TheOrc v1.13.0

TheOrc v1.13.0

Local-first, cloud-optional AI coding assistant — no subscriptions, no cloud model or inference provider required.

---

What's new in v1.13.0

Context Fabric closes its benchmark gate at GO — honestly. The CF-7 gate reached GO on cf-expanded-book-v1, a 128-segment un-marked corpus with 120 real held-out questions: Qwen3.5-9B answered 104/120 at 97.1% citation precision with 128/128 segments read, versus a best competing baseline of 52/120. Critically, this run replaces an earlier 2026-07-03 GO that was found to have scored against a corpus containing marked/leaked evidence lines — that result was retracted, the corpus rebuilt clean, and the gate re-run from scratch. Two independent bugs surfaced during live validation and were fixed to reach the honest verdict: the open-extraction reader was silently dropping facts on dense filler-heavy segments (now backed by a completeness-repair pass), and the exhaustive-leaf-coverage gate was checking a stale whole-corpus assumption left over from the old 16-segment fixture (now scoped to each question's own segment set).

A stricter NoKvSlot admission fix underneath the gate. The §7e prompt-overflow fix (exact-token admission) closed a class of Qwen3.5 KV-cache exhaustion crashes that had been corrupting benchmark scores — runs are now checked for NoKvSlot before their numbers are trusted, and Qwen3.5-Q8_0 is the new best-performing configuration on the gate.

Cost-tiered external code review. The grok-review tooling was reworked into four explicit modes — quick (default, cheap, latest commit), diff (pre-commit uncommitted check), full (PR-scope with repo reads and project conventions), and adversary (a red-team pass that hunts what the prior reviewer missed) — with a real -PR <n> flag, corrected tool-policy enforcement, and a verdict parser hardened against both narration-glued findings and fail-open false-CLEAN results. The tool proved itself during its own review, catching three real regressions in its own pull request before merge.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.13.0-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.13.0-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🐝 Warband (headless HIVE node, no GUI)

Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.12.0...v1.13.0

Built from v1.13.0 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.12.0

TheOrc v1.12.0

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.12.0

TheOrc Foundry ships its first specialist model. theorc-toolcaller — Qwen2.5-1.5B fine-tuned to propose the correct tool call (or correctly refuse) from a worker's role, its available tools, and a natural-language request — went from spec to a promoted, benchmarked, deployed model in one release cycle. The full pipeline landed in the Training Pit: a synthetic dataset generator with decision-type balance guarantees, a gated LoRA training runner (config-driven, immutable run manifests, GPU exclusivity enforced against every other Foundry consumer), and a new Stage 4 ARENA panel that benchmarks decision accuracy, tool precision, and per-class F1 live against a sealed 260-example held-out set.

A statistically honest refusal benchmark, not a vibe check. Alongside the sealed Arena set, a new Refusal Gauntlet generates thousands of deterministic adversarial cases across six failure families — foreign tools, out-of-role requests, near-miss tools, prompt injection, missing arguments, and ordinary no-tool conversation — and scores them with exact Clopper-Pearson confidence bounds and paraphrase-consistency checks, so the reported number is the defensible lower bound, not an optimistic point estimate. The gap it found got closed: retraining from r2 to r3 raised sealed-eval decision accuracy 97.3% → 98.5%, and gauntlet safety (never fabricating a tool call) 90.3% → 98.3% on held-out phrasings the model never trained on.

The trained specialist is live, opt-in, and learning from real use. theorc-toolcaller:qwen25-1.5b is deployed via Ollama and wired into the Swarm worker loop as an opt-in repair lane — when a worker's response contains no parseable tool call, the specialist gets one shot at proposing one before the turn falls through to today's behavior, still gated by the same deterministic tool-policy engine as every other call. Real usage now feeds the next training round from two organic sources, both off by default under one settings toggle: Swarm tool-call decisions, and — new this release — OrcChat single-agent chat decisions, captured under a wider "v1" tool inventory that's a deliberate sibling to Swarm's frozen six-tool set, not an edit to it.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.12.0-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.12.0-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🐝 Warband (headless HIVE node, no GUI)

Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.11.3...v1.12.0

Built from v1.12.0 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.11.3

TheOrc v1.11.3

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.11.3

Context Fabric's NoKvSlot mystery is finally closed — and it isn't TheOrc's bug. A months-old class of infrastructure crashes on Gemma-4-12B runs turned out, after direct A/B elimination of every leading theory (cross-conversation KV exhaustion, SWA cache sizing, force-recycle timing), to be a Gemma-4-specific limitation in upstream llama.cpp, not a defect in TheOrc's own runtime. Meta-Llama-3.1-8B and qwen2.5-coder-7b show zero NoKvSlot occurrences across 500+ combined benchmark questions on three independent machines — the crash is isolated to one model family, and Context Fabric's own retrieval/reduction mechanics are proven sound underneath it.

Retrieval quality climbs from 31% to a validated 48%+ pass rate through three deterministic fixes, not guesswork. Each tier targets a specific, measured failure mode found by categorizing every B3 miss against the evidence actually supplied to the model: Tier 1 fixes multi-word entity names dissolving into common-word noise (bag-of-words scoring couldn't tell "Station Alpha" from any card containing "station" and "alpha" separately); Tier 1.5 adds proximity-pair matching for paraphrased questions that invert entity word order; Tier 2 stops silently rejecting oversized evidence cards and truncates them instead. Cumulative effect on the 100-question suite: pass rate 31→45→56, pure retrieval misses 49→21→10, B3 now clearly beats a conventional top-k RAG baseline for the first time.

Tier 2.5 closes two more real gaps and finds the honest edge of what's left. Multi-hop chain questions were failing because the greedy evidence-fill spent its whole token budget on distractor segments that share an entity name with the question but belong to an unrelated fact chain — a reference-chasing pass now follows the chain's own shared identifiers into the linked segments, with 30% of the evidence budget reserved specifically so it always gets a turn. Measured live: full-retrieval MultiHop cases rose 6→9 out of 24. Separately, root-caused and fixed the long-standing boundary-stitch failure (Meta-Llama emits its linkedFacts field as JSON objects instead of strings — a tolerant parser now handles both) and closed a citation-precision shortfall, both now passing their gates and holding stable across five independent live validation runs. What's left is honestly reported, not hidden: the remaining MultiHop gap is a model-instruction-compliance ceiling (the model under-citing a multi-part chain answer), not a retrieval defect — logged with three scoped next options rather than claimed as solved.

Model Benchmark window (Phase 1) and safer benchmark tooling. A new read-only panel under Models → Model Benchmark… surfaces every model's CF-7 benchmark history — GO/NO-GO verdicts, question pass rate, citation precision, segment coverage — scanned automatically from .orc/adversarial/ artifacts. The benchmark CLI also gained --model/THEORC_CF_MODEL pinning so a shared model depot changing mid-run (a model toggled on or off by unrelated work) can no longer silently swap which model a benchmark run actually measures, plus verbose preflight diagnostics that print every candidate model's admission verdict and any disabled GGUFs on rejection.

Training Pit tab redesign. The dataset/training dashboard now scrolls as one page (a right-side scrollbar reaches every section regardless of window size), and the three pipeline stages — Generate Dataset, Orc Academy, The Foundry — are laid out side by side as numbered, color-coded cards instead of stacking as individually-expandable accordion sections. Inventory tiles (datasets/adapters/models) now read as stat cards with a colored accent bar and a large count.
Known issue: opening a workspace folder while the Training Pit tab is active can trigger a native stack-overflow crash under investigation — isolated reproduction attempts (real window, real repo data, multiple layout widths) have not yet caught it live; tracked for a follow-up patch.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.11.3-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.11.3-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🐝 Warband (headless HIVE node, no GUI)

Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.11.2...v1.11.3

Built from v1.11.3 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.11.2

TheOrc v1.11.2

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.11.2

Context Fabric becomes real product surface, not just architecture. CF-0 through CF-8 are now landed on master: deterministic ingestion, document graph retrieval, native readers/reducers, OrcChat Library attachment, verified citation labels, citation popup/source opening, distributed HIVE readers, benchmark-gate contracts, and hard-ingestion support for structured formats, OCR contracts, immutable versions, cache policy, vectors, and cross-links. The honest caveat stays intact: unattended million-token/LongBench runs and full multimodal page understanding are future benchmark/product work, not release claims.

Phase 3B native campaign engine ships. The HIVE MIND can now coordinate native-runtime campaign work instead of pretending distributed shell access is the product. This release lands the first full campaign-engine slice: typed campaign/work-unit contracts, capability-aware leasing, content-addressed model and artifact storage, worker-side native execution plumbing, verifier-oriented result metadata, and the first showcase packs including Native AI Eval Factory and Alien Signal Search.

OrcChat grows into a real modern chat surface. Chat now has a first-class tool pack built around the workflows people actually use in web chat: web search, page fetch, URL fetch, workspace browse/search/read/write, outline, test runs, and markdown document generation. It also gains file attachments and image attachments: images are carried through the multimodal message payload, text-like attachments are inlined into the prompt, and generated markdown docs come back as clickable links right in the conversation.

Release-quality native/runtime plumbing behind the scenes. The shared native runtime now sits in its own cross-platform project, headless agent-loop work can run without Ollama, and campaign/worker execution uses the same native path instead of an Ollama fallback. The release pipeline and tests were expanded around this path, and the final pass verified the OrcChat attachment/tooling slice plus the Phase 3B unit suite before tagging.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.11.2-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.11.2-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🐝 Warband (headless HIVE node, no GUI)

Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.11.1...v1.11.2

Built from v1.11.2 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.11.0

TheOrc v1.11.0

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.11.0

The HIVE MIND screen, redesigned. The constellation is now a living neural-swarm view drawn by a new immediate-mode renderer: each node is a role-shaped silhouette — a crowned hexagon for the Warchief, diamond for Coder, circle for Researcher, rounded square for UIDeveloper, triangle for Tester — with glow, a breathing core, and signal particles (all frozen by Lite Mode for thin machines). A new metrics rail shows nodes online, aggregate VRAM, and a role legend (live where the data exists, clearly-marked demo where the telemetry backend is still to come). Left-click a node for a detail panel; right-click → ⬡ Set role to assign HIVE roles and worker lanes. The rail is resizable (drag the splitter) and can be moved to either side (⇄), and your layout choice persists.

Remote task dispatch to a Warband. You can now dispatch a task to a headless HIVE node from anywhere via a new authenticated POST /hive/tasks/submit endpoint, with a configurable WarchiefUrl so a Warband can pull from a remote Warchief. Found and fixed the blocker that made this impossible: a worker polling its own queue had no way to authenticate to itself — same-machine callers are now trusted (the established req.IsLocal pattern), so a submitted task actually gets claimed and run. Verified end-to-end on a real Raspberry Pi.

Cleaner, self-healing constellation. A machine reached over both LAN and Tailscale now shows as one node, not two (dedup by identity, with automatic address fallback so it stays reachable whether you're home or roaming). The Tailscale scan only adds devices actually running TheOrc, so phones and other tailnet devices stop appearing as phantom nodes. Paired headless nodes now show up automatically.

HIVE repair that actually un-sticks a split fleet. When machines ended up in separate hives, there was no way to merge them — and no escape. Now the repair wizard shows your current hive up front and offers "Leave the current hive and join" in one click (keeping your keys and paired peers), plus a standalone 🚪 Leave current hive. This is how you pull a machine out of its own hive and into your main one.

Warband CI + Docker. The release pipeline now publishes the headless theorc-warband binary for linux-x64/osx-arm64, and ships a warband.compose.yml Docker template.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.11.0-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.11.0-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🐝 Warband (headless HIVE node, no GUI)

Deploy theorc-warband on any machine you want to join your HIVE without opening the
full app — a cloud box, a home-lab server, a Raspberry Pi. Download the binary for your
OS below, chmod +x theorc-warband, then either run it directly or set it up as a
systemd service (Linux) / launchd agent (macOS) so it survives reboots. See
--show-identity and --pair --target <host> --expect-fingerprint "<phrase>" for
joining an existing HIVE headlessly.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.10.0...v1.11.0

Built from v1.11.0 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.10.0

TheOrc v1.10.0

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.10.0

OrcChat: uncensored multi-backend chat, built from scratch in C#. A new chat surface — model-agnostic backend routing, streaming, user-controlled generation params, no frontend content filtering, no injected system prompt by default. Three uncensored Dolphin-line models added to the model catalogs with an UNCENSORED badge (opt-in only, never auto-recommended). Date/time grounding, a persisted system prompt across restarts, a live context-window usage indicator, HIVE node routing to run a chat on a paired machine, and inline image rendering in markdown output (![alt](src), http(s)/data:/local-file, background-thread decode).

Native runtime: real, working, and now actually reachable. The in-process LLamaSharp runtime, ModelDepot/SessionManager/AdapterManager/RuntimeOrchestrator/OrcScheduler VRAM-aware admission control — all of it was already implemented in prior releases, but two real opt-in paths (the llama.cpp server backend, and an experimental native main-chat mode) had zero Settings UI to actually turn them on. Both are now exposed, with automatic fallback to Ollama on any native failure. Verified on real hardware: a genuine CUDA build hit 67.7 tok/s on an RTX 4060, vs. ~6 tok/s CPU-only.

Found and fixed a real OrcChat bug on the first real end-to-end test against the new backend: tool definitions were serialized in the wrong wire shape — Ollama silently tolerated it, llama.cpp's stricter OpenAI-compatible server rejected it outright with a 500. Fixed; verified both single-turn and multi-turn conversations now work correctly against a local llama.cpp server with zero Ollama involvement.

Model downloader hardening. Downloads now auto-retry with resume on a dropped connection instead of silently stalling. SHA-256 verification — previously implemented but never actually wired up, since nothing fetched a hash to check against — now runs for real using HuggingFace's own LFS metadata, deleting a corrupted download before it gets registered as a usable model.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.10.0-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.10.0-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.9.5...v1.10.0

Built from v1.10.0 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.9.5

TheOrc v1.9.5

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.9.5

The installer now genuinely targets three OSes, not one. OrchestratorSetup was rewritten from a Windows-only WPF wizard to a cross-platform Avalonia app (Phase 1), with every OS-coupled action — hardware detection, firewall, launchers, uninstall — moved behind one IPlatformInstaller interface with real Windows, Linux, and macOS implementations (Phases 2, 4, 5). This release closes the gap those phases left open: nothing upstream of the installer's own logic could actually hand a non-Windows machine a real binary to install. release.yml now publishes a macOS (osx-arm64) build alongside Windows; the model manifest, the llama.cpp runtime resolver, and three separate spots in the running app (the update checker, self-updater, and llama-server launcher) all needed their own fixes — each only ever recognized Windows binary names, which would have shipped a Mac install that completes successfully and then can't update itself or find its own runtime.

Headless HIVE nodes are real now too. OrchestratorIDE.Daemon (theorc-warband) is the cross-platform, no-GUI HIVE node — first deployed to actual ARM64 hardware this release (a Raspberry Pi 4), running as a systemd service. It gained --pair/--show-identity CLI modes (same fingerprint-gated safety contract as swarmcli's) so a headless box can join an existing HIVE without ever needing a display.

Found and fixed a HIVE reachability gap no existing diagnostic caught: a node can have its URL ACL reservation and firewall rules all correctly in place and still be completely unreachable to peers, because the network interface a peer actually connects through was classified "Public" by Windows — the firewall rules are deliberately Private-profile-only, so they silently never apply there. Added HiveNetworkEnroller, shared between the installer and the app itself, with the missing diagnostic (FindPublicInterfacesAsync) and a one-click fix.

---

🪟 Windows — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

Portable: download TheOrc-1.9.5-win-x64-portable.zip instead — contains both OrchestratorIDE.exe and OrchestratorSetup.exe, extract both to the same folder, run OrchestratorIDE.exe.

🍎 macOS (Apple Silicon)

1. Download OrchestratorSetup below (no extension)
2. In Terminal: chmod +x OrchestratorSetup && ./OrchestratorSetup
3. First launch will show a Gatekeeper warning ("cannot be opened because the developer cannot be verified") — this build isn't code-signed yet. Right-click the file → Open, or run xattr -d com.apple.quarantine OrchestratorSetup first.
4. The wizard detects your hardware, downloads the right runtime and model, configures everything

Portable: download TheOrc-1.9.5-osx-arm64-portable.tar.gz instead — tar -xzf it, both binaries are already executable.

Intel Macs (osx-x64) aren't built yet — only Apple Silicon.

🖥️ System requirements

OS	Windows 10/11 (x64) · macOS (Apple Silicon)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · Apple Silicon (Metal) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.9.4...v1.9.5

Built from v1.9.5 · .NET 10 · Source · Support the Orc ☕

TheOrc v1.9.4

TheOrc v1.9.4

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.9.4

v1.9.0 — WPF deleted, CodeGraph v1, Native Runtime + HIVE MIND ready for multi-machine testing

The biggest cutover since the Avalonia migration started: WPF is gone. OrchestratorIDE/OrchestratorIDE.csproj and every WPF-only window, dialog, panel, and control were deleted outright — not archived, not stubbed, deleted. Avalonia is no longer "primary," it's the only desktop shell. ModelWikiWindow/ModelCompareWindow were retired rather than ported (their data layer stays; a from-scratch, data-bound rebuild is a real future feature, not a blocker). Everything operators actually use day to day — ask_user, the first-run wizard, sandbox bypass, self-update, model library/downloader, workspace/global agent rules — already had a real Avalonia home going into this release. Shared service code (Core/, Services/, Models/, Trust/) is untouched; only the WPF-exclusive UI layer is gone, which is what made deleting an entire desktop framework in one night tractable at all.

CodeGraph v1 — a Roslyn + SQLite code knowledge graph lets the agent query structure (callers, callees, complexity hotspots, architecture) instead of grepping files for every question. Five tools wired into the swarm; lifecycle-managed (background re-index on workspace open); ships with a codegraph-query dev skill for low-token structural lookups against the underlying SQLite DB directly.

Native Runtime, two layers, both real:

• The installer already defaults fresh installs to a local llama.cpp server instead of Ollama (IModelRuntime → LlamaCppServerRuntime) — this isn't new in v1.9, but it's worth being explicit about: most new installs of TheOrc have never required Ollama at all.
• New this release, experimental and opt-in: an in-process LLamaSharp runtime (no server process), AdapterManager for per-role persistent LoRA contexts with hot-swap, RuntimeOrchestrator tying both runtimes together, and OrcScheduler doing real VRAM-budget admission control so concurrent roles don't blow out a GPU. The first live path is an experimental HIVE-worker / main-chat opt-in (ExperimentalNativeHiveWorkerEnabled / ExperimentalNativeMainChatEnabled in Settings) with automatic fallback to Ollama on any runtime fault. Main chat, research chat, and SwarmSession stay on the configured default runtime unless explicitly opted in.

HIVE MIND — fixed a release-blocking startup bug. A pre-release smoke test caught HiveNodeServer.Start() silently failing on every normal-user install: a failed wildcard HttpListener bind (needs admin rights / a URL ACL reservation no normal process has) left the listener disposed internally, and the fallback cleanup code's own property access threw a second exception that masked the first inside an unobserved background task. Net effect: enabling HIVE MIND did nothing — no error, no log line, nothing listening — on every machine that wasn't running elevated. Fixed and verified live (/hive/info returns 200, UDP beacon confirmed listening). This is exactly the kind of bug dotnet test can't catch, since it requires a real socket bind; found by actually launching the built app before shipping it.

Training Pit suitability gate — a deterministic pre-training check that blocks write-task examples mislabeled into TESTER-lane roles before they reach the trainer — the exact contamination pattern that regressed ORC ACADEMY v2 (51.3% of v2's examples had this mislabeling).

Infrastructure fixes that came out of actually shipping this release:

• .github/workflows/ci.yml was still building the now-deleted WPF project — every push had been failing CI since the WPF deletion commit landed. Fixed, and CI now runs both test suites on every push instead of just build-checking.
• .github/workflows/release.yml was still publishing and packaging the WPF build as the actual downloaded artifact. Fixed to publish Avalonia with the same output filename, so every downstream consumer (installer, self-updater) keeps working unchanged.
• OrchestratorIDE.Daemon (the headless cross-platform HIVE node, theorc-warband) had been failing to build since before this release cycle started — a dependency on the heavy native LLamaSharp stack that a lightweight daemon shouldn't need. Decoupled via a new IHiveNativeRoleExecutor interface; the daemon now builds clean without pulling in native-runtime dependencies at all.
• Internal dev-only docs and scratch files (.grok/ specs, prompts, spike code; loose planning notes) stopped being published to GitHub — they're still on disk for development, just not part of the public repo going forward. README.md, SECURITY.md, LICENSING.md, CLA.md, docs/ROADMAP.md, docs/ARCHITECTURE.md, and .grok/PROJECT_TRUTH.md stay public.

v1.9.1 — HIVE MIND actually reachable across machines now

v1.9.0's HIVE fix made the node server start without crashing; it didn't make it reachable. Real multi-machine testing (NewcorePC ↔ HARDCOREPC, both on v1.9.0, one over Tailscale) found that the fix's own fallback path — binding localhost only when the wildcard bind fails — was itself the unfixed problem: a localhost-only listener is invisible to every other machine, even though the app shows no error and looks like it started fine.

Root cause: binding the wildcard prefix (http://+:port/, all interfaces) as a normal, non-admin process requires an http.sys URL ACL reservation (netsh http add urlacl). Nothing — not the installer, not the app — ever created one. So on every non-elevated install, the wildcard bind was silently denied and the app fell back to loopback-only, with no indication to the user that the node had effectively not started for HIVE's actual purpose.

Fixed:

• The installer's HIVE enrollment step now reserves the URL ACLs for both HIVE ports (7078, 7079) alongside its existing firewall rules. Everything is tried unelevated first (no-op if already granted); anything still missing is batched into a single UAC prompt instead of one popup per item.
• HiveTaskQueue (the Warchief's distributed-task-queue listener, used by Phase 3 Distributed Swarm) had the identical listener-reuse bug as HiveNodeServer did in v1.9.0 — fixed the same way, found by code review of the parallel class rather than waiting to hit it live.
• The gated Phi-4 Mini boss-model download (HTTP 401 on every fresh install since that HuggingFace repo got gated after the manifest was written) now points at a working, non-gated mirror.

If you installed v1.9.0 and HIVE MIND only ever seemed to discover other nodes one-directionally (or not at all), this is why — update to v1.9.1.

v1.9.2 — pairing actually works now

Validating v1.9.1's reachability fix across three real machines (not just a single dev box) surfaced the next real gap: pairing — the step where two machines agree to trust each other — never had a way to actually start. The "approve" side was fully built (the responder endpoints, the approval prompt), but nothing in either UI, current or the now-deleted WPF one, ever called the endpoint that initiates a pairing request. It wasn't a bug so much as an unfinished feature nobody had hit yet, because nothing could reach it.

Added: a "Pair with this node" action on any reachable HIVE node card, building the missing initiator side end to end — request, approval polling, shared-secret derivation.

This one went through real adversarial review (two independent AI reviewers, multiple rounds) before shipping, and it caught genuine problems with the first draft, not nitpicks:

• The approval-polling endpoint is unauthenticated by design, and the first version trusted a new peer as soon as it got an "approved" response — before the one real check (comparing a human-readable fingerprint between the two machines) ever happened. Fixed: trust is no longer written until the operator explicitly confirms the fingerprint matches what the other machine displays.
• A newly-paired peer was being granted enough standing to become eligible for real Warchief authority later, regardless of whether it should be. Capped to a safer default — becoming Controller-eligible is now a separate, deliberate decision, not a side effect of pairing.
• A few smaller correctness/cleanup fixes alongside those.

Also: the code-review tooling used throughout this project's development kept falsely flagging legitimate security-related code as a risk and refusing to review it. Switched the underlying model, confirmed by direct comparison that the new one doesn't have this problem.

Looking ahead to v2.0

v2.0's defining change: Native Runtime becomes the default, Ollama becomes fully optional. That flip is explicitly gated on multi-machine HIVE MIND validation of this release's native opt-in path across a real LAN/Tailscale network — not a fixed date. Also planned, not yet started:

• Promoting the experimental RuntimeOrchestrator/AdapterManager/OrcScheduler layer out of opt-in status once the v1.9 HIVE testing round validates it under real concurrent multi-role load.
• HIVE MIND Phase 3B — full multi-step AgentLoop-style tool execution on remote workers (file writes, shell commands, web search running on the worker machine itself), not just single-pass LLM calls.
• A cross-platform CI publish matrix for linux-x64/osx-arm64 Warband (daemon) binaries — the daemon itself is already cross-platform; nothing currently builds/ships those binaries automatically.
• A from-scratch, data-bound Avalonia rebuild of the Model Wiki/catalogue browsing experience retired in this release.

---

---

##...

TheOrc v1.9.2

TheOrc v1.9.2

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.9.2

v1.9.0 — WPF deleted, CodeGraph v1, Native Runtime + HIVE MIND ready for multi-machine testing

The biggest cutover since the Avalonia migration started: WPF is gone. OrchestratorIDE/OrchestratorIDE.csproj and every WPF-only window, dialog, panel, and control were deleted outright — not archived, not stubbed, deleted. Avalonia is no longer "primary," it's the only desktop shell. ModelWikiWindow/ModelCompareWindow were retired rather than ported (their data layer stays; a from-scratch, data-bound rebuild is a real future feature, not a blocker). Everything operators actually use day to day — ask_user, the first-run wizard, sandbox bypass, self-update, model library/downloader, workspace/global agent rules — already had a real Avalonia home going into this release. Shared service code (Core/, Services/, Models/, Trust/) is untouched; only the WPF-exclusive UI layer is gone, which is what made deleting an entire desktop framework in one night tractable at all.

CodeGraph v1 — a Roslyn + SQLite code knowledge graph lets the agent query structure (callers, callees, complexity hotspots, architecture) instead of grepping files for every question. Five tools wired into the swarm; lifecycle-managed (background re-index on workspace open); ships with a codegraph-query dev skill for low-token structural lookups against the underlying SQLite DB directly.

Native Runtime, two layers, both real:

• The installer already defaults fresh installs to a local llama.cpp server instead of Ollama (IModelRuntime → LlamaCppServerRuntime) — this isn't new in v1.9, but it's worth being explicit about: most new installs of TheOrc have never required Ollama at all.
• New this release, experimental and opt-in: an in-process LLamaSharp runtime (no server process), AdapterManager for per-role persistent LoRA contexts with hot-swap, RuntimeOrchestrator tying both runtimes together, and OrcScheduler doing real VRAM-budget admission control so concurrent roles don't blow out a GPU. The first live path is an experimental HIVE-worker / main-chat opt-in (ExperimentalNativeHiveWorkerEnabled / ExperimentalNativeMainChatEnabled in Settings) with automatic fallback to Ollama on any runtime fault. Main chat, research chat, and SwarmSession stay on the configured default runtime unless explicitly opted in.

HIVE MIND — fixed a release-blocking startup bug. A pre-release smoke test caught HiveNodeServer.Start() silently failing on every normal-user install: a failed wildcard HttpListener bind (needs admin rights / a URL ACL reservation no normal process has) left the listener disposed internally, and the fallback cleanup code's own property access threw a second exception that masked the first inside an unobserved background task. Net effect: enabling HIVE MIND did nothing — no error, no log line, nothing listening — on every machine that wasn't running elevated. Fixed and verified live (/hive/info returns 200, UDP beacon confirmed listening). This is exactly the kind of bug dotnet test can't catch, since it requires a real socket bind; found by actually launching the built app before shipping it.

Training Pit suitability gate — a deterministic pre-training check that blocks write-task examples mislabeled into TESTER-lane roles before they reach the trainer — the exact contamination pattern that regressed ORC ACADEMY v2 (51.3% of v2's examples had this mislabeling).

Infrastructure fixes that came out of actually shipping this release:

• .github/workflows/ci.yml was still building the now-deleted WPF project — every push had been failing CI since the WPF deletion commit landed. Fixed, and CI now runs both test suites on every push instead of just build-checking.
• .github/workflows/release.yml was still publishing and packaging the WPF build as the actual downloaded artifact. Fixed to publish Avalonia with the same output filename, so every downstream consumer (installer, self-updater) keeps working unchanged.
• OrchestratorIDE.Daemon (the headless cross-platform HIVE node, theorc-warband) had been failing to build since before this release cycle started — a dependency on the heavy native LLamaSharp stack that a lightweight daemon shouldn't need. Decoupled via a new IHiveNativeRoleExecutor interface; the daemon now builds clean without pulling in native-runtime dependencies at all.
• Internal dev-only docs and scratch files (.grok/ specs, prompts, spike code; loose planning notes) stopped being published to GitHub — they're still on disk for development, just not part of the public repo going forward. README.md, SECURITY.md, LICENSING.md, CLA.md, docs/ROADMAP.md, docs/ARCHITECTURE.md, and .grok/PROJECT_TRUTH.md stay public.

v1.9.1 — HIVE MIND actually reachable across machines now

v1.9.0's HIVE fix made the node server start without crashing; it didn't make it reachable. Real multi-machine testing (NewcorePC ↔ HARDCOREPC, both on v1.9.0, one over Tailscale) found that the fix's own fallback path — binding localhost only when the wildcard bind fails — was itself the unfixed problem: a localhost-only listener is invisible to every other machine, even though the app shows no error and looks like it started fine.

Root cause: binding the wildcard prefix (http://+:port/, all interfaces) as a normal, non-admin process requires an http.sys URL ACL reservation (netsh http add urlacl). Nothing — not the installer, not the app — ever created one. So on every non-elevated install, the wildcard bind was silently denied and the app fell back to loopback-only, with no indication to the user that the node had effectively not started for HIVE's actual purpose.

Fixed:

• The installer's HIVE enrollment step now reserves the URL ACLs for both HIVE ports (7078, 7079) alongside its existing firewall rules. Everything is tried unelevated first (no-op if already granted); anything still missing is batched into a single UAC prompt instead of one popup per item.
• HiveTaskQueue (the Warchief's distributed-task-queue listener, used by Phase 3 Distributed Swarm) had the identical listener-reuse bug as HiveNodeServer did in v1.9.0 — fixed the same way, found by code review of the parallel class rather than waiting to hit it live.
• The gated Phi-4 Mini boss-model download (HTTP 401 on every fresh install since that HuggingFace repo got gated after the manifest was written) now points at a working, non-gated mirror.

If you installed v1.9.0 and HIVE MIND only ever seemed to discover other nodes one-directionally (or not at all), this is why — update to v1.9.1.

v1.9.2 — pairing actually works now

Validating v1.9.1's reachability fix across three real machines (not just a single dev box) surfaced the next real gap: pairing — the step where two machines agree to trust each other — never had a way to actually start. The "approve" side was fully built (the responder endpoints, the approval prompt), but nothing in either UI, current or the now-deleted WPF one, ever called the endpoint that initiates a pairing request. It wasn't a bug so much as an unfinished feature nobody had hit yet, because nothing could reach it.

Added: a "Pair with this node" action on any reachable HIVE node card, building the missing initiator side end to end — request, approval polling, shared-secret derivation.

This one went through real adversarial review (two independent AI reviewers, multiple rounds) before shipping, and it caught genuine problems with the first draft, not nitpicks:

• The approval-polling endpoint is unauthenticated by design, and the first version trusted a new peer as soon as it got an "approved" response — before the one real check (comparing a human-readable fingerprint between the two machines) ever happened. Fixed: trust is no longer written until the operator explicitly confirms the fingerprint matches what the other machine displays.
• A newly-paired peer was being granted enough standing to become eligible for real Warchief authority later, regardless of whether it should be. Capped to a safer default — becoming Controller-eligible is now a separate, deliberate decision, not a side effect of pairing.
• A few smaller correctness/cleanup fixes alongside those.

Also: the code-review tooling used throughout this project's development kept falsely flagging legitimate security-related code as a risk and refusing to review it. Switched the underlying model, confirmed by direct comparison that the new one doesn't have this problem.

Looking ahead to v2.0

v2.0's defining change: Native Runtime becomes the default, Ollama becomes fully optional. That flip is explicitly gated on multi-machine HIVE MIND validation of this release's native opt-in path across a real LAN/Tailscale network — not a fixed date. Also planned, not yet started:

• Promoting the experimental RuntimeOrchestrator/AdapterManager/OrcScheduler layer out of opt-in status once the v1.9 HIVE testing round validates it under real concurrent multi-role load.
• HIVE MIND Phase 3B — full multi-step AgentLoop-style tool execution on remote workers (file writes, shell commands, web search running on the worker machine itself), not just single-pass LLM calls.
• A cross-platform CI publish matrix for linux-x64/osx-arm64 Warband (daemon) binaries — the daemon itself is already cross-platform; nothing currently builds/ships those binaries automatically.
• A from-scratch, data-bound Avalonia rebuild of the Model Wiki/catalogue browsing experience retired in this release.

---

---

##...

TheOrc v1.9.1

TheOrc v1.9.1

100% local AI coding assistant — no cloud, no subscriptions, no data leaving your machine.

---

What's new in v1.9.1

v1.9.0 — WPF deleted, CodeGraph v1, Native Runtime + HIVE MIND ready for multi-machine testing

The biggest cutover since the Avalonia migration started: WPF is gone. OrchestratorIDE/OrchestratorIDE.csproj and every WPF-only window, dialog, panel, and control were deleted outright — not archived, not stubbed, deleted. Avalonia is no longer "primary," it's the only desktop shell. ModelWikiWindow/ModelCompareWindow were retired rather than ported (their data layer stays; a from-scratch, data-bound rebuild is a real future feature, not a blocker). Everything operators actually use day to day — ask_user, the first-run wizard, sandbox bypass, self-update, model library/downloader, workspace/global agent rules — already had a real Avalonia home going into this release. Shared service code (Core/, Services/, Models/, Trust/) is untouched; only the WPF-exclusive UI layer is gone, which is what made deleting an entire desktop framework in one night tractable at all.

CodeGraph v1 — a Roslyn + SQLite code knowledge graph lets the agent query structure (callers, callees, complexity hotspots, architecture) instead of grepping files for every question. Five tools wired into the swarm; lifecycle-managed (background re-index on workspace open); ships with a codegraph-query dev skill for low-token structural lookups against the underlying SQLite DB directly.

Native Runtime, two layers, both real:

• The installer already defaults fresh installs to a local llama.cpp server instead of Ollama (IModelRuntime → LlamaCppServerRuntime) — this isn't new in v1.9, but it's worth being explicit about: most new installs of TheOrc have never required Ollama at all.
• New this release, experimental and opt-in: an in-process LLamaSharp runtime (no server process), AdapterManager for per-role persistent LoRA contexts with hot-swap, RuntimeOrchestrator tying both runtimes together, and OrcScheduler doing real VRAM-budget admission control so concurrent roles don't blow out a GPU. The first live path is an experimental HIVE-worker / main-chat opt-in (ExperimentalNativeHiveWorkerEnabled / ExperimentalNativeMainChatEnabled in Settings) with automatic fallback to Ollama on any runtime fault. Main chat, research chat, and SwarmSession stay on the configured default runtime unless explicitly opted in.

HIVE MIND — fixed a release-blocking startup bug. A pre-release smoke test caught HiveNodeServer.Start() silently failing on every normal-user install: a failed wildcard HttpListener bind (needs admin rights / a URL ACL reservation no normal process has) left the listener disposed internally, and the fallback cleanup code's own property access threw a second exception that masked the first inside an unobserved background task. Net effect: enabling HIVE MIND did nothing — no error, no log line, nothing listening — on every machine that wasn't running elevated. Fixed and verified live (/hive/info returns 200, UDP beacon confirmed listening). This is exactly the kind of bug dotnet test can't catch, since it requires a real socket bind; found by actually launching the built app before shipping it.

Training Pit suitability gate — a deterministic pre-training check that blocks write-task examples mislabeled into TESTER-lane roles before they reach the trainer — the exact contamination pattern that regressed ORC ACADEMY v2 (51.3% of v2's examples had this mislabeling).

Infrastructure fixes that came out of actually shipping this release:

• .github/workflows/ci.yml was still building the now-deleted WPF project — every push had been failing CI since the WPF deletion commit landed. Fixed, and CI now runs both test suites on every push instead of just build-checking.
• .github/workflows/release.yml was still publishing and packaging the WPF build as the actual downloaded artifact. Fixed to publish Avalonia with the same output filename, so every downstream consumer (installer, self-updater) keeps working unchanged.
• OrchestratorIDE.Daemon (the headless cross-platform HIVE node, theorc-warband) had been failing to build since before this release cycle started — a dependency on the heavy native LLamaSharp stack that a lightweight daemon shouldn't need. Decoupled via a new IHiveNativeRoleExecutor interface; the daemon now builds clean without pulling in native-runtime dependencies at all.
• Internal dev-only docs and scratch files (.grok/ specs, prompts, spike code; loose planning notes) stopped being published to GitHub — they're still on disk for development, just not part of the public repo going forward. README.md, SECURITY.md, LICENSING.md, CLA.md, docs/ROADMAP.md, docs/ARCHITECTURE.md, and .grok/PROJECT_TRUTH.md stay public.

v1.9.1 — HIVE MIND actually reachable across machines now

v1.9.0's HIVE fix made the node server start without crashing; it didn't make it reachable. Real multi-machine testing (NewcorePC ↔ HARDCOREPC, both on v1.9.0, one over Tailscale) found that the fix's own fallback path — binding localhost only when the wildcard bind fails — was itself the unfixed problem: a localhost-only listener is invisible to every other machine, even though the app shows no error and looks like it started fine.

Root cause: binding the wildcard prefix (http://+:port/, all interfaces) as a normal, non-admin process requires an http.sys URL ACL reservation (netsh http add urlacl). Nothing — not the installer, not the app — ever created one. So on every non-elevated install, the wildcard bind was silently denied and the app fell back to loopback-only, with no indication to the user that the node had effectively not started for HIVE's actual purpose.

Fixed:

• The installer's HIVE enrollment step now reserves the URL ACLs for both HIVE ports (7078, 7079) alongside its existing firewall rules. Everything is tried unelevated first (no-op if already granted); anything still missing is batched into a single UAC prompt instead of one popup per item.
• HiveTaskQueue (the Warchief's distributed-task-queue listener, used by Phase 3 Distributed Swarm) had the identical listener-reuse bug as HiveNodeServer did in v1.9.0 — fixed the same way, found by code review of the parallel class rather than waiting to hit it live.
• The gated Phi-4 Mini boss-model download (HTTP 401 on every fresh install since that HuggingFace repo got gated after the manifest was written) now points at a working, non-gated mirror.

If you installed v1.9.0 and HIVE MIND only ever seemed to discover other nodes one-directionally (or not at all), this is why — update to v1.9.1.

Looking ahead to v2.0

v2.0's defining change: Native Runtime becomes the default, Ollama becomes fully optional. That flip is explicitly gated on multi-machine HIVE MIND validation of this release's native opt-in path across a real LAN/Tailscale network — not a fixed date. Also planned, not yet started:

• Promoting the experimental RuntimeOrchestrator/AdapterManager/OrcScheduler layer out of opt-in status once the v1.9 HIVE testing round validates it under real concurrent multi-role load.
• HIVE MIND Phase 3B — full multi-step AgentLoop-style tool execution on remote workers (file writes, shell commands, web search running on the worker machine itself), not just single-pass LLM calls.
• A cross-platform CI publish matrix for linux-x64/osx-arm64 Warband (daemon) binaries — the daemon itself is already cross-platform; nothing currently builds/ships those binaries automatically.
• A from-scratch, data-bound Avalonia rebuild of the Model Wiki/catalogue browsing experience retired in this release.

---

---

⚡ Quick start — one-click install (recommended)

1. Download OrchestratorSetup.exe below
2. Run it — the wizard detects your GPU, downloads the right runtime and model, configures everything
3. Launch from the Desktop shortcut or Start Menu

📦 Portable install — works on fresh machines

Download TheOrc-1.9.1-win-x64-portable.zip — contains both OrchestratorIDE.exe and OrchestratorSetup.exe.

1. Extract both files to the same folder
2. Run OrchestratorIDE.exe
3. If no AI runtime is detected, the app offers to launch the setup wizard automatically — no prior AI tools required

🖥️ System requirements

OS	Windows 10 / 11 (x64)
.NET	Not required — fully self-contained
GPU	NVIDIA RTX (CUDA 12) · AMD (Vulkan) · CPU-only (AVX2 / baseline)
RAM	8 GB minimum · 16 GB recommended

---

Full changelog: v1.9.0...v1.9.1

Built from v1.9.1 · .NET 10 · Source · Support the Orc ☕