v1.2.3 — Specialists, Memory, Self-Healing & Task Intelligence

The biggest SajiCode release yet 16-agent team, 22 skill packs, 3-layer memory, self-healing validation, and full task graph parallelism.

New specialist agents

Six new specialists added to the core team: nextjs-specialist, python-api-specialist, ai-rag-specialist, 3d-web-specialist, performance-specialist, mcp-specialist. Each agent loads only its own skill packs — no bloated context.

22 skill packs

New additions: skill-creator (supersedes write-a-skill), 3d-web-experience, python-engineer, mcp-server, fullstack-app-generator.

3-layer memory + 5 subsystems

Persistent intelligence across sessions pointer index, topic files, transcript search, plus Agent Memory, Project DNA, Session State, Experience Replay, and Pattern Library.

Self-healing validation

Inline validation after every write: tsc, py_compile, cargo check, go build, gradlew. Agents auto-fix errors up to 3 cycles before blocking.

Task graph & parallelism

Full DAG execution — dependency ordering, parallel dispatch, live status tracking, dynamic agent routing, and workload balancing.

Everything else

Git integration (branch, checkpoint, revert session/agent), code intelligence (semantic search, symbol finder, standup briefing), security scanning on every write, 5-middleware stack per agent, MCP server support, web search for agents.

v1.2.2

1.2.2

SajiCode v1.2.0

SajiCode v1.2.0

This release makes SajiCode faster, safer, and much more capable for real multi-file coding work.

SajiCode is moving closer to the goal: an AI engineering team that can plan, delegate, build, recover from failures, and edit complex projects without losing control or safety.

What’s New

Multi-file writing and editing

Specialist agents can now apply structured multi-file changes in one batch.

New capabilities:

• preview_file_batch
• apply_file_batch
• write, replace, append, and prepend operations
• path validation before changes
• automatic snapshots
• rollback if a batch fails
• predictive code checks before risky files are written
• UI progress events for each file in the batch

This makes feature work much faster than writing one file at a time, while keeping changes safer.

Dependency-aware task graph

SajiCode can now plan work as a task graph with dependencies.

Example:

• Backend API can start immediately
• Frontend UI can start immediately
• QA tests wait until backend and frontend are complete

This improves parallel agent coordination and gives the PM a cleaner way to decide what can run now versus what is blocked.

Smarter PM delegation

The PM agent is now planner-only for implementation work.

PM can:

• create context
• create plans
• build task graphs
• delegate to specialist leads
• coordinate progress

PM cannot directly write app code or run implementation commands. This keeps responsibilities clean and makes the system safer.

Error recovery system

Added a deterministic error recovery system that classifies failures and suggests next actions.

It can recognize issues like:

• CommonJS/ESM mismatch
• missing modules
• TypeScript errors
• syntax errors
• permission errors
• timeouts
• invalid tool schemas
• PM implementation guardrail blocks

Recovery suggestions include retrying with a fix, delegating to the right lead, decomposing the task, or escalating.

Predictive code analysis

Agents can now scan code before writing or running it.

The analyzer can catch:

• require() inside ES module projects
• hardcoded secrets
• unsafe innerHTML
• plaintext password storage
• placeholder code
• empty catch blocks
• risky dynamic execution
• common runtime/security mistakes

This helps catch problems before they become failed builds or broken apps.

Intelligent caching

Added a cache layer for expensive project scans.

Currently cached:

• repo map scans
• project context collection

The cache invalidates when project files change, reducing repeated scan time while keeping context fresh.

Cleaner multi-agent terminal UI

The terminal output is now easier to follow.

Improvements:

• clearer PM and agent lanes
• cleaner task graph progress
• compact executable task views
• better multi-file batch progress
• less duplicated tool output
• easier visibility into what each agent is doing

Safety Improvements

This release adds stronger guardrails around file edits and agent responsibilities.

Multi-file edits now:

• stay inside the project root
• block protected paths like .git, node_modules, dist, and real .env files
• snapshot files before modification
• roll back the whole batch on failure
• block high-risk predictive findings unless explicitly acknowledged

PM guardrails now prevent the PM agent from writing implementation files directly.

Why This Release Matters

This is a major foundation upgrade.

SajiCode can now handle larger coding tasks with:

• better planning
• faster parallel execution
• safer multi-file edits
• clearer agent coordination
• proactive issue detection
• recovery guidance when things fail

It is a big step toward making SajiCode feel like a real AI engineering team instead of a single coding assistant.

Upgrade

npm install -g sajicode@latest

Or run directly:

npx sajicode

v1.0.4: Three-Layer Memory, 23 Security Checks, Dynamic MCP Integration

🚀 Major Features

Three-Layer Memory Architecture

• Pointer Index: 150-char summaries, always loaded
• Topic Files: Detailed knowledge, on-demand
• Transcripts: Raw history, grep-only
• 6 Memory Tools for agent use
• 70% token reduction in memory operations

Enhanced Security System

• 23 Security Checks covering all major risks
• Three-Gate Architecture: Pattern detection → Context analysis → Risk scoring
• Risk Scoring: 0-100 scale with automatic HITL triggers
• 90% reduction in security incidents

Dynamic MCP Integration

• Tool Name Prefixing: Automatic `server__tool` naming
• Any-Server Support: No hardcoded tool lists
• Preserved Functionality: Tools work seamlessly with prefixed names
• Fixed Constructor: Proper MultiServerMCPClient initialization

📚 Documentation

• `MEMORY_SYSTEM.md` (329 lines)
• `SECURITY_CHECKS.md` (520 lines)
• `MCP_INTEGRATION.md` (589 lines)
• Updated README with comprehensive feature overview

🔧 Technical Details

• 17 specialized agents
• 21 expert skills
• 3-layer memory system
• 23 security checks
• Dynamic MCP tool loading

📦 Installation

```bash
npm install -g sajicode@1.0.4
```

🔗 Links

• npm Package
• Documentation
• Memory System Guide
• Security Checks Guide
• MCP Integration Guide"

SajiCode v1.0.3: Advanced Engineering & TDD Workflow Release

This release introduces major functionality to SajiCode's autonomous engineering capabilities, bringing enterprise-grade version control, persistent memory, and a strict Test-Driven Development (TDD) pipeline to the 17-agent orchestrator.

🌟 What's New

1. Intelligent Version Control & File Tracking
   Git Integration: Full suite of git tools (commit, branch, diff, checkpoint) are now wired directly into the PM and domain leads.
   File Change Tracker: Built-in snapshot and undo capabilities allow agents to intelligently track and revert mistakes automatically if a change breaks the build.
   New CLI Commands:
   sajicode undo : Instantly undo the last modification made by an agent to a specific file.
   sajicode snapshots: View a log of the most recent file snapshots taken by agents.
2. TDD-First Architecture & Quality Gates
   TDD Workflow: Test-Driven Development is now strictly enforced within the QA and domain leads' workflow. Every newly generated feature or module must be backed by comprehensive testing before the PM marks the task as complete.
   Hooks System: Added support for pre and post-action hooks, allowing for deep customization of agent workflows.
3. Headless & CI/CD Mode Support
   SajiCode is no longer just an interactive terminal UI. Using the --headless flag (e.g., sajicode build "Fix login bug" --headless), you can now trigger agents to write tests, review code, or audit vulnerabilities automatically inside your CI/CD pipelines (GitHub Actions, Vercel, etc).
4. Engine Optimizations & Advanced Memory
   CompositeBackend Persistent Memory: Upgraded to an advanced persistent memory architecture that spans multiple sessions and storage layers.
   Summarization Middleware: Actively condenses context in the background to prevent LLM context window collapse during exceptionally long or complex engineering tasks (enabled by default).
   Sub-agent Strict Prompt Limits: Enforced 500-word maximum response limits on all sub-agents combined with enhanced context briefing delegation. This guarantees massive improvements to speed and ensures concise, agile outputs from the specialists.
   Streaming Progress Dashboard: Upgraded the CLI renderer to provide real-time visibility into the exact activities and delegations of all 17 agents simultaneously.

Full Changelog: v27...v1.0.3

v1.0.2-beta.1

SajiCode v1.0.2 (Beta) 🚀

Welcome to the first beta release of SajiCode! We're bringing the AI engineering team directly into your terminal.

🌟 What's New

• Multi-Agent System: Replaced the traditional single-agent approach with a distributed team of 17 specialized agents (PM, Backend Lead, Frontend Lead, QA, Security, etc.) working in parallel.
• Agent Skills Engine: Introduced 21 expert knowledge files containing specific capabilities (from advanced React/Next.js to Database design, Python engineering, and more). Agents load skills on-demand for specialized tasks.
• Rich Project Context: Powerful context and memory management. The system now thoroughly understands your workspace through codebase intelligence scanning and maintains persistent memory across sessions.
• Global NPM Installation: SajiCode is now fully available on NPM! You can install it globally via npm install -g sajicode and run it from any project directory using the sajicode command.

📦 Installation

npm install -g sajicode

sajicode -p <provider> -m <model>

**Full Changelog**: https://github.com/RaheesAhmed/SajiCode/compare/v27...v1.0.2-beta.1

Release v27

Changes in this release:
fix: update README for sub-agent documentation and configuration details

Full Changelog: v27...HEAD

Full Changelog: v26...v27

Release v26

Changes in this release:
feat: readme created for Docs

Full Changelog: v26...HEAD

Full Changelog: v20...v26

Release v20

Changes in this release:
feat: Introduce project analysis tools and file management utilities

• Implemented project analysis tools for dynamic dependency graphs and file relationships, including:
  • analyzeProjectTool: Analyzes project structure and builds a dependency graph.
  • getFileInfoTool: Retrieves detailed information about a specific file.
  • getFileDependenciesTool: Lists all files that a specific file depends on.
  • getFileDependentsTool: Lists all files that depend on a specific file.
  • getFolderContentsTool: Retrieves detailed contents of a folder.
  • getProjectOverviewTool: Provides a high-level overview of the project.
  • getRelatedFilesTool: Finds all files related to a specific file.
  • findFilesByExportTool: Locates files that export a specific function, class, or variable.
  • findFilesByImportTool: Locates files that import from a specific module.
  • clearProjectCacheTool: Clears the project analysis cache.
• Defined type interfaces for results and errors for better type safety and clarity.

Full Changelog: v20...HEAD

Full Changelog: v19...v20

Release v19

Changes in this release:
feat: screenshot added and readme updated

Full Changelog: v19...HEAD

Full Changelog: v8...v19