Mightling v1.6.0

Mac desktop app (preview, unsigned)

Mightling-1.6.0-arm64-preview.dmg (Apple silicon) and Mightling-1.6.0-x64-preview.dmg (Intel) are a preview of the desktop app for macOS 12 or later, a client of a Mightling node on your network. They are signed ad hoc only, with no Apple Developer ID and no notarization, so macOS refuses to open the app the first time. Drag Mightling to Applications, try to open it once, then allow it in System Settings → Privacy & Security → Open Anyway (on macOS 14 and earlier, Control-click the app and choose Open instead). Or, in Terminal: xattr -dr com.apple.quarantine /Applications/Mightling.app. Check the dmg against SHA256SUMS first; steps in docs/desktop.md.

Search your own documents. Mightling now keeps a private index of your files and the agent can search it. ~/Documents and ~/Downloads are indexed by default; add other folders with ling docs add <folder> --name <name>, and remove any of them, the defaults included, with ling docs remove <name>. In a session the agent searches with docs_search, reads a passage with docs_read, and cites the file and the page or lines it used. From a shell: ling docs search "notice period acme", ling docs read <id> --page 7, ling docs status.

• What it reads in this release: PDF (text layer), Markdown, reStructuredText, Org, LaTeX and plain text. Scanned PDFs are recorded as needing OCR and counted in status; Word, HTML, mail and OCR come in a later release.
• Every language. The embedding model is multilingual (snowflake-arctic-embed-m-v2.0), keyword search covers Chinese and Japanese, and right-to-left PDFs are read in order.
• Nothing leaves the machine. The index opens no network connection at any step, and ling-admin audit egress --docs proves it. /airgapped on does not turn it off.
• Safe to point at a Downloads folder. Installers, archives, disk images and videos are never opened; secret-looking files and folders are skipped; each PDF is read in a sandbox with no network and a 1 GB memory cap, so a malformed or hostile file fails alone. Indexing shares the code index's memory budget and yields to the model server.
• What it found on our test set: the right passage in the first 10 results for every one of 56 questions about 96 documents, in under 20 ms a query; 50,000 passages still answer in under 100 ms.
• On a GB10 node the installer downloads what the index needs (PDFium, ONNX Runtime and the embedding model, about 320 MB, each checked against a pinned checksum). On a node installed earlier, run ling-admin docs setup once after updating. Linux clients receive ling-docs but not these files yet, so for now the file index is a node feature; macOS and Windows clients do not have it.
• To switch it off: mightling_docs = false in dreamference.toml, or DREAMFERENCE_MIGHTLING_DOCS=0.

A new web UI: ling web. ling web now serves Mightling's own UI, built into every ling (until now it served a placeholder page). It opens on Ask, for questions that are not about a repository, with Work, the agent sessions, one click away.

• Ask threads: "New question" starts one; each thread has a folder of its own under ~/.mightling/ask/ and the agent's sandbox. Threads can be searched, renamed and archived, and every answer has a copy button.
• Attachments: the clip button, or paste a screenshot. Images go to the model (Qwen3.8 reads them); other files are placed in the thread's folder for the agent to open. Up to 20 MB an image and 100 MB a file.
• On a phone: below 720 px the thread list becomes a drawer, and the page fits without zooming or scrolling sideways. Sign a phone in with ling web pair, which prints an 8-digit code (one use, ten minutes).
• Signing in: ling web open opens a one-time link on this machine. Nothing a command run by the agent can read is a credential.
• On a node shared with your network, ling-admin node enable now starts ling web on the LAN (port 3100) and advertises it; ling-admin node enable --no-web keeps it on loopback.
• The desktop app's Chat is now Ask. On Linux and macOS, the app's Chat entry (menu label Ask) loads this UI from ling web on the same machine, starting ling web serve itself when none is running and stopping it on quit. On a client, Ask runs on the client against the node's model server. ling app opens it (--ask; --chat still works). On Windows, which has no ling web yet, Ask opens Work.
• The previous web chat is still there. It keeps running on port 3000 and is managed with ling-admin chat … as before. The image gallery, voice input and the Apps settings page are not in the new UI yet.

Ask from your phone: Signal, Matrix and Telegram (preview, off by default). A message you send becomes an Ask thread on the node, and the answer comes back in the same chat. Every messenger is off until you run its setup on the node: installing or updating Mightling turns none of them on, starts nothing and opens no port. All three need ling web running on the node (ling web start).

Messenger	Privacy	Turn it on	Turn it off
Signal	End-to-end encrypted; your own account, in Note to Self	ling signal setup	ling signal remove
Matrix	Your own homeserver on the node, reached over your tailnet	ling-admin matrix start, then ling chat start	ling chat stop, ling-admin matrix stop
Telegram	Less private: messages pass through Telegram's servers unencrypted	ling chat telegram setup, then ling chat start	ling chat telegram off

• Signal links the bridge to your own account as one more device, the way Signal Desktop is. You write in Note to Self; every other conversation is dropped unread. The bridge runs as a system account of its own (mightling-signal), so the agent can never read your Signal keys; signal-cli and its Java runtime are downloaded at setup, pinned by SHA-256. ling signal setup --dry-run prints every change first. A linked device holds keys to your whole account: read the Signal section of the messengers guide before linking.
• Matrix runs a private homeserver on the node, on a Docker network with no route out, and your phone reaches it through Tailscale (which you install and sign in to yourself). Use Element X.
• Telegram shows a warning and asks you to type yes: answers can carry what the agent read for you, such as mail, files and code, through Telegram's servers.
• The air gap applies. At /airgapped on, Signal sends nothing, not even a read receipt, and Telegram pauses; Matrix keeps working, because its homeserver never leaves the node.
• The egress audit says so. While a bridge is on, every ling-admin audit egress report names it as a declared exception.
• Why preview: the Matrix and Telegram bridge is tested against stand-ins of Telegram's Bot API and a Matrix homeserver, end to end through a real ling web; the Signal bridge is tested against a scripted signal-cli and a stand-in ling web; none has yet been exercised against a real Signal account, a real Telegram bot or a real homeserver with Element X, which is why each keeps its preview label. Full guide: the Phone messengers page of the documentation.

Refine mode stays opt-in. ling --refine (also ling exec --refine, [night] refine = true, or mightling_refine = true in dreamference.toml) first studies the task in a session that changes nothing, then does it in a fresh one. On a fixed 100-task sample of SWE-bench Verified, run on one DGX Spark, it resolved 69/100 against 68/100 without it (the two arms each won tasks the other lost: 10 against 9, McNemar p = 1.0), at about 2.2 times the time per task (a median of 13.9 minutes against 6.5), so it remains off by default. It can still help on tasks whose description is thin or misleading.

Measured: 68 of 100 SWE-bench Verified tasks. With the default configuration, 1.5.1's agent and Qwen3.8-27B resolved 68 of a fixed 100-task sample of SWE-bench Verified on one DGX Spark, offline, one attempt per task. That corresponds to an estimated 69% on all 500 tasks (95% band 63.9–73.4%), estimated from the per-task results of every public submission (scripts/swe_bench_compare.py). It is our own measurement, not a leaderboard entry. The same sample with refine mode on resolved 69, one more, at 2.2 times the time (above). On 50 fresh SWE-bench Verified tasks, a test-discipline prompt rule resolved 33 against the default's 34 (no measurable difference), so the default task prompt stays unchanged.

The installer installs the two new programs on Linux: ling-docs, linked onto your PATH, and ling-signal, placed beside ling with no link and not started. On a GB10 node it also runs ling-admin docs setup (skipped with --no-model). The unattended node install is otherwise as in 1.5.1.

Mightling for macOS, as a preview. The release now carries the desktop app for Macs: Mightling-1.6.0-arm64-preview.dmg for Apple silicon and Mightling-1.6.0-x64-preview.dmg for Intel, for macOS 12 or later. On a Mac the app is a client of your GB10: Ask and Work both run the ling bundled in the app against the node's model. It finds the node on your network by itself, as ling does (run ling-admin node enable on the GB10 first). The file index is not available on a Mac.

It is not signed with an Apple Developer ID and not notarized (signed ad hoc only), so macOS will not open it the first time. Check the dmg against the release's signed SHA256SUMS, drag Mightling to Applications, then:

• try to open it once, and click Open Anyway in System Settings → Privacy & Security;
• on macOS 14 and earlier you can instead Control-click (or right-click) the app, choose Open, and confirm;
• or, in Terminal: xattr -dr com.apple.quarantine /Applications/Mightling.app.

Allow it when macOS asks whether Mightling may find devices on your local network; that is how it reaches the GB10. The preview does not update itself, and ling app in a Mac terminal does not open it yet. Full steps: the Desktop app page of the documentation.

Faster decoding: 12 draft tokens. The DFlash2 drafter now proposes 12 tokens per step instead of 16. Measured on replayed agent sessions: +7% single-stream decode (48.3 against 45.1 tokens/s, acceptance 5.23 against 4.99 per step) and neutral with two agents at once; the output is unchanged, because speculative decoding preserves the model's distribution.

A benchmark run now has priority over the model server. ling-admin server start puts a small gate container in front of the engine on the public port. While a SWE-bench run holds the gate, a request that is not the run's is answered 503, naming the run and its time left, instead of slowing it down; ling-admin night pause [--for 2h] lets you through meanwhile. With no run, nothing changes.

Code Mode is on, and on this server it changes nothing yet. ling offers the model Code Mode's exec tool (a program that calls several tools in one turn) beside the ordinary tools. The model server renders only ordinary function tools into the prompt, so the model never sees exec and works through the ordinary tools, as it did in 1.5.1 (measured on 2026-10-10: eleven turns, no call to exec). Nothing breaks, nothing to configure; Code Mode becomes real when a server passes custom tools through, which ling-engine now does.

Paired nodes are found again after the rename, and always where they are now. A node paired under Puffin "did not answer over the pairing" after the switch to Mightling: its host key was stored under the old alias. The one-time migration now moves it. And no connection to a paired node uses its remembered address any more: every one browses the network for the node's id, keeps looking for up to a minute when the network does not answer, and then refuses rather than trying an address that may belong to another machine by now (a DHCP move had cost a benchmark night its second lane).

Fixed:

• The code index's memory budget saw no running index. Since the rename the code index looked for its running scopes under a folder that does not exist, so a new index run was admitted without counting the ones already running. Index runs again count each other, and the file index's, against one budget.
• The Puffin-to-Mightling move runs only for an installed Mightling, never from a source tree or a development build.

Upgrading from 1.5.1

1. Update: ling update on every machine (or the one-line installer again). Update the desktop app on your other computers at the same time.
2. On a GB10 node, once: ling-admin docs setup, which downloads what the file index loads (about 320 MB). Until then, ling docs status names what is not installed and says to run it. The installer does this step for new node installs.
3. ~/Documents and ~/Downloads are indexed by default on a node from the first session on. To keep them out: ling docs remove documents / ling docs remove downloads, or switch the index off entirely with mightling_docs = false in dreamference.toml.
4. To reach the new web UI from your phone or other computers, run ling-admin node enable again on the node: it then starts ling web on the LAN (port 3100) and advertises it. On the node itself, ling web open is enough.
5. The desktop app's Chat opens Ask (the new UI) instead of the previous web chat. The previous web chat still runs at http://localhost:3000 and is managed with ling-admin chat ….
6. Nothing else turns on by itself. The Signal, Matrix and Telegram bridges stay off until you run their setup; your settings, sessions and paired machines are unchanged.

Full Changelog: v1.5.1...v1.6.0