MakuluLinux
v2026.07.03-r8
Status Report — July 3 2026 · Electra Shell Roadmap
Project Status · Competitive Analysis · Roadmap

Electra AI — The Linux Jarvis

A fully integrated AI operating layer for Linux. One compiled binary, zero cloud dependencies, 80+ models, every OS action, a full IDE, an autonomous finance engine, and deep desktop integration that no SaaS competitor can replicate — because it requires owning the OS.

Competitive Scoring

Where Electra Stands vs. Competitors

Electra AI
98
Linux-native · Full OS control · 80+ models · IDE + Finance + Jarvis · Stable Coder
Cursor
91
Strong IDE · No OS integration · SaaS lock-in · macOS/Windows focus
Antigravity
84
Browser-based · Good UX · No local file access · Limited OS reach
Claude Code
82
Terminal-only · No GUI · No finance · No desktop integration
VS Code + Copilot
80
Great ecosystem · No AI routing · No OS control · Electron overhead
ChatGPT Desktop
71
Chat only · No code execution · No Linux integration · Cloud-only
OS & Desktop Integration
Electra
100
Cursor
15
Claude Code
20
Antigravity
8
VS Code
30
Coder / IDE Capability
Electra
98
Cursor
97
VS Code
90
Claude Code
80
Antigravity
72
Multi-Model AI Routing
Electra
100
Cursor
30
Antigravity
20
Claude Code
5
Autonomous Finance / Earn Online
Electra
100
Everyone else
0
Offline / Privacy Mode
Electra
95
VS Code
60
Claude Code
40
Cursor
20
Antigravity
5
Feature Comparison Matrix

What Electra Has vs. the Competition

Feature Electra Cursor Claude Code VS Code Antigravity
Multi-model routing (80+ models)~
Full GUI IDE (GTK3, GtkSourceView)~
Terminal / CLI mode~
Agentic tool-calling coder~
Structured planning pipeline (plan → approve → code → verify)~
Auto-compact with plan context restore
Agent Swarm (parallel workers)
LSP integration (pyright, tsserver…)
Ghost text inline completions~~
Streaming diff preview~
Open VSX extension support
AT-SPI screen awareness
D-Bus two-way desktop control
Voice input / TTS output
Offline / local models (Ollama)~
Telegram bridge
Home Assistant integration
Finance bot / earn online
Heartbeat background task agent
Discord / Reddit / RSS / Spotify agents
Git panel (built-in)~
Image generation & analysis
Project knowledge base (/pk)
ChromaDB vector memory
Patreon tier system
Single compiled binary (Nuitka)
Source code Q&A agent
Codeberg integration
Novel / long-form writing mode
Live Code Map (workspace awareness)~
Cinnamon / GNOME / KDE / XFCE support~
Right-click context actions (Nemo)

★ = Electra-unique or category leader · ✓ = present · ~ = partial · ✗ = absent

Shipped Features

What's Built & Working

🧠
Multi-Mode AI Routing
Routing agent analyzes every prompt and dispatches to Chat, Coder, Command, Writer, Novel, or Finance mode automatically. Intent-aware — picks the narrowest tool subset per task type (DEBUG / REFACTOR / QUESTION).
SHIPPED
🖥️
GUI IDE (electra_gui.py)
Full GTK3 + GtkSourceView IDE with file tree, tabbed editor, chat panel, activity log, git panel, diagnostics tab, minimap, image previews, split editor, voice input, command palette (Ctrl+Shift+P), quick file open (Ctrl+P), symbol breadcrumb, @-mention file context.
SHIPPED
🐝
Agent Swarm (Major/Commander)
Orchestrator decomposes task into per-file sub-tasks → up to 6 parallel worker agents each running 8-round agentic loops → Reviewer checks cross-file consistency and import conflicts. Live swarm status panel in GUI.
SHIPPED Major+
👁️
AT-SPI Screen Awareness
Reads the Linux Accessibility Tree of any focused app — browser, terminal, LibreOffice, Electron apps. Injects screen content into AI prompts automatically. No copy-paste needed. "What does this error mean?" just works.
SHIPPED UNIQUE
🔌
D-Bus Two-Way Desktop Control
Any app, script, or keyboard shortcut on the desktop can send queries to Electra via org.makululinux.Electra on the session D-Bus and receive responses. Methods: Ping, ExecuteIntent, GetContext, Notify, TileWindows.
SHIPPED UNIQUE
🧩
Open VSX Extension Browser
Browse, search, install, and uninstall .vsix extensions from Open VSX Registry. Auto-applies snippets, TextMate grammars, VS Code themes, and registers bundled LSP servers. Snippet palette + Tab expansion included.
SHIPPED
💰
Finance Bot (/finance)
Autonomous income streams: affiliate content generation, print-on-demand product creation & upload (Redbubble), freelance job scanner. SQLite ledger for earnings tracking. Telegram notifications. GUI center at /finance or /hustle.
SHIPPED UNIQUE
💓
Heartbeat Background Agent
Always-on task scheduler: scheduled and predictive tasks now run on real OpenAI-style function calling against /v1/chat/completions — the same mechanism interactive Command mode uses — instead of the old fragile [ACTION: arg] text-tag protocol. 21 tools: run_command, install_deps, snapshot/restore_packages, create_systemd_service, check_journal, git_command, kill_process, notify_user, and more, each gated by the existing shell blocklist. complete_task now requires a real verification result, same mandate as interactive Command mode. Tasks pulled from local SQLite DB or server. Telegram alerts on completion.
SHIPPED UNIQUE JULY 2
📱
Telegram Bridge
Full Telegram bot integration compiled into the binary. Bi-directional: send prompts from your phone, get AI responses back. Heartbeat agent sends notifications. Finance bot posts updates. Per-conversation history via conversation_id.
SHIPPED
🏠
Home Assistant Integration
Control smart home devices from chat. Natural language: "turn off the living room lights" → Home Assistant API call. Supports all HA entity types.
SHIPPED
🔊
Voice Input / TTS Output
Piper TTS (offline) for voice responses. Vosk / faster-whisper STT for voice input. edge-tts cloud fallback. Voice button in GUI chat panel. Complete offline voice pipeline.
SHIPPED
🔭
LSP Integration
pyright, tsserver, rust-analyzer, clangd, gopls all managed by electra_lsp.py. 4 AI tools: lsp_goto_definition, lsp_find_references, lsp_rename_symbol, lsp_hover. Bundles auto-detected from installed extensions.
SHIPPED
👻
Ghost Text Inline Completions
Dedicated fast-model chain (nemotron-nano → mis-mistral-small → qc-qwen-coder-flash → ms-qwen2.5-coder → llm7-deepseek-coder). Separate from main agent — never competes for tokens. Stale-result cancellation via monotonic seq counter.
SHIPPED
🗺️
Live Code Map
The AI knows exactly which files exist in the workspace and when they were last touched. Intent-aware tool routing based on task type (DEBUG/REFACTOR/QUESTION). Code map badge in activity log shows tracked file count.
SHIPPED
🔀
Streaming Diff Preview
write_file tool_calls intercepted in SSE loop before execution, pushing streaming_diff to GUI instantly. Diff tab opens while AI is still generating. Full hunk view. Auto-approve and reject buttons.
SHIPPED
🎨
Full Theme Isolation
Three-layer strategy: Adwaita base, priority-800 widget reset CSS, expanded universal reset block. Completely immune to system GTK theme (Breeze, Mint-Y, Arc, Yaru, etc.). Works identically on any Linux distro.
SHIPPED
📦
Single Nuitka Binary
Everything compiled into one self-contained executable: ai_terminal, electra_gui, all agents, TTS/STT, all services. Docker-based build (Ubuntu 22.04) for maximum distro compatibility. Distributed via GitHub.
SHIPPED UNIQUE
🐙
GitHub + Codeberg Agents
github_agent.py: repo management, issues, PRs, releases via natural language. codeberg_agent.py: Forgejo/Gitea API integration for the privacy-first Git alternative. Both wired to /github and /codeberg slash commands.
SHIPPED
🌐
Multi-Service Integrations
Discord bot (discord.py 2.x, slash commands, per-user history), Reddit agent, RSS agent, Spotify agent, Google services agent (gmail, calendar, drive), ISO agent for MakuluLinux ISO management.
SHIPPED
🧬
ChromaDB Vector Memory
Server-side MemPalace via ChromaDB for persistent semantic memory. Local .electra_memory.md for quick facts. Per-project .electra_project.md knowledge base with /pk commands. Stable conversation_id per project session.
SHIPPED
💬
Source Code Q&A Agent
/v1/source/ask FastAPI endpoint + floating chat widget on electra.html. Uses StepFun → NVIDIA NIM → LLM7 fallback chain. Queries private GitHub repo intelligently. Never reveals actual source code or API keys.
SHIPPED
⌨️
Electra Bar Widget
Floating Super+E hotkey bar on Cinnamon desktop. User types a prompt, it's forwarded to the AI terminal routing agent. Supports all modes. System-wide access without opening the full IDE.
SHIPPED UNIQUE
🔐
Patreon Tier System
Free (50/day) → Private (500) → Corporal (1500) → Sergeant (4000) → Major (10000, pays ds-/bt-/ol-) → Commander (∞). Tier-gated features. Live request counter. Upgrade prompts. GUI model picker enforces locks.
SHIPPED
🏗
Unified Coder — Intelligent Routing
Single coder mode replaces the old vibe-coder/architect split. Intent classifier routes every message automatically: questions get conversational replies, simple tasks get direct implementation, complex or architectural tasks auto-trigger the full PRE_FLIGHT → ARCHITECT → DEVELOPER → REVIEWER pipeline. No button, no slash command — it just does the right thing. GUI and terminal now identical.
SHIPPED JUNE 29
📋
Structured Planning Pipeline
PRE_FLIGHT asks 1–3 clarifying questions. ARCHITECT explores (read limit 12), outputs a discovery summary, then writes PROJECT_PLAN.md — a full plan listing files to create/modify/leave alone, edge cases, test strategy, and rollback plan. User approves or rejects with feedback. DEVELOPER implements step-by-step with narrated progress. REVIEWER runs tests + lint + git diff + security scan before sign-off. Plan survives context compaction.
SHIPPED JUNE 29
🩺
Self-Healing Stability System
50+ bugs fixed June 2026. _sanitize_messages() before every API call. _safe_append_user() prevents 400 errors. Auto-compact at 78% context with plan-context restore. CODER_FALLBACK_CHAIN rotates models on stream failure. AST validation on all patches. 14 critical coder pipeline bugs fixed in June 29 sprint.
SHIPPED
🚀
Background Task Queue — /bg
"Kick off X, keep chatting with me." Command mode can now hand a long task to its own detached process (ai_terminal.bin --command "<task>" --detached --bg-task-id <id>) instead of blocking the conversation — same self-relaunch pattern already used by /cmdide, so no in-process threading around ai_terminal.py's global state, zero race conditions. Own command_history, own conversation_id, full tool + verification mandate. confirm_dangerous/plan_task/ask_user already fail-safe on a closed stdin (dangerous ops auto-cancel, plans auto-approve with an audit trail + notification). New tool: run_in_background_task — the AI can decide to background something itself. Commands: /bg <task>, /tasks, /task show <id>, /task kill <id>.
SHIPPED UNIQUE JULY 2
📋
System Changelog — /changelog
"What have you changed on my system this week?" Command mode was already writing a timestamped Markdown entry to ~/.electra/command_log.md on every complete_task and auto-approved plan — summary, commands run, files modified — it just had no read path back. New query_changelog tool + /changelog [today|week|month|N|all] command read and filter that existing audit log by time window, so the agent answers from real tracked data instead of guessing from memory. Zero new logging code — pure query surface over data Electra was already collecting.
SHIPPED JULY 2
AI Provider Network

80+ Models Across 11 Providers

Free tier models available
Composite 60% capability + 40% speed routing
sf- prefix permanently banned from auto-routing
step- (StepFun)
StepFun AI
Primary coder/command/chat backbone. step-3.5-flash most stable. Highest cross-run reliability score.
FREE
qc- (QwenCloud)
Alibaba DashScope
2 keys, 600–15000 RPM. Web search + function calling. qwen3-coder models for coding tasks.
FREE
mis- (Mistral Direct)
Mistral AI
5 keys × 2000 req/day = 10,000/day budget. mis-mistral-large promoted for writer/novel modes.
FREE
cb- (Cerebras)
Cerebras Wafer-Scale
Ultra-fast inference (~3000 tok/s). cb-gpt-oss-120b and cb-zai-glm-4.7. Priority Pass 1 for command mode.
FREE
llm7- (LLM7)
LLM7 Community
Free inference endpoint. GPT-o4-mini (reasoning), deepseek-coder-v2-lite for ghost text chain.
FREE
ms- (ModelScope)
Alibaba ModelScope
1 RPS cap. Excluded from auto-routing. User-selectable only. Deepseek variants available.
FREE
opr- (OpenRouter)
OpenRouter
~1900 req/day across 2 keys. User-selectable only (daily-limit issues). Access to many frontier models.
FREE
ol- (Ollama Cloud)
Ollama Cloud Hosted
20+ large models (Qwen3-Coder-480B, Nemotron-Ultra, Devstral, MiniMax, Kimi-K2, etc.). Free hosted inference.
Major+
ds- (DeepSeek Direct)
DeepSeek Direct API
Paid per-token. DeepSeek V4-Pro, V4-Flash, V3.2, R1-0528. Top-priority in paid-tier routing pre-pass.
Major+
bt- (ByteDance Ark)
ByteDance Ark
Paid per-token. Access to ByteDance frontier models. Part of paid-tier routing pre-pass for Major/Commander.
Major+
sf- (SiliconFlow)
SiliconFlow
Permanently banned from auto-routing. User-selectable only. Unreliable cross-run stability. Still available manually.
MANUAL ONLY
Specialist Agents

The Agent Network

_travel_agent
Travel planning, itineraries, flight/hotel research via web search tools.
_app_agent
Install, remove, update, and manage desktop applications via package managers.
_nvidia_agent
GPU management, driver info, CUDA queries, nvidia-smi interaction.
_troubleshoot_agent
System diagnostics, log analysis, error resolution, step-by-step troubleshooting.
_docker_agent
Docker container management: build, run, stop, inspect, compose operations.
_service_agent
systemd service management: start, stop, enable, disable, status, logs.
_pkg_agent
Smart package management across apt/dnf/pacman with dependency resolution.
_schedule_agent
Cron-like task scheduling via Heartbeat agent. Add, remove, list scheduled tasks.
_weather_agent
Current conditions, forecasts, and weather-related queries via web integration.
_wallpaper_agent
Set, rotate, and manage desktop wallpapers. Integration with Cinnamon/GNOME settings.
_image_gen_agent
Generate images via Pollinations.ai and other providers. Analyze images via multimodal models.
_audio_agent
Audio playback control, Spotify integration, now-playing display via spotify_agent.py.
_ocr_agent
Extract text from images and screenshots using OCR. Integrates with AT-SPI screen context.
_screenshot_agent
Capture full desktop, window, or region screenshots. Auto-injects into AI context.
_clip_agent
Clipboard read/write operations. Context injection from clipboard content.
_ssh_agent
Remote server management via SSH. Key management, connection profiles, remote command execution.
_monitor_agent
Display configuration: resolution, refresh rate, multi-monitor arrangement via xrandr/wlr-randr.
_qr_agent
Generate QR codes from URLs, text, or contact info. Display inline or save to file.
_plugin_forge_agent
Auto-generates plugins on error encounters. Classifies errors, attempts auto-fix, files GitHub issues.
_pixabay_video_agent
Search and download stock video footage from Pixabay for content creation.
source_code_agent
Q&A about Electra itself. Queries private GitHub repo via /v1/source/ask. Never reveals source or keys.
github_agent / codeberg_agent
Full repo management for both platforms. Issues, PRs, releases, notifications, search — via natural language.
Subscription Tiers

Patreon Access Tiers

Tier Daily Requests Premium Providers Agent Swarm Priority
Free50 / dayStandard poolBase
Private500 / dayStandard poolBase
Corporal1,500 / dayStandard poolMedium
Sergeant4,000 / dayStandard poolHigh
Major — $40/mo10,000 / dayds- bt- ol- unlockedTop
Commander — $100/moUnlimited ∞ds- bt- ol- unlockedTop

Major+ subscribers get paid providers auto-prioritized in Coder and Command mode routing. GUI model picker sorts premium models to top.

Linux Distribution Support

Supported Platforms

Debian / Ubuntu / Mint / Pop!_OS / Zorin / Elementary / Kali / MakuluLinux / MX Linux
Fedora / RHEL / CentOS / Rocky / AlmaLinux / Nobara / Ultramarine
Arch / Manjaro / EndeavourOS / ArcoLinux / Garuda / CachyOS
Cinnamon · GNOME · KDE/Plasma · XFCE · MATE · LXQt · Budgie
X11 & Wayland supported
Compiled in Ubuntu 22.04 Docker → maximum glibc compatibility

The binary targets the oldest practical glibc (2.34 via Ubuntu 22.04 base) so it runs on essentially any distro released since 2021. MakuluLinux (Cinnamon on Ubuntu 24.04) is the primary development and showcase platform with the deepest integration (right-click context actions, Super+E bar, AT-SPI, D-Bus).

Development Roadmap

The Path to Full Jarvis

Foundation Complete
Terminal AI Center — base chat/coder/command routing SHIPPED
Core routing agent, multi-mode dispatch, Telegram bridge, binary compilation pipeline, Patreon auth.
Multi-provider model routing (11 providers, 80+ models) SHIPPED
Composite 60%/40% speed+capability scoring, CODER_FALLBACK_CHAIN, ban lists, tier-gated providers.
Specialist agent network (20+ agents) SHIPPED
App, Docker, SSH, weather, wallpaper, OCR, screenshot, clip, schedule, service, pkg, monitor, travel, and more.
Finance Bot & Heartbeat Agent SHIPPED
Affiliate content, print-on-demand, freelance scanner. Background task scheduler with Telegram alerts.
Service integrations — Discord, Reddit, GitHub, Spotify, RSS, Home Assistant, Google SHIPPED
Phase 1 Complete
GUI IDE (electra_gui.py) — full launch SHIPPED
GTK3 + GtkSourceView + WebKit2. File tree, tabbed editor, chat panel, activity log, minimap, git panel, diagnostics tab, voice input, image preview, split editor.
Command Palette, Quick File Open, @-mention context SHIPPED
Workspace session restore, symbol breadcrumb, deferred tree indexing SHIPPED
Light Ink default theme + multi-theme system SHIPPED
SSL/TLS cert repair for Nuitka onefile — never deletes /tmp/electra SHIPPED
Phase 2 Complete
LSP integration — pyright, tsserver, rust-analyzer, clangd, gopls SHIPPED
Ghost text inline completions with stale-cancellation SHIPPED
Agent Swarm (Major/Commander) — parallel multi-file editing SHIPPED
Streaming diff preview (Cursor-style) SHIPPED
Open-tab context auto-injection SHIPPED
Test generation button (Ctrl+Shift+G) SHIPPED
Inline Edit (Ctrl+K) — complete rewrite with ±30 line context SHIPPED
Filesystem tools full GUI integration (delete, move, mkdir, copy) SHIPPED
Token efficiency — LRU cache, open-tab fingerprint, pre-flight estimator SHIPPED
Phase 3 Complete
AT-SPI Screen Awareness — reads any focused app's accessibility tree SHIPPED HIGH
D-Bus Two-Way Desktop Control — org.makululinux.Electra on session bus SHIPPED HIGH
Open VSX Extension Browser — install .vsix, themes, snippets, LSP servers SHIPPED HIGH
Full GTK theme isolation — immune to any system GTK theme SHIPPED MED
Patreon tier enforcement in GUI model picker + terminal SHIPPED
Live Code Map + structured error parsing + intent-aware tool routing SHIPPED HIGH
Next-Step Suggestions — 3 clickable post-task actions in GUI SHIPPED
Command Mode GUI — dynamic 3-pane layout with execution log pane SHIPPED
Deep stability hardening — 40+ bugs fixed, _sanitize_messages, _safe_append_user, depth sentinels SHIPPED HIGH
Codeberg agent integration (Forgejo/Gitea API) SHIPPED
Source Code Q&A agent (/v1/source/ask) with web chat widget SHIPPED
Discord bot rewrite — discord.py 2.x, slash commands, per-user conversation_id SHIPPED
Phase 4 NOW ACTIVE
OVERDRIVE — Small Model Boost System SHIPPED HIGH
Client-side prompt engineering layer that automatically activates for models ≤35B effective parameters. Five layers: (1) auto-detection from MODEL_DISPLAY_INFO params tag — zero maintenance, (2) temperature auto-tune (0.25 for tool-calling/code, up to 0.70 for creative), (3) format anchor appended to end of system prompt exploiting recency bias, (4) think-before-answer CoT prefix for TASK intent, (5) context budget trim to model's effective attention window. Covers all modes (coder, command, chat, writer) and the offline 4B fallback. Toggle: /overdrive · Badge: ⚡ OD in GUI info bar.
True Loop Engineering — Self-Refinement & Reflection Loops SHIPPED HIGH
Elevates Electra from execution loops (AI runs tools until done) to reflection loops (AI evaluates its own output, rewrites its own prompts, and iterates before returning to the user). Inspired by Boris Cherny's (Anthropic / Claude Code) concept of "loop engineering" — the next wave where AI agents generate, refine and coordinate prompts autonomously.
WHAT WE ALREADY HAD (Current Loop Engineering):
Multi-turn agentic loops — coder & command agents run up to 320 turns (80 turns × 3 checkpoint segments), autonomously calling tools and reacting to results.
Multi-agent orchestrator — breaks complex prompts into agent-assigned steps (CODER/COMMAND/GITHUB/WRITER/DOCKER) and runs them in sequence with dependency injection.
Routing agent — classifies every user message and generates the correct routing decision automatically (chat / coder / command / writer).
Parallel swarm workers — spawns multiple coder sub-agents on different files simultaneously for large refactors.
GPS self-navigation — agents inject a per-turn GPS block summarising what's done and what's left, writing their own next-step plan.
Multi-intent splitting — compound prompts ("fix X and install Y") are split into sequential focused tasks automatically.
TRUE LOOP ENGINEERING — THREE SUBSYSTEMS + THREE GAP FIXES (June 2026):
Subsystem 1 — Critic Pass — after every coder/command task completes, a critic model scores the output 1-10 against the original request. If below threshold (7/10), the task is re-queued with a refined prompt incorporating the critic's specific objections. One refinement pass prevents infinite recursion. Gap A fix: critic now reads actual files written to disk (up to 3 files, 300 lines each) rather than only the agent's self-reported summary — evaluates real output.
Subsystem 2 — Prompt Decomposition — vague/complex requests are intercepted before routing, decomposed into 3–5 specific sub-tasks by a decomposer model, shown to the user for confirmation (with inline editing), then executed as focused agent calls. The AI writes its own prompts from the user's vague description. Gap B fix: two-stage detection — keyword pre-filter (free, no API call) + AI judge model that confirms the request is genuinely vague before decomposing. Eliminates false positives like "implement this specific 3-line fix."
Subsystem 3 — Dynamic Plan Revision — multi-agent orchestration plans are no longer static. When a step fails, the remaining plan is sent back to the planner model with full context of what succeeded and what failed. The plan is revised in-place before continuing.
Gap C — GUI Integration — all three subsystems now surface in the GUI activity log: decomposition plan with per-step rationale, auto-confirm in GUI mode (no terminal input required), per-step progress, critic score badge, and completion status. GUI users see everything terminal users see.
⚙️ Toggle: /loopeng on|off · individual subsystems: /loopeng critic on|off · /loopeng decomp on|off · /loopeng replan on|off
LSP autocomplete popup (symbol dropdown) SHIPPED HIGH
Wire GtkSourceView CompletionProvider to LSP completions endpoint. Real method names, signatures, and docstrings from pyright/tsserver as the user types — not AI guesses. LSP manager already running; the bridge to GtkSourceView's CompletionProvider API is the missing piece.
Agent plan approval UI (before coding starts) SHIPPED HIGH
Implemented as the structured planning pipeline (June 29 2026). ARCHITECT writes PROJECT_PLAN.md listing every file to create/modify/leave alone, then calls request_approval — a GUI dialog shows the full plan with a feedback text entry. User approves or rejects with typed notes. AI revises if rejected. Only after approval does DEVELOPER start implementation. More powerful than the original spec: the AI plans autonomously before presenting, not just showing a file list.
Unified Coder — Complete Overhaul & Stable Release SHIPPED HIGH
June 29 2026. Coder mode declared stable for the first time since initial design. Complete rebuild of the planning pipeline across patches r1–r5 (r1: PRE_FLIGHT phase + ARCHITECT overhaul; r2: discovery summary fix + transition guard; r3: GUI pipeline fix + 5 features; r4: 10 critical bug fixes; r5: Unified Coder mode).

Key deliverables: PRE_FLIGHT clarifying questions before any file access · ARCHITECT_READ_LIMIT=12 with discovery summary turn · PROJECT_PLAN.md as persistent plan artifact · GUI approval dialog with rejection feedback · DEVELOPER step narration · REVIEWER with mandatory tests+lint+diff+security · Auto-compact plan context restore · REVIEWER tool_choice=required enforcement · 14 bug fixes including the critical GUI MESSAGE intercept bug (plan pipeline was silently bypassed in GUI mode) · Unified Coder mode: single intelligent router replaces vibe-coder/architect split — questions route conversationally, simple tasks code directly, complex tasks auto-trigger full pipeline.

Result: GUI and terminal now identical. No manual mode switching. The AI decides automatically based on what you typed.
GitHub Actions / CI integration SHIPPED HIGH
TRUE LOOP ENGINEERING Subsystem 4. Heartbeat agent polls GitHub Actions every 5 minutes for failed runs across all watched repos. On new failure: fetches the run logs, calls the AI to diagnose the root cause (structured 4-part diagnosis: cause, files, fix, confidence), sends a Telegram alert with the analysis. With --auto-fix: dispatches the coder agent to implement the fix autonomously. Commands: /ci watch owner/repo [--auto-fix] [--chat <id>] · /ci status · /ci unwatch · /ci fix owner/repo <run_id>. State persisted in ~/.electra/ci_watch.json and ci_state.json.
Issue-to-PR automation SHIPPED HIGH
TRUE LOOP ENGINEERING Subsystem 5. Electra reads a GitHub issue, plans a fix (branch name, PR title, coder instructions, confidence level), creates the branch, runs the coder agent to implement the fix, and opens a PR — fully autonomous end-to-end. Low-confidence issues (design questions, needs-info) are skipped with explanation. Commands: /gh fix issue <number> (specific) or /gh fix issue (interactive list). Telegram notification at each stage. New GitHub API primitives added: get_issue(), create_branch(), create_pull_request(), list_open_issues(), get_run_logs().
Command Center Upgrade — Background Task Queue, System Changelog, Heartbeat Real Tool-Calling SHIPPED HIGH
July 2 2026. Command mode audit identified that the "control your PC" half of Jarvis was capable but had three concrete gaps versus Coder mode's polish: scheduled/predictive tasks ran on a weaker tool loop than interactive tasks, everything was single-threaded and blocking, and there was no way to query what Electra had actually changed on the system. All three closed this sprint:

Heartbeat Real Tool-Calling — heartbeat_agent.py's ReAct loop replaced the old [ACTION: arg] regex-tag protocol with real OpenAI-style function calling (confirmed against the live backend: non-streamed tool_calls work correctly with mode="none", though the model leaks <think> into content even alongside a tool call — stripped same as elsewhere). 21 tools, same verification mandate as interactive Command mode.

Background Task Queue/bg <task> and the run_in_background_task tool spawn a task as its own detached process via the existing --command --detached CLI path (same self-relaunch pattern as /cmdide), avoiding any in-process threading around ai_terminal.py's non-thread-safe globals. /tasks, /task show <id>, /task kill <id> round out the surface.

System Changelog/changelog [today|week|month|N|all] and the query_changelog tool read back the Markdown audit log (~/.electra/command_log.md) Command mode was already writing on every completed task, filtered by time window.

All three delivered as anchor-validated patch scripts (backup + ast.parse + py_compile before write), test-applied against a sandbox copy before hand-off.
BREAKTHROUGH FEATURES SHIPPED THIS SPRINT (relocated from the former Phase 5 slot, now reassigned to Electra Shell):
Project Brain — Persistent Semantic Memory SHIPPED HIGH
The compounding moat. A background agent runs after every coder session and maintains ~/.electra_project_brain.md: every file and what it does (updated on write), every architectural decision and why it was made, every bug fixed and how, user preferences learned over time. The coder, command, and loop engineering agents all read this document — dramatically improving context accuracy with zero extra API calls. No SaaS AI tool can replicate this because they have no persistent local storage.
Predictive Heartbeat — Electra Does Things Before You Ask SHIPPED HIGH
From reactive assistant to proactive digital employee. The heartbeat agent learns weekly start-of-day patterns per weekday from real activity history and auto-schedules a prep task the night before. An end-of-day commit nudge fires once per day if the user was active but never committed. Before every git push, Electra checks that branch's real CI failure history and prints an advisory warning if it has failed repeatedly in the last 21 days. Command: /predictive status|analyze|on|off.
Visual Feedback Loop — Screenshot → Act → Verify SHIPPED HIGH
The see-act-verify loop, closed. /vloop <goal> captures a baseline screenshot, dispatches a real fix attempt to the coder agent, takes an independent "after" screenshot, and asks a vision model to reply VERIFIED or NOT_VERIFIED with a reason — feeding that reason into the next attempt automatically, up to 3 tries. Built on scrot/gnome-screenshot plus vision models, fully native to the Linux desktop.
/me — AI Self-Knowledge Profile SHIPPED MED
Total transparency into what Electra has learned about you. /me gathers explicit /profile preferences, server-side and local MemPalace excerpts, Predictive Heartbeat's learned activity patterns, scheduled automations, and account tier, then hands it to the AI to write a warm, plain-English, second-person summary — never raw data, file paths, or internal jargon. Credential-pattern scrubbing runs before anything reaches a prompt.
Electra Pets — Living AI Desktop Companion SHIPPED HIGH
A small animated AI companion (electra_pet.py, ~2,800 lines, standalone Nuitka binary) living on the desktop as a transparent always-clickable window. Seven pets, four hand-built vector shapes, zero image assets. Persistent happiness/energy stats, real weather awareness, live system awareness, cursor awareness, seasonal decorations, 25+ animation states. Chat input forwards to a full ai_terminal.bin instance via the same terminal-detection and --chat routing electra_bar.py uses. push_pet_event() gives every other agent a one-line hook to make the pet react to real system events — wired into Heartbeat and GitHub agents. This is the clearest existing proof that Electra Shell (Phase 5) is viable — a native, always-on desktop citizen already ships and works today, just not yet as the actual window manager layer.
Phase 5 NOW ACTIVE ✦ ELECTRA SHELL
ARCHITECTURAL CONCEPT
✦ INTERACTIVE CANVAS DESKTOP
A traditional desktop is a pretty wallpaper, a grid of icons, a panel, and a menu — and outside those, the entire screen does nothing. The desktop should be one thing: a giant interactive canvas that the AI can turn into anything on request. No wallpaper, no desktop icons, no panel, no menu — the wasted space becomes the product. Ask for a live chessboard with pieces you can actually move, and it becomes that. Ask for a Stardew-Valley-style NPC farmer living out its day in real time, and it becomes that. Ask for a plain-English live system log — "you just opened Firefox, that error over there is a failed apt update, here's what happened" — and it becomes that. Ask for a categorized folder-style app menu, an analogue clock, or nothing but a clean wallpaper, and it becomes that instead. Whatever's on screen is just the currently-loaded canvas mode — swappable in real time, in response to a single prompt.

The scoping decision that makes it viable: don't write a compositor. Electra Shell rides on top of whatever compositor is already installed (Muffin on Cinnamon/MakuluLinux, or Openbox+picom as a generic X11 fallback) via standard EWMH/D-Bus. What replaces the panel + app-menu parser + Nemo desktop-icon layer is a single WebKit2GTK canvas — HTML/CSS/JS/<canvas>, not hand-built GTK widgets, because HTML/JS is exactly the medium the AI is already good at generating on demand, and electra_gui.py already carries WebKit2GTK as a dependency. A slim auto-hide dock (Mac-Dock-style, position-configurable) replaces the panel and app-menu entirely, and electra_bar.py's floating-widget role folds into it as the dock's AI input field.

The staging decision that makes it safe: this is the highest-stakes feature on the whole roadmap, because every other Electra feature is an app the user can close if it misbehaves — this is the thing they log into. So the canvas engine ships first as an ordinary, closeable, resizable window (Milestone 5A) with zero desktop-layer wiring at all. Generation quality, mode-swap reliability, and the Python↔JS bridge all get proven out there, where a bug costs a closed window. Only once that's stable at a satisfactory bar does the same engine get wired into the actual login session as the desktop layer (Milestone 5B onward) — and even then, behind the crash-safe fallback and "classic desktop" safety-net mode described below. Electra Pets already proved a native, always-on, AI-powered desktop citizen ships in production — the canvas generalizes that pattern to the whole screen, but earns its way onto the login screen in stages rather than all at once.
MILESTONE 5A — WINDOWED CANVAS PROTOTYPE (de-risk first, ships before any desktop-layer wiring):
Canvas engine as a normal, closeable app window NEXT HIGH
The actual first deliverable — deliberately not the desktop layer yet. A standard resizable/maximizable GTK window hosting the same WebKit2GTK canvas the final desktop version will use, launched from the terminal or GUI like any other Electra surface. Proves the canvas engine end-to-end (generation → mode-swap → render) with zero risk to anyone's login session if something breaks.
"Change my desktop to..." — natural-language canvas generation HIGH
A prompt like "make my desktop a chessboard I can play on" or "show me a live system log instead" routes to a dedicated canvas-generation agent: it writes (or edits) a self-contained HTML/CSS/JS mode folder under ~/.electra/canvas/modes/<name>/, validates it renders without a blank screen, then hot-swaps the window's WebView to it. Reuses the same streaming-diff-preview and visual-feedback-loop infrastructure Coder mode already has rather than a blind one-shot reload — a first attempt at something like a legal-move chessboard or a farm-sim with pathing is expected to need iteration, the same way a Coder-mode file edit sometimes does.
Python↔JS bridge, built as a strict allowlist from day one HIGH
A WebKit2.UserContentManager script-message bridge (register handler in Python, window.webkit.messageHandlers.electra.postMessage() in JS) so canvas modes that need real system data — the live log mode, a system-aware farm sim — can subscribe to specific, named events (subscribe_to_log, get_open_windows, etc.). No raw exec, filesystem, or shell access exposed to JS, ever — designed this way from the very first prototype, not retrofitted later once a canvas-mode marketplace (Phase 6) makes AI-generated JS from other people's prompts a real trust boundary.
Long-running WebView memory/stability soak test HIGH
WebKit2GTK left running fullscreen and always-on for days is a known memory-growth risk — fine for a browser tab closed nightly, not fine for something that is the desktop. Validated here, in the windowed prototype, under multi-day soak testing, before it's anywhere near login-session criticality: a scheduled WebView process recycle (reload the current mode into a fresh process on an interval) rather than relying on crash-recovery alone.
Starter canvas mode library MED
A handful of modes to validate generation quality against and give early users something to switch between: a plain wallpaper mode, an interactive category-folder app launcher (click a folder, it opens, click an app, it launches), a live system log/activity feed, an analogue + digital clock, and a simple playable chessboard. Each one also doubles as a worked example for the AI to pattern-match against when generating new modes.
Promotion gate — go/no-go before desktop-layer wiring HIGH
Explicit checkpoint, not a formality: Milestone 5B (session bootstrap) does not start until the windowed prototype has run stable across the soak test, generation quality is consistently acceptable for the starter mode library, and the bridge allowlist has held up under adversarial prompts (asking the AI to generate a mode that tries to read arbitrary files, etc.). If any of those aren't satisfied, iterate here — not in a login session.
MILESTONE 5B — SESSION BOOTSTRAP & LOGIN MANAGER INTEGRATION (gated on 5A's promotion checkpoint):
Custom xsession — LightDM/GDM/SDDM Entry Point HIGH
A .desktop file in /usr/share/xsessions/electra.desktop pointing to a launch script that execs Muffin (muffin --replace) followed by electra_shell.py as the desktop-layer process. LightDM (MakuluLinux's default), GDM, and SDDM all just exec whatever the xsession script points to — none of them care what's inside. Only begins once the windowed canvas prototype from 5A has cleared its promotion gate.
Compositor backend abstraction — shell_backend.py HIGH
A thin backend interface (MuffinBackend, OpenboxPicomBackend) so the shell process talks to "whatever WM is running" through one API instead of hardcoding Muffin calls everywhere. Cinnamon/MakuluLinux ships Muffin out of the box — zero extra install. Non-Cinnamon distros fall back to Openbox (trivially scriptable, EWMH-compliant) + picom for compositing effects. Same shell code runs on both; only the backend implementation differs.
Session recovery, crash safety & classic-desktop safety net HIGH
If electra_shell.py crashes, the user must not be dropped to a blank screen. Launch script wraps the shell in a restart loop with backoff, and a "safe mode" fallback that relaunches stock Cinnamon/Openbox if the shell fails 3 times in a row. The starter mode library's plain-wallpaper mode is also promoted to a proper "classic desktop" mode here — real file icons, real drag-and-drop — so users who want the familiar layout back always have it as one command away, not a UX dead end. Non-negotiable before this ever reaches a real user's login screen.
MILESTONE 5C — THE CANVAS ENGINE, WIRED INTO THE DESKTOP LAYER:
Fullscreen desktop-layer canvas window HIGH
Reuses the exact engine validated as a window in Milestone 5A — a single borderless, fullscreen GTK window set to _NET_WM_WINDOW_TYPE_DESKTOP and stacked below every normal app window, occupying the exact space a wallpaper used to. No new canvas logic gets written here; this milestone is purely "move the proven engine into the desktop layer" plus input-isolation hardening below.
Input isolation — dock and hotkey survive a hung canvas HIGH
Canvas and dock run as separate processes/windows, and the global "revert to safe wallpaper" hotkey is handled entirely outside the WebView — never routed through it. If a bad piece of AI-generated JS hangs the render thread, the dock and its recovery hotkey must still respond. This is the desktop-layer-specific hardening step that the windowed prototype didn't need, because a hung window there was just a window to close.
Canvas mode format — self-contained HTML/CSS/JS bundles HIGH
Each mode is a folder under ~/.electra/canvas/modes/<name>/ (an index.html plus assets — no build step, no framework requirement). "Swap the desktop" means telling the WebView to load a different mode's index.html — instant, no window manager involvement, no restart.
Open-window strip / running-app awareness on canvas MED
Canvas modes can optionally render a strip of currently-open apps (via the same python-wnck/EWMH window list the dock uses) docked to a screen edge inside the canvas itself — useful for modes like the log view or launcher where "what's actually open right now" is part of the content, not just the dock's job.
MILESTONE 5D — THE DOCK (replaces panel, app menu, and the floating electra_bar):
Auto-hide dock — GTK3, macOS-Dock-style HIGH
A single always-on-top GTK3 strip, auto-hiding until the cursor hits its screen edge, position configurable (top/bottom/either side) in Settings. Docked via EWMH strut hints only while visible, so it never permanently steals canvas space the way a fixed panel would. Replaces the panel, the app menu, and the taskbar in one component.
Electra AI input lives in the dock, not a floating widget HIGH
electra_bar.py's floating-window role retires — the AI input becomes a permanent dock element instead, reachable the same way on every canvas mode. Same routing-agent forwarding behavior as today (chat / coder / command / "change my desktop to..."), just structurally anchored instead of a separate window that can be lost or covered.
Running-app list & window switching — python-wnck HIGH
Window list, focus switching, minimize/maximize/close via python-wnck (the same library GNOME/Cinnamon's own taskbars use) or raw EWMH client messages via python-ewmh as a fallback. Read-and-send-messages against the compositor's own window state, not reimplemented window-management logic.
App launcher folded into the launcher canvas mode, not a dock menu MED
/usr/share/applications/*.desktop parsing still happens (freedesktop.org standard, mature Python parsers available) but feeds the interactive folder-launcher canvas mode from Milestone 5A/5C rather than a dock dropdown — keeping the dock itself minimal: AI input, running apps, and nothing else.
Workspace switcher & multi-monitor MED
xrandr wrapper (or python-xlib direct) for monitor detection and arrangement. Each monitor can run its own canvas instance and its own dock instance — well-documented territory every X11 panel project has solved before.
MILESTONE 5E — SYSTEM SERVICES LAYER:
Notification daemon — org.freedesktop.Notifications HIGH
Implements the standard D-Bus notification spec via dbus-python/pydbus — well-documented, mature territory. Owning this means AI-generated notifications (heartbeat alerts, CI failures, finance bot events) can render as toasts on the dock or feed straight into a live-log canvas mode, instead of routing through whatever notification daemon happened to already be running.
Systray — StatusNotifierItem protocol MED
D-Bus-based systray icon protocol (the modern replacement for the old XEmbed systray), surfaced in the dock. Needed for third-party apps (Discord, Spotify) that still expect a tray.
Settings manager — GTK app, writes to ~/.electra/ HIGH
Owns dock position/auto-hide behavior, theme, keybindings, AI model routing preferences, tier/account info, and canvas mode management (install/remove/switch/pin a favorite mode). A GTK3 app in the same style as Electra Pets' existing in-app Preferences window — no recompile needed for changes, everything live-reloads.
Session & power management — logind/UPower MED
Logout/suspend/hibernate/shutdown and battery status via systemd-logind and UPower D-Bus interfaces — mature Python bindings, not a hard problem, just necessary plumbing for a complete session.
MILESTONE 5F — AI-NATIVE INTEGRATION (the actual point of building this):
Command mode gets direct EWMH/D-Bus window control HIGH
Because the shell process already holds live EWMH window-list and D-Bus session state, Command mode's tool set gains genuinely first-class primitives instead of shelling out: "move this window to workspace 2," "close Spotify," "what's open on my screen right now," "snap this to the left half" — all become direct window-manager calls instead of scripted wmctrl shell-outs. Meaningfully more reliable and much faster.
Heartbeat agent as a session-start service MED
Heartbeat launches as part of session startup (via the xsession script) instead of a separate autostart/cron entry, and can push events straight into an active live-log canvas mode via the Python↔JS bridge if the user has one loaded — one less moving part, and it's alive from the moment the user logs in.
Electra Pets as a canvas-aware desktop citizen MED
Electra Pets already runs as an independent transparent always-on-top window today — under Electra Shell it becomes a properly registered desktop-layer client (correct stacking, correct workspace-follow behavior) and can optionally live inside a canvas mode instead of floating above it, for modes where that fits (the farm-sim mode is the obvious first case).
System-wide AI hotkey from any context MED
Global keybinding (owned outright now, not requesting one from Cinnamon's keybinding daemon) that surfaces the dock's AI input from anywhere — any app, any workspace, any window — with zero focus-stealing weirdness because the shell process controls focus policy directly.
MILESTONE 5G — MULTI-DISTRO EXPANSION (extends the existing docker.sh compatibility strategy):
Ubuntu 24.04 / Cinnamon / MakuluLinux — primary target HIGH
First fully-supported combination. Muffin backend, LightDM xsession, ships bundled with the existing compile.sh/docker.sh pipeline (Ubuntu 22.04 container for broader glibc compatibility, same as ai_terminal.bin today).
Generic X11 fallback — Openbox + picom MED
Covers Debian, Fedora, Arch, and any distro without Muffin preinstalled. Openbox is trivially scriptable and EWMH-compliant, so Milestone 5C/5D's code needs minimal changes to run against it instead of Muffin.
Build pipeline extension — compile.sh / docker.sh MED
Extend the existing Nuitka onefile Docker build to include the new shell modules plus bundled starter canvas modes (--include-module flags per satellite, same pattern as electra_pet.py, telegram_bridge.py, etc.), producing a single distributable shell binary alongside ai_terminal.bin.
MILESTONE 5H — WAYLAND TRACK (explicitly deferred, not blocking):
Wayland feasibility evaluation FUTURE
Deliberately not built now. X11 + EWMH + wnck is mature, well-documented, and exactly what Cinnamon/Muffin itself still runs on primarily. A pure-Python Wayland compositor is not realistic — the Python Wayland bindings are far thinner than wlroots, and layer-shell (the protocol the canvas and dock would need) has no equivalent maturity in Python yet. Newer GNOME/Fedora defaulting to Wayland means this gap is real for the multi-distro goal — revisit once Cinnamon/Muffin's own Wayland session matures upstream rather than building ahead of the ecosystem.
Phase 6 Planned ✦ IMPRINT
Everything below was on the active roadmap as of July 2 2026. None of it is cancelled — it's reprioritized behind Electra Shell (Phase 5), which is now the top strategic priority. This phase holds the deferred IDE polish items, the deferred breakthrough features, the Electra Pets expansions, the deferred platform/ecosystem items, and IMPRINT — all still planned, just queued behind the desktop-environment push.
DEFERRED FROM PHASE 4 — IDE POLISH:
Open VSX — remaining activation layers HIGH
Complete tabstop navigation (Tab $1→$2→$3), full tokenColors syntax highlighting from VS Code themes, and actually launching detected LSP servers registered from extensions. Snippet engine partially done — tabstop jumps and TM_ vars shipped, full nesting remains.
Visual Debugger (DAP panel) HIGH
Breakpoints, call stack, watch expressions, step-over/into/out via Debug Adapter Protocol. Biggest remaining IDE gap vs VS Code and Cursor. Python debugpy is the first target; Node.js inspector second. Large effort — 3–6 weeks properly scoped.
Linter integration → agent feedback pipe MED
Run pylint/eslint/cargo clippy output directly into the coder agent as structured context on every file save. Tighter loop between static analysis and AI-suggested fixes.
Analytics dashboard MED
Model usage statistics, per-session token consumption, routing decisions, and error rates visualized in a GUI panel. Helps users understand which models are fastest for their workloads.
Rolling context summarisation MED
Auto-compact already fires at 78% context — upgrade it to a proper rolling summary so the effective context never falls off the edge. Current hard-trim loses nuance. A rolling NLP summary preserves intent across very long sessions.
Windows + macOS support MED
platform_compat.py shim approach identified. GTK3 available on both platforms via MSYS2/Homebrew. AT-SPI and D-Bus are Linux-only — needs graceful degradation stubs. Expands addressable audience significantly. Linux remains the native home. Lower priority now that Electra Shell (Phase 5) is Linux-native by design.
DEFERRED FROM PHASE 5 — BREAKTHROUGH FEATURES NOT YET SHIPPED:
Speculative Execution — Electra Already Started HIGH
Perceived speed beats actual speed. While the user is still typing in the bar or GUI input, Electra analyses the partial prompt every 2–3 seconds. If it looks like a coder task, it silently pre-fetches the workspace snapshot and reads the most likely relevant files in the background. By the time the user hits Enter, the first 2–3 tool calls are already done. The AT-SPI daemon already monitors input fields; the bar widget already captures keystrokes. Effort: low. Impact: the app feels dramatically faster without any model change.
Specialist Agent Panels — Expert Councils HIGH
Instead of one agent reviewing output, three specialists do it in parallel. Extends the Loop Engineering Critic Pass into a panel system: a Security Agent, a Performance Agent, and a Test Agent. The orchestrator spins all three simultaneously via threading, collects their verdicts, and synthesises a combined quality report. Effort: low — the Critic infrastructure already exists.
ELECTRA PETS — PLANNED EXPANSIONS:
Deeper cross-agent event hooks MED
Extend the push_pet_event() pattern already wired into Heartbeat and GitHub agents to Discord, RSS, Telegram, and Spotify — the pet becomes the ambient notification layer for the whole AI Center instead of just two agents.
Wayland layer-shell + native KDE Plasma wallpaper embedding MED
Current window-manager-hint approach (DOCK type) works well on X11/Cinnamon, MakuluLinux's default. True wlr-layer-shell support for Sway/Hyprland and a native Plasma wallpaper plugin would extend proper background-layer embedding to compositors that support it. Naturally converges with Phase 5's Milestone 5G Wayland track once that lands.
Achievement-tier celebration system FUTURE
The "milestone" reaction state already exists for real achievements (100th commit, a long coding streak) — needs the actual milestone-detection logic wired into Heartbeat Agent's activity tracking to fire it automatically instead of only via manual push_pet_event() calls.
Community palette + pet sharing FUTURE
Let users export/import custom color palettes, and eventually community-built pet shapes, through the same GitHub-based plugin auto-sync mechanism the app already uses for command plugins.
Multi-monitor "follow the active window" mode FUTURE
Preferences already lets users pin the pet to a chosen monitor. A smarter mode would have it roam whichever monitor currently has focus instead of a fixed home screen.
DEFERRED FROM PHASE 5 — PLATFORM & ECOSYSTEM:
AI-OS deep integration — Cinnamon shell layer HIGH
Largely superseded by Electra Shell (Phase 5) — once Electra is the shell, this item is mostly complete by construction. Retained here for anything Electra Shell doesn't fully absorb: deeper applet/desklet compatibility for users who stay on stock Cinnamon.
Electra Cloud Sync — cross-machine memory + config MED
Sync memory, pinned files, project configs, and conversation IDs to makululinux.us. MemPalace already handles server-side memory — extend to configs and active sessions. Multiple machines, same Electra brain.
Community plugin marketplace MED
Extend the existing GitHub plugin auto-forge system into a curated marketplace at makululinux.com. Revenue share to incentivise community contributions. Third-party slash commands as first-class citizens.
Finance Bot v2 — dependency bump automation MED
Auto-detect outdated dependencies in monitored repos, generate PRs with bumped versions and passing tests. Combine with the heartbeat agent for fully autonomous repo maintenance revenue generation.
Multi-user / team mode FUTURE
Shared project memory, shared pinned files, collaborative coder sessions over Telegram bridge or WebSocket. MemPalace architecture extended to team-scoped memory namespaces.
Electra Marketplace — Vibe Publishing FUTURE
/v1/vibe/publish endpoint already exists. Build the frontend marketplace for user-created AI-generated web apps, prompts, and workflows.
ARCHITECTURAL CONCEPT
✦ IMPRINT
Most AI systems can retrieve memory. Very few carry behavioral continuity. The distinction: remembering information is retrieval — the AI knows more. Accumulating operational experience is continuity — the AI behaves differently. IMPRINT closes this gap: persistent behavioral pressure, confidence decay that survives session boundaries, contradiction carried over time, adaptive trust that responds to outcomes, and internal state that evolves through experience — not through prompts, not through retrieval, but through continuity. Every session leaves a mark. The system that wakes up tomorrow is not the same system that ran today.
Persistent Behavioral State — electra_state.json HIGH
The substrate for everything else in Phase 6. A structured state file at ~/.electra/state.json written after every session and read before any prompt is constructed. Tracks: per-domain operational confidence scores (command_mode: 0.87, file_regex_ops: 0.61, etc.), autonomy trust dial (0.0–1.0), contradiction log, domain failure rates, and last-updated timestamp. This file is the behavioral substrate — it changes what Electra does, not just what it knows. The critical difference: current memory systems inject text into prompts. This system mutates operational parameters before prompts are built.
Confidence Decay Engine — Cross-Session Confidence Scoring HIGH
Confidence resets to zero every session today — that ends here. A background thread runs after every session and scores outcomes per domain: success nudges confidence up (+0.02, capped at 1.0), failure decays it (−0.05 × severity), sustained stuck state decays further (−0.03). These scores persist across reboots via electra_state.json and directly shape system prompt language: confidence > 0.85 → "Proceed autonomously"; confidence 0.6–0.85 → "Proceed but verify key steps"; confidence < 0.6 → "Confirm before destructive operations." The existing _COMMAND_FAILURES_FILE already captures per-command failure history — this extends it to a scored, decaying, cross-session confidence model per domain class.
Autonomy Trust Dial — Replace Binary DRY_RUN with Graduated Trust HIGH
From on/off to a continuous spectrum. Replace the boolean DRY_RUN_MODE with a floating AUTONOMY_LEVEL (0.0–1.0) that adjusts dynamically based on session outcomes. A clean session (all tools succeed, user accepts output) raises trust slightly. A session with rollbacks, user corrections, or repeated failures lowers it. At 1.0: full autonomy, zero confirmations. At 0.5: confirm before destructive operations. At 0.0: confirm every tool call. Users can manually anchor the dial via /trust 0.7 or let it self-calibrate. This is the difference between Electra acting like a junior dev that needs approval and a senior dev you trust to run with a task.
Contradiction Detector — Carrying Conflict Across Time MED
No system exists today that flags when Electra is about to contradict itself. On every Project Brain update, a lightweight diff pass checks the incoming approach against existing Decisions and Preferences entries. If a contradiction is detected ("User said always use AST validation but just skipped it"), it is flagged to the user, logged to ~/.electra/contradictions.json, and injected into future sessions as context tension — not as an instruction, but as a named unresolved conflict the AI can reason about. This transforms Project Brain from a knowledge store into a consistency guardian.
Behavioral Pressure Layer — Preferences That Change Behavior, Not Just Prompts MED
The gap between knowing a preference and acting on it differently. Today Project Brain and MemPalace store user patterns as text and inject them as hints. The Behavioral Pressure Layer translates accumulated patterns into actual operational parameter mutations: verbosity budget adjusted (if the user repeatedly edits long outputs down, reduce generation length target), tool aggressiveness tuned (if the user frequently rejects first-attempt file writes, add a read-before-write gate), model routing shifted (if a specific model consistently fails on this user's codebase, lower its routing weight for this user). Experience → internal state mutation → behavioral change. Not retrieval.
Operational Experience Log — The Delta Between Sessions FUTURE
A structured record of what changed, not just what happened. After each session, a diff is computed between the pre-session and post-session behavioral state: which confidence scores moved and by how much, which contradictions were introduced or resolved, how the autonomy dial shifted, and what new patterns were recorded. This delta log becomes the foundation for user-facing "Electra self-awareness" reporting (/state report): "Since last week, your command-mode confidence has risen from 0.71 to 0.89. Regex operations remain your weakest domain at 0.58. Two unresolved contradictions detected." The system surfaces its own internal evolution explicitly.
Jarvis 6–18 months
Full Jarvis Control Layer — "do anything on my PC" HIGH
Every OS action accessible via natural language from any input: voice, typing, Telegram, phone app, desktop widget, or automation script. The D-Bus foundation is in place. The vision: Electra is the OS-level AI that makes Linux the most powerful personal computing platform ever built.
Electra Phone App (Android / iOS) HIGH
Native mobile frontend connecting to makululinux.us:2007. Telegram bridge already provides a path; a proper app adds voice, camera (image analysis), and push notifications. Control your Linux machine from your phone.
Autonomous Revenue Engine v3 — self-managing income HIGH
Finance bot + heartbeat + coder agent + github agent combined into a fully autonomous income loop: find freelance jobs → write proposals → deliver code → monitor repos → handle maintenance → collect payments. Jarvis earns money while you sleep.
Electra for Developers — API & SDK MED
Public API so third-party apps can integrate Electra's routing, memory, and tools. D-Bus is the local interface — add an HTTP SDK for remote integration. Build the ecosystem.

The Bottom Line — July 3 2026

Phase 4 complete. Electra AI is a full-stack Jarvis-style AI platform: a compiled binary, 80+ models, a complete IDE, autonomous background agents, finance engine, deep OS integration, and service connections that no SaaS competitor can replicate. The June 29 2026 sprint delivered the Unified Coder overhaul — Coder mode is now stable for the first time since its initial design. July 2 2026 delivered the Command Center Upgrade: Heartbeat Real Tool-Calling (full function-calling parity with interactive Command mode), Background Task Queue (/bg, /tasks), and System Changelog (/changelog). All shipped as anchor-validated, backup-protected patch scripts, test-applied before hand-off.

July 4 2026 correction: Electra Shell is Phase 5, top priority — reframed as an interactive canvas desktop, not a rebuilt traditional shell. Every item that was queued for Phase 5 as of July 2 — Speculative Execution, Specialist Agent Panels, the Electra Pets expansion list, AI-OS Cinnamon shell deepening, Cloud Sync, the plugin marketplace, Finance Bot v2, multi-user mode, and Electra Marketplace — moves to Phase 6 alongside IMPRINT. None of it is cancelled; all of it is queued behind a strategic bet that pays off bigger: building Electra Shell, where the entire screen becomes one fullscreen AI-generated canvas instead of a wallpaper-plus-icons-plus-panel layout. No desktop icons, no fixed panel, no static app menu — a single WebKit2GTK-hosted canvas that the AI can turn into a chessboard, a live farm sim, a plain-English system log, an interactive folder launcher, or anything else on request, plus a slim auto-hide dock (Mac-Dock-style) that replaces the panel/menu and houses the AI input. The scoping call that makes it viable: still don't write a compositor (Muffin, already installed on Cinnamon/MakuluLinux, or Openbox+picom elsewhere, keeps doing that job) — build the canvas and dock on top of it instead. The first concrete deliverable is a custom LightDM/GDM/SDDM xsession that boots Muffin + a Python shell process hosting the canvas — the smallest possible proof the architecture works end-to-end. Once that boots, "change my desktop to..." becomes a real command, and Command mode gains direct EWMH/D-Bus window control instead of shelling out to wmctrl. Electra Pets — already a native, always-on, AI-powered desktop citizen shipping in production — is the existence proof this generalizes.

Phase 6 (deferred + planned): next-level unconventional features + IMPRINT. Speculative Execution (pre-fetch while the user is still typing) and Specialist Agent Panels (security/performance/test experts in parallel) remain high-impact, low-effort items ready whenever Phase 5 frees up cycles. IMPRINT remains the architectural gap no competitor has closed — a persistent behavioral state substrate (~/.electra/state.json) that survives session boundaries: per-domain confidence scores that decay and recover through experience, a graduated autonomy trust dial, a contradiction detector, and a behavioral pressure layer that mutates operational parameters rather than prompts. Today Electra changes what the AI knows per session; IMPRINT makes Electra change how it behaves across sessions.

The OS moat keeps compounding, and Electra Shell is the deepest expression of it yet. Every sprint so far added capabilities that require owning the Linux distro and running locally — AT-SPI screen reading, D-Bus desktop control, persistent semantic memory, Telegram alerts from a background task. Electra Shell goes one level deeper: it doesn't just run on Linux, it becomes the desktop. No cloud AI competitor — Cursor, Copilot, ChatGPT, Antigravity — can ever follow this move, because none of them own an OS.

GitHub: github.com/raymerjacque/Electra_AI_Center

Electra AI Center: https://makululinux.us/electra.html

AI-OS: https://makululinux.us/ai-os.html

Full Guide: https://makululinux.us/ai-terminal-guide.html

Daily Updates/Community: https://www.patreon.com/makululinux

Source Code: https://makululinux.us/electra_agent.html