Sonderr Docs
The Sonderr glossary
Every term this product uses, defined the way we mean it — including the ones that sound grander than they are. If a word in the workspace ever feels fuzzy, its honest definition lives here.
Sonderr tries to use plain words for plain things. Where a term could oversell — personas, for example — the product says so in the interface, and this glossary says so here. Definitions are one honest paragraph each, grouped by the part of the product they belong to, with links to the guide that goes deep.
1 · Core surfaces & modes
| Term | What it means |
| Workspace | The browser interface served by your local runtime at 127.0.0.1:4173. The tab is the product; the terminal command only starts the engine underneath. |
| Task mode | What your composer message is allowed to cause: Ask (explain, analyze), Plan (shape work, touch nothing), Build (make changes), Vision (image understanding plus edit_image). Mode and access are separate controls. |
| Vision mode | Image understanding — OCR, screenshots, charts, error dialogs — plus image generation and editing via edit_image. Attached images switch it on automatically. |
| Live task card | The chat-native card that carries active work: plan, progress, evidence and deliverables attached to the step that produced them. |
| Studios | The goal-to-deliverable surface: seven tracks, milestone boards, and a definition of done per craft. Release-candidate since v1.6; where the v1.7 workflow release lands first. Capital-S Studios is the surface; a lowercase studio elsewhere in this site means DXN1-Studios, the company. |
| Crew | The two track-matched specialists that field every real Studios turn — for example Site Architect + Interface Designer on the website track. Phase-aware: plan, build, review and advise turns hand each worker a different specific task. |
| Collaboration room | A research crew in your chat: one lead with write access, joined by read-only workers who investigate and report. Six workers since Swarm 2.0, saved rooms restore intact. |
| Persona | A labeled fictional AI character on a room roster — a role with a brief, not a second model with special powers. The UI says this and the docs insist on it. |
| Synthesis pass | The structured verdict a lead publishes at the end of every room: agreements, contradictions, gaps, and up to three recommended next steps. An AI assessment, not a permission slip. |
2 · The workflow engine
| Term | What it means |
| Checkpoint | A durable resume point written by substantial Build work: goal, current milestone, verified evidence, decisions made, and one exact next action. Powers resume-after-crash and the Resume interrupted task offer. |
| Compaction | Automatic shrinking of an oversized provider conversation: the original request, recent activity and the latest checkpoint carry forward; bulky old tool outputs drop. Compactions are untrusted hints — the agent re-checks reality instead of trusting the summary. |
| VECT memory | The editable memory system behind Sonderr’s persistence — what lets work span days instead of dying with the context window. |
| Session | A conversation with its workspace state: task cards, uploads, task notes. Stored under ~/.sonderr/ on your machine. |
| Marathon task | A long-running Build task — up to 72 hours — held together by checkpoints and bounded provider chunks rather than one continuous context. |
| Brief | What you type when you’re doing it right: outcome, audience, constraints, references. See Prompting & task briefs. |
| Definition of done | Five professional rules per Studios craft, applied while working and re-run as the final acceptance check. The difference between “the agent stopped talking” and “the work meets the bar.” |
| Milestone board | The Studios view of a project: goal, milestone slices, current phase, deliverables as they land. Mutations respect task-mode gates; checkpoints tie into it. |
3 · Capabilities & doors
| Term | What it means |
| Skill | One of 65 markdown playbooks in the runtime’s skills/ directory — when it applies, the rules to follow, the definition of done. Matched to your task and loaded at the moment they matter, counted by a manifest so the runtime never has to guess. |
| Artifact | A deliverable produced by the agent: a file, presented in chat as a download card and previewed in the artifact panel. Real file on your machine, not a cloud blob. |
| Download card | The chat representation of a finished deliverable, served by the sandboxed local API. Clicking it opens the panel; the card is backed by the actual file or it doesn’t open. |
| Artifact panel | The side column (full-screen on mobile) with live previews: rendered markdown, syntax-styled code, pretty-printed JSON, CSV/TSV as a table, images on a checkerboard, binaries download-first. Copy and download in the header; ESC closes. |
| Upload cap | The 20 MB per-file limit on uploads, stored per-session under .sonderr/uploads/<session>/ with name de-duplication. Binaries attach as readable references, never garbled context. |
| MCP | Model Context Protocol — the standard way to extend the agent with external tools. Local and remote servers, stdio and HTTP transports, DNS-pinned endpoints, present-turn authorization for anything that mutates. |
| Plugin Hub | The surface for connectors: Gmail, Notion, Slack, Drive, GitHub and Linear — plus send-side email that drafts in the open before anything is sent. |
| Connector | One of the six integrations above, wired through the Plugin Hub. Read and draft are the default verbs; sending is always your call. |
| Wallet profile | The opt-in Web3 configuration for EVM and Solana. Actions are prepare-then-confirm; keys never render in chat. |
| Free-tagged model | A model you can use without an API key — the fastest path from install to a running task. Model discovery tags capabilities so the picker filters honestly. |
4 · Trust & operations
| Term | What it means |
| Local-first | The architecture rule that your sessions, files and credentials live on your machine, and the heavy lifting happens there too. See Your data, your machine. |
| Data map | The complete list of files Sonderr touches: ~/.sonderr/ for your data, the app folder under ~/.local/share/, the launcher link in ~/.local/bin/. Fully drawn in the CLI reference. |
| Update gate | The fail-closed rule since v1.5.10: a managed install checks GitHub Releases on start, and a newer stable blocks the workspace until it’s installed. No “skip this version.” The check sends GitHub your IP and request metadata — nothing else. |
| Sandboxed local API | The workspace’s own HTTP interface that serves artifact previews and download cards. Local by construction — it is not a tunnel to anywhere. |
| Workspace sandbox | The boundary around what agent actions can touch on your machine, enforced alongside task modes and the explicit-request rule for shell commands. |
| Sonderr-v1 | The optional local model: a SHA-256-verified Q4_0 GGUF plus a bundled llama.cpp CPU runtime, about 370 MB, installed under ~/.sonderr only after asking. |
| Secret scrubbing | The pass that keeps credentials out of places they shouldn’t render — wallet keys never appear in chat, credentials stay owner-only on disk. |
| Present-turn authorization | The rule that a permission is granted in the message where it’s used, not carried forward from an older one. A stale “yes” doesn’t authorize anything. |
5 · Why the vocabulary is strict
Half of these definitions contain a disclaimer, and that’s deliberate. An agent product earns trust by saying exactly what a thing is — personas are labeled fictional, synthesis is an assessment, compactions are hints — instead of letting the names promise more than the machinery delivers. The glossary is where that honesty gets enforced in writing. If you ever find a term used loosely in the product or in these docs, that’s a bug worth reporting like any other: it goes to the same inbox, and it gets treated the same way.