1
0
  • v0.2.5 1bea2dbc60

    NIG 0.2.5 Stable

    8ball released this 2026-11-10 02:54:28 +00:00 | 0 commits to main since this release

    What changed

    The status line says what a request is really waiting for. When a model takes a while, NIG used to show "Waiting for the model" for the whole wait, then "The model is writing a tool call" while nothing had been written yet. Now it says who has the request:

    • "Waiting for the router" while the router prepares the conversation and recalls memory, before any model is asked.
    • With a cloud model, it names the provider and the step: "The router is sending the request to Lutheos (192 KB)", "Waiting for Lutheos to answer · 25 s · 12 s of it was the router's memory recall", or "Lutheos: connection failed; the router is trying again (attempt 2)".
    • With the fleet's own engines: "Flash Next (4090s) gpu3 is reading the context · 181K tokens".
    • "The model is writing a tool call" now only shows while a tool call is actually being written.

    The more detailed lines need the router update that goes live alongside this release; with an older router NIG shows the provider's short name and the same kinds of messages.

    Stop really stops sub-agents. Pressing Stop on a chat now also holds its sub-agents that had been idle long enough to be unloaded from memory. Before, such a sub-agent could start working again on its own after the Stop.

    followup_task only goes to your own sub-agents. An agent can no longer hand a new task to an agent that is not below it (a sibling or its lead). It gets a clear message to use send_message for those instead, so tasks no longer land on agents that were not meant to take them.

    Downloads
  • v0.2.4 1bea2dbc60

    NIG 0.2.4 Stable

    8ball released this 2026-11-08 21:36:36 +00:00 | 0 commits to main since this release

    See the WeKnora indexing working

    • Progress bar in the project dialog. Under "Knowledge base" a bar shows how far the project's documents are, with a line saying what happens now: looking for documents, uploading 12 of 40, WeKnora indexing 3 of 40 (and how many documents are ahead in its queue), done, or why it stopped.
    • Progress in the sidebar. While a project's documents are being indexed, its row shows a thin bar; hover it for the same line.
    • "Re-index documents" shows activity at once.
    Downloads
  • v0.2.3 1bea2dbc60

    NIG 0.2.3 Stable

    8ball released this 2026-11-08 18:06:00 +00:00 | 0 commits to main since this release

    WeKnora memory for every project

    The router now keeps every chat in WeKnora, the fleet memory and knowledge store on the ASUS server. NIG connects to it by itself:

    • A weknora tool server. NIG asks the router for your account's WeKnora endpoint and registers it as the MCP server weknora, starred in Settings > Core the first time. The token stays in ~/.fable/secrets/weknora. If you turn the server off, it stays off.
    • Knowledge bases per project. Each project gets three: history (its chats, summaries and documents), facts (durable facts and your notes) and wiki (pages written every night from the history). They are created when the project is, and at the first chat of a run for existing projects.
    • Agents search their own project. A weknora search that names no knowledge base searches the chat's project. The project prompt tells agents to look there before asking you something an earlier chat may have settled.
    • Your documents go in too. The project folder's documents (Markdown, text, PDF, Office, HTML; not code) are uploaded when they change. Use "Re-index documents" in the project dialog to do it now.
    • Your notes are mirrored. Notes you edit in the project dialog also go to the project's facts knowledge base.
    Downloads
  • v0.2.2 1bea2dbc60

    NIG 0.2.2 Stable

    8ball released this 2026-11-07 22:08:30 +00:00 | 0 commits to main since this release

    NIG v0.2.2 — the harness and models reach your existing chats

    This release fixes why favourited skills, the new harness and the cloud models' real context never
    seemed to take effect, especially in chats you had already been using.

    Fixes

    • Your system prompt is rebuilt from the current settings at the start of every turn, so a skill or
      MCP server you add to Settings > Core, a model switch, and a newer build's harness all reach a chat
      you already had — not only brand-new ones. Until now the prompt was frozen when the chat was first
      created, which is why context-mode and the codebase-memory graph were "never used once" in old
      chats, and why a long chat could still be running an old prompt that named the wrong model.
    • Model menu shows each cloud model's real context window (Qwen 3.8 Max 1M, GLM 1M, DeepSeek 1M),
      and a chat on one of them may now use nearly that whole window instead of ~98K. [router]
    • Every RouterHub model is offered again — Claude, GPT, Kimi and the uncensored/abliterated ones —
      not only the uncensored set. Grok stays out. [router]
    • A cloud model's request that holds an image stays on that model instead of being answered by the
      local Flash Next. [router]
    • A model is no longer wrongly hidden by the health check (it now asks for a normal output length,
      which RouterHub requires). [router]

    The three router fixes are already live for every client; this build is what makes the desktop app
    show and use them.

    Downloads
  • v0.2.1 1bea2dbc60

    NIG 0.2.1 Stable

    8ball released this 2026-11-07 18:33:39 +00:00 | 0 commits to main since this release

    NIG 0.2.1

    A provider-first model menu

    • Providers first: the model picker lists providers (Recent, Local, LAN · DGX Spark, Lutheos, RouterHub, Qwen Cloud), each with its model count and a check if your chat's model is in it.
    • Models to the side: hover a provider, click it or press → and its models open beside it. Moving the mouse diagonally toward them doesn't close them. ↑/↓/Enter and ←/Esc work, and in Persian the menu mirrors.
    • Search: typing searches every provider at once, as one list.
    • Providers that are down stay in the list but greyed out, with the reason when you hover.
    • The local model shows its real name: an old default name ("Qwen3.8-27B (local)") no longer covers the router's label. A name you set yourself still wins.

    A stop for repeated commands

    When an agent runs the exact same command again and again, NIG now steps in, even when the output changes every run (for example a screenshot of an animated page).

    • At 6 in a row: NIG tells the model to stop repeating and change approach.
    • At 10 in a row: NIG ends the turn and shows you why, with the command.
    • Nothing restarts it: no automatic retry, goal or restart starts it again, and a sub-agent that hits it reports back as failed, with the reason.
    • What doesn't count: waiting on running work, checking a log and other polling.
    • Settings: the numbers are in Settings › Agent › Work, and 0 turns either off.
    Downloads
  • v0.2.0 1bea2dbc60

    NIG 0.2.0 Stable

    8ball released this 2026-11-07 16:09:48 +00:00 | 0 commits to main since this release

    NIG v2

    A quieter app

    • A calmer layout in the style of the Claude desktop app. The sidebar, chat and right panel are cut by thin lines only. Screens stay nearly empty until you point at something. Counts and costs appear on hover.
    • New interface font, Geist, with sharper small text. Inter and IBM Plex Sans are in Settings › Appearance › Interface font.
    • The model and effort pickers sit under the message box. The model menu has search, recent models, and groups you can open and close.
    • The right panel is one scrolling list: plan, goal, agents, jobs, context, files changed, outputs and MCP servers.
    • Copy, retry and branch sit under each answer, next to the time and how long it worked.

    Smooth streaming at any speed

    Text appears evenly however the model sends it, whether in bursts from speculative decoding, very fast or very slow. The display stays within about a quarter of a second of the model. Code blocks and formatting never show half-typed, and Persian appears word by word. Settings › Appearance › Motion turns it down or off.

    Agents that work together reliably

    • Reports: a sub-agent ends with a report (status, summary, files changed, what it checked), and its lead gets it as a notice, never as text that looks like yours.
    • Waiting: spawn_agent waits for the report by default, and run_in_background is still available.
    • Receipts: every message says whether it was delivered, queued while the agent is busy, held for you, or not delivered and why.
    • Restarts: sub-agents, background commands and queued messages survive a restart of NIG. Each is picked up again, or ends with a notice. On Linux and macOS, background commands keep their exit code and output.
    • Resume: message a finished agent to continue it with its context.
    • Linked chats: they report back when they finish, and you can ask to be told when one goes idle.
    • Messaging between chats: each project decides whether its chats may message each other (on by default) and whether chats in other projects may message them (off by default). The switches are in the project settings file for now.
    • Your rules still hold: a Stop is final, and idle chats only wake when open or with "Auto-start agents" on.
    • Tested by breaking it on purpose: NIG was killed mid-turn, connections were cut, and agents were overloaded and nested. Across 6,000+ events NIG broke none of its rules, against 10.8 per 100 events for v0.1.68.

    Every request on record

    Each chat keeps an append-only session log of what was sent to the model and what came back, so any step can be replayed exactly (fable log replay | verify | export).

    For testing only

    A new core loop with client-side compaction is included but off. It does not yet resume after a restart. Turn it on with v2_core = true in config.toml.

    Downloads
  • NIG 0.1.68 Stable

    8ball released this 2026-11-07 05:38:31 +00:00 | 0 commits to main since this release

    Changes

    • The code graph gets used. When the codebase-memory server is installed, NIG now offers its main tools (search, trace callers and callees, code snippets) up front with clear when-to-use notes, indexes the repository in the background when a chat starts there, fills in the project name by itself, and tells the agent what to do when a call fails. Structure questions ("who calls X", "where is this read") now go to the graph before text search. A disabled_mcp_servers entry still turns it off.
    • Ripple 65B (abliterated, on the DGX Sparks) appears in its own "Ripple (DGX Spark)" section of the model menu when your account may use it.
    Downloads
  • NIG 0.1.67 Stable

    8ball released this 2026-11-07 00:18:42 +00:00 | 0 commits to main since this release

    Fixes

    • The sign-in page always appears when the gateway needs it. If NIG started while the gateway was restarting or briefly unreachable, it could open the normal app with only a "Sign in required" row, and chats failed. NIG now asks the gateway again every 15 seconds until it answers, and any "sign in first" refusal brings up the full sign-in page.
    Downloads
  • NIG 0.1.66 Stable

    8ball released this 2026-11-06 23:42:35 +00:00 | 0 commits to main since this release

    Changes

    • New model menu. The model picker at the top of a chat now has a search box, a Recent group with your last five models, and one foldable section per provider with a model count. Models are sorted by when you last used them, then alphabetically; folds and the order are remembered per account.
    • Qwen Cloud models. The gateway now offers the Qwen Cloud (Singapore) models, such as Qwen 3.8 Max and Flash, GLM 5.3 and DeepSeek V4, in their own "Qwen Cloud" section. Admins choose who may use them in Settings > Users and access.
    Downloads
  • NIG 0.1.65 Stable

    8ball released this 2026-11-06 16:46:12 +00:00 | 0 commits to main since this release

    Changes

    • LarpRouter is now Lutheos. LarpRouter moved to Lutheos; the gateway now sends those models there, and NIG labels the provider "Lutheos (was LarpRouter)". Existing chats and saved model choices keep working.
    Downloads