eigencode

we haven't forgotten about you guys (if anyone is still left lmao) real token utility as long as the sun rises every day $CODE will never die

Code CLI v1.2.0: - run your shell tasks in background by yourself or using agents - give your agents more than context using datasets - run pre-built context-less Actions as Agents (AaA) - new auto-compression algorithm to keep your agents running forever without needing a break for compaction - new public Code CLI plugin registry - 16 new commands and stability fixes github.com/Polysystems/co…

Code CLI v1.1.0 - office integration for context share, knowledge graph and integration - new context engine, achieving x37.2 - x515.51 token saving on project context management with sub-100ms runs - poly agent-first for better token and agentic quality - internal usage tracking per provider - upgraded mainframe system with new options and much powerful tools - deeper rust integration, increasing agentic development speeds - new agent-to-agent communication system, saving ~93% of tokens than usual, mainly for multi-agent swarms - 23 new commands - fixed vulnerabilities and bugs

codecli v1.0.0 is now available in the video you'll see me demo the computer use feature asking it to test and debug and introduce new visual features all at the same time to a game i'm working on try out computer use simply by just asking in the chat for it to use it for something release notes: CodeCLI v1.0.0: The Agent Development Platform CodeCLI v1.0.0 pushes the beginning of our codecli transformation from a command-line coding companion into a full-stack agent development CLI Performance & Core Architecture CodeCLI now ships with a native Rust performance layer that replaces JavaScript implementations on critical path enabling: - Exact token counting - Line-level diff engine - Real-time security scanning - Prompt injection detection - Space-optimized algorithms - Wire-compatible RPC protocol Multi-Agent Orchestration System The platform introduces a completely rebuilt sub-agent system with mandatory parallel delegation and structured workflows: - 5-phase planning workflow (Explore → Design → Review → Finalize → Exit) with full Kanban integration - Capability-based routing across 23 typed capabilities with historical success tracking - Mandatory parallel delegation requiring 3+ independent items to run simultaneously - Multi-agent orchestration with 5-phase protocol and coordinator-worker architecture - Domain expert agents (physics, math, Rust, security, architecture) with automatic delegation - Council system where 3–7 agents debate and vote on critical decisions Plugin Ecosystem & Automation A complete plugin system and new automation tools extend agent capabilities: - Plugin marketplace compatible with anthropics/claude-plugins-official format - Two installation scopes (global and project-specific) with custom marketplace support - Hook system with 20 event types for pre/post-execution commands - Desktop GUI automation via computer_use tool with cross-platform support (Linux X11/Wayland, macOS) - Headless browser automation with 12 actions for testing and web interaction - Sandboxed Python environment with 30+ safe modules for numerical computation Enterprise Security & Reliability Version 1.0.0 delivers production-grade security: - Directory sandboxing - Permission hierarchy with glob pattern matching and automatic classification - Critical safeguards for XSS vulnerabilities, prototype pollution, and command injection vectors - SSRF protection blocking private IPs and cloud metadata endpoints - SMB credential leaks prevention for Windows - JSON corruption prevention - Doom loop detection with primary and secondary safeguards Context Management & Token Economy Intelligent context management reduces token consumption while maintaining session quality: - 61% system prompt compression across all providers with minimized tool descriptions - Micro-compaction, Reactive-compaction, and automatic context persistence - Vectored project context management - Cost tracking optimization Developer Experience & Tooling The TUI receives a complete redesign with new themes and enhanced workflows: - 3 new UI themes (Cherry, Crimson, Solar) joining existing Poly, Matrix, and Eigen - HUD aesthetic with scanline animations and enhanced status indicators - Command palette overhaul with grouped categories, persistent history, and favorite pinning - 50+ built-in tools including new clipboard operations, network diagnostics, and Jupyter notebook editing - Desktop notifications with native OS integration across macOS, Linux, and Windows - LaTeX rendering with Unicode math support and optional matplotlib bridge CodeCLI v1.0.0 represents not just an update, but a fundamental rearchitecture—transforming from a coding assistant into an integrated platform tool where teams of specialized agents collaborate, where security is baked into every layer, and where enterprise-scale development becomes accessible from the terminal. The future of AI-assisted development isn't a single model answering questions; it's an orchestrated ecosystem of specialized agents working in concert, and that future is now available with a single installation command.

omm 0.3.0 is now live - OMM v0.3.0: Native GPU Inference & Dual-Engine Architecture We are announcing the release of OMM v0.3.0, a major update that fundamentally transforms local AI inference by introducing native GPU execution and a new high-performance architecture. This release delivers dramatic speed improvements, expanded hardware support, and greater operational control, all while maintaining full backward compatibility with existing workflows. Key Features - Native GPU Inference * Models now load directly onto your GPU, eliminating the need for external service proxies. * Automatic VRAM management and model eviction ensure efficient resource utilization. * Real-time thinking token streaming provides immediate feedback during generation. - Dual-Engine Architecture * llama.cpp: The reliable default engine for broad hardware compatibility. * rvllm: A new, high-performance CUDA engine optimized for NVIDIA GPUs, featuring continuous batching, speculative decoding, and tensor parallelism across multiple GPUs. * Engine selection is automatic (NVIDIA GPUs default to rvllm) or manually specified via the `--engine` flag. - Enhanced Model & Tooling Ecosystem * HuggingFace Hub Integration: Directly download SafeTensors models with `omm pull --hf `. * Built-in Benchmarking: Measure throughput and latency with the new `omm benchmark` subcommand. * SIMD-Optimized Sampling: Token sampling is now accelerated with AVX2 (x86) and NEON (ARM) instructions. Expanded Control & Monitoring - New CLI & Configuration Options -- Gain precise control over inference with new `omm serve` flags: * `--engine `: Select the inference engine. * `--tp `: Enable tensor parallelism to split models across multiple GPUs. * `--kv-cache-dtype`: Optimize memory usage by setting the KV cache data type. * Persistent settings can be defined in the new `[engine]` section of `~/.omm/config.toml`. - Comprehensive Hardware Support -- OMM v0.3.0 provides optimized support across a wide range of hardware: * NVIDIA: Full support from Volta (Compute 7.0) through Blackwell (Compute 10.0). * Apple Silicon: Native execution on M1–M4 via Metal. * AMD: Support for MI and Radeon series via ROCm. * Vulkan: A cross-platform fallback for any GPU supporting Vulkan 1.2+. Upgrade to v0.3.0 Existing installations upgrade seamlessly with no breaking changes to compatible APIs (Ollama, OpenAI, Anthropic).

new build - improved IDE - improved agents, parallelization and swarms - built-in tools and complex systems to enhance coding flow - new commands and components for increased productivity - create web/mobile/desktop apps and platforms using any framework it's gonna be paywalled for premium users for now due to all the features and the costs for running so many agents in the web and so on

$CODE

fixed all the bugs in research and made it work way more integrated and seamless now refactored plans so now plans contains several embedded research jobs within it, basically you can run several jobs within one plan for more comprehensive research designer got more advanced features and you can access any of the previous verticals from within it fixed the feed so that it automatically sends any of the articles off into a research job if you press research this cooked the knowledge section so that you can easily search through all your past research data point and transform them into knowledge trees for comprehensive research reports and stuff everything should work pretty smoothly now

cooked research up it's way more balanced and comprehensive now the standard research is now a quick research where you're more interacting with the research agent in an iterative process and get quick results and don't have to wait and you can do easy follow up questions, visualize results into infographics, and leverage the quick research to send off a deep research job if you want, etc etc i also fixed small bugs here and there, fixed the rate limits all the verticals and standard research is now consolidated into the research designer you can see if any research runs failed or only partially worked i play around and experiment with all the different features here even get it to do some novel research and theory crafting on some theories i shoot at it for fun like "cognition, brain wave oscillations, coherence, resonance, entanglement - recursive attractors are resonance sequencers for orchestrating consciousness at large scales" and it comes with a bunch of potential stuff to build on it, talking about neural oscillations and synchronization and resonance and neural systems and so on blah blah blah it's getting pretty close

overhauled research and basically made it a comprehensive research hub that has a bunch of stuff now created 11 intelligence engines that run on top of the research to increase the quality: - cross reference - entity-extractor - citation-network - source-verification - trend-analyzer - contradiction-detector - consensus-mapper - evidence-grader - temporal-causality - sentiment-stance - comparative-scorer improved the standard research page with a bunch with more rich features, the standard research has methodology now along with additional advanced configurations, and for the verticals you can now also do comparative research flows connecting verticals in one research flow there's also more comprehensive research designs now, you can use the plans tap to use preconfigured research templates with multiple steps, or you can create your own plan and configure it with your own steps as you like and then for a more advanced research mode there's the designer tab where you can configure things like reasoning, token budget, sources, which intelligence engines it should use, source preferences, domains it should prioritize, domains it should exclude, any vertical routes you want incorporated, etc then there's also a knowledge index page for connecting saved completed research jobs and transforming them into a structured knowledge system where you can chain jobs into higher research designs and what not for academic and corporate work and so on there's also a dashboard page now i've made it so that you can run several research jobs at the same time, you don't have to sit and stare at it while it's loading any more, you can just set up the research flow and then you'll see the job running in the dashboard, and you can run several research jobs at the same time you can also see all your previous completed research jobs and go back and review them also on the dashboard page there's another new feature, and this might be my favorite one it's basically a custom intelligent live feed, for aggregating articles, news, github repos, X posts, market updates, papers, etc basically anything NO MORE DOOM SCROLLING you can now curate your own high quality feed that aggregates sources from all around the internet so you only get high quality information streams without any of the slop you can customize it so you can choose exactly what you want it to aggregate and what sources you want it to pull from, you can even combine it so it pulls specific twitter users with arxiv papers with ai lab releases with github repos for example (!) you also can select articles or sources from your feed aggregator and then run research on them or incorporate them into your research flows (!!) and i did all this in under 24 hours (!!!) sometimes i surprise myself and tbh research still isn't done, i have more ideas i want to integrate, and i need to work on the rate limits for running multiple concurrent research jobs and improve the knowledge indexing and some other stuff i have a vision for it so i'm gonna keep cooking it until it gets there

new github scanner and market research tools

been expanding and overhauling our research feature to make it more powerful and incorporate various different research styles - Academic Research: with literature reviews, research gap analysis, and methodology comparison - Competitive Intelligence: with market positioning, competitor analysis, and strategic planning - Product Discovery: find and compare products, services, and tools - Hypothesis Testing: explore evidence for and against a claim - Market Research: with market sizing, trend analysis, and opportunity mapping - Investigative Deep Dive: with fact-checking, timeline reconstruction, and entity mapping - Github Scanner: search and analyze open source repositories in the video here i demo the hypothesis testing tool i'm still not fully satisfied with where it's at tbh, so further updates coming soon

first you expand then you contract folded all the psyche stuff into a more clean integrated interface that's like a user journal

implemented psychological testing and comprehensive talk therapy into psyche the therapists each emulate a famous psychologist and have access a large amount of exercises and tools they can suggest to the user they also have contextual understanding of the user within the poly platform - being able to see the user's documents and similar to provide suggestions for example in the second screenshot here i'm testing carl rogers, where the therapist picks up on my patterns, suggests looking at some of my work on office to cheer me up, and wants to talk deeper about some of my work to help encourage a mental shift this could really help a lot of people tbh poly will save the world

new updates for studio i use and experiment with a lot of creative tools myself, and feel like with this update we've finally breached the threshold for reaching parity with or exceeding many of the top creative tools on the web like adobe express or similar canvas has had several improvements and been debugged, it actually feels a lot more like an image editing canvas now and has become a lot more comfortable to use the video editor has had a complete rewrite and overhaul, but now we have a quality video editor available in the web, a lot of people thought you can't get a good video editor to work in a web app, but we're basically there the video editor has most standard video editing features you might find in something like premiere pro, along baked in ai tools like text to video, image to video, caption generation, and so on all those things along with our built in transition library makes it seamless to use the video editor to create high quality schizo edits or whatever you want the audio editor has been removed and just consolidated into the video editor the animation studio has been removed completely (we're thinking of baking it back in later on in a smarter way) the batch editor has been debugged the workflow editor has gotten some additional templates added for automating creative pipelines overall i'm generally pretty happy with how far the studio has come in such a short amount of time and the potential it still has next up in studio we'll be working on creating a custom film studio for one-shotting very high quality commercials and short-films with brand persistence across generations and a bunch of other stuff, but i won't go into too much detail about that for now

working on improving all the games, some of them aren't challenging enough and so on - already overhauled cognitive switch and memory matrix will have them all overhauled by tomorrow along with other updates basically the plan for the psyche section of chat is to integrate a comprehensive suite of psychological tools you can see some of them here from the last rewrite i was working on before this integration basically still need - chat therapy, CBT, and other therapy styles - testing - more games - journaling tools and then it'll all get integrated with the comprehensive behavioral scoring system that gives the user a personal map of their psyche and the games will also be overhauled and split into 2 categories basically all the games that are available right now are gonna be categorized as minigames instead and then there's gonna be larger game experiences for example i demo a bit of a game i'm working on here called fathom fathom is basically gonna be a whole narrative style rpg - where the user goes through a long game interacting with different characters and archetypes and their choices continuously effect the game and narrative options creating multiple different consequence chains and various endings and so on and the whole game is it's own kind of psychological reading, gauging and mapping the user's psyche throughout depending on the choices they make and there's gonna be other games like that basically and once i nail the behavioral scoring integration with all the games and the user profile then i can also use it for future non-psychological games and turn them into psychological experiences like imagine integrating the behavioral scoring system in the pokemon game i made with codecli the other day, you could have kids and people play these games and gauge their psyche without them even knowing it, giving far more accurate results because they're being observed in natural behavioral responses instead of knowing that they're being observed and performing for a desired result lots of potential for the future here lots of potential for the future everywhere in poly tbh still so early still getting warmed up still just the beginning etc

new updates to psyche: - improved fusion reading - added games section for psychological & cognitive games, with 11 initial games on a more personal note, i've actually been working on the idea for psyche as a side project over the course of many years, and still have a lot of stuff to integrate as i've essentially built out a comprehensive behavioral scoring system that covers all aspects of the human psyche by splitting it up in over a thousand attributes that relate to hundreds of sub-domains which are clustered under a manageable number of primary domains it took me a long time to even think of that system, and the first time i wrote it down was as powerpoint slides in my free time back in 2022, now i've vibe coded and rewritten the app for it over 5 times, and now i'm at my final rewrite where it's not its own app anymore but going to be an integrated part of poly - the scoring system still hasn't been implemented yet, but it'll come after integrate the rest of the games and tools and stuff and our new platform user section is completed anyway one of the things i want to do with psyche once poly is built out - is to offer it to danish social services to use in school programs, prisons, and other social programs in general to help disenfranchised youth and mentally fragmented people understand their mental patterns and heal themselves so that they can integrate back into society and become productive and happy from my own observations as a teenager and a young man hanging out with criminals what i learned is that it is always due to some kind of loaded past i don't believe anyone in this world is truly a criminal out of inherent malevolence it is always due to some kind of childhood trauma or unfortunate circumstance or mental illness that they become a criminal, it's always born out of a frustration and anger towards society which is a projection of this trauma or circumstance which plagues them - i.e. the boy who felt betrayed or abandoned by society will in turn reflect that by eventually choosing to betray and abandon society himself but nobody truly wants to be a criminal, it's idealized in rap music and hbo drama series, but the reality is that 'the gangster lifestyle' is one full of misery - any true criminal knows this people generally want to find collective acceptance and meaning, the reason why criminals struggle to find it is because they have a lot of underlying mental and emotional baggage in the form of trauma, neglect, abandonment, and similar - thus they cope with this baggage by reflecting these negative feelings back at society and being a criminal i have known many gangsters throughout my life, and the only people who really hype up the 'gang gang' lifestyle are the hang around types, the plastic gangsters, the pretenders who aren't really gangster but just listen to rap music and do drugs and hang out with some gang affiliates to feel cool and tell people about it the 'real gangsters' know that it is a life of trauma and misery - i have known gang leaders, drillers, robbers, runners, pushers, shooters, etc they all have the same misery deep inside - the violence, drug abuse, rejection of society, etc - it's all a projection of this misery in various forms hopefully one day psyche can help with that - by helping people genuinely understand themselves and where their misfortune arises from, and giving them deeper self-awareness so that they can heal their mental-emotional afflictions and actualize their true self

new codecli updates! CodeCLI v0.10.0: - Superpowers to drastically increase agents powers in specific skills - Councils for debating and voting to increase accuracy in multi-agent decision making - Stronger Multi drastically increases development speed by better leveraging swarms (1.3x-1.7x increase) - Built-in memory, better indexing, web search and package verification - Better UI/UX and recognize your terminals using tags when multi-tasking - Upgraded Mainframe and RC, serve Code CLI as MCP and integrated Hooks more deeply - Serve Code CLI and integrate into your apps - Raw Mode for cutting distractions or allowing other agents to use Code without deeper integration - Better token usage and consumption economy (11-15% less token usage)

new codecli updates we now have all our tools baked into codecli itself (cargo install our polymcp tool package for full integration) we also have a bunch of new commands, updated memory, session persistence, preconfigured skills and plugins to add, and a lot more in the demo here i use codecli to almost oneshot a pokemon-like game - i use two inputs - first i give codecli the initial instructions - then i ask it to audit it's implementation for any gaps for the demo i'm using claude code inside codecli why claude code? - Opus 4.6 still remains the best coding model available, meaning in terms of quality this is the best model to use - Anthropic subsidize a lot of costs via the claude 20x max subscription, meaning in terms of price efficiency this remains the best option available on the market - thus quality/costs running opus through claude code is the best option available as it outcompetes everything else why use claude code inside codecli rather than in claude code? well we've optimized codecli as a more effective coding cli harness, meaning: - token optimization, you get more out of your tokens in codecli than in claude code, thus you'll get more value for your money - speed, codecli is a lot faster than claude code - tool availability, we have over 81 tools baked directly into codecli, making claude code even more powerful - memory and session persistence, codecli gives claude code greater memory than what it has in its own cli - swarms, multi, and sub-agents, our multi-agent system isn't just sub-agents, we spawn several opus agents working parallel together offering a far more powerful multi-agent experience - loops, we don't just have an 'auto system' we have a comprehensive loop system that allows us to effectively one shot very high quality work with minimal user input all in all if i have to do the same task using claude code in codecli and claude code cli, then it will be done 5x faster in codecli and it will be far more accurate due to our increased tools and more powerful memory system

RT @DionysianAgent: go vote for poly right now or you're gay and your entire family is getting replaced by robots bags.fm/hackathon/apps…

we've also made major updates to studio honestly there's so much in studio i struggle to even list it all out but the video gives you a little taste - major canvas overhaul - added custom imagine page with: -- product showcase for easily curating high quality product images for product branding -- ASCII Art studio for converting images and videos into ascii -- Image to 3D -- QR Code Art -- Pixel Art Studio for creating game assets -- Comic and Storyboard generator -- Cinemagraph for animating images into moving images - batch editor for editing a large amount of files at the same time - workflow node canvas for creating complex ai workflows across an extensive amount of ai functions the workflow node canvas is probably one of our highest quality features right now and you can basically one shot high quality ads, videos, etc just by stitching together nodes of different functions and running it and even though studio is already at such a feature rich level it's still nowhere close to where it will be in a short amount of time our vision is to enable people and companies to be able to one shot high quality commercials, short-films, comics, and a variety of content types with studio essentially transforming the entire meaning of creative work and how fast you can create and iterate