Skip to content

VivanRajath/gitagent-hackathon

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

repo-sandbox-agent

Built for the gitagent Hackathon. Defined using the gitagent standard. Brought to life with gitclaw.

A multi-agent system that scans any Git repository to detect its runtime and run instructions, then routes natural-language code-edit requests to tiered sub-agents to keep LLM costs proportional to task complexity. It autonomously generates themed Next.js websites from a single prompt using a five-stage AI pipeline, with images generated via Gemini Imagen 3, Pollinations.ai, and Puter.js. After a site is generated, voice commands can edit individual CSS variables in real time via a dedicated voice-edit pipeline — no full rebuild needed.


What It Does

You give it a prompt like:

you> create a hulk themed website

And it orchestrates 23 specialized AI agents across two modes:

Website Builder — an architect plans the design, a researcher discovers the design DNA (including intent-driven color palettes), an art director curates image themes and produces a refinedPalette, a copywriter fills every content zone, and an image agent generates AI visuals via Gemini Imagen 3 → Pollinations.ai → Puter.js. The architect then generates five design variants, marks one recommended:true, and writes the chosen palette to globals.css. The result is a fully themed, live-reloading Next.js site on localhost:3001.

Code Editing — a guardrails squad blocks credential leaks and unsafe diffs, a research team finds prior art, and three developer tiers (jnr / snr / architect) apply edits with proportional model cost. A QA tester screenshots the result and dispatches fixes.

Voice Editing — speak a change to the running site ("change navbar text to white", "make the background darker") and a 3-stage voice-edit pipeline applies only the requested CSS variable — no full rebuild. Zone-specific CSS variables (--color-nav-text, --color-hero-text) ensure that targeting "navbar text" never accidentally changes hero or body text.

Then you can also edit the generated site by clicking elements and speaking, or by showing hand gestures to your webcam:

  • Click any element → it gets selected (orange highlight)
  • Hold 🎙 → speak a natural-language edit ("make this text more dramatic")
  • Peace sign ✌️ → move your hand left/right to shift the site's color theme
  • Pinch a selected element → context menu appears (edit text, redesign, change color)

Editing Any Local Repository

You can point the agent at any file on your machine:

you> C:\Users\Vivan Rajath\Desktop\repo\sandbox-test-python\main.py update this feature

The orchestrator routes this to the developer squad, loads the pure source via file_read, applies the minimal requested change, and safely writes back via file_write.


The gitagent Standard

This project is a fully compliant gitagent repository. Every agent follows the standard:

<agent-name>/
├── agent.yaml       # spec_version, name, model, skills, tags
├── SOUL.md          # identity & personality
├── RULES.md         # must-always / must-never constraints
└── skills/
    └── <skill-name>/
        └── SKILL.md # YAML frontmatter + execution instructions

Full repo structure:

repo-sandbox-agent/
├── agent.yaml
├── SOUL.md / RULES.md / ARCHITECTURE.md
├── skills/                        # 19 orchestrator-level skills
│   ├── architect-website/         # triggers website-builder pipeline
│   ├── generate-images/           # Gemini → Pollinations → Puter.js
│   ├── route-intent/
│   ├── apply-edit/
│   ├── design-ui/
│   ├── find-resources/
│   ├── run-guardrails/
│   ├── run-research/
│   └── ... (19 total)
├── tools/
│   ├── file-read.yaml
│   ├── file-write.yaml
│   ├── search.yaml
│   ├── shell-exec.yaml
│   └── scripts/
└── agents/
    ├── code-editor/               # edit orchestrator
    │   └── agents/
    │       ├── jnr-developer/
    │       ├── snr-developer/
    │       ├── architect/
    │       ├── uiux-designer/
    │       └── site-tester/
    ├── guardrails/                # safety orchestrator
    │   └── agents/
    │       ├── policy-enforcer/
    │       ├── secret-sentinel/
    │       ├── diff-auditor/
    │       └── scope-validator/
    ├── research-agent/            # search orchestrator
    │   └── agents/
    │       ├── github-searcher/
    │       ├── devpost-searcher/
    │       ├── web-searcher/
    │       └── solution-proposer/
    ├── website-builder/           # website generation pipeline
    │   ├── research-agent/        # design DNA + intent-driven palette
    │   ├── resourcer-agent/       # visual art direction + refinedPalette
    │   └── uiux-agent/            # copy + content agent
    ├── voice-intent/              # voice-edit NLU — maps speech → CSS edits
    ├── scope-validator/           # voice-edit safety — validates CSS field list
    ├── image-gen-agent/           # AI image + animation agent (shared)
    └── resourcer-agent/           # stock image + font resource agent

Agent Architecture (23 Agents)

Squad Agent What It Does
Orchestrator repo-sandbox-agent Top-level router. Classifies user intent and delegates to the correct squad.
Code Editor code-editor Edit orchestrator. Dispatches to tier sub-agents based on complexity.
jnr-developer Single-file edits. Typo fixes, isolated bug patches. Cheapest model tier.
snr-developer Multi-file features. New routes, cross-component integration. Receives fix reports from site-tester.
architect Full-codebase scope. System refactors. Read-only shell access.
uiux-designer All UI/UX changes. Translates canvas sketches and hand gestures into React components.
site-tester Read-only QA. Screenshots the live site, harvests console errors, dispatches fixes. Max 3 cycles.
Guardrails guardrails Safety interceptor. Invoked before every file write, shell exec, and sub-agent dispatch.
policy-enforcer Issues ALLOW/BLOCK on every action based on RULES.md. Cannot be overridden.
secret-sentinel Blocks access to credentials. Scans paths and diffs for .env, *.pem, API keys.
diff-auditor Pre-write safety review. Catches eval() injection, mass deletions, auth removal.
scope-validator Validates tier routing. Prevents complex tasks being sent to cheap agents.
Research research-agent Search coordinator. Fans out to searcher sub-agents.
github-searcher Searches GitHub repos, code, and issues.
devpost-searcher Searches Devpost hackathon projects.
web-searcher General web search fallback.
solution-proposer Aggregates all search results into an actionable implementation proposal.
Website Builder website-builder Pipeline orchestrator for AI website generation (5 stages).
wb/research-agent Extracts design DNA: site type, intent-driven palette, layout variants, image keywords.
wb/resourcer-agent Produces image theme phrase, per-zone seeds, font, and refinedPalette override.
wb/uiux-agent Writes all site copy: headlines, nav, cards, features, CTA, footer.
wb/architect Generates 5 design variants, sets recommended:true on best match, writes palette to globals.css.
Voice Edit voice-intent NLU layer for live voice editing. Maps natural language → { edits: [{field, value}] }.
scope-validator Safety gate for voice edits. Validates each CSS field against the allowed variable list.
Creative image-gen-agent Crafts image prompts → Gemini Imagen 3 → Pollinations.ai → Puter.js config. Writes CSS @keyframes animations.
resourcer-agent Curates Unsplash/Picsum images, WCAG-compliant color palettes, Google Font pairings.

Skills (19)

Skills are composable instruction modules. The agent reads the skill's SKILL.md and executes its steps using registered tools.

architect-website · route-intent · apply-edit · design-ui · detect-runtime · test-site · code-review · run-guardrails · enforce-policy · scan-secrets · audit-diff · validate-scope · find-resources · generate-images · run-research · search-github · search-devpost · search-web · propose-solution

Type /skills in the REPL to list all available skills.


Tech Stack

Layer Technology
Agent Standard gitagent v0.1.0
Agent Runtime gitclaw SDK
LLM Provider Groq (Llama 3.3 70B primary · Llama 4 Scout 17B fallback · Llama 3.1 8B cheap tier)
Generated Sites Next.js 15 + Tailwind CSS
Gesture Tracking MediaPipe Vision (GestureRecognizer, CDN-loaded at runtime)
AI Image Gen (primary) Google Gemini Imagen 3 (imagen-3.0-generate-002)
AI Image Gen (fallback) Pollinations.ai (free, no API key required)
AI Image Gen (client) Puter.js puter.ai.txt2img() via puter-image-config.js
Stock Images Unsplash Source, Picsum Photos

CSS Variable System

The generated site uses zone-isolated CSS custom properties so voice edits are surgical:

Variable Zone Changed by
--color-bg Page background "change background"
--color-primary Buttons, accents, highlights "change primary color"
--color-secondary Hero headline, section titles "change secondary / hero title color"
--color-text Body / card text "change all text" (global only)
--color-nav-text Navbar links & labels "change navbar text to …"
--color-hero-text Hero subtext & body copy "change hero text to …"
--font-display Heading / display font "change the font to …"

Saying "change navbar text to white" patches only --color-nav-text — hero and body text are unaffected.


Getting Started

See the Runbook for setup and usage instructions.

See the Architecture for the full technical deep-dive into every agent, skill, tool, and data flow.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors