Scarf v2.17.0

Scarf v2.17.0

The headline: local models. Scarf can now run against Ollama, LM Studio, vLLM, llama.cpp, or any OpenAI-compatible endpoint — with a proper picker, live model discovery, and guardrails so the model you pick actually works. Alongside it, a deep chat reliability overhaul that fixes four ways a chat session could silently stall, and a settings fix that was quietly resetting your dropdowns. This is a feature release, not a patch — but it stays fully compatible with your current Hermes (no version bump required).

Local models: pick a model running on your own hardware

The model picker (Settings → the model row, and the chat's "pick a model" sheet) gains a Remote | Local filter. The Local tab lists Ollama, LM Studio, vLLM, llama.cpp, and Custom endpoint, each writing exactly the config Hermes needs — including the base URL Ollama requires but Hermes doesn't default (without it, Hermes silently falls through to a cloud provider).

• Live model discovery. For a local server, Scarf asks the daemon what's actually installed and lists it — with each model's size and quantization. It distinguishes "the daemon isn't running" from "no models installed yet," so you're never left guessing. This works for a remote server's daemon too, over the same SSH connection — Scarf probes the Hermes host's local endpoint, not your Mac's.
• Context-window gating. Hermes requires a model with at least a 64K token context window, and rejects anything smaller at chat time with an opaque error. Scarf now reads each local model's real context window and blocks sub-64K models in the picker with a plain-language explanation (and a nudge toward the llama3.1 family), so you find out before you pick, not after a cryptic failure. Errors that do come back from Hermes now show the real reason instead of "Internal error."
• Image heads-up. If you attach an image to a model that can't see images, Scarf now warns you in the composer — including for local models, where it reads the daemon's own capability report. (Otherwise Hermes quietly routes the image to a lossy text description, which is where the confusing "this model can't read images / requires a paid subscription" replies came from.)
• Clean switching. Changing provider — local↔cloud or between locals — now scrubs stale endpoint keys so a leftover base_url can't silently misroute your next chat.

One honest caveat: on Hermes versions before 0.18 (and until an upstream fix ships), Hermes's auxiliary features — automatic chat titles, context compression, and image-vision routing — don't yet recognize local providers and fall back to cloud providers or degrade. Your main chat works correctly against the local model; these secondary features are the rough edge. We've filed the upstream fix (hermes-agent #64144); updating Hermes once it lands resolves it.

Chat sessions that don't silently stall

Dogfooding the local-model work surfaced four long-standing ways a chat could wedge or lose output — all now fixed:

• Phantom sessions. Opening a session the server could no longer restore left Scarf attached to a session that didn't exist, so your next message came back as a bare refusal. Scarf now detects the not-restorable response and cleanly starts a fresh session instead.
• The deaf transcript. Re-sending a message identical to one already in your history could cause the entire reply to stream in invisibly — the turn ran on the server, but nothing painted. Fixed.
• The stuck spinner. A chat that wedged on start would spin forever with the sidebar frozen — restart-only recovery. Session starts are now bounded by a watchdog that surfaces a retry, and clicking another chat always supersedes a slow start instead of locking the UI.
• Mid-turn cleanup. Switching away from or deleting a session mid-reply now cancels the turn cleanly instead of orphaning the process — including when you delete the active session from the Sessions tab.

Settings dropdowns that remember your choice

Some Settings dropdowns — Web Tools search/extract backends, a few Advanced controls — saved your selection but reloaded showing blank or the default. The Mac app was carrying a duplicate config parser that had drifted out of sync with the shared one and didn't understand newer keys. We deleted the duplicate (the app now uses one parser everywhere) and added a build-time gate that fails CI if any settable key ever lacks a reader again, so this class of bug can't come back.

Also new

• "Choose model…" on the mismatch banner. When your configured model and provider don't line up, the banner now always offers a one-click path to just pick a valid model from the full picker — not only the two guessed fixes.

Under the hood

• New LocalModelProviders / LocalModelEnumerator / LocalModelConfigPlan in ScarfCore, source-verified against the Hermes v0.17 runtime and covered by an extensive test suite; every provider-switch write path routes through one config planner with crash-safe ordering.
• The session-layer fixes moved the ACP pipe readers off the Swift concurrency pool (they could starve it), added generation-based supersede to the start path, and made session/load failure handling and reconnect load-only (which also stops Hermes from accumulating orphan sessions on reconnect).
• ScarfCore suite: 978 tests; Mac app test bundle: 264. Two upstream Hermes contributions filed this cycle (issue #64144 / PR #64146, plus a review comment on an in-flight alias PR).

Upgrade notes

• Sparkle will offer the update automatically, or use Scarf → Check for Updates. macOS 14.6+ deployment target unchanged.
• No Hermes version floor. Everything here works against Hermes 0.17 (verified). The local-model auxiliary-feature caveat above resolves by updating Hermes to 0.18+ once the upstream alias fix ships.
• iOS / ScarfGo: no new build needed for this release. The Mac local-model and session-layer work compiles in shared ScarfCore but doesn't change iOS runtime behavior. (A separate iOS-only fix — routing the ScarfGo model-preflight write through the same config planner so it scrubs stale keys — is queued for the next TestFlight build.)