Dipralix v0.3.2 — Alive

Dipralix v0.3.2 — Alive

Dipralix becomes the developer's mirror: it carries an identity, remembers across sessions, and resumes one ongoing relationship instead of starting from zero.

Highlights

• Comes alive on first run — give it a nickname and tell it how to be; it researches an operating approach, you approve it, and it carries that identity into every session.
• Cross-session memory — on every start it reads the verified-outcome ledger, memory, and prior sessions and tells you where you left off. No relearning, no wasted tokens.
• More connectors — Slack and Notion join GitHub, Discord, Gmail, Drive, and MCP. /connect shows what's wired.
• Self-evolution — /evolve folds token-budgeted industry research into memory, so running Dipralix over time compounds.

New commands

/alive · /resume · /evolve · /connect · /ledger

Install — every platform (auto-detected)

macOS (Intel + Apple Silicon) · Linux (x86_64 + ARM64) · Windows (x86_64 + ARM64)

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/Zyferon/dipralix/main/install.sh | bash

# Windows
irm https://raw.githubusercontent.com/Zyferon/dipralix/main/install.ps1 | iex

Verified

168 tests pass · cargo clippy clean · cargo fmt --check clean.

Full changelog: see CHANGELOG.md.

Dipralix v0.3.1 — Verified Outcome Ledger + proof-of-work

Added

• Verified Outcome Ledger (src/ledger.rs): append-only, per-repo log at .dipralix/ledger/outcomes.jsonl recording each task's verdict, calibrated confidence, files touched, and machine-checked proof. Temporal supersession when a past fact stops holding. Injected into the system prompt each session.
• Proof-of-Work contract: an outcome is only verified when its verification command exited 0.
• record_outcome tool (core tools 17 → 18) and /ledger view command.
• v3 system prompt: persist / verify / be-honest, explicit tool-routing decision tree, hard verification gates, prompt-injection resistance.

Fixed

• Install on Intel Macs: the macos-x86_64 binary now builds (cross-compiled on the arm64 runner; the retired macos-13 runner had stalled, so the asset was missing and install.sh 404'd).

Verified

159 tests pass; clippy, cargo audit, and cargo deny all clean.

v0.3.0 — realtime team sync (E2E encrypted, serverless mesh)

dipralix v0.3.0 — realtime team sync

A multi-model terminal coding agent (Gemini, Claude, OpenAI) that now syncs
your AI context across a team — end-to-end encrypted, with a serverless
LAN mesh or a self-hosted relay. Two small Rust binaries, no Node, no Python,
no telemetry, no login server.

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/pratikacharya1234/dipralix/main/install.sh | bash

# Two devs, same network, no server — encrypted P2P sync of .dipralix/
dipralix-cli --sync --mesh --room myproject --secret "team-passphrase" --user alice

● What's new

• Realtime team sync. Share .dipralix/ — memory, plans, skills, approval
  policy — live across terminals. Your source code never leaves your machine:
  an allowlist rejects source, .env, secrets, and keys before the wire.
• Serverless P2P mesh. Peers find each other over mDNS on the LAN
  (_dipralix._tcp.local.) and sync over direct TCP. No server, no STUN, no TURN.
• End-to-end encryption. Noise NNpsk0 (X25519 + ChaCha20-Poly1305) on every
  link — ephemeral per-session keys (forward secrecy), the room secret as the
  pre-shared key (mutual auth). A captured frame is ciphertext.
• Self-hosted relay (dipralix-server) for cross-network teams: JWT-per-room
  auth, SQLite persistence, replay-on-reconnect.
• Team UX: presence/heartbeat, append-only terminal chat, and a 2-of-N
  approval quorum for high-risk commands.

◆ Why it's different

Self-hosted and zero-trust by design. If dipralix disappears tomorrow, your
team's AI memory still lives in .dipralix/ on your machines and your server.
You adopt it for the collaboration, not the brand.

▲ Honest scope

• Mesh is LAN-only (mDNS is link-local). Across the internet, use the relay
  server or --peer host:port. WebRTC/NAT-traversal is intentionally not in this
  build.
• No CRDT yet — concurrent edits to one file are last-write-wins by content
  hash, not a character-level merge.

□ Quality

155 tests pass — including a two-node mesh that converges a real file over
encrypted TCP. clippy -D warnings, cargo fmt --check, cargo audit, and
cargo deny check all clean.

Install

Platform	Command
macOS / Linux	curl -fsSL https://raw.githubusercontent.com/pratikacharya1234/dipralix/main/install.sh | bash
Windows	irm https://raw.githubusercontent.com/pratikacharya1234/dipralix/main/install.ps1 | iex
Source	git clone https://github.com/pratikacharya1234/dipralix && cd dipralix && cargo build --release

Prebuilt binaries for Linux x86_64, macOS x86_64/arm64, and Windows x86_64 are
attached below (both dipralix-cli and dipralix-server).

Full changelog: see CHANGELOG.md.

Dipralix v0.1.0

Breaking

• Renamed FORGE → Dipralix. Binary is now dipralix-cli. Config directory .forge/ → .dipralix/. Env var FORGE_API_KEY → DIPRALIX_API_KEY (legacy GEMINI_API_KEY still honored). CI workflow renamed .github/workflows/dipralix-ci.yml. History file moved to ~/.dipralix-history.

Added — Phase 2A: Core Intelligence

• Memory Core (src/memory.rs): Persistent project decisions in .dipralix/memory/decisions.md and cross-project patterns in ~/.dipralix/patterns/. New AI tools memorize_decision and memorize_pattern. Memory injected into the system prompt.
• Lazy Context (src/context.rs): ContextAssembler dynamically loads skills from .dipralix/skills/ (project) and ~/.dipralix/skills/ (global) based on detected ProjectDna. Reduces wasted tokens on irrelevant context.
• Peer Review Engine (src/debate.rs): Multi-model Red Team vs Blue Team debate on high-risk bash commands (Confirm/Deny). Arbitrator decides; rejections short-circuit execution and feed back to the main agent.

Added — Phase 2B: Developer Experience

• Living Documentation (src/living_docs.rs, /docs sync): Auto-syncs ARCHITECTURE.md against the current codebase using the configured backend.
• Comment Protocol (src/comment_protocol.rs, /tasks): Scans the workspace for DIPRALIX: directives in source comments across Rust, JS/TS, Python, Go, YAML, TOML, MD. Commands: /tasks list, /tasks execute N, /tasks dismiss N (rewrites the marker to DIPRALIX-DONE:).
• Plan Visualizer (src/plan_visualizer.rs, /plan): Terminal-native ASCII plan view with status glyphs, dependency annotations, risk badges (safe/review/danger), and a progress bar. Parses .dipralix/plans/current.md.
• Code Fingerprinting (src/fingerprint.rs): New dipralix --init scaffolds .dipralix/{project,conventions}.md, safety.toml, and approval.toml from detected stack. New dipralix --fingerprint and /fingerprint print a quality score (0–100) with prioritized suggestions.

Added — Phase 2C: Power User Features

• Approval Matrix (src/approval.rs, /approval): Four-level per-action approval policy (Auto / Notify / Confirm / Deny) loaded from .dipralix/approval.toml. Sub-classifies bash by command shape (sudo, rm, git push, docker run, curl). /approval speed fast and /approval speed safe flip the matrix in bulk.
• Infra Awareness (src/infra.rs, /infra): Static analysis for Dockerfile (latest tag, root user, ENV secrets, single-stage), Kubernetes manifests (resource limits, probes, runAsNonRoot, privileged), and Terraform (open ingress, public-read ACL, encryption disabled). Pure text scan — no cloud API calls.
• Browser Engine, lite (src/browser.rs, /fetch <url>): reqwest-based fetcher with simple HTML→Markdown extraction and on-disk cache at ~/.dipralix/cache/web/. Headless Chromium deferred to a later release.

Changed

• dipralix init now creates .dipralix/memory/, .dipralix/skills/, and .dipralix/plans/ alongside the config scaffolding.
• Stale core_tool_count_is_correct assertion updated from 14 to 17 to reflect the new memory tools.

Fixed

• /docs handler compile error — now constructs a local BackendClient via active_config(...) instead of referencing an out-of-scope client.
• Cleaned unused imports in context.rs, debate.rs, living_docs.rs.

FORGE v0.0.2 — NULLVOID UI · Domain Bootstrap · Production Hardening

FORGE v0.0.2 — NULLVOID · Domain Bootstrap · Zero Emoji

What's New

🧬 Domain Bootstrap — 15 domain blueprints with real-time web search
Pick what you're building (mobile, web, AI, blockchain, cybersecurity...) and FORGE pre-loads the latest tech stack, architecture, security patterns, and best practices. Custom domain input for anything else. DuckDuckGo search fetches current recommendations live.

🎨 NULLVOID Terminal Theme
Spectral ghost logo with 5-color FORGE gradient. Protocol header with process info. Unicode geometric glyphs — zero emoji anywhere. Box-drawn response blocks, amber streaming thoughts, pipe-separated stats bar. Phantom-hacker aesthetic.

🛡️ Production Hardening

• 11 emoji instances purged, replaced with nullvoid glyphs
• EMBER voice AI disabled for v0.0.2 (under dev, shipping v0.0.3)
• CI build fixed (ALSA optional, no API key = graceful skip)
• diff_view match-arm case bug fixed
• 28 clippy auto-fixes applied
• Branch protection on main
• 0 warnings, 28/28 tests passing

Install

curl -fsSL https://raw.githubusercontent.com/pratikacharya1234/forge/main/install.sh | bash

Quick Start

export GEMINI_API_KEY='***'
forge-cli

Full Changelog

FORGE v0.0.1 — First Public Release

Changelog

All notable changes to FORGE are documented in this file.

[0.0.1] — 2026-04-28

First Public Release

FORGE v0.0.1 — the open-source, multi-model terminal coding agent. 1M token context. Free. Previously developed as GeminiX, rebranded to FORGE to reflect universal multi-model support and independence from any single AI provider.

Multi-Model Support

• Gemini 2.5 Pro/Flash/Lite, 2.0 Flash
• Claude 4 Opus/Sonnet, Claude 3.5 Sonnet (Anthropic API)
• GPT-4.1, GPT-4o, o3, o4-mini (OpenAI API)
• Backend abstraction layer with canonical message format
• SSE streaming per provider with tool call round-trips
• Provider auto-detection from model name

Auto Model Routing

• /model auto analyzes task complexity and picks the best model
• Complex (refactor, architecture, security) → reasoning model
• Simple (find, read, explain) → fast/cheap model
• Everything else → balanced model
• Provider-aware: uses available API keys intelligently

Explain-Before-Execute

• /explain on|off toggles pre-execution summaries
• Shows planned file changes, shell commands, expected impact
• Asks for confirmation before running tools
• --explain CLI flag

Test-Fix Loop

• /test-fix <command> [max-cycles]
• Runs tests, detects failures, feeds errors to model
• Model fixes code, re-tests, loops until passing
• Handles truncated output for large test suites

Persistent Memory

• /memorize <fact> saves to .forge/memory.md
• /forget <keyword> removes matching entries
• /memory displays all memorized facts
• Auto-injected into system prompt each turn

Agentic Loop

• Streaming Gemini API with real-time token display
• Thinking/reasoning token visualization (Gemini 2.5+)
• Parallel tool execution via Tokio
• Configurable iteration limits with pause/resume
• Auto-apply mode for non-interactive use
• Single-prompt mode for scripting

Built-in Tools (16)
read_file, write_file, edit_file (fuzzy matching + occurrence parameter), append_file, bash (streaming + safety classification), list_files, search_files (regex), glob, create_directory, delete_file, move_file, copy_file, url_fetch (cached), git_snapshot

Safety System

• 4-level risk classification: Allow, Warn, Confirm, Deny
• Pipe-to-shell detection and blocking
• Per-project .forge/safety.toml with category-level overrides
• Trusted/blocked command lists

Diff & Undo

• Unified diff preview before file writes
• Per-hunk interactive review (accept/reject/skip per change)
• In-memory undo stack with /undo and /undo N
• Git snapshot creation and rollback

Context Management

• Token usage display per turn (prompt + output + thinking)
• Context window progress bar with configurable warnings
• Auto-compaction at threshold (summarizes history via model)
• Manual /compact command

Session Persistence

• Binary save/restore of full conversation history
• /session save, load, list, delete
• Auto-save after each successful turn

Cost Tracking

• Per-model pricing with USD display
• Session cost accumulation
• Daily budget support with 80% warning

MCP Support

• Full JSON-RPC 2.0 MCP client over stdio
• Protocol 2025-03-26 compliance
• Auto-discovers tools from any MCP-compliant server
• Parallel server startup with timeout safety

Native Integrations (33 tools)

• GitHub: 12 tools (issues, PRs, repos, code search, branches)
• Discord: 7 tools (messages, channels, guilds, embeds)
• Google Drive: 7 tools (files, folders, upload, search)
• Gmail: 7 tools (send, list, search, labels, read status)

Named Profiles

• Configurable profiles in ~/.forge/config.toml
• Per-profile model, thinking, grounding, auto_apply, budget
• /profile command for switching

Security Sweep

• Cargo audit + npm audit integration
• Static secret scanning (API keys, tokens, passwords)
• Gemini-powered CVE analysis

Additional Commands

• /learn: Clone and load public git repos for Q&A
• /load: Load directory tree into context
• /screenshot: Vision-based bug analysis
• /pr: Auto-create pull requests
• /models: List available Gemini models
• /debug: Debug information dump
• /history: Show conversation history
• /cost: Show session cost breakdown