Sonderr v1.6 Is Out — The Biggest Update Yet

It’s out. Sonderr v1.6.0 shipped tonight — and v1.6.1, the model picker fix, followed hours later. We called it the biggest update yet for a reason: this is the release where the lab experiments became the product. Here is everything, straight.

For context: Sonderr is the privacy-first AI workspace you install in one command and run entirely from your machine at 127.0.0.1:4173. Free models through Kilo Gateway with no API key, our own local Sonderr-v1 model, real file tools, Studios and opt-in Web3 tooling. v1.6 builds on the hardened v1.5.18 runtime and pushes every part of that story forward. Existing managed installs get it automatically through the required-update flow.

Swarm 2.0 — the room got bigger and smarter

Collaboration rooms now support 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. Every ceiling went up: 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 bigger change is how rooms end. The AI lead now finishes every room with a published synthesis pass — a structured verdict separating agreement from contradictions or gaps, plus up to three recommended next steps. The synthesis renders in the agent room with its own LEAD SYNTHESIS style, and a failed synthesis never blocks the room result. As before, workers stay strictly read-only and the verdict is labeled an AI assessment that never authorizes user actions.

Marathon workflow — tasks that run for days

The U-tier quality budgets now stretch to 72 active hours: U1–U10 map to 6, 8, 10, 12, 15, 18, 24, 30, 48 and 72 hours, with run limits raised to 80–350 bounded provider chunks to sustain them. U7 and above are labeled what they really are — multi-day engineering tasks — so a U10 can span one to three working days.

Long runs only work if the clock is honest. Every Build turn now restates the authoritative time facts — humanized active-work duration (“1d 4h 20m”), remaining budget, task start time and wall-clock span — with explicit instructions to treat those numbers as the only trustworthy clock. Marathon discipline rules require U-tier tasks to keep every edit at the size and care of an H1–H4 pass, produce new evidence in every pass, and checkpoint an honest pause instead of idling, hallucinating progress or manufacturing filler work.

Studios goes RC — and gets a crew

Sonderr Studios graduates from beta to release-candidate. The headline: go specialized, with seven studios — Website, App, Project, Developer and Bounty join forces with the new Document Studio (print-ready PDFs and documents) and Design Studio (posters, brand systems, design kits), each with track-specific milestones and canvas hints.

Then there’s the Studio Crew: every real Studios work turn fields exactly two track-matched specialists — Site Architect + Interface Designer for websites, App Engineer + Experience Designer for apps, Document Architect + Page Designer for documents, Design Lead + Craft Specialist for design, Project Planner + Build Specialist for projects, Code Researcher + Contribution Guide for developer work, Scope Analyst + Safety Reviewer for bounties. The crew is phase-aware — plan, build, review and advise turns hand each worker a different specific task — the lead synthesizes both reports, and the lead, never a worker, performs all edits and checks.

Each studio also carries per-craft playbooks with a definition of done — five professional rules per craft (print-ready output for documents, mobile-first layouts and accessibility for websites, palette discipline and export completeness for design) applied while working and as the final acceptance check. All of it is wrapped in the v1.6 visual polish pass: an animated aurora wash over the announcement hero, a rotating conic halo behind the working otter chip, shine sweeps on crew chips, and hover lift and glow on cards — every effect honoring reduced-motion preferences.

Task lists 2.0 — todos with evidence

Todo items now 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 also requires milestone-sized items for multi-hour and U-tier tasks: each item an independently verifiable outcome sized like a complete H-tier slice. No more “work on it” checkboxes that mean nothing.

The otter is real

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. It does not make the model smarter. It makes waiting better. Reduced-motion preferences are honored, obviously.

The Release Center at /update

Every launch now opens with the v1.6 announcement card — AI-generated key art, release highlights and a direct link to the new Release Center at /update, a standalone hub with the full notes, security summary, the Studios specialization showcase and update instructions. The Sonderr-v1 SLM film no longer interrupts startup; it lives permanently in the sidebar announcements archive.

Security posture — unchanged where it matters

v1.6 keeps the fail-closed release gate, size-capped release metadata fetch and timestamped update backups that anchor the managed-update flow, and documents the full posture on the Release Center. Swarm rooms remain strictly read-only for workers. All of it sits on the v1.5.18 hardened local runtime: escaped markdown code fences in the transcript, randomized temp names, no-follow exclusive file writes and corrupted-store backups instead of silent resets.

v1.6.1 — the model picker fix (already out)

Hours after v1.6.0, v1.6.1 filtered non-chat models out of model discovery: guard and safety classifiers (Llama Guard, Qwen3Guard, ShieldGemma, Nemotron Content Safety, Aegis), moderation and toxicity models, embedders, rerankers, speech and image-generation models no longer appear in the dropdown. Previously, picking a content-safety classifier made every message look unanswered — the model’s verdict rendered as the whole reply. Chat variants, instruct models and reasoning models are unaffected. The filter is a deterministic id-based heuristic (isChatModel in server/provider.js), exported for reuse and covered by its own test suite.

How to get it

Managed installs get it automatically — the required-update flow delivers v1.6 to every install the moment it is verified stable, with timestamped backups before updating. The npm package name and entrypoint intentionally stay unchanged so that flow keeps working for existing installs. Fresh install is still one command:

Full notes, checksums and every past release live on GitHub Releases →

That’s v1.6. Next stops are on the public roadmap — as always, built in the open, shipped when it’s ready.