Episode 112: Google Antigravity, ChatGPT Mac with WebMCP, and MiniMax-Text-01's 4M Context Engine

I'm Nova. I'm Alloy, and this is AgentStack Daily. ClaudeCode, the terminal-based AI coding agent, shipped version 0.267 on NPM alongside Hermes Agent 9.7. The update improves command line execution and tightens memory usage during multi-directory refactors, while Hermes hardens candidate state isolation so persistent background session state survives plugin repairs. Meanwhile, Google is unifying desktop, terminal, and programmatic agent development under the Google Antigravity ecosystem, OpenAI is rebuilding ChatGPT for Mac into a unified workspace with computer history and WebMCP site tools, and Minimax is delivering 4 million tokens of context with lightning attention. In the wider news cycle, Mercury 2.5 brings diffusion-style parallel reasoning to OpenRouter, Meta asks users to trust Muse with email, calendars, and payments, Cognition reaches a $48 billion valuation, and 1Password measures a 21% velocity lift from codecs. From dedicated agent desktop canvases to unified operating system integration and OpenWeight's reasoning architectures, there's plenty of concrete technology to unpack today. Let's get right into the releases. Clawed code arrived on NPM with targeted improvements for engineers running autonomous coding agents directly from the command line. The tool has become an essential harness for developers who prefer terminal-first workflows instead of working through an IDE chat panel. In this update, Anthropic focused on how the agent plans and executes multi-directory file modifications. When an agent touches multiple files across nested directory hierarchies, preserving difference hygiene on active GIT branches is critical to avoid unintended state mutations and broken builds. Clawed code now structures its file change plans more conservatively, inspecting working tree status before committing file modifications. In addition, sub-agent task delegation has been refined to keep prompt contexts compact during deep codebase exploration, ensuring that subsidiary research tasks don't bloat the primary conversation context. The runtime architecture also manages memory usage more predictably across extended coding sessions, reducing the risk of token serialization bottlenecks during long-running background tasks. Terminal output mechanisms have also received Polish, rendering structured difference previews and tool execution status codes cleanly in the console. On a parallel release track, Hermes Agent released version 9.7 with a strong focus on persistent gateway lifecycle management. Hermes now isolates candidate state during maintenance operations, which means background gateway processes can survive configuration changes, plug-in upgrades, and dependency repairs without severing active client connections or dropping in-flight sessions. The update also tightens memory reclamation during prolonged agent sessions, reclaiming resources aggressively after heavy tool use runs. Configuration options have been expanded to allow operators to define strict evaluation boundaries, while background watchers report latency and throughput metrics directly to system logs. For teams running persistent autonomous agents in production, these operational refinements mean background daemons can self-heal and update smoothly without stranding running workflows. That combination highlights where autonomous engineering tooling is maturing. The terminal agent has to handle dirty local GIT branches, uncommitted workspace changes, and unexpected exit codes without corrupting the repository. At the same time, the persistent gateway daemon running in the background needs to absorb updates and environment changes without forcing a manual reboot of the entire infrastructure. Both sides of the developer loop are getting noticeably more dependable, which builds real confidence for long-running workflows. Exactly. Neither tool is chasing cosmetic redesigns here. Cloud code is making command line file editing safer and less prone to context sprawl, while Hermes is making the long-running daemon more resilient against unexpected upgrade failures. If you're operating autonomous coding agents in your day-to-day development loop, both of these updates provide immediate stability improvements that keep your developer sessions online and your code trees clean. Google has unified its autonomous developer tooling under the Google Antigravity platform, establishing an agent-first development environment spanning desktop, terminal, and programmatic SDKs. Rather than treating artificial intelligence as an inline autocomplete widget inside a conventional code editor, Antigravity 2.0 provides a dedicated desktop workspace designed specifically for parallel agent orchestration. At the center of the interface is the auxiliary pane, which tracks live background tasks, sub-agent execution hierarchies, modified file artifacts, and interactive browser pages in real-time. Developers can assign distinct tasks to specialized sub-agents, monitor their step-by-step progress, and inspect intermediate diffs without losing their place in the primary code base. Honestly, for developers who work primarily in the terminal, the Antigravity CLI, known as AG and implemented in Go, delivers near-instant startup times and rapid command execution. The CLI acts as a native harness for running autonomous tasks, invoking sub-agents, and executing scripted repository maintenance directly from shell scripts or continuous integration pipelines. Complementing both tools is the Google Antigravity Python SDK, which allows teams to lease agents programmatically, connect external model context protocol servers, and design custom multi-agent workflows with fine-grained control over skills, rules, and workspace boundaries. The platform enforces a strict separation between planning and tool execution, prompting agents to formulate structured implementation plans before applying code edits. Telemetry streams report real-time latency and token throughput across active sub-agent networks, providing full auditability for automated development. Look, the strategic shift here is treating the agent as a primary collaborator rather than a glorified search bar. By providing native desktop management, a fast-go CLI, and a Python orchestration SDK, Antigravity establishes an end-to-end foundation for autonomous software engineering. It's a complete ecosystem redesign that rethinks how developers interact with parallel intelligence. OpenAI has delivered a comprehensive architectural rebuild of its macOS desktop application, merging chat, work, and codecs into a unified native client designed for deep workflow automation. The update rethinks how desktop assistants interact with the host operating system. The most significant new capability is computer history, an opt-in feature that constructs a searchable local timeline of user activity across supported applications and browser tabs. Crucially, OpenAI designed computer history without capturing invasive screenshots, continuous screen recordings, or audio feeds. Instead, the system records structured application metadata, enabling the model to reference past work context accurately while preserving user privacy. It's an intelligent approach to desktop memory that doesn't sacrifice security or expose personal screen contents to cloud storage. That's wild, and simultaneously the desktop app's built-in browser has gained support for WebMCP site tools across Chrome, Brave, Edge, Opera, and Vivaldi. WebMCP extends model context protocol principles to web pages, allowing websites to publish declarative action schemas that the assistant can invoke directly. This enables workflows like leaving comments in shared documents, modifying task statuses in issue trackers, or querying dashboard data without fragile web scraping. Paired with native Google Drive integration in the composer, the redesigned Mac OS client shifts chat GPT from a conversational sidebar into an active operational workspace. All timeline events are encrypted locally, and browser tool execution runs within strict sandbox boundaries to prevent unauthorized actions. Moving from passive screen scraping to structured metadata and standardized web tool schemas is a massive architectural upgrade that gives the desktop assistant genuine operational leverage without turning your workstation into a privacy compromise. It sets a new standard for desktop agent integration that prioritizes local control. Minimax has released Minimax Text01, a 456 billion parameter mixture of experts architecture activating 45.9 billion parameters per token. Engineered specifically for complex reasoning, autonomous software engineering, and long-horizon agent workflows, the model features a native 4 million token context window powered by Minimax's proprietary lightning attention mechanism. In conventional transformer architectures, standard quadratic self-attention causes compute overhead and memory consumption to balloon dramatically as sequence lengths enter the millions. Lightning attention resolves this bottleneck by utilizing chunked linear recurrence, ensuring constant memory requirements during pre-fill and sustaining high token generation throughput across millions of tokens. Minimax Text01 The training pipeline for Minimax Text01 incorporated massive mathematical reasoning datasets alongside synthetic execution traces from complex software projects. The resulting weights demonstrate remarkable proficiency on multi-step reasoning evaluations, coding agent tasks, and needle-in-a-haystack retrieval across exhaustive technical documentation. By releasing the model weights openly and supporting flexible deployment across high-performance-serving frameworks, Minimax provides developers with an open alternative to proprietary long-context APIs. For teams building autonomous agents that need to ingest entire code repositories, architectural diagrams, and extensive debugging logs in a single prompt, a 4 million token working memory eliminates the need for lossy chunking strategies and complex vector databases. It's a huge win for open weights that lets builders run long-context reasoning locally or in private clouds without recurring API costs. Right. An open weights model capable of reasoning across 4 million tokens without quadratic slowdown fundamentally changes the memory economics for autonomous engineering agents. You can keep an entire multi-repository project in active memory without losing subtle crossfile references. That opens up entirely new possibilities for repo-scale debugging and architecture-level refactoring that previously required brittle retrieval heuristics. Inception has made Mercury 2.5 available on open router, featuring a 260,000 token context window and a 4,096 token output limit. What distinguishes Mercury 2.5 from conventional auto-aggressive models is its diffusion-style decoding mechanism. Rather than predicting text strictly from left to right, one sequential token at a time, Mercury generates an entire token draft simultaneously and iteratively refines it across multiple denoising passes. This parallel refinement borrows the mathematical intuition of image diffusion and applies it directly to language generation. Inception positions Mercury 2.5 as its fastest reasoning model, arguing that parallel token refinement allows the architecture to solve complex multi-step reasoning problems without leaving accelerator hardware idling on sequential token prediction. On standard GPUs, generating long reasoning chains token by token often underutilizes compute cores. By updating multiple token positions in parallel, Mercury aims to drastically reduce time-to-first token and overall generation latency on intensive reasoning tasks. But wait, does parallel token refinement actually hold up when the model faces strict logical deductions? That's the real test. The 260,000 token context window accommodates substantial code repositories, technical documentation, and long conversational histories in a single call, but developers will want to see verified proof on coding benchmarks before swapping out auto-aggressive models in production stacks. Look, that's exactly the right question to ask. If parallel decoding cuts generation latency without degrading accuracy, it transforms interactive agent response times. If it struggles with exact syntax, it remains an experimental demo. We'll be watching closely to see how independent third-party benchmarks evaluate its reasoning fidelity across real programming tasks. It's an architecture worth tracking closely as parallel decoding models mature. Meta has launched Muse, an ambitious personal artificial intelligence agent that asks for direct access to user email, calendars, payment systems, and health services. Meta positions Muse as an active operational assistant that can complete tasks across applications, moving beyond conversational chatbots that simply suggest text in a chat window. The launch generated significant discussion on Hacker News, with more than 500 upvotes and intense debate over privacy boundaries. The pitch is obvious convenience. An assistant that notices a rescheduled flight, automatically updates calendar entries, notifies meeting attendees, and manages booking adjustments delivers genuine practical value. It's the vision of an assistant that truly handles daily logistics without endless manual intervention from the user. Here's the thing, handing over credentials for email, calendars, financial payment rails, and sensitive health records places an enormous amount of private infrastructure inside a single permissions boundary. Skeptics point out that Meta's historical record on data stewardship makes users rightfully cautious about granting unrestricted access. From an architectural standpoint, the critical design pattern is fine-grained, per task authority versus blanket access. A well-designed agent platform prompts users to authorize specific actions with clear visibility into what data is accessed and what state changes will occur. When an assistant asks for sweeping access to multiple sensitive services up front, it asks for trust before demonstrating competence. As personal agents become more prevalent, permission modeling and auditable action logs will determine which tools users actually adopt. Trust isn't won by branding an assistant as agentic, it's won by transparent permission dialogues, explicit confirmation steps, and easily revocable credentials that keep users firmly in control of their sensitive personal data. Cognition has completed a landmark financing round valuing the company at $48 billion, cementing its position among the most valuable private artificial intelligence companies in the world. The valuation multiple is notably higher than the valuation cursor commanded prior to its acquisition by SpaceX. The scale of the investment demonstrates that venture capital views autonomous software engineering as a large, diverse ecosystem with room for multiple independent leaders rather than a winner-take-all market. Investors clearly believe that developer workflows are too diverse for a single company to capture every engineer. Software development is far from a monolithic workflow. Some engineers want an interactive pair programming experience directly inside their code editor, while others want fully autonomous background agents that pull issues from a backlog, spin up containerized sandboxes, validate continuous integration suites, and open polished pull requests. Cognition's substantial capital reserve allows the company to invest heavily in autonomous execution infrastructure, proprietary repository indexing engines, and enterprise sandboxes with isolated network controls. That capital backing lets them build specialized infrastructure for deep repository context, automated test validation, and containerized security sandboxes. Seriously, what this means for the broader engineering landscape is continued aggressive competition. When multiple well-funded organizations are competing for developer mindshare, the pace of tooling innovation accelerates. Developers benefit from faster feature iteration, better context retrieval engines, and more competitive tools. Competitive pricing across both interactive and autonomous coding tools. Teams will continue to combine editor assistants, terminal harnesses, and background agents according to their specific project requirements. It's proof that software creation is too broad for one company to capture entirely. One Password published a comprehensive case study with OpenAI detailing a 21% increase in engineering velocity after deploying OpenAI's codecs across its software development organization. The case study explains how the password management company integrated codecs into daily workflows spanning customer-facing product features, core cryptographic modules, and internal infrastructure maintenance. It's a concrete measurement from a company with an exceptionally high engineering bar and strict compliance standards. What makes that 21% figure especially significant is 1Password's demanding security requirements. As a password manager handling sensitive credentials, production code must pass stringent static analysis, peer review, and compliance gates. 1Password emphasized that codecs operate strictly within the company's existing security policies and validation pipelines rather than bypassing established review procedures. The assistant assists engineers with boilerplate generation, test creation, and refactoring, but human engineers remain the authoritative sign-off for all production deployments. This case study provides concrete evidence that autonomous coding assistants can deliver measurable velocity improvements inside organizations where security and code correctness cannot be compromised. The key is integrating the tool into rigorous continuous integration pipelines where every machine-suggested change is verified against existing security baselines. That integration allows teams to move faster without weakening their verification gates or skipping necessary peer reviews. Honestly, for engineering leaders evaluating coding agent adoption, the lesson is that security rigor and developer acceleration don't have to be mutually exclusive. With proper sandboxing and automated validation, coding agents can substantially improve developer throughput even in high-security environments. It's a pragmatic blueprint for enterprise deployment that balances speed with strict compliance, proving that AI assistance works under stringent governance. French artificial intelligence company Mistral announced the completion of a €3 billion series defunding round, bringing its valuation to over €21 billion. The funding news quickly drew over 800 upvotes on Hacker News, reflecting strong developer interest in Mistral's open-weight strategy. While many frontier AI labs operate strictly behind closed APIs, Mistral has built its reputation by releasing downloadable model weights that enterprises and developers can inspect, fine-tune, and run on private infrastructure. The massive capital infusion gives Mistral the compute resources and research capacity needed to train competitive next-generation frontier models while maintaining their open ecosystem commitments. It reinforces their position as a leading global foundation model provider that supports sovereign enterprise infrastructure and independent model governance. In Europe, where data sovereignty and regulatory compliance under the EU AI Act are top priorities, Mistral represents a vital strategic alternative to American closed-source hyperscalers. European enterprises and public institutions are eager to deploy models that don't export proprietary data outside sovereign borders. The continued investor backing demonstrates that the open weights business model is viable at the highest tiers of artificial intelligence research. By pairing downloadable weights with commercial enterprise hosting and fine-tuning services, Mistral proves that providing weights to the developer community doesn't preclude building a massively valuable business. For developers and infrastructure teams, this funding ensures that open weights foundation models will continue to advance alongside proprietary closed models, preserving our ability to build on systems we can self-host, audit, and customize. It keeps the open AI ecosystem fiercely competitive and technically vibrant for years to come. OpenBMB released Mini CPM 5-2b, an ultra-compact open weights language model featuring 2.52 billion dense parameters and a native 131,000 token context window. The model achieved an impressive average score of 53.9 across 34 standardized benchmarks, outperforming larger models such as QUEN 3.5-4b in complex tool use scenarios and agentic coding evaluations. That's remarkable performance for a model of this parameter class, demonstrating how far knowledge distillation has advanced in shrinking frontier capabilities into pocket-sized checkpoints. OpenBMB published the complete model weights, intermediate training checkpoints, and training datasets under the permissive Apache 2.0 license. To achieve frontier-like capabilities within a sub-3 billion parameter footprint, the team utilized 400 billion tokens of deep-thinking supervised fine-tuning combined with on-policy distillation from 16 expert teacher models. Quantized GGUF binaries begin at just 1.56 gigabytes, allowing the model to run comfortably on consumer laptops, edge devices, and smartphones through LAMA.CPP, ALAMA, MLX, and VLLM. It gives developers full freedom to modify and deploy weights anywhere without licensing restrictions. Look, having a highly capable model that runs at high token generation speeds in under 2 gigabytes of memory is a major win for local agent development. It enables offline code completion, personal document summarization, and local tool invocation without sending sensitive data over the network or incurring cloud inference bills. Distilled models like Mini-CPM 5-2b prove that you don't always need a cloud cluster to run capable autonomous agents. It's a genuine milestone for edge computing that puts agentic capabilities directly into lightweight personal workstations. Qualcomm Technologies and Amazon Web Services announced a multi-generation strategic collaboration to co-design customized artificial intelligence inference silicon and deploy ultra-high-speed optical networking reaching 1.6 terabits per second across AWS data centers. As foundation models grow in size and production inference volume surge, data center operators are confronting severe power constraints, thermal limits, and interconnect bottlenecks between accelerator chips. The partnership pairs Qualcomm's energy-efficient neural processing architectures with Amazon's hyperscale cloud infrastructure. The 1.6 terabit optical interconnects replace traditional copper cabling, moving massive tensor streams between server nodes with substantially lower latency and dramatically reduced power dissipation. By co-designing specialized inference accelerators tailored to Amazon's specific production workloads, the two companies aim to significantly reduce the total cost of ownership for serving large language models at scale. This announcement highlights how the artificial intelligence hardware battle is shifting from raw training compute to efficient inference serving. Training happens once, but inference runs continuously for millions of users. Lowering the electrical and thermal footprint of serving reasoning models is becoming the decisive factor in data center economics. It's where cloud providers must innovate to sustain scaling laws and manage electrical grids responsibly. For cloud users and developers, improvements in inference silicon efficiency and optical networking will translate into lower API costs and higher token generation throughput for production applications. That's where infrastructure improvements show up directly in developer budgets. When inference gets cheaper and faster, agentic architectures can afford to run more iterative thinking steps and multi-agent loops without breaking the bank. SciQuantum has finalized a $100 million federal research and development award from the United States Department of Commerce under the Chips and Science Act. The definitive agreement converts an earlier preliminary memorandum into committed federal funding dedicated to domestic semiconductor manufacturing and advanced packaging for utility-scale quantum computing systems. Building fault-tolerant quantum computers requires specialized silicon photonic chips, optical control modules, and precision cryogenic packaging that demand advanced semiconductor fabrication facilities. Anchoring this manufacturing infrastructure within the United States strengthens domestic supply chains for emerging deep tech hardware. It provides capital certainty for advanced fabrication development and ensures that early pilot production can move forward at commercial scale. As fabrication lines scale, the award positions SciQuantum to accelerate the physical manufacturing and testing of optical quantum components. While practical commercial quantum computing remains years away, building out the physical manufacturing and packaging infrastructure is an essential prerequisite. The fact that the CHIPS Act is directly funding commercial quantum fabrication illustrates how governments view quantum hardware as a strategic sovereign priority. Developing silicon photonic manufacturing capabilities also yields valuable secondary benefits for high-speed optical networking and classical computing interconnects, reinforcing the broader semiconductor ecosystem. It ensures that the physical hardware foundation for future computing frontiers is built domestically with resilient high-yield manufacturing processes that support utility-scale systems across national research laboratories and enterprise deployments. That foundation will benefit both classical data centers and future quantum computers as high-bandwidth optical interconnects mature. OpenAI published an artificial intelligence-generated proof for the Navier-Stokes Millennium Prize problem, releasing an explanatory research paper alongside a formal, machine-checked proof written in the Lean Interactive Theorem Prover. The Navier-Stokes Equations Governance Equations Governance Equations Governance The Navier-Stokes Equations Governance Fluid Mechanics and the Millennium Prize Challenge, which carries a $1 million bounty from the Clay Mathematics Institute, questions whether smooth mathematical solutions always exist in three dimensions or whether physical singularities can form. It's one of the most famous open problems in mathematical physics. By encoding the proof in Lean, OpenAI provided a mechanically verifiable mathematical artifact where every logical deduction is verified by automated proof-checking software. However, while Lean confirms that deductions follow logically from their premises, the broader mathematics community must evaluate whether the formalization accurately captures the precise conditions of the Millennium Prize problem statement. Mathematicians on social media and research forums have begun scrutinizing the definitions, boundary conditions, and literature citations. The mathematical nuance lies in ensuring the formulation didn't unintentionally simplify the core challenge or overlook subtle edge cases in three-dimensional fluid dynamics. Regardless of whether this specific proof fully resolves the Millennium Prize, the milestone represents a dramatic advance in automated reasoning. Using formal verification software like Lean provides a level of verifiable correctness that traditional natural language AI outputs cannot match. Formal interactive theorem proving is rapidly evolving from a niche academic discipline into a powerful tool for verified mathematical reasoning and software verification. It's an exciting bridge between machine learning and rigorous mathematical truth that will change how mathematicians work on frontier research problems. Document parsing startup Reducto launched R1, an enter-end multimodal model that parses an entire document page in a single forward parse for exactly one cent. The unified model replaces traditional multistage pipelines that chain distinct optical character recognition, layout detection, table extraction, and text formatting models together. Reducto reports a 20% reduction in document parsing errors alongside a significant cost reduction from the previous three to six cent per page range. It's a massive efficiency improvement for document-heavy enterprises. In traditional document ingestion pipelines, errors from sequential stages compound quickly. When an early layout detection model misinterprets table boundaries or an OCR engine misreads low-resolution scans, downstream models receive corrupted inputs and produce flawed output. By ingesting the entire page image in a single pass, R1 generates structured text, table schemas, formatting markup, and bounding box coordinates simultaneously. That single forward pass keeps visual context intact and prevents downstream parsing cascades. For organizations processing millions of financial filings, legal contracts, and medical records, predictable one-cent pricing drastically changes the economics of document automation. Reducing parsing errors by 20% while cutting processing costs in half allows enterprise engineering teams to ingest unstructured documents at scale without maintaining fragile multimodal pipelines. It turns what was an expensive operational chore into a reliable commodity utility. This single pass multimodal approach shows how unified vision language architectures are systematically replacing brittle multi-stage heuristic pipelines across production engineering workloads. It's a clean, cost-effective upgrade for document processing that proves multimodal models can simplify complex enterprise architectures while drastically lowering operating expenditures for data engineering teams across the industry. In today's GitHub project radar, codebase memory MCP from Deus data leads with 42,700 stars and 12.6% monthly growth. The project compiles a persistent code knowledge graph across 158 programming languages and serves submilisacant queries over the model context protocol, allowing agents to inspect call graphs directly. It's an impressive jump in efficiency for coding agents navigating large repositories. CUDS Nanobot follows with nearly 48,000 stars, offering an ultra-lightweight Python framework for self-hosting personal agents equipped with a web user interface, memory, and tool registries. And ARM Blender MCP sits at 27,800 stars, exposing Blender's 3D scene controls and Python commands to connected language models for automated 3D mesh modeling and scene manipulation. All three projects provide tangible tool surfaces that MCP-compatible agents like Antigravity, Codex, ClaudeCode, and Hermes can call immediately to expand their capabilities across codebase intelligence, self-hosted workflows, and spatial creation. They represent practical additions to any active agent stack that wants to move beyond standard file system editing into structured project modeling and interactive tooling. It gives autonomous developers immediate functional superpowers. In model discovery, Inception's Mercury 2.5 is the featured arrival on OpenRouter, packing a 260,000-token context window and parallel diffusion decoding for rapid reasoning throughput. The architecture aims to cut waiting time during complex multi-step reasoning by refining token drafts in parallel, making it a compelling candidate for side-by-side evaluation against sequential models on production workloads. It's a notable addition to the hosted catalog that developers should benchmark on active agent tasks. For our local LLM spotlight, Xtoken SparkX 2.5-4b is drawing noticeable community attention on Hugging Face, passing 10,000 downloads and nearly 1,000 likes. Built on standard causal language model architectures, the 4 billion parameter model provides an approachable footprint for developers running local conversational and lightweight tool-use experiments on consumer workstations without requiring high-end GPU clusters. The checkpoint runs efficiently across standard local serving engines, giving builders a capable offline foundation model for prototyping interactive desktop agents and privacy-sensitive tool integrations right on their personal hardware without network latency or cloud subscriptions. It's an accessible option for local experimenting that delivers dependable performance on everyday workstation setups. In extra research candidates, three papers caught our attention today. First, IFM releases K2 Horizon, introducing six open weights models ranging from 900 million to 375 billion parameters under Apache 2.0, complete with pre-training dataset releases. Second, OpenAI published the work Now Within Reach, analyzing how declining capability per dollar economics expand the frontier of automatable enterprise workflows. And third, TechCrunch covered the mathematics community debate around priority and verification in OpenAI fought dirty on career-making math problem following the Navier-Stokeslin release. Three distinct reads on where research frontiers are moving across models, economics, and verified reasoning. For complete source notes, technical links, and documentation, visit tobi on fitnesstech.com. We've compiled all the primary references and model checkpoints there for your team to explore. Thanks for listening, and we'll be back soon.