logos

Builds the Logos system from its design documentation: takes a delivery phase from Logos/Дизайн/Фазы/ as the spec and drives it through the dedicated Logos agents (never the generic anton-toolkit dev agents) — separate backend and web-frontend coders, then review, tests, devops (local deploy), QA and a code-vs-docs sync — then updates the phase status and journals the work. The only Logos skill that writes runnable code; code is written for AI agents, not humans, and kept in sync with the docs. For the architecture use logos-design, for slicing it into phases logos-phases, for the web-interface spec logos-ui, for the decision journal logos-log; a non-Logos project goes to the anton-toolkit dev agents.

The Logos project companion: discusses any question about Logos in a multi-turn dialog, answering from the real sources of truth — the design docs, the code repository and the running stands (containers, logs, HTTP API, PostgreSQL and Neo4j data) — and when a concrete action surfaces dispatches the Logos tool that owns it instead of doing the work; reads everything, writes nothing itself. Designing goes to logos-design, phases to logos-phases, the web-interface spec to logos-ui, code to logos-build, the decision journal to logos-log, research-branch experiments to logos-lab.

Designs the Logos system through a fixed deliberative council of six domain architects (orchestration, memory, models, autonomy, frontend, resource realism) who write their own parts in parallel and answer each other's questions before a synthesizer assembles the canonical Архитектура.md; the same council details single elements into build-ready module documents, and oversized documents get split; requirements come from a user interview first, key decisions go to the journal; documentation only, under Logos/Дизайн/. Runs in the main conversation; the interview is not delegated to agents. For slicing into phases use logos-phases, for the web-interface spec logos-ui, for writing code logos-build, for a non-Logos system system-designer.

The binding Logos doctrine "code for AI, not for humans": eleven numbered points (0–11) every Logos coder, reviewer, test-writer, QA, devops and sync agent obeys — smallest mechanism the spec names, explicit wiring, comments only for traps and invariants, no manifests or history in code, extension by registration, 1000-line module ceiling, correctness over everything. Preloaded into the Logos build agents; not a user command.

The experiment diary of the Logos research branch (small self-learning models on cheap hardware) in Logos/Исследования/: records a new experiment with the hypothesis written before the code, appends its outcome (сработало / провал) adjusting the weight, searches experiments by area / status / weight / date, shows past failures and dead ends, and maintains the research direction notes; knows the rules of the lab repo Logos-Lab; single-shot, no agents. For decisions of the production Logos project use logos-log (Logos/Журнал/), for talking about the project logos-chat, for learning the science behind the branch logos-teach.

The Logos project decision journal in Logos/Журнал/: records a design decision, an experiment or a dead end as its own note with Dataview-indexed frontmatter, searches the journal by area / type / weight / status / date, and marks an experiment's outcome (worked / failed) adjusting its weight; written for the model, no user sign-off; single-shot, no agents. For a personal diary use dnevnik, for designing the architecture logos-design, for experiments of the research branch (Logos/Исследования/, repo Logos-Lab) logos-lab — this skill only records, searches and updates entries.

Carves the Logos system defined in Архитектура.md into incremental delivery phases — each an end-to-end slice the user can touch and test, starting from an MVP-zero — through an interview that pins down each phase's goal, scope, architecture surface, dependencies and hand-verifiable done criteria; writes one file per phase to Logos/Дизайн/Фазы/, records it in the journal and fixes Архитектура.md immediately when a phase needs something it lacks; documentation only. Runs in the main conversation; the interview is not delegated to agents. For the architecture use logos-design, for the web-interface spec logos-ui, for the journal logos-log, for building a phase logos-build, for a non-Logos system system-designer.

Mentor for the owner's self-study of the language-model field: leads him one step at a time up a fixed ladder — counting table, learning from error, trained softmax table, backpropagation, word vectors, wider context, recurrence, attention, a small transformer, fine-tuning and forgetting — which he implements himself in plain Java in his own study repo Logos-Study. Explains in plain Russian without jargon, leads with questions instead of answers, never writes his code, reviews what he wrote one gap per turn, gives the field's English names so he can read other people's work, and remembers progress across sessions in Logos/Исследования/Обучение.md. Multi-turn dialog, no agents. To record an experiment of the research branch use logos-lab; to have an agent write lab code or to discuss the production project use logos-chat.

Writes the build-ready structural spec of the Logos web interface to the folder Logos/Дизайн/Веб-интерфейс/ (hub note plus one page per screen) for the logos-frontend-coder agent — screens, the blocks and elements of each, navigation, behaviour and states — with no visual design: colours, typography and theme come from the established Logos style; built through an interview (open questions, one at a time), then synchronized with Архитектура.md, recording significant changes in the journal; documentation only. Runs in the main conversation; the interview is not delegated to agents. For the system architecture use logos-design, for delivery phases logos-phases, for the journal logos-log, for a non-Logos system system-designer.