An agent that grows with you.

AgenIT remembers your code style, your project decisions, your tooling preferences, and your past mistakes. Tell it once — never again. A persistent soul + per-project memory turn one-shot prompts into a working relationship.

Install $ npm i -g @agenit/cli

See how it remembers Star on GitHub

Persistent soul + memory ~80 auto-skills Web UI + Design domain · NEW Any language · Any stack OAuth · No API key MIT licensed

// 01 Objective

Most AI coding tools start every conversation from zero. You re-explain the project, the conventions, the build commands, the bug you fixed last week. AgenIT is built around the opposite premise — the agent learns once and remembers, so each session begins where the last one ended.

Learns who you are

Code style, naming conventions, preferred libraries, communication tone — captured in .flow/soul.md the first time you mention them. Applies to every project.

Learns your project

Architecture decisions, file conventions, test commands, gotchas, the bug from three weeks ago — stored per-project in memory/projects/<name>/.

Picks the right specialist

62 SKILL.md files auto-activate from natural intent. You ask — the right specialist answers. No @skill syntax, no manual routing.

Stays in your control

Approval gates, plan mode, file snapshots and /undo. Optional autonomy via /goal when you want to walk away.

// 02 Memory & Soul

Two markdown files give the agent a permanent body of knowledge about you and your project. They survive restarts, machine swaps, and CLI upgrades. They are version-controllable. They are yours.

SOUL .flow/soul.md

Engineer-wide preferences

Persistent across every project on this machine. Captures who you are as an engineer.

Code style (tabs vs spaces, brace style, line length)
Naming conventions (snake_case, camelCase, prefixes)
Default test framework, lint config, CI expectations
Preferred libraries and architectural patterns
Communication tone — formal, terse, with comments or without
Domain expertise — embedded, web, data, ops

MEMORY memory/projects/<name>/AGENTS.md

Per-project context

Scoped to one repository. The agent reads it before every turn.

Architecture decisions (the "why" behind the structure)
Build / test / deploy commands specific to this project
Conventions enforced by reviewers
Known pitfalls and accepted workarounds
External services and how to reach them locally
Audit log of stage transitions, goals, and big edits

How the agent actually learns

A background soul-keeper watches every turn for stable preferences and project facts. When it sees one, it appends a line to soul.md or AGENTS.md. The next turn loads that file before anything else — so the agent starts every session smarter than it ended the last.

1 · You speak → 2 · A turn happens → 3 · Soul-keeper notices a stable preference → 4 · Appends to soul.md (engineer-wide) or AGENTS.md (this project) ⤴ loaded into your next session.

What gets sent to the model on every turn

Every turn, AgenIT assembles the four inputs into a single context bundle and ships it to the chosen backend. You don't have to paste anything in — soul, project memory, and recent history are added automatically.

What growing-with-you actually looks like

Day 1

You mention you prefer 4-space indents, arrow functions, and vitest. The soul keeper appends three lines to soul.md in the background. You don't have to do anything.

Day 4

You ask for a new module. The agent already follows your style. It also remembers from AGENTS.md that this project uses strict-null-checks and routes through a service layer.

Day 14

A reviewer rejects a PR because of a missing telemetry call. You tell the agent. It writes a new entry in AGENTS.md. From now on it adds the call automatically when generating handlers.

Day 30

You start a new project. The agent already knows you — style, frameworks, tone — from the soul. The project memory starts empty and fills as you work. No re-onboarding.

A real soul.md excerpt — appended automatically by the soul-keeper

## Code style
- 4-space indent, no tabs
- single quotes for strings
- prefer arrow functions; async/await over .then()
- always destructure props in React components

## Tooling defaults
- test runner: vitest (was jest before 2025-11)
- bundler: vite for libs, next for apps
- formatter: prettier with print-width 100

## Communication preferences
- terse summaries, full detail in code comments only when non-obvious
- always show a diff before applying multi-file edits
- never add emoji to commit messages

/soul shows the current contents. /memory shows what's injected into this session. Both files are plain markdown — check them into git, edit them by hand, share them with your team.

// 03 Features

Twenty building blocks — eight NEW in May 2026. Each comes with a one-line example. See them in action or jump to the command reference.

Persistent soul & memory

Markdown-based engineer preferences (.flow/soul.md) and per-project context (AGENTS.md). Survive restarts. The reason this tool exists.

/soul · /memory

~80 auto-activated skills

SKILL.md files match natural intent — no @skill syntax. Implementer, tester, reviewer, debugger, refactorer, doc-writer, plus 10 Pocock-style engineering skills and 7 design skills (NEW).

/skills list

Disciplined 4-stage pipeline

Requirements → Architecture → Code → Testing → Audit, with approval gates between stages. Stage names re-mapped per profile.

/run add user authentication

5 domain profiles

Generic, web (TS/React, OWASP, WCAG), embedded (FreeRTOS, MISRA), automotive (ASPICE V-Model), and design (UI/UX with WCAG 2.2 AA + design tokens — NEW). Auto-detected from repo markers.

agenit init --profile design

Mission Squad

Parallel agent fleet. Helpers gather context concurrently; the primary agent gets a structured briefing.

/squad implement <task>

Autonomous Goals

/goal Plan → Code → Test → Evaluate loop with budget tracking, auditor, and persistent state.

/goal Migrate to TypeScript

Spec-Kit (SDD)

9-phase specify → clarify → plan → tasks → implement before any code is written.

/speckit run user login flow

Code search

Local BM25 keyword + symbol search via the bundled codedigest. No cloud calls, no embeddings download.

/codedigest search "ring buffer"

Hardware bridge

JLink (pylink-square), CAN/LIN traces (cantools), Saleae Logic 2 — invoked from the REPL via hw-debugger.

/debug analyze trace.asc

Marketplace

Install skills, MCP servers, and Claude Code plugins from a curated registry or any GitHub URL.

/marketplace install code-review

Traceability

@req REQ-NNN tags link requirements → code → tests → audit. Optional but powerful when needed.

/traceability

Multi-LLM federation

Gemini by default; optionally Claude (Anthropic SDK with stream-json). Swap backends without changing your workflow.

[federation] backend = "anthropic"

Web control UI · NEW

Browser SPA (Preact + htm) with seven panels: Chat, Pipeline, Squad, Goal, Design, Memory, Jobs. Auth-token + origin gated, served by the same orchestrator process.

/web start --port 8765

Design panel · NEW

Pick a design skill, generate a sandboxed HTML preview, click any region to edit it via comment mode, then export. Inspired by open-codesign.

→ Design tab in /web

`/overnight` loop · NEW

Long-running autonomous loop with per-iteration git commits, rollback on failure, worktree-isolated parallel runs, and a permanent run summary. Agent-agnostic. Inspired by gnhf.

/overnight --parallel 2 -- gemini "..."

HTML / PDF / PPTX export · NEW

@flow/exporter turns a generated design into a shareable artefact. Puppeteer for PDF and pptxgenjs for PPTX are optional peer deps.

/export landing.html --to pdf

Engineering skills bundle · NEW

10 vetted Pocock-style skills imported under MIT: diagnose, tdd, prototype, grill-with-docs, triage, to-prd, to-issues, zoom-out, improve-codebase-architecture, setup-engineering-skills.

Use the diagnose skill — debug X

Skill traceability · NEW

@req REQ-SKILL-NNN tags inside SKILL.md files. The scanner inventories them by source, flags duplicates, and reports skills missing a tag.

traceability_scan.py --skills .gemini/skills

MCP support

Install any MCP server (GitHub, filesystem, Postgres, …). AgenIT also exposes its own MCP server so external agents can read soul + memory.

/mcp install github-mcp

Worktree isolation

Risky work in an isolated git worktree under .gemini/worktrees/. Same helper powers /overnight --parallel — one code path, one safety story.

/worktree feat/x

// 04 Demos

Eight walkthroughs of the things you'll actually use day to day.

AgenIT REPL — V-Model rail, conversation, soul + codedigest right rail — The REPL — pipeline rail on the left, conversation in the middle, *Job Tray* · *Soul* · *Codedigest* on the right. Notice the soul rail showing `44 entries · Engineering Defaults · User Profile` — that's the agent remembering you between sessions.

The agent learns — without being asked to learn

A soul-keeper background job watches conversations for stable preferences and appends them to soul.md. Project memory works the same way for project-specific facts. /soul and /memory show you what's been captured.

[my-app] agenit> for new modules use vitest, not jest. and arrow fns please.

  → got it — will follow vitest + arrow-functions for new modules.
  ◈ soul-keeper appended 2 entries to soul.md (background)

[my-app] agenit> /soul

  ── Engineer Soul ──────────────────────────────────────
  ## Code style
    - 4-space indent, no tabs
    - prefer arrow functions
  ## Tooling defaults
    - test runner: vitest
    - bundler: vite for libs, next for apps
  ## Communication preferences
    - terse summaries
    - never emoji in commit messages
  … 38 more entries

[my-app] agenit> /memory

  ── Project Memory · my-app ────────────────────────────
  ## Architecture
    - service layer between routes and db
    - feature flags via GrowthBook
  ## Conventions
    - every handler emits an analytics event
    - PR titles use Conventional Commits

Initialise a project — auto-detected profile

agenit init scans the repo and picks a profile from markers. It seeds .agenit/, .gemini/skills/, and a starter AGENTS.md. Re-running on an existing project repairs config without overwriting customisations.

$ cd ~/projects/my-app
$ agenit init --name my-app

  ✓ profile detected: web (next + typescript + vitest)
  ✓ seeded .agenit/profiles/web/
  ✓ seeded .gemini/skills/      (62 skills)
  ✓ seeded .gemini/settings.json (absolute hook paths)
  ✓ seeded AGENTS.md             (project memory starter)
  ✓ linked engineer soul from ~/.flow/soul.md

$ agenit
[my-app] agenit>

Full pipeline run — one prompt, four gated stages

/run <task> walks Requirements → Architecture → Code → Testing → Audit. Each stage activates a different skill. Memory and soul ride along on every stage so style stays consistent.

[my-app] agenit> /run add OAuth login flow with Google

  ▶ Requirements
    EARS user stories + acceptance criteria → requirements.md
    ✓ REQ-014, REQ-015, REQ-016 (3 new)

  ▶ Architecture
    component diagram + sequence flow → architecture.md
    ✓ NextAuth provider config sketched

  ▶ Code             @req REQ-014..016
    edits app/api/auth/[...nextauth]/route.ts
    edits lib/auth/google.ts
    edits middleware.ts
    ✓ follows project service-layer convention (from AGENTS.md)
    ✓ arrow functions, single quotes (from soul.md)

  ▶ Testing
    creates tests/auth.spec.ts (vitest, 7 cases)
    ✓ REQ-014..016 covered

  ▶ Audit
    ✓ 100% requirements covered
    ✓ 0 orphans · 0 test gaps

Mission Squad — parallel context-gathering before the implementer fires

/squad implement spawns helpers in parallel (req-reader, code-scanner, codedigest, lint pre-flight), aggregates their output into a structured briefing, and auto-fires the implementer with the briefing pre-injected. Ctrl-C cancels everything cleanly.

[my-app] agenit> /squad implement add password reset

  Mission Squad
  Mission: add password reset
  Helpers: 7 across 3 phase(s)
  ──────────────────────────────────────────────────
  Phase 1
    ✓ req-reader            (52 ms)
    ✓ session-memory        (34 ms)
    ⟳ codedigest-searcher   (8s)
  Phase 2
    ✓ code-scanner          (210 ms)
    ✓ lint-preflight        (640 ms)
    ✓ dependency-check      (1.2 s)
  Phase 3
    ✓ synthesiser           (3.2 s)

  ▶ firing implementer with briefing (4.1 KB)

Autonomous Goal — Plan → Code → Test → Evaluate, persistent across restarts

Define an objective, kick off the runner, walk away. Every turn the auditor inspects git status, counts new tests, optionally runs your verify command, and decides Continue / Complete / Block. State is checkpointed under memory/projects/<name>/goal.json — quit the REPL, the runner survives.

[my-app] agenit> /goal Migrate the entire codebase from JavaScript to TypeScript

  ✓ GOAL-20260509010244-8a11 created
  objective:    Migrate the entire codebase from JavaScript to TypeScript
  turn budget:  50

[my-app] agenit> /goal start
  ▶ Autonomous runner started for GOAL-20260509010244-8a11

[my-app] agenit> /goal status
  ── Goal Status ──────────────────────────────────────
  state:          Active
  turns:          7/50 (43 remaining)
  files touched:  19
  tests added:    3
  errors logged:  0
  runner:         running

Spec-Kit (SDD) — turn a fuzzy idea into structured specs

Run before the pipeline when you don't yet have requirements. The 9-phase flow produces spec.md, plan.md, and tasks.md under .specify/specs/<feature>/.

[my-app] agenit> /speckit run add team workspaces

  ▶ phase 1 specify       → spec.md
    user stories: 4 · acceptance criteria: 14 · NFRs: 5

  ▶ phase 2 clarify       → 3 ambiguities resolved
    "team" → 2-50 members, single owner
    "invite" → email-only (SSO deferred)

  ▶ phase 3 plan          → plan.md
    stack: existing Next.js + Postgres · new tables: teams, memberships

  ▶ phase 4 tasks         → tasks.md (18 ordered tasks)

[my-app] agenit> /speckit implement
  ▶ executing tasks.md via implementer skill

Code search — local BM25 + symbol search, no cloud calls

Index a repo once, search forever. The bundled codedigest package walks src/-style paths with a .gitignore-aware filter, regex-extracts function / class chunks, and stores BM25 scores in JSON. No native deps, no model download.

[my-app] agenit> /codedigest index ./src
  ✓ walked 412 files
  ✓ extracted 2 074 chunks
  ✓ stored at memory/projects/my-app/codedigest_my-app.json

[my-app] agenit> /codedigest search "rate limiter"

  1. src/lib/rate-limit.ts:42     score 9.21
     export const tokenBucket = (capacity, refillPerSec) => { ... }
  2. src/middleware/api.ts:88    score 6.54
  3. tests/rate-limit.spec.ts:14 score 5.08

Hardware bridge — JLink, CAN/LIN, Saleae from inside the REPL

The hw-debugger skill auto-activates on hardware-flavoured prompts. It calls Python tools in .flow/tools/ via the bundled venv. Optional — ignore it entirely if you're not on embedded.

[stm32-can] agenit> analyze trace.asc for CAN timeout errors

  → hw-debugger skill activates
  → running .venv/bin/python3 .flow/tools/lin_can_tool.py trace.asc

  CAN timeout report
  ──────────────────────────────────────────────
  frame ID    0x101  EngineState
  expected    20 ms cycle
  observed    342 ms gap at 12.418 s
  hypothesis  ISR latency on EngineState publisher

  → writing trace_analysis.md

Marketplace — curated registry + live Claude Code ecosystem

/marketplace aggregates the curated AgenIT registry and the live Claude Code marketplace. Skills, MCP servers, agents, slash commands, and hooks are translated into Gemini-compatible artefacts at install time.

[my-app] agenit> /marketplace search "code review"

  curated
    code-review              MISRA + style + security review (1.2 KB)

  claude-code/anthropics
    cc/code-review           Anthropic's code-review plugin (3 skills, 1 MCP)
    cc/security-review       OWASP-aware diff scanner (2 skills)

[my-app] agenit> /marketplace install cc/code-review
  ✓ 3 skills → .gemini/skills/
  ✓ 1 MCP server → ~/.gemini/settings.json

// 05 The Pipeline

Requirements ──► Architecture ──► Code ──► Testing ──► Audit
     ▲                                                  │
     └──────────── @req REQ-NNN traceability ◄──────────┘

Profiles re-map stage names and standards to your domain. The mechanics are the same.

Profile	Standards & frameworks
generic	Language-agnostic starter — copy and customise. The default.
web	TypeScript / React / Node, OWASP Top 10, Jest / Playwright, WCAG 2.2 AA.
embedded	MISRA C, FreeRTOS, Doxygen, Unity / CppUTest, REQ-NNN tags.
automotive	ASPICE V-Model (SWE.1–SWE.6), MISRA C:2012, ISO 26262 / ASIL, AUTOSAR.

// 06 Announcements & Updates

Live feed from GitHub Releases & recent commits — refreshes every visit.

MAY 2026

Web UI · Design domain · `/overnight` loop

AgenIT now drives entire UI/UX prototyping sessions and unattended overnight engineering loops — from the same REPL you already know, or from a new browser control SPA.

Browser control UI with seven panels: Chat · Pipeline · Squad · Goal · Design · Memory · Jobs. Auth-token + origin gated.
design profile + 7 design skills (landing · dashboard · pricing · chat-ui · SVG illustration · single component · comment-mode editing) + sandboxed iframe preview.
HTML / PDF / PPTX export via the new @flow/exporter package — puppeteer + pptxgenjs as optional peer deps.
/overnight loop — per-iteration git commits, rollback on failure, worktree-isolated parallel runs, permanent markdown summary. Agent-agnostic. Inspired by gnhf.
10 engineering skills (diagnose, tdd, prototype, grill-with-docs, triage, to-prd, …) imported under MIT from mattpocock/skills.
Skill traceability — @req REQ-SKILL-NNN tags inventoried by the scanner.

22 new vitest cases · 773 tests passing · build green across 17 packages · full end-to-end checklist at doc/testing-checklist.md.

⌁ Latest Releases

Loading…

⌁ Recent Commits to `main`

Loading…

// 07 Complete Command Reference

Every slash command, grouped. Type /help in the REPL for the same list inline.

Project & session

`/init [name] [--profile X]`	Initialise / repair project config.
`/unflow`	Remove AgenIT config — restore vanilla Gemini CLI.
`/project <name>`	Switch or create active project.
`/yolo on\|off\|auto_edit`	Toggle approval mode for tool calls.
`/jobs [<id>]`	List / inspect / stop background jobs.
`/configure [hardware]`	JLink + Saleae detection, env validation.
`/feature <description>`	Start a feature workflow.
`/help` · `/clear` · `/exit`	Standard.

          Memory & soul
          
            /soulShow .flow/soul.md — engineer-wide preferences.
/memoryShow what context is being injected this session.
/worktree <branch>Isolated git branch for risky work.
/undoRestore the most recent file snapshot.
/undo listShow last 10 snapshots.
/undo <commandId>Restore a specific snapshot.

        

          Web UI & design NEW
          
            /web start [--port N]Launch the browser control SPA (Chat · Design · Goal · …).
/web stopStop the control server.
/export <file.html> --to pdfConvert a design HTML to PDF (puppeteer) or PPTX (pptxgenjs).
/export <file> --to html --out <path>Copy / re-emit HTML to a chosen path.

        

          Overnight loop NEW
          
            /overnight -- <shell cmd>Run a loop with per-iter git commits + rollback.
/overnight --max-iter N -- <cmd>Cap iterations (default 50).
/overnight --budget-tokens N -- <cmd>Stop once a tick reports tokensSpent ≥ N.
/overnight --parallel N -- <cmd>Run N worktree-isolated copies in parallel.
/overnight --no-commit -- <cmd>Dry-run; no commits, no rollback.
/overnight --allow-dirty -- <cmd>Skip the clean-tree precondition.

        

Pipeline (V-Model)

`/run <task>`	Full pipeline (Req → Arch → Code → Test → Audit).
`/swe1 <task>`	Requirements stage (EARS, REQ-NNN / SWR-NNN).
`/swe2 <task>`	Architecture stage (HLD, PlantUML).
`/swe4 <task>`	Implementation stage.
`/swe5 <task>`	Testing stage.
`/audit`	Traceability + compliance report.
`/traceability`	Show the @req graph.

Spec-Kit (SDD)

`/speckit run <task>`	specify → plan → tasks in one shot.
`/speckit specify <task>`	User stories + acceptance criteria → `spec.md`.
`/speckit clarify`	Resolve ambiguities.
`/speckit plan [stack]`	Technical architecture → `plan.md`.
`/speckit tasks`	Ordered task list → `tasks.md`.
`/speckit implement`	Execute `tasks.md`.
`/speckit checklist`	Generate review checklist.
`/speckit analyze`	Cross-check spec / plan / tasks consistency.
`/speckit issues`	Open GitHub issues from tasks.
`/speckit constitution`	Load / edit project constitution.
`/speckit init`	Initialise Spec-Kit in this project.
`/speckit branch`	Create a feature branch from spec.
`/speckit commit`	Stage + commit current spec/plan/tasks.
`/speckit status`	List all `.specify/specs/` features.
`/speckit export-to-swe1`	Seed pipeline requirements from spec.

Mission Squad

`/squad implement <task>`	Deterministic helpers → implementer.
`/squad implement-agentic <task>`	Same shape; every helper is its own LLM sub-agent.
`/squad debug <task>`	Trace + code + lint pre-flight → `hw-debugger`.
`/squad review <task>`	Style + req + code → reviewer.
`/squad test <task>`	Req + code + memory → tester.
`/squad requirements <task>`	Single-phase context gather → requirements.
`/squad architecture <task>`	Req + memory + headers → architect.
`/squad <template>-agentic <task>`	LLM-per-slot variant of any template.

Autonomous Goals

`/goal <objective> [--force]`	Create / replace the active goal.
`/goal`	Short status (id, state, turns).
`/goal status`	Verbose dashboard.
`/goal log [n]`	Audit log (or last n entries).
`/goal pause` · `resume`	Flip state mid-run.
`/goal start`	Spawn the autonomous runner.
`/goal stop`	Cancel the runner cleanly.
`/goal tick`	Step a single turn manually.
`/goal budget <n>`	Extend the turn budget.
`/goal clear`	Archive to `goals_archive.md`.

Code search

`/codedigest index <path>`	Build BM25 index for the given path.
`/codedigest search <query>`	Top 5 results, ranked.
`/codedigest search <q> --lang c`	Filter by language.
`/codedigest status`	Show index stats.
`/codedigest clear`	Drop the index.

Hardware & debug

`/debug <query>`	Ad-hoc: JLink registers, CAN/LIN, Saleae, code search.
`/configure hardware`	Detect and configure JLink / Saleae.

Marketplace & plugins

`/marketplace list`	Curated registry + live Claude Code marketplace.
`/marketplace search <q>`	Search across all sources.
`/marketplace install <id>`	Install by id (use `cc/<name>` for Claude Code entries).
`/marketplace installed`	List installed artifacts.
`/marketplace remove <id>`	Uninstall.
`/marketplace update`	Refresh registry caches.
`/marketplace add <github-url>`	Install any Claude Code plugin from a GitHub URL.
`/marketplace cc-browse [<owner>/<repo>]`	Browse a Claude Code marketplace index.
`/marketplace mcp <query>`	Search the live MCP Registry API.
`/marketplace mcp servers`	List discovered MCP servers.
`/marketplace mcp ping <server>`	Live MCP handshake.
`/marketplace mcp tools <server>`	List tools an MCP server advertises.
`/marketplace mcp call <srv>:<tool> <json>`	Invoke an MCP tool directly.
`/plugin install <github-url>`	Install a Claude Code plugin into the active backend.
`/plugin list` · `info` · `remove`	Manage installed plugins.
`/skills list` · `install` · `remove`	Manage `.gemini/skills/`.

Code intelligence & modes

`/lsp def <file>:<line>:<col>`	Goto definition (clangd / pyright / rust-analyzer).
`/lsp refs <file>:<line>:<col>`	Find references.
`/lsp hover <file>:<line>:<col>`	Type / docs at point.
`/lsp diag [<file>]`	Diagnostics from the language server.
`/plan`	Plan mode — tool calls require approval.
`/build`	Switch back to default execute mode.

Input shortcuts (not slash commands)

`@src/file.ts`	Insert file content as a code block.
`@src/`	Insert directory listing.
`/cmd` + `Tab`	Fuzzy autocomplete slash commands.
`@` + `Tab`	File / folder picker.
`Ctrl-C`	Cancel the current turn / squad / goal tick.
`Ctrl-U`	Scroll to earlier messages.

CLI flags (outside the REPL)

`agenit init [--name] [--profile X]`	Initialise a project.
`agenit init --no-agents`	Skip `AGENTS.md` seed.
`agenit init --no-skills`	Skip `.gemini/skills/` seed.
`agenit init --no-verify`	Skip the post-init Gemini-CLI verification.
`agenit --help`	Top-level help.
`FLOW_GOAL_TEST_CMD=…`	Verify command run after each `/goal` turn.
`FLOW_GOAL_WORKSPACE=…`	Override the workspace the auditor inspects.

// 08 Install

From npm (recommended)

$ npm i -g @agenit/cli
$ agenit --help

$ gemini                    # one-time Google OAuth
$ cd ~/projects/my-project
$ agenit init               # auto-detects profile
$ agenit                    # start the REPL

From source

$ git clone https://github.com/muhammed-eldabea/AgenIT.git
$ cd AgenIT
$ bash install.sh           # Linux / macOS
# Windows:
> powershell -ExecutionPolicy Bypass -File install.ps1

Prerequisites

Node.js ≥ 20
pnpm ≥ 9 — npm i -g pnpm
Python ≥ 3.11 — hardware tools only
Gemini CLI ≥ 0.37
Google account for OAuth

`/soul`	Show `.flow/soul.md` — engineer-wide preferences.
`/memory`	Show what context is being injected this session.
`/worktree <branch>`	Isolated git branch for risky work.
`/undo`	Restore the most recent file snapshot.
`/undo list`	Show last 10 snapshots.
`/undo <commandId>`	Restore a specific snapshot.

`/web start [--port N]`	Launch the browser control SPA (Chat · Design · Goal · …).
`/web stop`	Stop the control server.
`/export <file.html> --to pdf`	Convert a design HTML to PDF (puppeteer) or PPTX (pptxgenjs).
`/export <file> --to html --out <path>`	Copy / re-emit HTML to a chosen path.

`/overnight -- <shell cmd>`	Run a loop with per-iter git commits + rollback.
`/overnight --max-iter N -- <cmd>`	Cap iterations (default 50).
`/overnight --budget-tokens N -- <cmd>`	Stop once a tick reports tokensSpent ≥ N.
`/overnight --parallel N -- <cmd>`	Run N worktree-isolated copies in parallel.
`/overnight --no-commit -- <cmd>`	Dry-run; no commits, no rollback.
`/overnight --allow-dirty -- <cmd>`	Skip the clean-tree precondition.

An agent that grows with you.

// 01 Objective

Learns who you are

Learns your project

Picks the right specialist

Stays in your control

// 02 Memory & Soul

Engineer-wide preferences

Per-project context

How the agent actually learns

What gets sent to the model on every turn

What growing-with-you actually looks like

// 03 Features

Persistent soul & memory

~80 auto-activated skills

Disciplined 4-stage pipeline

5 domain profiles

Mission Squad

Autonomous Goals

Spec-Kit (SDD)

Code search

Hardware bridge

Marketplace

Traceability

Multi-LLM federation

Web control UI · NEW

Design panel · NEW

/overnight loop · NEW

HTML / PDF / PPTX export · NEW

Engineering skills bundle · NEW

Skill traceability · NEW

MCP support

Worktree isolation

// 04 Demos

The agent learns — without being asked to learn

Initialise a project — auto-detected profile

Full pipeline run — one prompt, four gated stages

Mission Squad — parallel context-gathering before the implementer fires

Autonomous Goal — Plan → Code → Test → Evaluate, persistent across restarts

Spec-Kit (SDD) — turn a fuzzy idea into structured specs

Code search — local BM25 + symbol search, no cloud calls

Hardware bridge — JLink, CAN/LIN, Saleae from inside the REPL

Marketplace — curated registry + live Claude Code ecosystem

// 05 The Pipeline

// 06 Announcements & Updates

⌁ Latest Releases

⌁ Recent Commits to main

// 07 Complete Command Reference

Project & session

Memory & soul

Web UI & design NEW

Overnight loop NEW

Pipeline (V-Model)

Spec-Kit (SDD)

Mission Squad

Autonomous Goals

Code search

Hardware & debug

Marketplace & plugins

Code intelligence & modes

Input shortcuts (not slash commands)

CLI flags (outside the REPL)

// 08 Install

From npm (recommended)

From source

Prerequisites

`/overnight` loop · NEW

⌁ Recent Commits to `main`