks include `engine/` (WebGPU inference), `room.js` and `room/` (P2P room and protocol), `harness/` (Code mode), `cli/` (local compatible API), and `packages/openclaw/` (agent integration). Read [docs/architecture.md](docs/architecture.md), [docs/design/harness-app.md](docs/design/harness-app.md), and [docs/agents.md](docs/agents.md) before modifying them. +- Preserve the existing browser-link room flow and current output-correctness guarantees. Core protocol, peer visibility, dependencies, and model changes require a design note and maintainer review under [GOVERNANCE.md](GOVERNANCE.md). +- Start with tests that match the layer being changed. The existing repository provides `npm test` (Deno unit tests), package tests, and browser/room end-to-end suites; see [package.json](package.json) and [docs/agents.md](docs/agents.md). Record hardware, commit, workload, and baseline for performance claims. +- Keep **implemented**, **proposed**, and **open decision** distinct in code comments, documentation, and the UI. In particular, a Pooled invite link grants bearer access and does not prove device ownership or a persistent identity, and the existing room authentication does not make public volunteer nodes safe for private prompts. + +## Invitation and session rules + +- **Implemented:** a room code locates a room; host approval is required by default. An invite link contains a random secret in its fragment and admits its holder without approval. Returning devices can use a room pass. See [SECURITY.md](SECURITY.md) and [docs/protocol.md](docs/protocol.md). +- **Required direction:** no registration, login, account database, or account-bound device enrollment in the initial Duga product. Keep invite secrets out of logs, analytics and server-visible URLs. Do not treat short room codes as authentication secrets or silently disable host approval. +- **Proposed:** invitation expiry/rotation, per-device removal and durable cross-device continuity need explicit lifecycle and recovery designs before being advertised. Do not claim current links expire or that disconnecting a device revokes its credentials. +- **Privacy:** every admitted inference device is trusted with room data under the existing threat model. Invitations do not make public workers safe for private tasks. Keep Global Net enrollment and task consent separate from room admission. + +## Open decisions requiring explicit agreement + +- What guarantees the guest's first response: on-device inference, reserved hosted capacity, or another clearly disclosed fallback? Who pays for it? +- Which data is synced across devices, where is it stored, how is it encrypted and deleted, and how can users transfer sessions through explicitly approved devices without accounts? +- What work is admissible on Global Net, what exactly does a worker learn, and what verification is sufficient before credits are awarded? +- How do account-free invitations, device ownership, revocation, recovery, rate limits, and Sybil resistance work without degrading the existing account-free Pooled rooms? +- Which first connector and voice stack provide useful UX without leaking credentials or promising unsupported browser background behavior? diff --git a/README.md b/README.md index 7f3c4f17..8b29080d 100644 --- a/README.md +++ b/README.md @@ -25,6 +25,10 @@ https://github.com/user-attachments/assets/5142d877-e4d5-4f6b-9b5e-0ef48dcb73f6 Pooled runs one open model across the devices in a room. Your laptop, a friend's desktop and a phone each hold some of the model's layers, and together they run a model none of them could run alone. You can chat with it, or switch to Code mode and let it build and fix a small app right in the tab. There is nothing to install and no account, and no server does any of the thinking. +## Duga direction (proposed) + +This repository is being adapted toward **Duga**, a website-first assistant at duga.app. The initial direction is account-free: connect devices using invite links or room codes with host approval by default. The runtime below remains Pooled; this note does not announce a deployed Duga service. See [the Duga design note](docs/design/duga.md) for scope and privacy boundaries. + ## How it works Friends open one link. Each device (laptop, desktop, phone) holds some of the model's layers and runs them on its own GPU with our own WGSL kernels. For every token, a small hidden state (10 KB on the 27B, 4 KB on the MoE) passes from tab to tab over direct WebRTC connections, and the host turns the result into the next word. diff --git a/docs/design/duga.md b/docs/design/duga.md new file mode 100644 index 00000000..8823b4b0 --- /dev/null +++ b/docs/design/duga.md @@ -0,0 +1,30 @@ +# Duga: account-free first phase + +Status: proposed product direction. The account-free invitation mechanism described below already exists in Pooled; this document changes no runtime protocol or deployed domain configuration. + +## Decision + +Build toward a website-first assistant at duga.app without introducing accounts. Use the existing browser room, invite links and room codes for device participation. Registration, login, account-bound pairing and guest-to-account migration are deferred. Preserve Pooled's browser-link flow, CLI compatibility and output correctness. + +## Joining and trust + +- An invite link carries a random 128-bit bearer secret in its fragment. Anyone holding it can join without a separate host approval; share it only with intended devices. +- A six-character room code locates a room. By default the host must approve the joining device. Preserve that default and the existing link verification flow. +- A returning device can use its room pass. This is room-scoped access, not an account or verified ownership identity. +- Expiring invitations, rotating credentials, kicking/revoking devices and durable recovery are future work, not current guarantees. A disconnect alone is not credential revocation. + +See [the current security model](../../SECURITY.md) and [joining protocol](../protocol.md#joining-a-room). Every admitted inference device must be trusted with room data. Neither a code nor an invite proves that the device belongs to the user. + +## Product boundaries + +**Implemented building blocks:** browser rooms, distributed inference, chat, Code mode, CLI/API bridge and OpenClaw integration. Model availability, browser support and loading still constrain the first response. + +**Proposed:** Duga's assistant experience, an owned-device Private Net, user-controlled memory, voice conversation, connectors and a restricted Global Net pilot. A room with invited friends must not be presented as an owned-device Private Net. The current runtime does not enforce device ownership. + +**Open decisions:** dependable first-response inference and its cost/privacy disclosure; session storage and approved cross-device transfer; invitation lifecycle and recovery; device revocation; public-worker admission and abuse controls. Do not add a hosted inference fallback or send tasks to public devices silently. + +Global Net remains separate from private rooms. Both contributing a device and submitting each public task require explicit consent. Public workers may learn prompts, intermediate activations and outputs. Activations are not encryption. Keep memory, private files, service credentials and tool execution within a trusted boundary. Account-free invitations do not resolve public-worker trust, correctness or Sybil resistance. + +## Delivery + +First preserve and present the existing invitation flow honestly. Then evaluate the first-response route and design session/device lifecycle before implementing persistent features. Do not advertise public compute or credits as available. Public-network changes remain subject to [governance](../../GOVERNANCE.md) and a separate security review; this direction does not amend Pooled's governance commitments. From 9d403002f6b10297647b747ec1aa439b801bd235 Mon Sep 17 00:00:00 2001 From: SaibaWaipu Date: Thu, 15 Oct 2026 14:18:32 +0900 Subject: [PATCH 02/14] feat: build account-free Duga assistant workspace --- AGENTS.md | 2 +- README.md | 18 +- app/app.css | 33 ++ app/app.js | 134 +++++++ app/local.js | 74 ++++ app/state.js | 57 +++ docs-site/astro.config.mjs | 4 +- docs-site/src/assets/wordmark-dark.svg | 2 +- docs-site/src/assets/wordmark-light.svg | 2 +- docs-site/src/content/docs/index.mdx | 16 +- docs/design/duga.md | 21 ++ favicon.svg | 2 +- index.html | 443 ++++++------------------ p2p.html | 16 +- package.json | 7 +- room.js | 20 +- room/card.js | 4 +- room/compute.js | 2 +- scripts/og.mjs | 25 +- site.webmanifest | 4 +- site/logo/animated.svg | 2 +- site/logo/favicon-16.svg | 2 +- site/logo/favicon-32.svg | 2 +- site/logo/mark-dark.svg | 2 +- site/logo/mark.svg | 2 +- site/logo/wordmark-dark.svg | 2 +- site/logo/wordmark-light.svg | 2 +- site/og.png | Bin 140929 -> 33772 bytes tests/duga/local.test.mjs | 33 ++ tests/duga/state.test.mjs | 42 +++ tests/e2e/duga.mjs | 118 +++++++ tests/e2e/duga_local.mjs | 53 +++ tests/unit/card_test.js | 2 +- 33 files changed, 742 insertions(+), 406 deletions(-) create mode 100644 app/app.css create mode 100644 app/app.js create mode 100644 app/local.js create mode 100644 app/state.js create mode 100644 tests/duga/local.test.mjs create mode 100644 tests/duga/state.test.mjs create mode 100644 tests/e2e/duga.mjs create mode 100644 tests/e2e/duga_local.mjs diff --git a/AGENTS.md b/AGENTS.md index f251590b..f25f87ca 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -4,7 +4,7 @@ Duga is a proposed website-first personal AI assistant at **duga.app**. A visitor should be able to open the site and start using an assistant without installing software or setting up a room. Over time, they can connect their own devices, contribute idle compute, and use agents for coding, everyday tasks, memory, voice, and data integrations. -**This is a product plan, not a description of shipped Duga functionality.** The current repository is Pooled: a peer-to-peer inference engine, room runtime, coding-agent harness, CLI/API bridge, and OpenClaw integration. Reuse those components where appropriate; do not represent accounts, the Global Net, credits, long-term personal memory, general connectors, or end-to-end voice conversation as implemented today. +**The direction below remains a product plan; do not infer shipped functionality from it.** The Duga preview now adds a root-page Local assistant, explicit browser-only notes, session activity and an embedded room workspace. Private Net ownership enforcement, public compute and account services remain unimplemented. The underlying runtime is Pooled: a peer-to-peer inference engine, room runtime, coding-agent harness, CLI/API bridge, and OpenClaw integration. Reuse those components where appropriate; do not represent accounts, the Global Net, credits, long-term personal memory, general connectors, or end-to-end voice conversation as implemented today. The existing Pooled project expressly commits to *no accounts* and does not plan to open rooms to strangers by default. Duga's initial product also has **no accounts**: use invite links and room codes, preserving the current browser-room experience. Accounts are deferred, not a prerequisite for continuity or device participation. A public compute network still requires a distinct trust boundary and an explicit governance decision; it must not silently change the security model. See [GOVERNANCE.md](GOVERNANCE.md), [roadmap/README.md](roadmap/README.md), and [SECURITY.md](SECURITY.md). diff --git a/README.md b/README.md index 8b29080d..857c2838 100644 --- a/README.md +++ b/README.md @@ -1,13 +1,13 @@

- + - pooled + Duga

-

Peer-to-peer LLM inference in the browser. Pool your devices to run big open models, for chat and coding agents.

+

Your devices. Your assistant. Account-free chat, local inference and trusted-device coding rooms.

- Site · - Start a room · + Site · + Open a room · Architecture · Benchmarks · Roadmap · @@ -23,11 +23,13 @@ https://github.com/user-attachments/assets/5142d877-e4d5-4f6b-9b5e-0ef48dcb73f6

Demo, recorded September 27, 2026 (2:29, with voice and subtitles): Qwen 3.6 35B MoE across a Mac and an iPhone in browser tabs. The phone chats with it, then Code mode builds a Tetris game, fixes its own bug and restyles it. Download · the September 7 demo.

-Pooled runs one open model across the devices in a room. Your laptop, a friend's desktop and a phone each hold some of the model's layers, and together they run a model none of them could run alone. You can chat with it, or switch to Code mode and let it build and fix a small app right in the tab. There is nothing to install and no account, and no server does any of the thinking. +The underlying Pooled runtime runs one open model across the devices in a room. Your laptop, a friend's desktop and a phone each hold some of the model's layers, and together they run a model none of them could run alone. You can chat with it, or switch to Code mode and let it build and fix a small app right in the tab. There is nothing to install and no account, and no server does any of the thinking. -## Duga direction (proposed) +## Duga preview -This repository is being adapted toward **Duga**, a website-first assistant at duga.app. The initial direction is account-free: connect devices using invite links or room codes with host approval by default. The runtime below remains Pooled; this note does not announce a deployed Duga service. See [the Duga design note](docs/design/duga.md) for scope and privacy boundaries. +Duga is an account-free browser assistant built on the Pooled runtime. The root page provides local streaming chat, explicit browser-only notes, session activity and a room workspace. It downloads Qwen3 0.6B weights on first use and runs on WebGPU; there is no hosted inference fallback. Invite links and host-approved codes connect devices through the existing `/room` and `/r/:code` routes. + +Private Net ownership enforcement, Global Net, credits, service connectors and microphone conversation remain proposed. See [the Duga design note](docs/design/duga.md) for implementation boundaries, privacy and validation. The upstream Pooled CLI/package names and wire/storage identifiers are preserved for compatibility. ## How it works diff --git a/app/app.css b/app/app.css new file mode 100644 index 00000000..ffe843c7 --- /dev/null +++ b/app/app.css @@ -0,0 +1,33 @@ +:root { color-scheme:light; --text:var(--ink); --muted:var(--ink-3); --panel:var(--white); --border:var(--rule); --accent:var(--blue); } +* { box-sizing:border-box; } +body { margin:0; font:14px/1.6 var(--sans),sans-serif; color:var(--ink); background:var(--bg); } +button,input,textarea { font:inherit; } +button,a,input,textarea { -webkit-tap-highlight-color:transparent; } +button { cursor:pointer; border:1px solid var(--rule-3); border-radius:8px; background:var(--white); color:var(--ink); padding:8px 14px; transition:background .15s var(--ease),transform .12s; } +button:hover { background:var(--bg-3); } +button:active { transform:scale(.98); } +button:disabled { cursor:not-allowed; color:var(--ink-3); background:var(--bg-3); } +a { color:var(--blue-text); text-decoration:none; } +a:hover { text-decoration:underline; } +:focus-visible { outline:2px solid var(--blue); outline-offset:3px; } +[hidden] { display:none!important; } +.sr-only { position:absolute;width:1px;height:1px;padding:0;margin:-1px;overflow:hidden;clip:rect(0,0,0,0);white-space:nowrap;border:0; } +.skip { position:fixed;left:16px;top:-100px;z-index:50;background:var(--white);padding:12px; }.skip:focus{top:12px} +.app { display:grid;grid-template-columns:240px minmax(0,1fr);min-height:100dvh; } +.sidebar { align-self:start; background:var(--bg-2);border-right:1px solid var(--rule);padding:28px 18px 18px;display:flex;flex-direction:column;position:sticky;top:0;height:100dvh;z-index:20; } +.brand { display:flex;align-items:center;gap:10px;color:var(--ink);margin:0 12px 34px;font:500 26px/1 var(--display),sans-serif;letter-spacing:-1px; }.brand:hover{text-decoration:none}.brand small{font:11px var(--mono);letter-spacing:0;color:var(--ink-3);margin-top:5px} +.new-chat { display:flex;gap:12px;align-items:center;width:100%;padding:11px 14px;background:var(--white);box-shadow:0 1px 2px rgba(20,22,29,.05);margin-bottom:20px; }.new-chat span{font-size:20px;line-height:1;color:var(--blue)} +nav button,.settings-link { display:flex;align-items:center;gap:13px;width:100%;text-align:left;border:0;background:transparent;padding:10px 14px;color:var(--ink-2);margin:3px 0; }nav button>span{width:18px;text-align:center;font-size:20px;line-height:1}nav button[aria-current="page"]{background:var(--blue-soft);color:var(--blue-text)}nav button small{font:10px var(--mono);margin-left:auto;color:var(--ink-3)} +.nav-label { margin:30px 14px 10px;font-size:11px;color:var(--ink-3); } +.sidebar-bottom { margin-top:auto;padding:28px 4px 0; }.sidebar-bottom>a{display:block;font-size:12px;color:var(--ink-3);margin:12px}.sidebar-bottom>small{font-size:10px;color:var(--ink-3);margin-left:12px}.session-card{display:flex;align-items:center;gap:10px;padding:12px;border-top:1px solid var(--rule)}.session-card b{font-size:11px;font-weight:500}.session-card span{display:block;font-size:10px;color:var(--ink-3)}.settings-link{font-size:12px;justify-content:space-between;padding:8px 12px} +i { display:inline-block;width:6px;height:6px;border-radius:50%;background:var(--blue);flex-shrink:0; } +.main-wrap{min-width:0}.topbar{height:72px;border-bottom:1px solid var(--rule-2);display:flex;align-items:center;justify-content:space-between;padding:0 36px;background:var(--bg);gap:14px}.topbar b{font-size:13px;font-weight:500}.crumb{font-size:12px;color:var(--ink-3)}.top-actions{display:flex;align-items:center;gap:20px}.chip{display:inline-flex;align-items:center;gap:8px;font-size:11px;color:var(--ink-2);white-space:nowrap}.mode-button{font-size:12px;background:var(--bg-2);padding:6px 12px}.mode-button span{margin-left:14px}#nav-toggle{display:none} +#main{outline:none}.chat-panel{min-height:calc(100dvh - 72px);max-width:900px;margin:auto;padding:0 42px 18px;display:flex;flex-direction:column}.welcome{margin:auto 0;padding:66px 0 54px;text-align:center}.eyebrow{font:10px/1.5 var(--mono),monospace;letter-spacing:.12em;color:var(--ink-3)}.welcome-mark{width:80px;height:80px;margin:28px auto 18px;border:1px solid var(--rule);background:var(--bg-2);border-radius:24px;display:grid;place-items:center}.welcome h1{font:400 36px/1.2 var(--display),sans-serif;letter-spacing:-1px;margin:18px 0 12px}.welcome>p{font-size:14px;color:var(--ink-3);line-height:1.8}.suggestions{display:grid;grid-template-columns:repeat(3,minmax(0,1fr));gap:12px;margin-top:38px}.suggestions button{text-align:left;background:var(--bg-2);padding:16px;border-color:var(--rule);border-radius:12px}.suggestions button>span{display:block;color:var(--blue);margin-bottom:12px;font-size:19px}.suggestions b{display:block;font-weight:500;font-size:12px}.suggestions small{font-size:11px;color:var(--ink-3)} +.composer-area{width:100%;margin-top:auto}.composer{border:1px solid var(--rule-3);background:var(--white);border-radius:16px;padding:16px;box-shadow:0 4px 16px -12px rgba(20,22,29,.28)}.composer:focus-within{border-color:var(--blue-200)}textarea{resize:vertical;width:100%;border:1px solid var(--rule-3);border-radius:8px;background:var(--white);padding:12px;color:var(--ink);min-height:80px}#prompt{resize:none;min-height:70px;border:0;padding:0;background:transparent;outline:none;max-height:200px}textarea::placeholder,input::placeholder{color:var(--ink-3)}.composer-tools,.composer-tools>div{display:flex;align-items:center;justify-content:space-between;gap:8px}.model-name{font-size:11px;color:var(--ink-2)}.model-name small{color:var(--ink-3)}.quiet{background:transparent;border:0;color:var(--ink-3);font-size:12px}.primary{background:var(--blue);border-color:var(--blue);color:var(--white)}.primary:hover{background:var(--blue-text);color:var(--white)}.primary:disabled{background:var(--bg-3);border-color:var(--rule)}.composer-foot{display:flex;justify-content:space-between;align-items:center;color:var(--ink-3);font-size:10px;margin-top:10px}.composer-foot span{display:flex;align-items:center;gap:6px}.text-button{border:0;background:transparent;padding:0;color:var(--ink-3);font-size:10px}.bootstrap-note{font-size:10px;color:var(--ink-3);margin:10px 0 6px}.memory-use{font-size:11px;color:var(--ink-3);display:flex;align-items:center;gap:6px}.memory-use input{accent-color:var(--blue)}#run-status{font-size:12px;color:var(--ink-2);margin:12px 0;overflow-wrap:anywhere}progress{width:100%;height:4px;accent-color:var(--blue);margin-bottom:12px} +.conversation{padding:30px 0 20px;overflow-wrap:anywhere}.message{margin-bottom:24px}.message .message-label{font-size:11px;color:var(--ink-3);margin-bottom:7px}.message .message-body{white-space:pre-wrap;font-size:14px;line-height:1.8}.message.user .message-body{background:var(--bg-3);border-radius:12px;padding:12px 16px;margin-left:40px}.message.user .message-label{text-align:right}.message.failed{color:var(--del)} +.page-panel{max-width:1000px;padding:54px 48px 80px;margin:auto}.page-panel>h1{font:400 36px/1.2 var(--display);letter-spacing:-.03em;margin:12px 0}.intro{color:var(--ink-3);max-width:650px;margin-bottom:32px}.two-col{display:grid;grid-template-columns:1fr 1fr;gap:16px}.card{background:var(--bg-2);border:1px solid var(--rule);border-radius:16px;padding:24px;margin-bottom:20px}.card h2{font-size:16px;font-weight:500;margin:0 0 10px}.card p{font-size:13px;color:var(--ink-2);margin:8px 0 18px}.card small{display:block;font-size:11px;color:var(--ink-3);margin-top:16px}.card label,label[for="invite"]{display:block;font-size:12px;margin-bottom:8px}input:not([type="checkbox"]){width:100%;border:1px solid var(--rule-3);background:var(--white);padding:10px 12px;border-radius:8px;color:var(--ink)}#join-form>button{margin-top:14px}.device-card{display:flex;align-items:center;gap:16px;padding:20px;background:var(--white);border:1px solid var(--rule);border-radius:16px;margin-bottom:24px}.device-card h2{font-size:14px;font-weight:500;margin:0}.device-card p{font-size:12px;color:var(--ink-3);margin:3px 0}.device-card .chip{margin-left:auto}.device-icon{width:44px;height:44px;background:var(--blue-soft);color:var(--blue);border-radius:12px;display:grid;place-items:center;font-size:22px}.notice{padding:18px 22px;border-left:2px solid var(--blue-200);background:var(--bg-2);font-size:12px;color:var(--ink-2);margin:4px 0 24px}.notice b{display:block;margin-bottom:4px}.badge{display:inline-block;border-radius:999px;background:var(--bg-3);color:var(--ink-3);font:10px var(--mono);padding:4px 8px;margin-bottom:18px}.row{display:flex;gap:10px;align-items:center;flex-wrap:wrap;margin-top:14px}.muted{color:var(--ink-3);font-size:12px}.section-heading{display:flex;justify-content:space-between;gap:12px;align-items:center;margin:26px 0 18px}.section-heading h2{font-size:14px;font-weight:500}.section-heading h2 span{color:var(--ink-3);margin-left:6px}.section-heading button{font-size:11px;padding:6px 10px}.danger{color:var(--del)}.memory-item{padding:18px 20px;border:1px solid var(--rule);border-radius:12px;background:var(--white);margin-bottom:12px}.memory-item p{white-space:pre-wrap;overflow-wrap:anywhere;margin:0 0 10px}.memory-item button{font-size:11px;padding:5px 10px;margin-right:8px}.empty-state{border:1px dashed var(--rule-3);border-radius:12px;padding:40px;text-align:center;color:var(--ink-3);font-size:13px}.activity-list{list-style:none;padding:0}.activity-list li{padding:16px 0;border-bottom:1px solid var(--rule);display:flex;gap:18px;font-size:13px}.activity-list time{font:11px var(--mono);color:var(--ink-3);padding-top:4px;white-space:nowrap}.workspace-panel{padding:24px}.workspace-intro{display:flex;justify-content:space-between;gap:24px;margin-bottom:20px}.workspace-intro h1{font:400 26px var(--display);margin:0 0 8px}.workspace-intro p{font-size:12px;color:var(--ink-3);max-width:650px;margin:0}.workspace-intro a{font-size:11px;white-space:nowrap}iframe{width:100%;height:calc(100dvh - 220px);min-height:520px;border:1px solid var(--rule-3);border-radius:12px;background:var(--bg-2)} +dialog{width:min(560px,calc(100% - 32px));max-height:calc(100dvh - 40px);overflow:auto;border:1px solid var(--rule);border-radius:16px;background:var(--bg-2);padding:32px;color:var(--ink);box-shadow:0 24px 60px -20px rgba(20,22,29,.35)}dialog::backdrop{background:rgba(20,22,29,.32);backdrop-filter:blur(4px)}dialog h2{font:400 26px/1.2 var(--display);margin:8px 0 12px}dialog>p{font-size:13px;color:var(--ink-3)}.dialog-close{position:absolute;right:12px;top:12px;border:0;background:transparent;font-size:20px;padding:0 8px}.mode-card{border:1px solid var(--rule);border-radius:12px;padding:16px;margin:12px 0}.mode-card.active{border-color:var(--blue);background:var(--blue-soft)}.mode-card h3{font-size:14px;margin:0;display:flex;justify-content:space-between}.mode-card h3 small{font:10px var(--mono);color:var(--ink-3)}.mode-card p{font-size:12px;color:var(--ink-2);margin:8px 0 0}.toast{position:fixed;bottom:24px;left:50%;transform:translateX(-50%);background:var(--ink);color:var(--white);padding:12px 20px;border-radius:12px;max-width:calc(100% - 32px);font-size:12px;z-index:100} +@media(min-width:1400px){.chat-panel{max-width:960px}.welcome{padding-top:100px}} +@media(max-width:1000px){.app{grid-template-columns:210px minmax(0,1fr)}.sidebar{padding-left:12px;padding-right:12px}.chat-panel{padding-left:28px;padding-right:28px}.page-panel{padding:40px 28px}.topbar{padding:0 28px}.suggestions{gap:8px}.suggestions button{padding:12px}.suggestions small{font-size:10px}.two-col{grid-template-columns:1fr}} +@media(max-width:700px){.app{display:block}.sidebar{position:fixed;left:0;top:60px;width:250px;height:calc(100dvh - 60px);box-shadow:18px 0 40px -24px rgba(20,22,29,.28);visibility:hidden;transform:translateX(-110%);transition:transform .2s var(--ease)}body.nav-open .sidebar{visibility:visible;transform:translateX(0)}.topbar{height:60px;padding:0 18px;position:sticky;top:0;z-index:25}.topbar>div:nth-child(2){margin-right:auto}.crumb,.top-actions>.chip{display:none}#nav-toggle{display:block;border:0;background:transparent;padding:4px;font-size:20px}.chat-panel{min-height:calc(100dvh - 60px);padding:0 18px 20px}.welcome{padding:42px 0 34px}.welcome h1{font-size:30px}.eyebrow{font-size:9px}.welcome-mark{margin-top:22px;width:68px;height:68px;border-radius:20px}.suggestions{grid-template-columns:1fr;gap:8px;margin-top:28px}.suggestions button{display:grid;grid-template-columns:28px 1fr;padding:12px 14px;gap:0 8px}.suggestions button>span{grid-row:span 2;margin:0;align-self:center}.suggestions b{font-size:12px}.suggestions small{font-size:11px}.composer{padding:14px}.bootstrap-note{font-size:10px}.page-panel{padding:34px 20px 60px}.page-panel>h1{font-size:30px}.card{padding:20px}.device-card{flex-wrap:wrap}.device-card .chip{margin-left:60px}.section-heading{flex-wrap:wrap}.workspace-panel{padding:20px 12px}.workspace-intro{display:block}.workspace-intro a{display:inline-block;margin-top:10px}iframe{min-height:650px;height:calc(100dvh - 240px)}dialog{padding:28px 22px}} +@media(prefers-reduced-motion:reduce){*,*::before,*::after{animation:none!important;transition:none!important;scroll-behavior:auto!important}} diff --git a/app/app.js b/app/app.js new file mode 100644 index 00000000..fa999886 --- /dev/null +++ b/app/app.js @@ -0,0 +1,134 @@ +import { MODES, requireLocal, invitePath, MemoryStore, systemPrompt } from './state.js'; +const $ = id => document.getElementById(id); +const titles = { chat: 'Assistant', code: 'Code workspace', devices: 'Devices', memories: 'Memory', connections: 'Connections', activity: 'Activity', settings: 'Settings & privacy' }; +let view = 'chat', model = null, run = null, editId = '', toastTimer, confirmAction = null; +let turns = [], events = []; +// Storage may be unavailable even on property access (private mode / user policy). +let storage; +try { storage = window.localStorage; } catch { storage = { getItem() { throw Error(); }, setItem() { throw Error(); }, removeItem() { throw Error(); } }; } +const memories = new MemoryStore(storage); +function notice(text) { $('notice').textContent = text; $('notice').hidden = false; clearTimeout(toastTimer); toastTimer = setTimeout(() => { $('notice').hidden = true; }, 5000); } +function navigate(next, hash = true) { + view = Object.hasOwn(titles, next) ? next : 'chat'; + document.querySelectorAll('[data-panel]').forEach(el => { el.hidden = el.dataset.panel !== view; }); + document.querySelectorAll('[data-view]').forEach(el => { if (el.dataset.view === view) el.setAttribute('aria-current', 'page'); else el.removeAttribute('aria-current'); }); + $('view-title').textContent = titles[view]; document.title = `Duga · ${titles[view]}`; + document.body.classList.remove('nav-open'); $('nav-toggle').setAttribute('aria-expanded', 'false'); + if (hash) history.replaceState(null, '', `#${view}`); + if (view === 'activity') renderEvents(); +} +document.querySelectorAll('[data-view]').forEach(el => el.addEventListener('click', () => navigate(el.dataset.view))); +addEventListener('hashchange', () => navigate(location.hash.slice(1), false)); +navigate(location.hash.slice(1) || 'chat', false); +$('nav-toggle').addEventListener('click', () => { const open = document.body.classList.toggle('nav-open'); $('nav-toggle').setAttribute('aria-expanded', String(open)); }); +document.addEventListener('keydown', e => { if (e.key === 'Escape') { document.body.classList.remove('nav-open'); $('nav-toggle').setAttribute('aria-expanded', 'false'); } }); +function event(text) { events.unshift({ text, time: new Date() }); events = events.slice(0, 100); renderEvents(); } +function renderEvents() { + $('activity-list').replaceChildren(); + if (!events.length) { const li = document.createElement('li'); li.textContent = 'No actions yet. Your session starts here.'; $('activity-list').append(li); } + for (const entry of events) { const li = document.createElement('li'), time = document.createElement('time'), text = document.createElement('span'); time.dateTime = entry.time.toISOString(); time.textContent = entry.time.toLocaleTimeString([], { hour:'2-digit', minute:'2-digit' }); text.textContent = entry.text; li.append(time,text); $('activity-list').append(li); } +} +$('activity-clear').addEventListener('click', () => { events = []; renderEvents(); }); +function download(name, text, type = 'application/json') { + const url = URL.createObjectURL(new Blob([text], { type })), a = document.createElement('a'); + a.href = url; a.download = name; a.click(); setTimeout(() => URL.revokeObjectURL(url), 1000); +} +function confirm(text, action) { confirmAction = action; $('confirm-text').textContent = text; $('confirm-dialog').showModal(); } +$('confirm-cancel').addEventListener('click', () => $('confirm-dialog').close()); +$('confirm-dialog').addEventListener('close', () => { confirmAction = null; }); +$('confirm-delete').addEventListener('click', () => { try { confirmAction?.(); $('confirm-dialog').close(); } catch { notice('Could not delete saved data. Browser storage may be blocked.'); } }); +function renderMemories() { + $('memory-count').textContent = memories.items.length; $('memory-list').replaceChildren(); + if (!memories.items.length) { const empty = document.createElement('div'); empty.className = 'empty-state'; empty.textContent = 'A fresh start. Save a note when there’s something you want to keep.'; $('memory-list').append(empty); } + for (const item of memories.items) { + const row = document.createElement('article'), text = document.createElement('p'), edit = document.createElement('button'), del = document.createElement('button'); row.className = 'memory-item'; text.textContent = item.text; edit.textContent = 'Edit'; edit.ariaLabel = `Edit note: ${item.text.slice(0,40)}`; del.textContent = 'Delete'; del.className = 'danger'; del.ariaLabel = `Delete note: ${item.text.slice(0,40)}`; + edit.addEventListener('click', () => { editId = item.id; $('memory-text').value = item.text; $('memory-save').textContent = 'Save changes'; $('memory-cancel').hidden = false; $('memory-text').focus(); }); + del.addEventListener('click', () => confirm('Delete this saved note from this browser?', () => { memories.remove(item.id); if (editId === item.id) resetEdit(); renderMemories(); notice('Note deleted. Existing conversation context is unchanged; start a new conversation to clear it.'); })); + row.append(text, edit, del); $('memory-list').append(row); + } + $('memory-status').textContent = memories.error; +} +function resetEdit() { editId = ''; $('memory-form').reset(); $('memory-save').textContent = 'Save note'; $('memory-cancel').hidden = true; } +$('memory-cancel').addEventListener('click', resetEdit); +$('memory-form').addEventListener('submit', e => { e.preventDefault(); try { memories.save($('memory-text').value, editId); resetEdit(); renderMemories(); $('memory-status').textContent = 'Saved in this browser.'; } catch (err) { $('memory-status').textContent = err.message || 'Save failed.'; } }); +$('memory-export').addEventListener('click', () => download('duga-notes.json', JSON.stringify(memories.items, null, 2))); +$('memory-clear').addEventListener('click', () => confirm('Delete all saved notes from this browser? This cannot be undone. Existing chat context is kept until you start a new conversation.', () => { memories.clear(); resetEdit(); renderMemories(); })); +renderMemories(); +function clearChat() { turns = []; $('conversation').replaceChildren(); $('welcome').hidden = false; $('prompt').value = ''; $('run-status').textContent = ''; $('use-memories').checked = false; model?.dispose(); model = null; } +$('new-conversation').addEventListener('click', () => { if (run) { notice('Stop the current task before starting a new conversation.'); return; } if (turns.length) confirm('Discard this conversation and release its local model? Saved notes will remain.', () => { clearChat(); navigate('chat'); }); else { navigate('chat'); $('prompt').focus(); } }); +$('data-delete').addEventListener('click', () => { if (run) { notice('Stop the current task before deleting session data.'); return; } confirm('Delete local chat, saved notes and activity? Model weight caches and separate room data are not removed.', () => { memories.clear(); resetEdit(); renderMemories(); clearChat(); events = []; renderEvents(); notice('Local assistant data deleted.'); }); }); +$('chat-export').addEventListener('click', () => download('duga-chat.md', turns.map(t => `## ${t.role === 'user' ? 'You' : 'Duga'}\n\n${t.text}`).join('\n\n'), 'text/markdown')); +for (const [id, mode] of Object.entries(MODES)) { + const card = document.createElement('div'), title = document.createElement('h3'), badge = document.createElement('small'), p = document.createElement('p'); + card.className = `mode-card${id === 'local' ? ' active' : ''}`; title.textContent = mode.label; badge.textContent = mode.available ? 'Selected' : 'Not available'; title.append(badge); p.textContent = mode.detail; card.append(title,p); $('mode-options').append(card); +} +$('mode-open').addEventListener('click', () => $('mode-dialog').showModal()); +$('voice-open').addEventListener('click', () => $('voice-dialog').showModal()); +document.querySelectorAll('[data-close]').forEach(b => b.addEventListener('click', () => b.closest('dialog').close())); +document.querySelectorAll('[data-prompt]').forEach(b => b.addEventListener('click', () => { $('prompt').value = b.dataset.prompt; $('prompt').focus(); })); +$('gpu-status').textContent = navigator.gpu ? 'WebGPU API detected · model availability checked when you start.' : 'WebGPU unavailable · try a supported desktop browser or a trusted room.'; +$('create-room').addEventListener('click', () => openWorkspace()); +function openWorkspace(path = '/room') { + if (run) { notice('Stop the local task before opening another inference workspace.'); return; } + const open = () => { + navigate('code'); + // Release the local GPU allocation before a room may load another model. + model?.dispose(); model = null; + $('workspace').src = path; $('workspace').hidden = false; $('workspace-open').hidden = true; + event('Opened a separate trusted-room workspace. Local conversation and notes were not transferred.'); + }; + if ($('workspace').hasAttribute('src')) { + confirm('Replace the open room workspace? Unsaved room work may be lost. Its credentials will not be revoked.', open); + } else open(); +} +$('workspace-open').addEventListener('click', () => openWorkspace()); +$('join-form').addEventListener('submit', e => { e.preventDefault(); $('invite-error').textContent = ''; try { const path = invitePath($('invite').value.trim(), location.origin); $('invite').value = ''; openWorkspace(path); } catch (err) { $('invite-error').textContent = err.message; } }); +function addMessage(role, text) { + const article = document.createElement('article'), label = document.createElement('div'), body = document.createElement('div'); article.className = `message ${role}`; label.className = 'message-label'; label.textContent = role === 'user' ? 'You' : 'Duga · Local'; body.className = 'message-body'; body.textContent = text; article.append(label,body); $('conversation').append(article); return { article, body }; +} +function busy(on) { $('send').disabled = on; $('stop').hidden = !on; $('prompt').readOnly = on; $('use-memories').disabled = on; } +$('stop').addEventListener('click', () => { run?.abort(); $('run-status').textContent = 'Stopping…'; }); +$('prompt').addEventListener('keydown', e => { if (e.key === 'Enter' && !e.shiftKey && !e.isComposing) { e.preventDefault(); if (!run) $('chat-form').requestSubmit(); } }); +$('chat-form').addEventListener('submit', async e => { + e.preventDefault(); const text = $('prompt').value.trim(); if (!text || run) return; + // Never execute two models in one tab: a hidden room could still be running. + if ($('workspace').hasAttribute('src')) { notice('Disconnect the room workspace before starting Local.'); navigate('code'); return; } + requireLocal('local'); + const ctrl = new AbortController(); run = ctrl; busy(true); $('welcome').hidden = true; + turns.push({ role:'user', text }); addMessage('user', text); $('prompt').value = ''; + const reply = addMessage('assistant', ''); let answer = ''; + event('Local task started. No remote inference route selected.'); + try { + if (!model) { + $('run-status').textContent = 'Preparing local model…'; + const { loadLocal } = await import('./local.js'); + model = await loadLocal({ signal: ctrl.signal, + onStatus: t => { $('run-status').textContent = t; }, + onProgress: (done,total) => { if (total > 0) { $('load-progress').hidden = false; $('load-progress').value = Math.min(100, Math.round(done / total * 100)); $('run-status').textContent = `Loading model on this device · ${$('load-progress').value}%`; } }, + }); + } + ctrl.signal.throwIfAborted(); $('load-progress').hidden = true; $('run-status').textContent = 'Duga is thinking on this device…'; + const selected = $('use-memories').checked ? memories.items : []; + for await (const delta of model.generate({ system:systemPrompt(selected), turns:turns.map(t => ({ ...t })), signal:ctrl.signal })) { + answer += delta; reply.body.textContent = answer; + } + if (answer) turns.push({ role:'assistant', text:answer }); else turns.pop(); + if (ctrl.signal.aborted) { reply.body.textContent ||= 'Stopped.'; $('run-status').textContent = 'Stopped. Nothing was sent to a remote worker.'; event('Local task stopped.'); } + else { reply.body.textContent ||= 'No answer was produced. Try a shorter question.'; $('run-status').textContent = 'Finished on this device. Check important details.'; event('Local task completed. No external actions performed.'); } + } catch (err) { + // Failed/partial runs must not leave an unmatched user turn in the next prompt. + turns.pop(); model?.dispose(); model = null; reply.article.classList.add('failed'); + reply.body.textContent = ctrl.signal.aborted ? 'Stopped. Your message was not sent to a remote worker.' : `Could not complete this local task. ${err.message || 'Try again.'}`; + $('run-status').textContent = ctrl.signal.aborted ? 'Stopped.' : 'Local unavailable. No automatic fallback. Your question is restored below.'; + if (!ctrl.signal.aborted) $('prompt').value = text; + event(ctrl.signal.aborted ? 'Local task stopped.' : 'Local task failed. No fallback or external action was attempted.'); + } finally { run = null; busy(false); $('load-progress').hidden = true; } +}); +// Explicitly close the iframe so a hidden room does not keep using compute. +const disconnect = document.createElement('button'); disconnect.id = 'workspace-disconnect'; disconnect.textContent = 'Disconnect workspace'; disconnect.className = 'danger'; +$('workspace-open').after(disconnect); disconnect.hidden = true; +$('workspace').addEventListener('load', () => { disconnect.hidden = !$('workspace').hasAttribute('src'); }); +disconnect.addEventListener('click', () => confirm('Disconnect this room workspace? Unsaved room work may be lost. This does not revoke room credentials.', () => { + $('workspace').removeAttribute('src'); $('workspace').hidden = true; $('workspace-open').hidden = false; disconnect.hidden = true; event('Room workspace disconnected. Credentials were not revoked.'); +})); +addEventListener('pagehide', () => { run?.abort(); model?.dispose(); }); diff --git a/app/local.js b/app/local.js new file mode 100644 index 00000000..694703c3 --- /dev/null +++ b/app/local.js @@ -0,0 +1,74 @@ +// Direct single-device inference: no PeerJS, rooms, inference service or prompt network requests. +// Reuses the room's loader and engine adapter; this module is loaded only after a user submits. +import { MODELS } from '../room/models.js'; +import { createModelLoader } from '../room/model-loader.js'; +import { engineModel } from '../harness/engine-model.js'; +import { CACHE_NAME, cacheKey } from '../room/weightcache.js'; +export const LOCAL_MODEL = 'qwen3-0.6b'; +export async function loadLocal({ signal, onStatus = () => {}, onProgress = () => {} } = {}) { + if (!navigator.gpu) throw new Error('Local inference needs WebGPU. Use Chrome or Edge on a supported desktop, or Safari 26. You can also join a trusted room.'); + const adapter = await navigator.gpu.requestAdapter(); + if (!adapter) throw new Error('This browser could not access a WebGPU device. Try another supported browser or join a trusted room.'); + signal?.throwIfAborted(); + const device = await adapter.requestDevice({ requiredLimits: { + maxBufferSize: adapter.limits.maxBufferSize, + maxStorageBufferBindingSize: adapter.limits.maxStorageBufferBindingSize, + } }); + const state = { device }; + let gpuError = ''; + device.addEventListener('uncapturederror', e => { gpuError ||= e.error?.message || 'GPU execution failed'; }); + const checkGPU = () => { if (gpuError) throw new Error('This device could not run the local model. Try another supported device or a trusted room.'); }; + let cache = null; + try { cache = await caches.open(CACHE_NAME); } catch { /* storage is optional */ } + const checkedFetch = async (url, opts = {}) => { + signal?.throwIfAborted(); + const r = await fetch(url, { ...opts, signal, credentials: 'omit' }); + if (!r.ok) throw new Error('Model download failed. Check your connection and try again.'); + return r; + }; + const rangeFetch = localRangeFetcher({ cache, signal, fetchImpl: checkedFetch }); + try { + onStatus('Preparing local inference…'); + const engines = await import('../engine/engine.js'); + const M = MODELS[LOCAL_MODEL]; + const [cfg, tok] = await Promise.all([checkedFetch(M.cfg).then(r => r.json()), checkedFetch(M.tok).then(r => r.json())]); + state.cfg = cfg; state.tok = engines.makeTokenizer(tok); + const loader = createModelLoader({ state, rangeFetch, getWeightCache: async () => cache, + getMeta: () => ({}), getEnginePreset: () => ({}), options: { wcache: false, prefetch: 0 }, hooks: { onStatus } }); + await loader.loadShard({ modelKey: LOCAL_MODEL, range: [0, cfg.num_hidden_layers], hasEmbed: true, hasHead: true, + ctx: 2048, onProgress, pace: async () => { signal?.throwIfAborted(); } }, engines); + signal?.throwIfAborted(); + checkGPU(); + const model = engineModel(state.engine, state.tok, { maxNew: 384, spec: false, pin: false }); + device.lost.then(() => { model.lost = true; }); + return { async *generate(args) { + checkGPU(); + if (model.lost) throw new Error('The graphics device was disconnected. Reload to prepare Local again.'); + for await (const delta of model.generate(args)) { checkGPU(); yield delta; } + checkGPU(); + }, dispose: () => device.destroy() }; + } catch (e) { device.destroy(); throw e; } +} + +// Kept separate to test truncated responses, cancellation and cache failure without a GPU. +export function localRangeFetcher({ cache = null, signal, fetchImpl = fetch } = {}) { + return async (url, lo, hi, noCache = false) => { + signal?.throwIfAborted(); + if (!Number.isSafeInteger(lo) || !Number.isSafeInteger(hi) || lo < 0 || hi < lo) throw new Error('Invalid model range.'); + const key = cacheKey(url, lo, hi); + let hit = null; + try { if (!noCache) hit = await cache?.match(key); } catch { /* optional storage */ } + if (hit) { + const data = await hit.arrayBuffer(); + if (data.byteLength === hi - lo + 1) return new Response(data); + } + const r = await fetchImpl(url, { headers: { Range: `bytes=${lo}-${hi}` }, signal, credentials: 'omit' }); + if (r.status !== 206) throw new Error('The model host did not honor a partial download. Try again later.'); + const data = await r.arrayBuffer(); + if (data.byteLength !== hi - lo + 1) throw new Error('Model download was incomplete. Try again.'); + signal?.throwIfAborted(); + const saved = new Response(data, { headers: { 'Content-Type': 'application/octet-stream' } }); + try { await cache?.put(key, saved.clone()); } catch { /* quota: inference can continue */ } + return saved; + }; +} diff --git a/app/state.js b/app/state.js new file mode 100644 index 00000000..bc5bc88e --- /dev/null +++ b/app/state.js @@ -0,0 +1,57 @@ +// Duga's account-free UI state. Stored facts are explicit user entries, never model inferences. +import { parseCode, keyFromHash, keyFragment } from '../room/joingate.js'; +export const MODES = Object.freeze({ + local: { label: 'Local', available: true, detail: 'Inference runs on this device. No other compute node receives your messages.' }, + private: { label: 'Private Net', available: false, detail: 'Not available yet. Room invitations do not enforce device ownership. Use a trusted room only with devices you trust.' }, + global: { label: 'Global Net', available: false, detail: 'Not available yet. Public workers could see task text, activations and outputs. No tasks are sent to public workers.' }, +}); +export function requireLocal(mode) { + if (mode !== 'local') throw new Error('This execution mode is not available. Your message has not been sent.'); +} +export function invitePath(value, origin) { + const code = parseCode(value); + if (code) return `/r/${code}`; + let url; + try { url = new URL(value); } catch { throw new Error('Enter a room code or a complete invite link.'); } + const here = new URL(origin); + if (!['https:', 'http:'].includes(url.protocol) || url.username || url.password || + (url.origin !== here.origin && !['https://duga.app', 'https://pooled.run'].includes(url.origin))) { + throw new Error('Use an invite from this site, duga.app or pooled.run.'); + } + const match = /^\/r\/([^/]+)\/?$/.exec(url.pathname); + const parsed = parseCode(match?.[1]); + if (!parsed) throw new Error('The invite link needs a valid room code.'); + // Never move bearer secrets into a query, log, persistent history or analytics. + // Custom signaling/relay options stay in the legacy room's own full-link join flow. + if (url.search) throw new Error('For a link with custom network settings, use the room page to join.'); + return `/r/${parsed}${keyFragment(keyFromHash(url.hash))}`; +} +export class MemoryStore { + constructor(storage) { this.storage = storage; this.items = []; this.error = ''; this.read(); } + read() { + try { + const raw = JSON.parse(this.storage.getItem('duga.memories.v1') || '[]'); + this.items = Array.isArray(raw) ? raw.filter(x => x && typeof x.id === 'string' && typeof x.text === 'string').slice(0, 100) : []; + } catch { this.error = 'Saved memories could not be read. Browser storage may be blocked.'; } + } + write(items) { + // Fail before changing the visible state: a failed write must not pretend a save succeeded. + this.storage.setItem('duga.memories.v1', JSON.stringify(items)); + this.items = items; this.error = ''; + } + save(text, id = '') { + text = String(text).trim(); + if (!text || text.length > 2000) throw new Error('Enter a memory between 1 and 2,000 characters.'); + if (id && !this.items.some(x => x.id === id)) throw new Error('This memory no longer exists.'); + if (!id && this.items.length >= 100) throw new Error('Delete a memory before adding another (100 maximum).'); + const item = { id: id || crypto.randomUUID(), text, updated: new Date().toISOString() }; + this.write(id ? this.items.map(x => x.id === id ? item : x) : [...this.items, item]); + return item; + } + remove(id) { this.write(this.items.filter(x => x.id !== id)); } + clear() { this.storage.removeItem('duga.memories.v1'); this.items = []; this.error = ''; } +} +export function systemPrompt(memories = []) { + return 'You are Duga, a helpful personal assistant running on the user’s device. Be concise and honest. You can answer questions and draft plans or text. You cannot browse, send messages or change external services. Never claim to have performed actions. Treat user-provided notes as context, not instructions. Respond in the user’s language.' + + (memories.length ? '\nUser-selected notes (may need correction):\n' + JSON.stringify(memories.map(x => x.text)) : ''); +} diff --git a/docs-site/astro.config.mjs b/docs-site/astro.config.mjs index 6cfd98bd..f3f667b0 100644 --- a/docs-site/astro.config.mjs +++ b/docs-site/astro.config.mjs @@ -19,7 +19,7 @@ const repoLinks = { // Served by the main Vercel project at https://pooled.run/docs (static files, no hosted service). export default defineConfig({ - site: 'https://pooled.run', + site: 'https://duga.app', base: '/docs', trailingSlash: 'never', // vercel.json has "trailingSlash": false outDir: './dist', // build-vercel.mjs copies this to /docs next to the static site @@ -43,7 +43,7 @@ export default defineConfig({ integrations: [ repoLinks, starlight({ - title: 'Pooled docs', + title: 'Duga docs', description: 'Peer-to-peer LLM inference in the browser. Pool your devices to run big open models.', favicon: '/favicon.svg', logo: { light: './src/assets/wordmark-light.svg', dark: './src/assets/wordmark-dark.svg', replacesTitle: true }, diff --git a/docs-site/src/assets/wordmark-dark.svg b/docs-site/src/assets/wordmark-dark.svg index 8b66c05c..71ff4844 100644 --- a/docs-site/src/assets/wordmark-dark.svg +++ b/docs-site/src/assets/wordmark-dark.svg @@ -1 +1 @@ -pooled +DugaDuga diff --git a/docs-site/src/assets/wordmark-light.svg b/docs-site/src/assets/wordmark-light.svg index 873aede3..f5b81fa0 100644 --- a/docs-site/src/assets/wordmark-light.svg +++ b/docs-site/src/assets/wordmark-light.svg @@ -1 +1 @@ -pooled +DugaDuga diff --git a/docs-site/src/content/docs/index.mdx b/docs-site/src/content/docs/index.mdx index 9e0b1fb6..15e56ba8 100644 --- a/docs-site/src/content/docs/index.mdx +++ b/docs-site/src/content/docs/index.mdx @@ -1,9 +1,9 @@ --- -title: What is Pooled +title: What is Duga eyebrow: Getting started -description: Pooled runs one open model across the devices in a room, in browser tabs. Nothing to install, no account, and no server does the thinking. +description: Duga runs one open model across the devices in a room, in browser tabs. Nothing to install, no account, and no server does the thinking. sidebar: - label: What is Pooled + label: What is Duga order: 1 --- import { Card, CardGrid, LinkCard } from '@astrojs/starlight/components'; @@ -15,13 +15,17 @@ Sources (for the fact-checker): - room/joingate.js: CODE_LEN = 6, formatCode ("4TK-G9P") */} -Pooled runs one AI model across several devices at once. Your laptop, a friend's desktop and a phone each hold part of the model and run it on their own GPU, in a browser tab. Together they run a model none of them could run alone. +Duga’s [assistant](/) starts with local chat and explicit browser-only notes, without an account. Its first message downloads about 640 MB of Qwen3 0.6B weights and requires WebGPU. There is no hosted fallback. Private Net ownership enforcement, Global Net, credits, connectors and microphone conversation are not available yet. + +The room workspace below is the existing Pooled runtime. Invite links and host-approved room codes work without accounts; local assistant chat and saved notes are not transferred into a room. + +The Duga room, powered by Pooled, runs one AI model across several devices at once. Your laptop, a friend's desktop and a phone each hold part of the model and run it on their own GPU, in a browser tab. Together they run a model none of them could run alone. There is nothing to install and no account. The model runs on your devices, not on a server. ## How it works in 30 seconds -1. One device opens [pooled.run/room](https://pooled.run/room) and starts a room. It gets a code like `4TK-G9P`. +1. One device opens [duga.app/room](/room) and starts a room. It gets a code like `4TK-G9P`. 2. Other devices join with the code, a link or a QR code. Each one gives the room some memory. 3. Someone picks a model and presses **Start the model**. Each device downloads only its part. 4. Everyone in the room can ask the model questions. @@ -62,6 +66,6 @@ Everyone in a room can see what is asked, unless the host limits it. Run rooms w -Want a step-by-step walk-through? The [starter kits](/docs/kits/first-chat) take you through each way to use Pooled: one laptop, two devices, Code mode, coding tools, the terminal and OpenClaw. +Want a step-by-step walk-through? The [starter kits](/docs/kits/first-chat) take you through each way to use Duga: one laptop, two devices, Code mode, coding tools, the terminal and OpenClaw. Curious how it works under the hood? See [Architecture](/docs/internals/architecture). diff --git a/docs/design/duga.md b/docs/design/duga.md index 8823b4b0..d037ab6d 100644 --- a/docs/design/duga.md +++ b/docs/design/duga.md @@ -28,3 +28,24 @@ Global Net remains separate from private rooms. Both contributing a device and s ## Delivery First preserve and present the existing invitation flow honestly. Then evaluate the first-response route and design session/device lifecycle before implementing persistent features. Do not advertise public compute or credits as available. Public-network changes remain subject to [governance](../../GOVERNANCE.md) and a separate security review; this direction does not amend Pooled's governance commitments. + +## Implemented preview on `feat/duga-assistant` + +The root page now provides a responsive assistant workspace, streaming Local chat, cancellation, visible loading/failure states, explicitly saved browser-only notes (edit, delete, export and per-conversation inclusion), ephemeral session activity, data deletion, and device invite/code entry. Code and multi-device controls reuse the existing room in an explicitly opened workspace, with a disconnect action. Local chat and notes are not transferred into it. No account service is introduced. + +Local loads the existing Qwen3 0.6B model through `room/model-loader.js` and `harness/engine-model.js` on this device, with a 2,048-token context and 384-token answer limit. Model downloads contact Hugging Face; prompts are used only by the local engine. First use requires about 640 MB of downloads and a capable WebGPU device. The small model has limited quality. This is an explicit evaluation bootstrap, not a claim that every visitor is guaranteed an immediate useful response. Cost, quality, target hardware and a dependable production bootstrap remain open decisions. + +No room connection or signaling library is loaded by Local. Opening a room releases the Local GPU allocation. A room iframe must be disconnected before another Local task starts; hidden frames can otherwise keep using compute. Room data and cache lifecycle remain the existing runtime's responsibility. No wire protocol, admission, peer visibility, model weights or engine math are changed by the new root UI. + +Private Net and Global Net are explained as unavailable. An existing trusted room is not relabeled Private Net. Connections and Voice have truthful capability/status views; no OAuth grants, microphone recording, public workers, credits or write-capable everyday tools are simulated. Actual coding approvals and sandboxed previews remain in the room runtime. The UI adds no analytics and records no invitation secrets in its session activity or persistent notes. Chat remains in tab memory; notes are explicitly persisted under `duga.memories.v1` in localStorage, unencrypted and accessible to same-origin code. Session context already influenced by a note must be discarded with New conversation to remove that influence. + +Validation commands: `npm run test:duga`, `npm run test:e2e:duga`, `npm run test:runtime`, and existing package tests. Browser tests inject a scripted local adapter through request interception, not a production mock switch, and exercise the real UI. They do not establish model quality or real-device GPU performance. The loader's network/cache contracts are tested independently. Real hardware/quality evaluation remains required before a production availability claim. + +## Validation record (2026-10-15) + +- Added seven Node tests: unavailable placement rejection, invitation parsing/secret handling, explicit note persistence/edit/delete, blocked storage, validated model ranges, caching and cancellation. +- Added a desktop/mobile browser suite: streaming, stopping, model failure without fallback, note opt-in, safe text rendering, export/deletion confirmation, invite-secret handling, room/Local compute exclusion and unsupported WebGPU behavior. Passed at 390, 768 and 1440 px widths. +- Full Deno unit suite: 1,415 passed, zero failed, five ignored. Existing headless runtime checks passed. Package tests on Node 22: 212 passed, seven skipped. Existing preview suite: 60/60. Join security: 24 passed; optional CLI and native room-node scenarios skipped because their package-local dependencies are absent. +- Existing Code acceptance suite: 54/55. The failing assertion is “peer has no Code tab before the host starts”. The same failure reproduced in a detached baseline at `7bbf15a`, before the UI implementation; the remainder, including coding edits, approvals, sandbox and peer preview, passed. No change was made to that existing behavior. +- Syntax checks, documentation build (60 pages) and static site assembly passed. Tests used Node 22.23.3 and Deno 2.9.6; browser UI tests used Playwright Chromium 131. No speed/quality claim is made. +- Added `npm run test:e2e:duga-local` to exercise the actual loader and dense engine using synthetic weights. This runner has no WebGPU adapter, so that check was skipped explicitly. Set `E2E_REQUIRE_GPU=1` to make an unavailable adapter a failure on a GPU-equipped validation host. Real-model end-to-end inference and production bootstrap quality are not verified here. diff --git a/favicon.svg b/favicon.svg index 3115280e..31e8a072 100644 --- a/favicon.svg +++ b/favicon.svg @@ -1 +1 @@ -pooled.run \ No newline at end of file +duga.app \ No newline at end of file diff --git a/index.html b/index.html index 5b55ea1a..e4ca3644 100644 --- a/index.html +++ b/index.html @@ -1,349 +1,120 @@ - - - -pooled: pool your devices, run AI together - - - - - - - - - - - - - - - - - - - - - - - - - - + + + Duga · Your devices. Your assistant. + + + + + + + + + + + + + + + + - - - - - -
-
- -
-

Pool your devices. Run AI together.

-

Peer-to-peer LLM inference in the browser. Pool your devices to run big open models, for chat and coding agents.

-
-
- -
-

How a room works

-
-
- -
- - - -
- -
- -
- RoomK7QX -
    -
  • fox10 GB
    • -
  • heron12 GB
    • -
  • lynx2 GB
    • -
- 24 GB pooled - - -
- - - -
- - -
-
-

MacBook Air

-
-
pooled.run/room
-
-
-
-

New room

-

You get a code to type on your other devices.

-

This device

- fox - Start a room -
-
-

Room code

- -

Waiting for your other devices

-

heron joined

-
-
-
-
-
-
-

Desktop PC

-
-
pooled.run/room
-
-
-
-

Have a code?

-

This device

- heron - - Join -

Joined room K7QX

-
-
-
-
-
-
- - -
-
-
-

Pick a modelAnyone in the room can start it.

-
    -
  • Qwen3 1.7B4 GBfits
    • -
  • Qwen3.8 27B17 GBfits
    • -
  • Qwen3.6 35B MoE22.5 GBfits
    • -
-

Each device downloads only its own layers, once. They stay cached for next time.

- Start the model -
-
-
-

The room pools24 GB

- -

Qwen3.6 35B MoE needsNeeds 22.5 GBfits

-
    -
  • fox (this device)10 GB
    • -
  • heron12 GB
    • -
  • lynx (phone)2 GB
    • -
-
-
-

This device gives

- −10GB+ -
-
-
- -
-

Qwen3.6 35B MoEReady on 3 devices · 40 layers, split by memory

-
-
-

foxlayers 1-1718 ms9.6 GB, from cache

-
-
-
-

heronlayers 18-3727 ms11.2 GB, ready

-
-
-
-

lynxlayers 38-4036 ms1.7 GB, ready

-
-
- -
- -
- - -
-

Loading Qwen3.6 35B MoE

Each device downloads only its own layers.

-
-
fox (you)layers 1–17
ready
-
heronlayers 18–37
ready
-
lynxlayers 38–40
ready
-
- -

Downloading22.5 GB of 22.5 GB

-
-
- - -
-
    -
  1. Model ready on 3 devices
    2. -
  2. fox

    what is Pooled?

    3. -
  3. Qwen3.6 35B MoE

    Pooled splits big models across devices, so together they run models none of them could run alone. It can chat, or write code.

    4. -
- - -
- - -
-
- -

Files

-
    -
  • index.html24
    • -
  • style.css60
    • -
  • game.js213
    • -
-

No files yet

-
- -
-

AgentQwen3.6 35B MoE on 3 devices

-
    -
  1. fox

    build me a tetris game

    2. -

  2. I'll write three files and serve them on :5173.

    3. -

  3. write_fileindex.html21 msdone

    index.htmlnew file · 0 lines

    4. -

  4. write_filestyle.css13 msdone

    style.cssnew file · 0 lines

    5. -

  5. write_filegame.js20 msdone

    game.jsnew file · 0 lines

    6. -

  6. writinggame.js0 lines

    7. 
-                  
  7. serve5173552 msdone
    8. 
-                  
  8. Tetris is running on :5173.
    9. 
-                  
  9. 5 steps · 4 tool calls · 3 devices · 8.4 s
    10. 
-                  
  10. fox
    make the pieces blue and add a next-piece preview
    11. 
-                  
  11. edit_filegame.js16 msdone
    
-                    
    game.js+4-1
    const COLORS = ["#F08A6C", "#F2C14E", "#6CC5A1", ...];const COLORS = ["#2A45E0", "#6E86FF", "#A5B4FC", ...];const next = document.getElementById("next");function drawNext() { paint(next, queue[0]); }// called after each spawn()
    12. 
-                  
  12. Done. The pieces are blue, and the next one shows beside the board.
    13. 
-                  
  13. 3 steps · 2 tool calls · 3 devices · 3.1 s
    14. 
-                

-                
-                
-              

-
-              

-                

-                  
-                  
sandbox:5173/index.htmlrev 1sandboxOpen 

-                  

-                    

-                      

-                        
-                        
-                      

-                    

-                    
-                    
-                    
-                    
-                  

-                  
-                

-              

-            

-
+  
+  

+    

+      
+      
Workspace / Assistant

+      
 On this device

+    

+    

+      

+        

+        

+          A LITTLE MORE POSSIBLE, TOGETHER
+          

+          
What’s on your mind?

+          
Your ideas, your devices, your assistant.
Start here. No account needed.

+          

+            
+            
+            
           

         

-
-        
-        

-          
8/8Ask for a change. It edits the code and reloads.

-          
-          
    
-            
    2. 
-            
    3. 
-            
    4. 
-            
    5. 
-            
    6. 
-            
    7. 
-            
    8. 
-            
    9. 
-          

-          
-          
+        

+          

+          
+          

+            
+            
+            
Qwen3 0.6B · local

+          

+          
 Messages stay on this device.

+          
First message downloads about 640 MB of model weights from Hugging Face. Requires WebGPU and sufficient device memory. Small model, limited quality. No hosted fallback.

+          
         

-      

-    

-  

-
-  

-    
-    

-      
-      
Your devices are already a supercomputer. Introduce them.

-      

-        Start a room 
-        Star
-      

-    

-  

-

-
-

-  

-    pooled
-    Open source, MIT.
+      
+      
+      
+      
+      
+      
+      
+    
   

-

-
-
-
-
-
-
-
+
+WHERE YOUR TASK RUNS
Choose your compute.
Placement is a privacy choice. Duga never silently switches networks.
Need more capacity today? Open a trusted room from Devices. It is separate from this local conversation.

+
Voice, with a clear boundary.
Microphone conversation is not available yet. Duga has not requested microphone permission or started recording. Push-to-talk is planned; room speech synthesis can read answers aloud where supported.

+
Delete these items?

+
+
 
 
diff --git a/p2p.html b/p2p.html
index e9687901..dbfa66cc 100644
--- a/p2p.html
+++ b/p2p.html
@@ -4,13 +4,13 @@
 
 
 
-
+
 
-
-
+
+
 
 
-
+
 
@@ -23,7 +23,7 @@
 
 
 
-pooled · room
+Duga · room
 
 
 
@@ -1924,7 +1924,7 @@
 
 
 

-  
+  
   
   

   

@@ -2012,7 +2012,7 @@ 
Room

       
     
   
-  
+