Sonderr v1.7.3 — the stop update

One automatic workflow — Ask, Plan, Build, and Vision are now one

• The mode picker is gone. The composer shows a single Auto chip: Sonderr answers simple questions directly, plans when you ask for a plan, builds when you ask for work, and handles images automatically — no mode to switch, ever (the Claude pattern). Attaching an image routes the turn through vision automatically, and the model menu prioritizes vision-capable models whenever an image is attached.
• Legacy modes still work through the API for older clients; internal read-only contexts (the built-in docs assistant, team rooms, helper runs) keep their boundaries. The real permission gate was always Tools & Access, and it stays exactly as strict as before.
• Every tool-gate error was rewritten to be mode-agnostic — nothing tells you to "switch modes" anymore.

A real Stop button

• While Sonderr works, the send button becomes a Stop button (square glyph, Esc works too — unless a modal has focus). Stopping aborts the live stream instantly and freezes the run UI at exactly what is on screen: running step cards flip to "Stopped", and an honest chip explains that the summary will appear in history.
• The server wraps the stopped turn up gracefully: one tool-less pass writes a brief honest summary of what was completed, verified, and what remains — saved to history even after the browser stopped listening. A persisted "Stopped · N steps done" chip renders in history too.
• The swarm honors Stop like every other work: workers that had not started are marked skipped (visible on the board with an amber dot and "Stopped · skipped" label), remaining phases are skipped, and the lead closes the room with what the team had gathered.

Tool calling, fixed and faster

• Parallel read-only rounds: when one round asks for several strictly read-only tools (file reads, searches, wallet reads, MCP reads, web research), they now run concurrently instead of one-by-one. Any round containing a side-effect tool stays strictly sequential, so mutations can never race.
• Malformed-argument repair: providers occasionally emit truncated or invalid tool-call JSON. One cheap structural repair (balance braces, drop trailing commas, escape raw control characters in strings) now saves the step instead of failing the call and burning a whole provider round on self-correction.
• Narration rides with the tools: the workflow prompt now instructs narration to be emitted in the same response as the next tool calls — never a whole round spent on narration alone, which was a hidden round-trip tax on every multi-step task. Independent read-only lookups are batched into one response.

Latency, made visible

• The worked chip now reports first-token latency ("Worked for 42s · 7 steps · first token 1.8s") so slow providers are visible instead of mysterious.

The official docs, one click away

• The app now links to the official DXN1 Studios docs site — dxn1-docs.vercel.app — from the topbar (Docs), the welcome screen, the built-in handbook, and the Release Center. /docs/online redirects straight there; the local handbook keeps covering the essentials offline.

Sonderr v1.7.2 — the swarm update

The swarm, top tier

• The inline team card is rebuilt into a live mission board. The title now tells the story in real time — "Team researching · 1/4 done" → "Cross-review underway · 2/4 reviewed" → "Commander is synthesizing…" → "Team run complete · 42s" — and a three-segment phase tracker (Research / Cross-review / Synthesis) advances with the run, pulsing on the active phase.
• Every worker renders as a live row with its avatar, role, current activity, status dot, and an honest elapsed-time chip that ticks once per second while the run is live. Team signal counters (shared findings, teammates helped, risk flags with a high-risk alert tint, votes recorded) fill in as the team works.
• The lead's synthesis lands on the board itself as a verdict preview card when the run completes — no need to open the room to see where the evidence landed. The poll summary stays one click from the full room (Alt+5).
• The room modal gains the same phase tracker under its header, a live elapsed chip in the subtitle while the run is active, and upgraded roster cards showing per-worker elapsed time, token usage, and finding/review counts.

A real room engine underneath

• Rooms now carry a coarse phase (research → review → synthesis → complete) plus honest timestamps: the room stamps startedAt/completedAt, and each worker's clock starts on its first research/review step and stops at completion or failure. The cross-review and synthesis transitions are announced on the lead's timeline in plain words ("First pass is in — the team is now cross-reviewing…"), so the board reads like a mission feed, not a status dump. Restored (archived) rooms backfill the complete phase so old runs render correctly.

Workflow finesse, continued (Claude-grade)

• Spawning a team renders as a proper step card — "Spawn AI team · 4 workers · team poll" — whose running state mirrors the live team ("3 workers · Researching"), instead of an untitled raw tool name.
• The first narration before any tool call is labeled PLAN, matching Claude's plan-first pattern, live and in history.
• A failed step no longer silently returns to "Thinking…": the status line says what happened — "Recovering from read file failure — continuing the plan…" — Claude-style honest recovery.
• The workflow system prompt now tells the lead to treat the live board as the presentation layer: reference the strongest evidence, name disagreements, and lead with the synthesis instead of restating every worker's raw findings.

Sonderr v1.7.1 — the companion polish

The otter pet, rebuilt

• The screen companion was redrawn from scratch: rounder "baby" proportions, a bigger head with large glossy eyes that carry double highlights and now genuinely blink (periodic, staggered per eye), blush cheeks that pop when it is happy, small rounded ears with inner shading, a fluffy two-tone curl tail that sways, a glossy pebble on its belly, a chest fluff tuft, and layered water with soft ripples plus three twinkling star sparkles. Bigger on screen (118x92), crisp vector at every size, all states re-tuned: faster paddling while Sonderr works, a wiggle-and-blush celebration, a slow-drift nap with z's, and full prefers-reduced-motion support.

The working animation is Claude-grade now

• The status line no longer runs a mascot scene. The otter swim scene is retired from the status line (the otter lives as the draggable pet only) and replaced with the Claude pattern: a small four-point spark in the brand gradient that gently pulses and sways beside the shimmer text. Minimal, calm, premium.

Workflow finesse (researched against Claude, Grok, and GLM)

• The status text now crossfades into each new activity instead of hard-swapping.
• Finished tool cards de-emphasize (like Claude's past steps) and their icon chip picks up the state color: green when done, red when failed, blue-tinted while running.
• Workflow turns end with a Claude-style footer chip — "Worked for 12s · 4 steps" — summarizing the run at a glance.

Sonderr v1.7.0 — the workflow update

Token-by-token streaming

• Responses now stream live as the model writes them, Claude-style: OpenAI-compatible SSE is parsed into the same completion shape (content, tool-call deltas reassembled by index, finish reason, usage), and text is pushed to the UI as throttled delta events (~90 ms batches). Deltas are live-only — they are never persisted into message events, so history stays lean; mid-loop narration is preserved separately as small segment events so reloads still show the story between tool cards.
• Bulletproof fallbacks, verified by tests: a gateway that ignores the stream flag and answers JSON is sniffed by content type and parsed directly (streaming remembered off for the rest of the turn, no wasted request); a provider that rejects stream:true with an HTTP 400 falls back to the classic path mid-turn; an unreadable streamed body retries once through the classic path. The local sonderr-v1 model keeps the classic path. The old 429/413/TPM/compaction machinery is untouched and fully re-verified.

Live step cards and narration

• Every tool call renders as a Claude-style step card: a pulsing running state while it executes, a duration badge when it lands (e.g. "2.3s"), and the existing collapsible input/output body. The status line gains a step counter ("Step 3 · Reading files…") and a live elapsed timer chip.
• The workflow system prompt now asks the model to narrate one or two short sentences between tool rounds — what it learned, what it is doing next — so a run reads like a story instead of a silent file dump, and never to paste raw tool output into the reply.

The self-healing loop

• A mid-loop provider failure no longer kills the turn ("lists files then dies" is fixed). The follow-up request retries once on a compacted conversation; if recovery is impossible, the turn ends with an honest progress summary built from the actual completed tool steps ("Completed read_workspace_file · index.html…"), marked incomplete so the checkpoint and conversation persist for resume.
• Hitting a safe limit no longer dead-ends with "say continue": the turn immediately writes a real handoff (what was verified, what remains, the exact next action) via a tool-less synthesis request, with the classic notice as fallback.
• Safe ceilings doubled: up to 40 tool→model rounds (was 24) and 120 tool calls (was 80) per user message.

Announcements and docs

• New v1.7 launch announcement ("Watch Sonderr think. Literally.") with fresh key art, a v1.7 archive entry, an updated Release Center at /update with the workflow story, and version badges updated to v1.7.0.

Sonderr v1.6.2 — the companion update

The otter pet

• Add a ChatGPT-style screen companion, specialized: an otter. It floats on its own little pool anywhere on the workspace — drag it anywhere (pointer events, clamped to the viewport), click to pet it for a burst of hearts, and watch its moods: it paddles energetically while Sonderr works, celebrates when a task list completes, and naps (with z's) after four quiet minutes. Toggle it from the new "Otter pet" sidebar item; the on/off choice, position, and lifetime pet count persist in local storage. The pet is a hand-built vector (crisp at every size), ships on by default for new sessions, and honors prefers-reduced-motion.
• The app now emits sonderr:pet events (working, idle, celebrate, meet) from the send flow and task-list rendering; the v1.6.2 announcement's "Meet the otter" button turns the pet on and throws a celebration.

Working animation, rebuilt

• Replace the raster otter chip (a static PNG bobbing in a circle) with a hand-built vector swim scene: an otter gliding between two moving wave layers, tail flicking, bubbles rising — every part animates on smooth transform/opacity curves with proper easing. The status text gets a Claude-grade shimmer sweep. Both honor reduced-motion preferences.

Bulletproofed workflow

• Empty final responses no longer render as a silent blank bubble: the turn ends with an honest inline notice explaining what happened and how to fix it.
• The chat handler refuses classifier/guard models in the chat slot server-side (mirroring the v1.6.1 picker filter): the turn ends immediately with an actionable explanation instead of streaming a verdict like "User Safety: safe" that reads like silence.
• The model picker flags any model whose id looks like a classifier with a "classifier" badge, and choosing one shows a heads-up toast.
• A stuck-stream watchdog: if a running turn produces no events for 75 seconds, the status line says so honestly ("Still working — the model is taking longer than usual…") instead of leaving an indefinite spinner. The watchdog timer is cleared on every exit path and can never leak.

Announcements and docs

• New v1.6.2 launch announcement with generated pet key art; the v1.6.0 announcement moves to the announcements archive alongside the Sonderr-v1 film, and the Release Center at /update gains an "Otter pet" section with new generated art (pet scene and glide scene) plus the bulletproofing story.

Internal

• Bump the application version to 1.6.2. New test/companion.test.js suite pins the pet/animation/bulletproofing wiring (22 test suites total); npm test runs it as part of the standard suite.

Sonderr v1.6.1 — the model picker fix

The model picker fix

What happened: Sonderr's model discovery listed every catalog entry as a chat model — including safety classifiers and guard models. If you picked one (for example a Nemotron Content Safety model), every message looked unanswered: the classifier did its job and replied with its verdict — e.g. User Safety: safe — which rendered as the whole assistant message.

The fix: model discovery now filters non-chat models out of the chat picker:

• Safety and guard classifiers (Llama Guard, Qwen3Guard, ShieldGemma, Nemotron Content Safety, Aegis)
• Moderation, toxicity, and sentiment models
• Embedders, rerankers, and classic encoders (BERT, CLIP)
• Speech models (Whisper, TTS, STT) and image-generation models (DALL-E, Stable Diffusion, Flux)

Chat-tuned models stay — including chat Nemotron instruct variants. The filter is a deterministic id-based heuristic (isChatModel in server/provider.js), exported for reuse and covered by a dedicated test suite.

Updating: existing 1.6.0 installs can use the in-app one-click updater (npm package name and entrypoint unchanged), or pull the tag directly. Full v1.6 notes live in the in-app Release Center at /update and in the v1.6.0 release.

Sonderr v1.6.0 — the biggest update yet

v1.6.0 — The biggest update yet

Announcements and visuals

• Every launch now opens with the Sonderr v1.6 announcement card (AI-generated key art, release highlights, and a direct link to the new Release Center); the Sonderr-v1 SLM film no longer interrupts startup and remains permanently available in the sidebar announcements archive.
• Add a dedicated Release Center page at /update — a standalone hub with the full v1.6 notes, security summary, Studios specialization showcase, and update instructions, linked from the launch announcement, the archive, and the README.
• Generate v1.6 art assets: key art with the otter companion, an otter mascot portrait, and Swarm 2.0 network art (web/assets/sonderr-v16-*.png), plus new Studios visuals — the Studios hero scene, the two-specialist crew art, and the four-craft board (web/assets/sonderr-studios-*.png).
• Add the otter companion: while Sonderr works — including through the final streamed response — the status line shows a gently swimming otter chip with a ripple animation and "Sonderr is working…" text, honoring reduced-motion preferences.
• Update the sidebar brand to v1.6.0 and refresh Studios branding to release-candidate status.

Swarm 2.0

• Collaboration rooms support up to six read-only workers (previously three) drawn from sixteen personas across eight roles; Builders, Testers, Documenters, and Strategists join the roster, and saved rooms restore their full-sized rosters.
• The AI lead now ends every room with a published synthesis pass — a structured verdict separating agreement, contradictions or gaps, and up to three recommended next steps; a failed synthesis never blocks the room result.
• Raise collaboration ceilings: six shared findings per worker, four risk flags per worker (sixteen per room), twelve live help threads, a 160-entry transcript, and two anonymous helper calls per room.
• The synthesis entry renders in the Alt+5 agent room with a distinct "LEAD SYNTHESIS" style, and the room subtitle mentions the synthesis pass.

Marathon workflow (S1–U10)

• Extend U-tier quality budgets to a 72-active-hour maximum (U1–U10 = 6, 8, 10, 12, 15, 18, 24, 30, 48, 72 hours); U7 and above are labeled multi-day engineering tasks, so a U10 can span one to three working days.
• Raise U-tier run limits to 80–350 bounded provider chunks to sustain marathon tasks.
• Restate authoritative time facts on every Build turn: humanized active-work duration (for example "1d 4h 20m"), remaining budget, task start time, and wall-clock span — with explicit instructions to treat these numbers as the only trustworthy clock and never estimate elapsed time from memory.
• Add marathon discipline rules: U-tier tasks keep every edit the size and care of an H1-H4 pass, must produce new evidence in every pass, and checkpoint an honest pause instead of idling, hallucinating progress, or manufacturing filler work.
• Update the long-running-work skill, quality_checkpoint schema text, and Build/Ask system prompts to match the 72-hour scale and the anti-idle, anti-hallucination rules.

Task lists 2.0

• Todo items support an optional evidence note: the next concrete action for open items or the verifying evidence for completed ones, persisted through the store sanitizer and rendered inside the live task card.
• Todo guidance now requires milestone-sized items for multi-hour and U-tier tasks — each item an independently verifiable outcome sized like a complete H-tier slice.

Studios

• Sonderr Studios graduates from beta to release-candidate branding across the sidebar and Studios home.
• Go specialized: add the Document Studio (print-ready PDFs and documents) and Design Studio (posters, brand systems, and design kits) alongside Website, App, Project, Developer, and Bounty studios — seven specialized starting points with track-specific milestones and canvas hints, plus a new generated craft board and Studios hero art.
• Add the Studio Crew: every real Studios work turn fields exactly two track-matched specialists (Site Architect + Interface Designer, App Engineer + Experience Designer, Document Architect + Page Designer, Design Lead + Craft Specialist, Project Planner + Build Specialist, Code Researcher + Contribution Guide, Scope Analyst + Safety Reviewer). The crew is phase-aware — plan, build, review, and advise turns hand each worker a different specific task for what needs to happen right now — the lead synthesizes both reports, and the lead — never a worker — performs all edits and checks. The studio rail shows the crew with their live phase chip, per-phase assignments, and a two-specialist phase art card, and the spawn_subagents tool is always supplied for Studios work turns.
• Add per-craft playbooks with a definition of done: each studio carries five professional rules (for example print-ready output for documents, mobile-first layouts and accessibility for websites, palette discipline and export completeness for design) that the lead applies while working and as the final acceptance check; the studio rail shows the craft standard chips and the definition of done inline.
• Add a "What the crew makes" showcase to the Studios home — generated deliverables flat-lay art (bound report, design posters, site on laptop and tablet) with the four output chips — plus new phase-aware crew art (two specialists advancing through plan, build, and review stages) and a cinematic v1.6 finale hero on the Release Center.
• Premium visual polish: animated aurora wash over the announcement hero, a rotating conic halo behind the working otter chip, shine sweeps on crew task chips and the announcement CTA, hover lift and glow on crew cards, craft chips, path cards, and hero art, and a sheen sweep on the deliverables showcase — all honoring reduced-motion preferences.

Security

• Keep the fail-closed release gate, size-capped release metadata fetch, and timestamped update backups that anchor the managed-update flow; document the v1.6 posture on the Release Center.
• Swarm rooms remain strictly read-only for workers, and the synthesis verdict is labeled an AI assessment that never authorizes user actions.

Internal

• Bump the application version to 1.6.0. The npm package name and bin/sonderr-1.5.js entrypoint intentionally stay unchanged so the managed one-click updater and install paths keep working for existing 1.5.x installs.

---

Install / update: run the installer or use the in-app updater — the npm package name and entrypoint are unchanged, so existing 1.5.x installs update in one click.

Full notes: see the in-app Release Center at /update.

v1.5.18 — hardened local runtime

v1.5.18 — hardened local runtime

Security

• Escape model and file content inside fenced markdown code blocks in the frontend renderer, closing a stored XSS path in the chat transcript and artifact previews.
• Randomize temporary file names for secret-store and session-data writes so symlink attacks cannot target predictable temp paths.
• Replace O_TRUNC with O_EXCL on temp file opens and add failure cleanup so partial writes cannot silently corrupt the data store.
• Back up corrupted data.json and credentials.json to timestamped .bak files before resetting, instead of silently dropping all sessions and settings.
• Fail-closed on platforms without fs.constants.O_NOFOLLOW instead of silently disabling symlink protection.
• Preserve existing temp-file safety patterns already used by the MCP config writer as the model for all other local-file writes.

Reliability

• Validate Uniswap V3 addresses at startup and in prepareSwap so misconfigured or stale contract pins fail fast with a clear error.
• Add NFC normalization to intent regexes in safety.js so Unicode homoglyphs cannot bypass authorization boundaries.

Frontend

• Replace the agent status spinner with Claude-style animated thinking dots in lightblue.
• Ensure the model delivers finished work as real workspace files with download cards instead of pasting long content into chat.

Tests

• Add test/web-ui.test.js covering renderMarkdown XSS regression, inline code safety, link rendering, esc edge cases, and safeExternalHref security boundaries.
• Expand test/store.test.js with corruption-handling and UUID temp-file naming coverage.

Sonderr v1.5.17 — Safer tools, sharper teams

Sonderr v1.5.17

More capable teams. Clearer boundaries. Safer work.

Sonderr is a localhost AI workspace built for engineering. This release brings the agent room into the conversation, makes the tools and mode boundaries clearer to every model, and adds another layer of checks around actions that can change data or run code.

▶ Watch the Sonderr-v1 launch film · Browse the source · Read the full changelog

---

A team, right inside the chat

Launch a research team without leaving the current conversation. The commander can coordinate up to three focused, read-only workers, compare their evidence, collect help requests and risk flags, and run a lead-authored poll. Open the live Agents room with Alt+5 to follow assignments, discussion, profiles, and votes. Worker votes stay separate from your vote; workers cannot edit files, run commands, call integrations, or take actions for you.

The model knows what it can actually do

Task mode and Tools & Access now have distinct meanings throughout tool selection, model guidance, and server-side checks:

• Ask answers, inspects supplied read-only sources, and researches the public web. It does not edit files or run project scripts.
• Plan creates a read-only implementation plan.
• Build handles requested workspace changes. Full PC access is also required for shell commands and project checks.
• Vision handles image understanding and, when enabled, image generation or edits.

The model uses the schemas supplied for the current request, explains the specific missing mode or access setting, and does not claim it can switch the selector for you. A denied action cannot be retried through another tool or shell route.

Security and reliability

This release tightens several boundaries across the app:

• Server-side mode and access checks back up schema filtering for workspace writes, scripts, MCP, wallet operations, Studio updates, email drafts, images, and task-private state.
• Actions that change data or execute commands require affirmative intent in the current message. Negated instructions and how-to questions do not authorize a file edit, email draft, saved-lead update, image edit, or command.
• Project verification routes to the bounded project-check runner; general shell access is exposed only for a direct command request with Full PC access.
• Workspace writes and reads reject symlinks into protected paths. Public web and MCP requests pin validated DNS answers and block unsafe destinations; local config and memory writes use safer file handling.
• Session/view guards reduce cross-chat UI races, and provider responses are constrained to the active tool schemas with structured arguments.

These safeguards reduce specific attack and failure paths; they are not a claim that any model or security boundary is infallible. See SECURITY.md for reporting guidance.

Sonderr-v1 stays small by design

Sonderr-v1 is our first model: a 0.6B-parameter SLM specialized for the Sonderr environment. The model remains small; tools, web research, and the workspace supply task-specific reach. We intend to keep Sonderr’s model line SLM-scale and improve it over time.

The verified Q4_0 model and compatible Linux CPU runtime remain available from v1.5.16’s model assets; this release does not replace or retrain the weights. The installer continues to verify those pinned assets before installing them locally.

Get the release

Install or update with the Sonderr installer. The managed app checks the latest stable release and, when required, installs the exact release tag before reopening the workspace. Source checkouts remain development checkouts.

Verification

• npm ci
• npm run check
• npm test
• npm audit --omit=dev
• npm pack --dry-run

Sonderr v1.5.16 — AI swarm and verified Q4 runtime

Sonderr 1.5.16 — The read-only AI swarm

Sonderr rooms now support an accountable lead and up to three task-focused workers. They can share findings, ask and answer directed help requests, flag evidence-backed risks, cross-review one another, and vote in the lead's poll. The human vote is shown separately. The Alt+5 team room includes live activity, profiles, and saved-room recovery. Worker calls remain read-only, and their conclusions are presented as AI assessments for review.

Workspace and static-file access now checks canonical paths to block symlink aliases into protected data or outside allowed roots. Skill loading and small-model tool routing are also improved.

Sonderr-v1 Q4 local runtime

Compatible Linux x64 devices can install a SHA-256-verified Q4_0 model and bundled llama.cpp CPU runtime (about 370 MB total). The installer checks the downloaded model and runtime against release-pinned SHA-256 values before installing. It requires glibc 2.34+, GLIBCXX 3.4.31+, OpenSSL 3, and libgomp. Other platforms use the existing merged Transformers model and CPU runtime.

The Q4 model is derived from the merged Qwen2.5-0.5B-Instruct checkpoint, distributed under Apache-2.0. The model card and upstream license are attached. The bundled llama.cpp runtime includes its MIT license. No general capability benchmark is claimed.

Verification

• npm ci
• npm run check
• npm test
• npm audit --omit=dev — 0 vulnerabilities
• Fresh-home Q4 installer verification using the exact pinned model and runtime assets
• Extracted llama.cpp runtime Q4 inference smoke test (HTTP 200)
• Agent-room coordination smoke test