Commit remaining workspace changes

8 files changed, 1618 insertions, 3 deletions

diff --git a/cursor-pricing-research.md b/cursor-pricing-research.md
new file mode 100644
index 0000000..fe8f114
--- /dev/null
+++ b/cursor-pricing-research.md
@@ -0,0 +1,187 @@
+# Cursor: Models & Pricing Research
+
+Research compiled 2026-06-03. Sources: cursor.com/pricing,
+cursor.com/docs/models-and-pricing, Morph, CloudZero, Vantage, Apidog,
+eesel AI, devtoolsreview, beyondtmrw.org (Composer 2.5 release coverage).
+
+---
+
+## 1. What Cursor is
+
+Cursor is an AI-native code editor (VS Code fork) made by Anysphere.
+It bundles frontier models (Claude, GPT, Gemini, DeepSeek, plus its own
+Cursor Composer) into agentic editing, multi-file Composer edits, Tab
+autocomplete, and cloud/background agents. As of early 2026 Anysphere
+reportedly crossed $2B ARR with >1M paying subscribers, used by 64% of
+Fortune 500.
+
+---
+
+## 2. Subscription plans (cursor.com/pricing, May 2026)
+
+| Plan        | Price            | Annual (~20% off) | Credit pool     | Notes                              |
+|-------------|------------------|-------------------|-----------------|------------------------------------|
+| Hobby       | Free             | —                 | None            | Limited Agent + Tab, ~1 wk Pro trial |
+| Pro         | $20 / mo         | ~$16 / mo         | $20             | All frontier models, Cloud agents  |
+| Pro+        | $60 / mo         | ~$48 / mo         | $60 (3× Pro)    | Heavy daily coding                 |
+| Ultra       | $200 / mo        | ~$160 / mo        | $400 (20× Pro)  | Power users, priority features     |
+| Teams       | $40 / user / mo  | ~$32 / user / mo  | $20 / seat      | SSO, admin, analytics              |
+| Enterprise  | Custom           | Annual only       | Pooled usage    | SCIM, audit logs, SLA              |
+
+Key shift: in June 2025 Cursor moved from "500 fast requests / month"
+to a credit-based system — your subscription buys a $ pool of API
+credits. After the pool is gone, overages are billed at the underlying
+model's API rate (no penalty markup), or you can upgrade.
+
+---
+
+## 3. Models Cursor exposes
+
+Per the docs and 2026 coverage, the model picker includes roughly:
+
+- **OpenAI**: GPT-5.4, GPT-5.4-mini, GPT-5.5, o3-mini, GPT-4.5
+- **Anthropic**: Claude 4 Sonnet, Claude 4 Opus, Claude 4.6 Opus
+- **Google**: Gemini 3 Pro
+- **xAI / DeepSeek / others** appear intermittently
+- **Cursor Composer 2.5** — Cursor's own agentic coding model
+  (built on Moonshot Kimi K2.5 with Cursor's own RL post-training),
+  released 2026-05-18, default in Agent mode
+- **Auto** — Cursor's router that picks a cost-efficient model for you
+
+Exact availability shifts frequently; the docs page
+(cursor.com/docs/models-and-pricing) is the live source.
+
+---
+
+## 4. Per-model cost (per million tokens, May 2026 list prices)
+
+Cursor charges the underlying provider's API rate. A representative
+slice (numbers from Cursor docs and third-party reports):
+
+| Model                       | Input $/MTok | Output $/MTok | Relative cost in Cursor pool |
+|-----------------------------|--------------|---------------|------------------------------|
+| Cursor Composer 2.5 Standard| $0.50        | $2.50         | cheapest frontier agent      |
+| Cursor Composer 2.5 Fast    | $3.00        | $15.00        | default in product, low-latency |
+| Cursor Auto (router)        | $0.25 cache read / $1.25 input / $6.00 output | — | "included" on paid plans — does **not** drain the credit pool |
+| GPT-5.4 (typical fast)      | $2.50        | $15.00        | 1× base                      |
+| GPT-5.4-mini                | ~            | ~             | 0.5× (cheapest non-Auto)     |
+| Claude 4 Sonnet             | ~$3 / $3-5   | ~$15          | ~1× base                     |
+| Gemini 3 Pro                | ~$1-2        | ~$6-12        | ~1× base                     |
+| o3-mini                     | ~$3          | ~$12          | ~2× (reasoning)              |
+| Claude 4 Opus               | $5.00        | $25.00        | 5–10× — drains the pool fast  |
+| GPT-4.5                     | ~$5-10       | ~$15-30       | 5–10×                        |
+
+These are public-list / community-derived numbers; Cursor does not
+publish a single tidy table, but Settings → Account → Usage shows your
+real per-request burn.
+
+Approximate requests per $20 Pro pool (community reports, late 2025 / 2026):
+
+- ~500 with GPT-5.4 / GPT-5.4-mini
+- ~225 with Claude 4 Sonnet
+- ~45–90 with Claude 4 Opus or GPT-4.5
+- **Unlimited** with Auto (no pool deduction)
+
+Cursor does **not** charge a markup on Auto on paid plans, but adds a
+"Cursor Token Rate" of $0.25 / MTok on top of API pricing for non-Auto
+agent requests on Teams plans.
+
+---
+
+## 5. The two modes (Normal vs Max)
+
+| | Normal mode | Max mode |
+|---|---|---|
+| Pricing | Fixed per-request, drawn from credit pool | Token-based: API rate + 20% margin |
+| Tool calls / interaction | 25 | 200 |
+| Context | Truncated to ~10–15K | Full model context (up to 200K, 1M for some) |
+| Slow fallback | Yes (10/day after fast is out) | No — requires usage-based billing |
+
+Max Mode is where the surprise bills come from: a single complex
+session with 150 tool calls, 200K input, 20K output can cost $3–8;
+three of those per day = $180–500 / month on top of the subscription.
+
+---
+
+## 6. How the price relates to Auto
+
+This is the single most important thing to understand about Cursor's
+pricing in 2026.
+
+**Auto mode is the only "free" path on a paid plan.** When you let
+Cursor pick the model, the request is included — it does not consume
+your credit pool. Auto is priced by Cursor at roughly:
+
+- $0.25 / MTok cache read
+- $1.25 / MTok input
+- $6.00 / MTok output
+
+…but on a paid plan those amounts are absorbed into the subscription
+(per the Vantage breakdown). You only pay from the pool when you
+**manually pin** a frontier model (Claude Sonnet, GPT-4.5, Opus,
+etc.) or when you switch to **Max Mode** (which is always metered).
+
+Practical effect:
+
+- **Auto-only workflow** → $20 Pro feels essentially "unlimited" for
+  chat/agent; same as the old 500-request plan, possibly more
+  generous.
+- **Hand-picked Sonnet for everything** → ~225 requests / month from
+  the $20 pool, then overages at API rate.
+- **Hand-picked Opus / Max Mode for heavy work** → $20 pool can be
+  gone in a single session; $200–500 / mo bills are reported.
+
+So the relationship is:
+
+> The credit pool is the *budget for explicit model choices*. Auto is
+> the *included, pooled* tier. The more you let Auto pick, the closer
+> Cursor behaves to flat-rate; the more you pin a frontier model or
+> use Max Mode, the more it behaves like raw metered API spend with a
+> small Cursor markup.
+
+---
+
+## 7. Hidden cost traps to watch
+
+- **Overage billing** — past the pool you can opt into pay-as-you-go
+  at API rate. $0.04 / "premium request" was the old rule; under the
+  credit system it's straight token cost. Set a hard cap.
+- **Max Mode** — turns off after fast requests run out; needs
+  usage-based billing. Big-billed silently.
+- **Tab completions** — unlimited on paid plans, but model-powered
+  ones still draw from the pool.
+- **Background / Cloud Agents** — metered per task, often Max-Mode
+  pricing.
+- **No credit roll-over** — unused pool evaporates each month.
+- **Composer vs Auto** — Composer 2.5 Standard at $0.50 / $2.50 is
+  dramatically cheaper per task than Opus ($5 / $25) at near-parity
+  coding intelligence (79.8% vs 80.5% on SWE-Bench Multilingual), so
+  pinning the cheapest agent model is often the best $/quality trade.
+- **BYOK** — Bring Your Own OpenAI / Anthropic key is supported;
+  bypasses the pool entirely, you pay the provider directly. Useful
+  for users with negotiated enterprise API pricing.
+
+---
+
+## 8. Quick "what should I expect to pay" guide
+
+| Usage pattern | Plan that fits | Realistic monthly cost |
+|---|---|---|
+| Hobby / evaluate | Hobby | $0 |
+| Individual dev, mostly Auto + occasional Sonnet | Pro | $20 |
+| Heavy agent / pinned-Sonnet | Pro+ | $60 |
+| All-day Opus / Max Mode | Ultra | $200 + likely overages |
+| 3-person team, mostly Auto | Teams | $120 ($40×3) |
+| 25-person engineering org | Enterprise | Custom, pooled; budget ~$12K–$30K/yr |
+
+---
+
+## TL;DR
+
+Cursor charges a subscription that includes (a) a dollar pool of API
+credits and (b) **unlimited Auto-mode usage**. Picking a frontier model
+by hand burns the pool at 1× (cheap models) to 10× (Opus / GPT-4.5) the
+base rate. Max Mode is always metered at API rate + 20% and is where
+unexpected bills hide. The cheapest competitive path is Auto + Cursor
+Composer 2.5 Standard ($0.50 / $2.50 per MTok), which is roughly 10×
+cheaper than Opus at near-parity coding benchmark scores.
diff --git a/nvidia-rtx-research.md b/nvidia-rtx-research.md
new file mode 100644
index 0000000..a785a6c
--- /dev/null
+++ b/nvidia-rtx-research.md
@@ -0,0 +1,428 @@
+# NVIDIA RTX Platform Research (June 2026)
+
+Compiled 2026-06-03. Sources: nvidia.com (GeForce, RTX AI PC, DGX
+Spark, RTX Spark, changelog), en.wikipedia.org/wiki/GeForce_RTX_50_series,
+Tom's Hardware, PCMag, The Verge, The Register, The Guardian,
+Engadget, MacRumors, TechSpot, runpod.io, videocardz, wccftech,
+hotHardware, ofzenandcomputing, bestvaluegpu, brave search summaries.
+
+---
+
+## The correction: RTX Spark ≠ DGX Spark
+
+This is important and I got it wrong in the first draft.
+
+**Two separate products that share the same superchip silicon:**
+
+| | **RTX Spark** | **DGX Spark** |
+|---|---|---|
+| **Audience** | Consumers, creators, gamers, AI PC users | AI developers, researchers, data scientists |
+| **OS** | **Windows 11 on Arm** | **DGX OS (Ubuntu 24.04 custom)** |
+| **Form factor** | **Slim laptops** (as thin as 14 mm) **+ small desktops** | Standalone 1.2 kg mini-PC |
+| **Power budget** | 80 W (laptop) / 100 W (desktop) | ~200 W (the GB10 at rated performance) |
+| **Launched** | Announced 2026-06-01 at **Computex 2026**; ships **fall 2026** | Announced CES 2025 as "Project Digits"; **shipping since Oct 2025** |
+| **Available from** | ASUS, Dell, HP, Lenovo, Microsoft Surface, MSI (Acer, GIGABYTE to follow) | NVIDIA + OEM partners |
+| **Initial price** | Not announced; partner SKUs (Microsoft Surface Laptop Ultra among them) TBA | $3,999 launch, ~$5,000 in 2026 due to memory shortage |
+
+The silicon is the same **GB10 Grace Blackwell Superchip** — Blackwell
+GPU + Grace Arm CPU + up to 128 GB LPDDR5x unified memory, NVLink-C2C
+chip-to-chip interconnect. NVIDIA's productization of that chip
+into two product lines (consumer Windows laptop vs Linux AI dev
+workstation) is the "new RTX platform" story in 2026.
+
+Sources for the split:
+- The Register, 2026-06-01: "The silicon may be the same but the
+  operating system isn't. While Nvidia's DGX Spark and GB10 partner
+  systems shipped with DGX OS, a lightly customized version of
+  Ubuntu 24.04, RTX Spark systems will ship with Windows."
+- PCMag: "The key difference is that RTX Spark is specifically
+  meant for consumers and the Windows 11 OS, whereas DGX Spark runs
+  a custom version of Ubuntu Linux."
+- NVIDIA Developer Forum (engineer response): "The CPU/GPU in the
+  RTX products is similar to the GB10 in terms of ARM cores and
+  tensors. The DGX spark can consume 200 W at rated performance,
+  while the RTX systems are 80 W (laptop) and 100 W (desktop)."
+- Tom's Hardware: "RTX Spark hasn't come out of nowhere; it's the
+  consumer-oriented sibling of the GB10 Grace Blackwell superchip
+  already shipping inside the Linux-based DGX Spark mini-PC."
+
+The user was right to push back. Let me make sure the rest of the
+research reflects this correctly.
+
+---
+
+## 1. The 2026 NVIDIA RTX platform: the full picture
+
+NVIDIA's RTX brand in 2026 covers three product lines built on the
+**Blackwell** architecture (TSMC 4N / 3N process):
+
+1. **GeForce RTX 50 Series** — discrete consumer GPUs (desktop +
+   laptop). Launched January 2025.
+2. **NVIDIA RTX Spark** — a brand-new Windows-on-Arm laptop /
+   compact-desktop platform, announced at Computex 2026 (June 1).
+   Consumer sibling of the GB10 superchip.
+3. **NVIDIA DGX Spark** — the Linux AI dev workstation cousin of
+   RTX Spark, launched 2025 (originally "Project Digits").
+
+The unifying message NVIDIA is pushing across all three in 2026 is
+"personal AI computer" / "agentic AI on device" — local LLMs, local
+agents (NemoClaw / OpenShell), local content creation, plus the
+gaming / creator / productivity workloads that already lived on RTX.
+
+---
+
+## 2. GeForce RTX 50 Series ("Blackwell for the rest of us")
+
+### 2.1 Announcement and release
+
+- **Announced:** CES 2025, January 6, 2025
+- **First cards on sale:** January 30, 2025 (RTX 5090, 5080, 5070)
+- **Full stack rolled out** through Q1–Q3 2025
+- **Fabrication:** TSMC custom node "4N" (5 nm-class)
+- **Interface:** PCIe 5.0 (first consumer GPUs to use it; the 5050
+  uses x8, everything else x16)
+- **Memory:** GDDR7 across the lineup (RTX 5050 is the odd one out
+  and still ships with GDDR6); first consumer GPUs with GDDR7
+- **Power connector:** mandated 16-pin 12V-2x6 (the safer revision
+  of the RTX 4090's 12VHPWR) on all AIB designs
+- **Display output:** DisplayPort 2.1b UHBR20 (80 Gbps) + HDMI 2.1b
+  — first GeForce to support 4K @ 480 Hz or 8K @ 165 Hz with DSC
+- **Media engine:** 9th-gen NVENC (3 on 5090, 2 elsewhere) +
+  6th-gen NVDEC (2 on 5080/5090, 1 elsewhere); first GeForce with
+  4:2:2 hardware encode/decode for pro video
+
+### 2.2 The lineup (MSRP, launch date, key specs)
+
+| Card           | Launch MSRP | Date         | CUDA   | VRAM          | Bus   | TDP  |
+|----------------|-------------|--------------|--------|---------------|-------|------|
+| RTX 5050       | $249        | Jul 2025     | 2,560  | 8 GB GDDR6    | 128b  | 130 W |
+| RTX 5060       | $299        | May 2025     | 3,840  | 8 GB GDDR7    | 192b  | 145 W |
+| RTX 5060 Ti    | $379 (8 GB) / **$429 (16 GB)** | Apr 2025 | 4,608 | 8/16 GB GDDR7 | 256b | 180 W |
+| RTX 5070       | $549        | Mar 2025     | 6,144  | 12 GB GDDR7   | 256b  | 250 W |
+| RTX 5070 Ti    | $749        | Feb 20, 2025 | 8,960  | 16 GB GDDR7   | 256b  | 300 W |
+| RTX 5080       | $999        | Jan 30, 2025 | 10,752 | 16 GB GDDR7   | 256b  | 360 W |
+| RTX 5090       | **$1,999**  | Jan 30, 2025 | 21,760 | 32 GB GDDR7   | 512b  | 575 W |
+
+**AI TOPS (the marketing number):** RTX 5090 — 3,352; 5080 — 1,801;
+5070 — 1,000-ish (per-card in NVIDIA's spec sheets).
+
+### 2.3 What's new vs RTX 40 series (Ada Lovelace)
+
+- **5th-gen Tensor Cores** with **FP4** precision. FP4 is the step
+  that made local LLMs (Llama 3.1 8B in int4, etc.) actually
+  runnable on a consumer GPU. RTX 50 quotes "4× faster LLM chat"
+  vs a non-RTX laptop.
+- **4th-gen RT Cores** "built for Mega Geometry" — much higher
+  ray-triangle throughput, enabling full path tracing in shipping
+  titles.
+- **DLSS 4** — a vision-transformer-based upscaling model (vs the
+  CNN model in DLSS 3) and the new **Multi Frame Generation (MFG)**,
+  exclusive to RTX 50. Where DLSS 3 interpolated 1 frame per
+  rendered frame, MFG generates up to 3 additional frames per
+  rendered frame; combined with the standard 1 that gives the
+  "4× MFG" / 6× MFG modes in the marketing. 75 titles had MFG at
+  launch.
+- **DLSS 4.5** (mid-2026) — adds **Dynamic MFG** (frame count
+  adapts to scene complexity) and a 2nd-gen transformer for ray
+  reconstruction.
+- **Reflex 2 with Frame Warp** — further input-latency reduction
+  by warping frames based on the most recent mouse input.
+- **NVIDIA ACE / NIM on RTX** — RTX 50 GPUs are positioned as
+  on-device AI PCs running NIM microservices for Llama, Riva
+  speech, FLUX vision, retrieval models. Local agents, local
+  LLMs, local Stable Diffusion.
+- **Max-Q updates for laptops** — Advanced Power Gating + an
+  "ultra" low-voltage GDDR7 state, claimed to give up to 40%
+  battery-life gain over RTX 40 laptops.
+
+### 2.4 Real-world pricing in 2026
+
+The MSRP column above is *launch list*. By mid-2026 the picture is
+messed up:
+
+- **RTX 5090**: per Wccftech, HotHardware, VideoCardz, Best Value
+  GPU, retailers in early 2026 had no stock at $1,999; mean eBay
+  sale price was ~$4,086 (Q1 2025) and the **MSRP card was still
+  basically impossible to buy at MSRP in early 2026**, with real
+  street prices $3,000–$5,000 driven by scalpers, AI demand, and
+  the 2025–2026 GDDR7 / VRAM shortage. One-year retrospectives
+  called it "nearly twice the MSRP." Best Value GPU's June 2026
+  tracker showed ~$4,199 baseline.
+- **RTX 5080** had somewhat better availability but was also above
+  MSRP for most of 2025.
+- **RTX 5070 / 5070 Ti** were the *actually attainable at MSRP*
+  cards through 2025; 5070 specifically was a $549 1440p workhorse.
+
+The reason for the squeeze: AI demand is siphoning Blackwell
+allocation, GDDR7 supply is constrained, and tariffs during the
+2024–2025 stockpiling window distorted inventories.
+
+### 2.5 Reception and controversies
+
+- **Jensen's "5070 = 4090 performance" claim** turned out to rely
+  on DLSS 4 + MFG (i.e., generated frames, not raw rasterization).
+  Wikipedia and several outlets flagged this as misleading.
+- **RTX 5090 power** jumped to 575 W (vs 450 W on the 4090) — the
+  highest of any consumer GeForce ever, requiring a 1,000 W PSU
+  recommendation and the new 12V-2x6 connector.
+- **Missing ROPs defect** in early 5090/5080 batches (some AIB
+  cards shipped with fewer than advertised render output units;
+  NVIDIA offered replacements).
+- **Laptop variants** ship from March 2025 starting at $1,299 for
+  the RTX 5070 laptop GPU, $2,599+ for the 5090 laptop tier.
+
+---
+
+## 3. NVIDIA RTX Spark (Computex 2026, the genuinely new platform)
+
+This is the actually new product category unveiled 2026-06-01 at
+Computex Taipei. NVIDIA's positioning: "the fusion of NVIDIA AI and
+RTX graphics in a single chip redefines Windows PCs."
+
+### 3.1 The chip: "RTX Spark Superchip"
+
+Same silicon family as the DGX Spark's GB10, but tuned for the
+80 W (laptop) / 100 W (desktop) power envelope:
+
+- **CPU:** 20-core NVIDIA Grace (Arm), built with **MediaTek**.
+  Same Cortex-X925 + Cortex-A725 layout as the GB10.
+- **GPU:** NVIDIA Blackwell RTX, **up to 6,144 CUDA cores**,
+  5th-gen Tensor Cores with FP4, RT Cores.
+- **Interconnect:** **NVLink-C2C** chip-to-chip between CPU and
+  GPU die (the same coherent interconnect as the GB10 in DGX
+  Spark).
+- **Memory:** **up to 128 GB LPDDR5x unified** (CPU + GPU share
+  one pool). Same ceiling as DGX Spark.
+- **AI performance:** **up to 1 PFLOP FP4** (with sparsity).
+- **Process:** **TSMC 3 nm**, ~70 billion transistors.
+- **AI-equivalent claim:** NVIDIA has said the integrated graphics
+  are equivalent to an **RTX 5070 laptop GPU**.
+
+Compared to the DGX Spark GB10:
+- Same CPU + GPU core topology
+- Same 128 GB unified memory ceiling
+- Lower power budget (80 W / 100 W vs ~200 W) → lower sustained
+  clocks
+- **No ConnectX-7 200 Gbps networking** in the consumer SKUs (that
+  stays as a DGX Spark / DGX Station data-center feature, although
+  OEMs may add it)
+- **No NemoClaw / NIM enterprise stack** — RTX Spark gets the
+  consumer agent runtime (Windows-native agents on the new OS
+  security primitives + NVIDIA OpenShell)
+
+### 3.2 Form factor and partners
+
+- **Laptops:** chassis as thin as 14 mm and as light as ~3 lb
+  (~1.36 kg), precision-machined aluminum. Claim: "the most
+  power-efficient RTX chip ever made, in a chassis so slim you'll
+  forget you're carrying it."
+- **Desktops:** small, ultra-efficient desktops, marketed for
+  "always-on AI agent use cases."
+- **Laptop launch partners (Computex 2026 reveal):** ASUS, Dell,
+  HP, Lenovo, **Microsoft Surface** (the **Surface Laptop Ultra**
+  is the most prominent), MSI. Acer and GIGABYTE to follow.
+- **Microsoft collaboration:** Windows 11 on Arm is the OS.
+  Microsoft and NVIDIA built new OS security primitives for
+  on-device agents. NVIDIA OpenShell provides the agent runtime.
+
+### 3.3 Confirmed RTX Spark laptops at Computex 2026
+
+(Per PCMag's "Every Nvidia RTX Spark Laptop Announced So Far" and
+the HP press release on their own site)
+
+- **Microsoft Surface Laptop Ultra** — flagship Windows-on-ARM
+  reference design, "signals a raw power revolution" per PCMag
+- **HP OmniBook Ultra 16** and **HP OmniBook X 14** — HP
+  claims the X 14 will be "the world's thinnest RTX Spark"
+- **ASUS ProArt P16** — creator-targeted
+- **Dell XPS 16** — premium thin-and-light
+- **Lenovo Yoga Pro 9n**
+- **MSI Prestige N16 Flip** — 2-in-1 convertible
+
+Pricing was **not announced** at the Computex keynote. Reddit /
+nvidia reaction: a 128 GB unified-memory SKU is widely expected to
+land in the $3,000–$5,000+ range.
+
+### 3.4 Software / agent story
+
+- **NVIDIA OpenShell** — open-source runtime for open-weight
+  agents, the consumer counterpart to the DGX Spark's
+  NemoClaw / OpenClaw enterprise stack.
+- **Project G-Assist** — on-device AI assistant for tuning /
+  controlling the PC.
+- **NIM microservices on RTX** — language, speech, vision,
+  retrieval, design models running locally.
+- **DLSS 4.5, Reflex 2, G-SYNC, OptiX, TensorRT, CUDA** — the
+  full NVIDIA stack runs natively on RTX Spark (the first time
+  the entire CUDA + RTX stack has been on a Windows-on-Arm SoC).
+- **Native 4:2:2 hardware encode/decode, AV1 encoders, NVIDIA
+  Broadcast** — for creator / streaming workflows.
+- **AI Blueprints** — pre-built agentic workflows
+  (PDF-to-Podcast, 3D object generation, 3D-guided generative AI).
+
+### 3.5 Why this matters strategically
+
+This is NVIDIA's first direct shot at the Apple Silicon /
+Qualcomm / AMD Strix Halo "AI PC" category, but with the
+full CUDA + RTX stack on the chip. Key competitive angles:
+
+- **vs Apple M-series:** Unified memory and per-watt efficiency
+  in the same ballpark, but you get CUDA, RTX graphics, full
+  Windows, and a "1 PFLOP FP4" AI claim.
+- **vs Qualcomm Snapdragon X (Windows on Arm):** Qualcomm's
+  exclusivity deal just expired; NVIDIA is the obvious next
+  serious Windows-on-Arm player. PCMag headline: "Welcome to the
+  Superchip Era: 6 Ways the Nvidia RTX Spark Will Upend the PC
+  Industry."
+- **vs AMD Strix Halo (Ryzen AI Max+ 395):** AMD already has
+  128 GB unified memory on a single chip in 2025, and Tom's
+  Hardware's DGX Spark review noted that the GB10 "beats out
+  AMD's Ryzen AI Max+ 395" in many AI workloads. RTX Spark is
+  the laptop counterpart.
+- **vs Intel:** per Yahoo Finance, "taking aim at Intel and AMD
+  with the debut of the RTX Spark superchip for Windows
+  laptops."
+
+---
+
+## 4. NVIDIA DGX Spark (the Linux AI dev workstation)
+
+This is the one that already shipped. Announced CES 2025 as
+"Project Digits," renamed to DGX Spark at GTC 2025, shipping
+since October 2025.
+
+### 4.1 What it is
+
+- **Form factor:** 150 × 150 × 50.5 mm, 1.2 kg — Mac Mini / Intel
+  NUC class chassis. Quiet (35 dB idle).
+- **Power:** 240 W external supply, GB10 TDP ~140 W.
+- **OS:** DGX OS (Ubuntu 24.04 with NVIDIA's stack).
+
+### 4.2 The chip: GB10 Grace Blackwell Superchip
+
+- **CPU:** 20-core Arm — 10× Cortex-X925 + 10× Cortex-A725
+- **GPU:** Blackwell with 5th-gen Tensor Cores (FP4) and 4th-gen
+  RT Cores
+- **Memory:** **128 GB LPDDR5x coherent unified**, 256-bit
+  interface, **273 GB/s bandwidth** (the bottleneck — much lower
+  than H100's HBM3)
+- **Storage:** 4 TB NVMe M.2 with self-encryption
+- **Networking:** 1× RJ-45 10 GbE **+** a **ConnectX-7 NIC at
+  200 Gbps** (the killer feature — you can link two Sparks to
+  work with 405-billion-parameter models)
+- **I/O:** 4× USB Type-C, Wi-Fi 7, BT 5.4, 1× HDMI 2.1a + up to
+  3× DisplayPort over USB-C, 1× NVENC, 1× NVDEC
+
+### 4.3 Performance claims
+
+- **Up to 1 PFLOP FP4 AI** (with sparsity)
+- 128 GB unified memory → **AI models up to 200B parameters** for
+  inference on a single unit
+- **Fine-tune models up to 70B parameters** on one unit
+- **Two units linked** via ConnectX-7 → 405B parameters
+- Pre-installed DGX OS with the full NVIDIA AI software stack:
+  NeMo, RAPIDS, NIM microservices, Isaac, Metropolis, Holoscan
+  for robotics / edge
+
+### 4.4 Pricing and availability
+
+- **Launch price:** **$3,999** (held through 2025)
+- **2026 real-world:** **~$5,000** due to the same GDDR7 / memory
+  shortage that pushed the RTX 5090 above MSRP
+- **Channel:** NVIDIA + OEM partners (Dell Pro Max, ASUS, MSI, HP
+  all ship GB10-based boxes; the Spark itself is the quietest of
+  the bunch per InsiderLLM's comparison)
+- **DGX Station** — a higher-end GB300 / Blackwell Ultra
+  workstation variant also announced in 2025 for users who need
+  more memory bandwidth than the Spark's 273 GB/s; a "DGX Station
+  for Windows" was teased by Microsoft at the RTX Spark launch
+
+### 4.5 Software story (Linux / AI dev focus)
+
+- **NVIDIA NemoClaw** — part of the Agent Toolkit, an open-source
+  reference stack that adds security/privacy guardrails to
+  **OpenClaw** (the local-agent runtime). Runs on RTX PCs, DGX
+  Station, and DGX Spark.
+- **NVIDIA OpenShell** — open-source runtime for open-weight
+  agents (the enterprise version).
+- Pre-installed DGX OS with the entire NVIDIA AI stack.
+- Designed to be the desktop counterpart to cloud-hosted agents;
+  "always-on, private, on-device" is the pitch.
+
+---
+
+## 5. RTX 50 series vs RTX Spark vs DGX Spark — how to think about them
+
+| | **RTX 50 Series (GeForce)** | **RTX Spark (laptop/desktop)** | **DGX Spark (mini-PC)** |
+|---|---|---|---|
+| Audience | gamers, creators, AI PC users | consumers, creators, gamers, AI PC developers | AI developers, researchers, robotics |
+| Form factor | discrete GPU in a desktop / laptop | slim laptop or small desktop | standalone 1.2 kg mini-PC |
+| OS | Windows / Linux | **Windows 11 on Arm** | **DGX OS (Ubuntu 24.04)** |
+| Memory | 8–32 GB GDDR7 (high bandwidth) | up to 128 GB LPDDR5x unified | 128 GB LPDDR5x unified |
+| Memory bandwidth | 320–1,792 GB/s | lower (273 GB/s class) | 273 GB/s |
+| AI throughput | 1,801–3,352 INT8/FP8 TOPS | 1 PFLOP FP4 (chip-level, ~RTX 5070 laptop equivalent graphics) | 1 PFLOP FP4 |
+| Power | 130–575 W (card) | 80 W (laptop) / 100 W (desktop) | 240 W (whole box) |
+| Sweet spot | gaming, on-device 8B–13B LLMs, SD/Flux | local 70B-class LLMs, agentic AI, slim Windows laptop | 200B inference, 70B fine-tuning, edge / robotics |
+| Price | $249 (5050) → $1,999+ (5090) | TBA — expected $2,000–$5,000+ | $3,999 launch, ~$5,000 in 2026 |
+| Shipped | Jan 2025 | Fall 2026 | Oct 2025 |
+
+They're complementary, not competing. Pick the right tool for the
+job:
+
+- **Gaming / mainstream creator** → GeForce RTX 50 series
+- **Slim AI-first Windows laptop** with massive unified memory →
+  RTX Spark laptop
+- **Linux dev workstation for serious local model work** → DGX Spark
+- **Gaming + light AI dev on one machine** → RTX 5090 + 5090
+  laptops are still the best mix
+
+---
+
+## 6. The rest of the 2026 RTX stack
+
+- **GeForce NOW** — cloud gaming, still positioned as the "any
+  device" RTX experience
+- **G-SYNC displays** — G-SYNC Pulsar is the current top tier
+- **NVIDIA Studio** — RTX-accelerated creative suite
+- **NVIDIA Broadcast** — AI webcam/mic for streamers (v2.2 in
+  2026)
+- **RTX Video** — RTX Video Super Resolution + HDR, browser/VLC
+- **RTX Remix** — modders' tool to remaster classic games with
+  full path tracing + DLSS
+- **Project G-Assist** — local AI assistant for tuning/optimizing
+  the PC
+- **DLSS 4.5** — Dynamic MFG + 2nd-gen transformer, available
+  across the 50 series
+- **Reflex 2 / Frame Warp** — competitive latency reduction
+
+---
+
+## 7. TL;DR
+
+- The "new NVIDIA RTX platform" in 2026 is **three** things, not
+  two:
+  1. **GeForce RTX 50 series** (Blackwell consumer GPUs) —
+     launched January 2025, 7 SKUs from RTX 5050 ($249) to
+     RTX 5090 ($1,999 MSRP, real-world $3k–$5k in 2026).
+     Headline features: 5th-gen Tensor Cores with FP4, 4th-gen
+     RT Cores, DLSS 4 Multi Frame Generation, GDDR7, PCIe 5.0,
+     12V-2x6 power.
+  2. **NVIDIA RTX Spark** — **brand-new** at Computex 2026, ships
+     fall 2026. A consumer Windows-on-Arm laptop and small-desktop
+     platform built on the "RTX Spark Superchip" (Blackwell GPU +
+     Grace Arm CPU + up to 128 GB unified memory + 1 PFLOP FP4 AI).
+     Laptops as thin as 14 mm. From ASUS, Dell, HP, Lenovo,
+     Microsoft Surface (Laptop Ultra), MSI. Built around Windows
+     11 on Arm + NVIDIA OpenShell for on-device agents.
+  3. **NVIDIA DGX Spark** — the Linux AI dev workstation cousin of
+     RTX Spark, shipped since October 2025. Same GB10 superchip
+     silicon, 128 GB unified memory, ConnectX-7 200 Gbps
+     networking for linking two units, runs DGX OS. $3,999 launch
+     (~$5,000 in 2026). For running 200B inference and 70B
+     fine-tuning locally.
+- All three share the **Blackwell** architecture and the **RTX**
+  brand, but they target very different users.
+- The strategic 2026 story is RTX Spark as NVIDIA's direct
+  answer to Apple Silicon and Qualcomm Snapdragon X in the
+  Windows-on-Arm AI PC category — and the first time the entire
+  CUDA + RTX stack has been on a single Windows-on-Arm SoC.
diff --git a/pi/agent/models.json.bak b/pi/agent/models.json.bak
new file mode 100644
index 0000000..f071691
--- /dev/null
+++ b/pi/agent/models.json.bak
@@ -0,0 +1,743 @@
+{
+  "providers": {
+    "hyperstack": {
+      "baseUrl": "http://hyperstack1.wg1:11434/v1",
+      "apiKey": "$EMPTY",
+      "api": "openai-completions",
+      "compat": {
+        "supportsStore": false,
+        "supportsDeveloperRole": false,
+        "maxTokensField": "max_tokens"
+      },
+      "models": [
+        {
+          "id": "Qwen/Qwen3.6-27B-FP8",
+          "name": "Qwen3.6 27B FP8 [vm]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 262144,
+          "maxTokens": 8192,
+          "compat": {
+            "thinkingFormat": "qwen-chat-template",
+            "supportsStore": false,
+            "supportsDeveloperRole": false,
+            "maxTokensField": "max_tokens"
+          }
+        },
+        {
+          "id": "cyankiwi/gemma-4-31B-it-AWQ-4bit",
+          "name": "Gemma 4 31B IT [vm]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 131072,
+          "maxTokens": 8192
+        },
+        {
+          "id": "cyankiwi/NVIDIA-Nemotron-3-Super-120B-A12B-AWQ-4bit",
+          "name": "Nemotron 3 Super 120B [vm]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 262144,
+          "maxTokens": 8192
+        },
+        {
+          "id": "Qwen/Qwen2.5-Coder-32B-Instruct-AWQ",
+          "name": "Qwen2.5 Coder 32B [vm]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 32768,
+          "maxTokens": 8192
+        },
+        {
+          "id": "QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ",
+          "name": "Qwen3 Coder 30B [vm]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 65536,
+          "maxTokens": 8192,
+          "compat": {
+            "thinkingFormat": "qwen-chat-template",
+            "supportsStore": false,
+            "supportsDeveloperRole": false,
+            "maxTokensField": "max_tokens"
+          }
+        },
+        {
+          "id": "casperhansen/deepseek-r1-distill-qwen-32b-awq",
+          "name": "DeepSeek-R1-Distill 32B [vm]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 32768,
+          "maxTokens": 8192
+        },
+        {
+          "id": "Qwen/Qwen3-32B-AWQ",
+          "name": "Qwen3 32B [vm]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 32768,
+          "maxTokens": 8192,
+          "compat": {
+            "thinkingFormat": "qwen-chat-template",
+            "supportsStore": false,
+            "supportsDeveloperRole": false,
+            "maxTokensField": "max_tokens"
+          }
+        },
+        {
+          "id": "cyankiwi/Devstral-Small-2507-AWQ-4bit",
+          "name": "Devstral Small 2507 [vm]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 32768,
+          "maxTokens": 8192
+        }
+      ]
+    },
+    "hyperstack1": {
+      "baseUrl": "http://hyperstack1.wg1:11434/v1",
+      "apiKey": "$EMPTY",
+      "api": "openai-completions",
+      "compat": {
+        "supportsStore": false,
+        "supportsDeveloperRole": false,
+        "maxTokensField": "max_tokens"
+      },
+      "models": [
+        {
+          "id": "cyankiwi/gemma-4-31B-it-AWQ-4bit",
+          "name": "Gemma 4 31B IT [vm1]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 131072,
+          "maxTokens": 8192
+        },
+        {
+          "id": "cyankiwi/NVIDIA-Nemotron-3-Super-120B-A12B-AWQ-4bit",
+          "name": "Nemotron 3 Super 120B 1M [vm1]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 1048576,
+          "maxTokens": 8192
+        },
+        {
+          "id": "Qwen/Qwen3.6-27B-FP8",
+          "name": "Qwen3.6 27B FP8 [vm1]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 262144,
+          "maxTokens": 8192,
+          "compat": {
+            "thinkingFormat": "qwen-chat-template",
+            "supportsStore": false,
+            "supportsDeveloperRole": false,
+            "maxTokensField": "max_tokens"
+          }
+        },
+        {
+          "id": "Qwen/Qwen2.5-Coder-32B-Instruct-AWQ",
+          "name": "Qwen2.5 Coder 32B [vm1]",
+          "reasoning": false,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 32768,
+          "maxTokens": 8192
+        },
+        {
+          "id": "QuantTrio/Qwen3-Coder-30B-A3B-Instruct-AWQ",
+          "name": "Qwen3 Coder 30B [vm1]",
+          "reasoning": true,
+          "input": [
+            "text"
+          ],
+          "cost": {
+            "input": 0,
+            "output": 0,
+            "cacheRead": 0,
+            "cacheWrite": 0
+          },
+          "contextWindow": 65536,
+          "maxTokens": 8192,
+          "compat": {
+            "thinkingFormat": "qwen-chat-template",
+            "supportsStore": false,
+            "supportsDeveloperRole": false,
+            "maxTokensField": "max_tokens"
+          }
+        },
+        {
+          "id": "casperhansen/deepseek-r1-distill-qwen-32b-awq",
+          "name": "DeepSeek-R1-Distill 32B [vm1]",
+          "reasoning": true,
+          "input": [
+            "text"