Pulse Documentation
Pulse - ADE (Agentic Development Environment): an environment in which code is written by AI agents, and the developer sets tasks, monitors the work and makes decisions.
New workflows: collaboration, project measurements, results of the functional audit, safety of input and reliability of states.
Where to start
| Document | What about |
|---|---|
| Installation and update | DMG, first launch, macOS requests, updates, uninstallation |
| Quick start | First 15 minutes: agent → task → graph → diff → commit |
| Tasks and worktree | Board, prompt, resources, skills, sessions, diff, terminal, archive |
Instructions
| Document | What about |
|---|---|
| Agents and models | Claude Code, Codex, Kimi, OpenCode; local model and GigaChat; agent options, instruction versions, context window |
| Agent swarm | Composition of roles, launching a repository, task protocol, worktree of participants, merging, timeouts |
| Control: permissions, policies, budgets, sandbox, audit | What an agent can do, how much he can spend and how to check it |
| External MCP servers | Connecting tools via Model Context Protocol, code memory |
| Library of skills and roles | Catalog, installation of skills, comparison, creation of agent options |
| UI Catalog | Ready-made interface blocks: previews and rules for humans, tools ui_catalog and ui_template for agent |
| Quality inspectors | Who checks the work of agents, six evaluation axes, auto-check, comparison table |
| Limits and session transfer | Limit chip in chat, which eats up the limit, continuation of work on another model with a report |
| Platforms | Builds for macOS, Windows and Ubuntu: what works everywhere, what only on macOS, what with the phone |
| Assembly protection | What is closed in the assembly (bytecode, minification, fuses), which Electron does not have, signature order |
| Access and licenses | Website pulse.avacloude.ru: registration by access code, issuance and revocation of licenses, personal account |
| Mobile remote control | Application plan for Android and iOS: remote control for the device, connection via QR, device registry, deadlines |
| Memory and context | Repository memory, task notes, history compression under the model window |
| Safety and authenticity | Session drafts, task stage in the graph, cost and rewind, unpublishing without a network, protocol for busy working copies |
| Recovery and work context | Register of requests and receipts, reasons for stopping, "Needs attention", saving space after restart, startup identity and "loop/restart" distinction |
| Full review cycle | Task contract with requirements version, comments on the version and file, package to the author, contract checks and inspector assessment, acceptance |
| Your code and app in one window | “Files” tab: working copy of a task or session, editor with file version control, application preview via a local proxy, selection of an element for the session draft |
| Environment and Schedule Commands | A single list of commands for the MCP server with the circuit version, receipts and log; schedules with pre-check, time zone, occupancy and absence rules |
| Keyboard shortcuts | All combinations by section |
| Pulse Terminal | Warp level terminal inside ADE: what is already there, vocabulary, entry points, open questions; series plan - terminal/plan.md |
| If something went wrong | Keychain, window did not open, GigaChat, local model, log and report to developers |
Directories
| Document | What about |
|---|---|
| Development | Stack, commands, tests, DMG build, release publication, code structure |
| Responsiveness to big data | Measurements on five window surfaces, corrected squares, what was left unrepaired and how it is customary to measure the growth form here |
| Agents from within | Agent registry, built-in API agent, swarm, memory, policies, MCP - how it works in code |
| Interface architecture | Target architecture, phases M0–M8, what has been done |
| UX interface audit (09.23.2026, 0.30.38) | Full run of the UI: application, remote control, cloud, website. Tokens, components, statuses, board, permissions, graph, keys; 67 pictures; finds from path:string; replay scripts |
| Coating with design system | Manual matrix covering screens and installers |
| Audit corrections 09/20/2026 | Tests for scenarios F01–F10 and associated fixes |
| UI audit fixes 09/20/2026 | Checks 14 groups of interface fixes |
| Module map | Origin of the generated map and its verification |
| GigaChat | Built-in agent design on top of GigaChat API |
| Harness plan | Scheme “Agent = LLM + Harness”, phases H1–H6 |
| CHANGELOG | Versions of what has changed, “Where we left off” |
Terms
- Problem - prompt + repository + your own git-worktree on the branch
pulse/<id>. Unit of work. - Session — one agent process inside a task. A task may have several of them.
- Turn — one request to the model and everything it did in response: text, tool calls, permissions.
- Agent - a program that works with code (Claude Code, Codex, Kimi Code, OpenCode) or a built-in Pulse agent on top of the model.
- Model — what the agent thinks on: local server (Ollama, LM Studio, LiteLLM, llama.cpp, vLLM) or GigaChat.
- Inspector - a role that does not write code, but evaluates someone else’s work: a strong model, evaluation along six axes, findings and corrections to the author.
- Swarm — a named set of roles (architect, analyst, front-end, back-end, security, product, tester, DevOps), which is launched on the repository with one button.
- Worktree — a working copy of the repository on a separate branch; Each has its own task.