v3.12.0 — Replit Integration + Lite Mode

What's New

Replit Integration

kbot now auto-detects Replit environments and adapts:

• Lite mode (--lite) — skips heavy tools (Docker, browser, local models, GPU) to fit cloud IDE resource limits
• Auto-detection — Replit env vars (REPL_ID, REPL_SLUG) trigger lite mode automatically
• Replit Secrets — API keys from Replit Secrets are auto-detected
• Public URL — serve mode shows your Repl's public URL on startup

Replit Template

Fork-ready template at templates/replit/ — gives you a full AI agent REST API in ~30 lines:

• GET / — status + tool count
• POST /chat — send a message, get an agent response
• POST /chat/stream — SSE streaming response
• GET /tools — list all available tools

Install

npm install @kernel.chat/kbot@3.12.0

Bug Fixes

• Prompt evolution auto-rollback now uses success rate as fallback when self-eval is disabled
• Tree planner best-path extraction simplified

v3.11.0 — 8 Research Breakthroughs

8 Research Breakthroughs — No Other CLI Agent Has Any of These

v3.10.0

1. Voyager Skill Library — successful tool chains auto-distill into reusable skills indexed by embedding
2. MAR Multi-Persona Reflection — 5 specialists critique failures, judge synthesizes lessons
3. Spec-Driven Development — `kbot spec` generates formal requirements before coding (EARS notation)
4. Three-Tier Memory Synthesis — observations → reflections → identity (Stanford Generative Agents)

v3.11.0

5. LATS Tree Search Planning — Monte Carlo Tree Search for branching plans with backtracking (`--tree`)
6. GEPA Prompt Self-Optimization — specialist prompts evolve from execution traces (ICLR 2026)
7. A2A v0.3 Protocol — agent-to-agent interop, Agent Cards, task lifecycle (`kbot a2a`)
8. MCP Apps — interactive HTML/JS UI in conversations (Chart.js, tables, diffs, Mermaid)

Also in this release

• `kbot status` — unified Kernel dashboard
• `kbot automate` — event-driven automations (file watch, schedule, git hooks, webhooks)
• `kbot --plan` — read-only exploration mode
• Collective learning fully wired with real signal values

Install / Upgrade

```bash
npm install -g @kernel.chat/kbot
```

v3.7.1 — Full Cognitive Stack + VS Code Extension + CI Pipeline

What's New

Full Cognitive Stack (v3.7.0)

All 11 cognitive modules now fire on every agent cycle:

• Free Energy, Predictive Processing, Reasoning, Intentionality, Temporal,
  Confidence, Strange Loops, Integrated Information — plus the existing
  Learning, Self-Eval, and Skill Rating.

Skills Auto-Discovery

Drop .md files in ~/.kbot/skills/ or .kbot/skills/ and kbot auto-loads
them as context. The equivalent of Copilot Agent Skills.

Autotelic Agent

New kbot autotelic command — self-purpose + self-agency. Senses the
highest-impact fix and decides what to do.

VS Code Extension

packages/vscode-kbot/ — MCP server registration + 6 commands + context menu.
Package: kbot-0.1.0.vsix (30KB).

CI Pipeline

GitHub Actions runs build + tests on web, kbot, and VS Code extension.
Branch protection enabled with required checks.

Surface Sync

19 files aligned to 290 tools, 23 agents, v3.7.1. GitHub description
rewritten. 20 high-traffic topics. 4 GitHub releases (was 0).

Install / Upgrade

npm install -g @kernel.chat/kbot

v3.7.0 — Full Cognitive Stack + Skills Auto-Discovery

The Intelligence Upgrade

kbot had 11 cognitive modules. Only 3 were wired into the agent loop. The other 8 were dead code — architecturally beautiful TypeScript implementations of peer-reviewed research papers that never actually ran.

v3.7.0 wires them all in.

Now Active in Every Agent Cycle

Module	Paper	What It Does
Free Energy	Friston, 2010	Explore vs exploit decisions
Predictive Processing	Clark, 2013	Anticipates user's next action
Reasoning	—	Auto-selects strategy, records outcomes
Intentionality	—	Intrinsic motivation drives quality
Temporal	—	Action sequences, identity evolution
Confidence	—	Estimates ability, records actual effort
Strange Loops	Hofstadter, 1979	Detects self-referential reasoning
Integrated Information	Tononi, 2004	Measures synthesis quality (Φ)

Previously only ran: learning, self-eval, skill-rating (3/11).
Now all 11 fire. kbot thinks with its full brain.

Skills Auto-Discovery

Drop .md files in ~/.kbot/skills/ (global) or .kbot/skills/ (project) and kbot auto-loads them as context. The equivalent of Copilot Agent Skills or Cursor Rules for AI.

Autotelic Agent

New concept: kbot autotelic — the fusion of Bootstrap (self-purpose) and Limitless Execution (self-agency). An agent that generates its own goals from observation and achieves them.

Install / Upgrade

npm install -g @kernel.chat/kbot

v3.6.1 — Surface Hidden Features

Documentation-only release. Both READMEs now surface 60% more features that existed in code but weren't documented:

• 9 new tool categories: Quality, Creative, Training, Sandbox, Notebooks, Build Matrix, LSP, Memory, Forge
• 9 new CLI commands: voice, watch, bootstrap, plugins, models, changelog, completions, cloud, pair/team/record
• 5 power-user flags: --architect, --thinking, --self-eval, --computer-use, -p

No code changes. Install/upgrade: npm install -g @kernel.chat/kbot

v3.6.0 — Limitless Execution + Forge Registry + Immune Agent

What's New in v3.6.0

Limitless Execution (5 new patterns)

• Tool Discovery — smart error messages guide AI to mcp_search → mcp_install → forge_tool
• Cost-Aware Model Routing — trivial tasks use fast model, complex tasks use full model
• Goal Decomposition — routeStepToAgent() assigns specialist agents per plan step
• Fallback Chains — URL fails → web search. Command not found → npx. Auto-recovery.
• Self-Extension — forge_tool creates tools at runtime, sandboxed, persisted

Forge Registry

• Publish tools you create. Other users install them.
• kbot forge publish / kbot forge install

Autopoietic Health

• kbot vitals — live system viability dashboard
• Self-monitoring: memory, API health, tool pipeline, learning engine

Immune Agent

• Self-audit system that detects configuration drift, stale patterns, broken tools
• Auto-heals safe issues, flags destructive ones for human review

Security Hardening

• forge_tool blocklist expanded from 13 → 29 rules
• 56 new security tests for forged tool sandbox
• Reserved name protection (can't overwrite built-in tools)

Stats

• 290 tools | 23 agents | 20 providers
• 392 tests across 16 test files
• Build passes clean on Node.js 20+

Install / Upgrade

npm install -g @kernel.chat/kbot

Or run directly:

npx @kernel.chat/kbot