-
-
Notifications
You must be signed in to change notification settings - Fork 3
Changelog
CortexPrism edited this page Jun 17, 2026
·
1 revision
All notable changes to CortexPrism.
Format: Keep a Changelog · Versioning: SemVer
- 11 new LLM providers — Cerebras, Fireworks, Perplexity, NVIDIA NIM, Moonshot (Kimi), Novita AI, LM Studio, LiteLLM, Hugging Face Inference Router, Alibaba (Qwen), Venice AI (total: 24 providers)
-
Plugin update system — automatic version checking via GitHub Releases/Tags API with
autoUpdatesupport - Provider-specific LLM settings — reasoning effort, repetition penalty, recency filter, num_ctx, keep_alive, citations, and more surfaced in the Web UI
- Provider card summary badges — active extra settings displayed inline in Settings
- Dynamic provider + model selects in agent modal — auto-fetches models for chosen provider
- Quartermaster unified page — merged QM + Model Intel into a single page with Tool Orchestration and Model Intelligence tabs
- QM Settings panel — enable/disable Model Intelligence, pin dedicated QM provider, choose strategy
- QM patterns tab rework — real pattern rows from database
- First-time password setup — Settings Security tab handles both first-time and change scenarios
- Request-flow architecture doc — Mermaid flowchart of the full request lifecycle
-
Memory health CLI —
cortex memory healthwith per-tier stats and color-coded decay -
Memory heuristics CLI —
cortex memory heuristicstriggers heuristic learning cycle -
20+ migration —
updated_aton episodic_memory, graph_relation_types seed -
Agent loop debug tracing —
[loop]prefixed console.log for every tool round
- Plugin auto-update HTML parse crash — uses archive tarballs for GitHub-sourced plugins
- Daily semantic decay was a no-op — corrected column name
last_accessed - Heuristic cycle metrics always reported zero — now use real
rowsAffected - Reflection patterns duplicated on every turn — now deduplicates and merges confidence
- Memory injection missing metadata — now surfaces category, tags, topics, entities
- Noisy knowledge graph entities — filtered common English words
- Bare JSON tool calls leaked to chat UI — fixed with brace-depth walker
- Raw LLM response leaked to chat UI — buffered streaming with tool detection
-
<tool_result>XML leaked to chat UI — removed, with client-side safety net - Duplicate tool call execution — XML regions stripped before bare JSON scan
- LLM hang on tool follow-up rounds — 90-second
AbortSignaltimeout - Model looping without final answer — escalating per-round instructions
- Voice & TTS system — full pipeline: STT via Whisper, TTS via OpenAI/ElevenLabs, VAD
- Voice WebSocket protocol — real-time audio streaming with server-side transcription
- Voice API routes and Auto-TTS pipeline hook
- Web UI file upload — drag-and-drop PDF, images, documents
- Multimodal LLM support — Anthropic, OpenAI, Google, Ollama, Bedrock
- PDF text extraction — integrated into file_read tool and chat uploads
- Session resume on page refresh — seamless continuation across reconnects
-
web_fetchtool — fetch URL content as cleaned plain text -
file_globtool — glob pattern file matching -
shelltool wired in — previously unregistered - Service management —
cortex startandcortex restartcommands
- Agent loop — multimodal support, tool follow-up instruction embedding
- System prompt augmentations — File Context and Environment sections
-
file_readtool — workspace parameter, PDF extraction cap -
code_exectool — explicit sandbox isolation warnings
- PDF extraction silent failure — Uint8Array for Deno compatibility
-
code_execDocker filesystem blindness — agent wasting rounds in sandbox - Tool output displayed without analysis — follow-up instruction in tool results
- Unified service installation —
cortex installandcortex uninstall - Server service installation —
cortex serve install - Cross-platform service helper — systemd (Linux), launchd (macOS), NSSM/Task Scheduler (Windows)
- Extended Windows installer with daemon-only/server-only flags
- Windows path resolution —
fromFileUrlthroughout codebase - Windows process management — cross-platform kill helpers
- Windows shell execution — PowerShell fallbacks
- Server stability — error handler on
Deno.serve()
- Cross-platform support — macOS & Windows alongside Linux
- Desktop automation — macOS (screencapture, osascript, cliclick), Windows (PowerShell/.NET)
- Daemon service installation — systemd, launchd, NSSM
- CI/CD expansion — multi-platform test matrix
- Package distribution manifests — Homebrew, Chocolatey, Scoop, winget
- Code signing guide for desktop app distribution
- Dashboard as default landing page
- Navigation consolidation — Status merged into Dashboard
- Three new Dashboard widgets
-
dashboard_manageLLM tool for agent CRUD on widgets - Custom HTML widget type — agent-created widgets
- Soul system expansion — 7 personality templates,
cortex soulCLI - Expanded DEFAULT_SOUL — 10 sections with behavioral guidance
- Model Quartermaster (MQM) — learning-based model selection engine
- 6-signal prediction engine with adaptive EMA learning
- Three arbiter strategies (conservative/balanced/aggressive)
- Server UI Quartermaster dashboard
- Pipeline hook stages: pre-llm, post-llm
- Quartermaster — tool orchestration learning system
- 5-signal prediction engine, three-mode decision system (automate/suggest/defer)
-
cortex qmCLI with 9 subcommands
- Model configuration CLI —
cortex models - Reasoning effort / extended thinking for Anthropic, Google, OpenAI
For the complete history, see the full CHANGELOG.md in the repository.
CortexPrism — Open-source agentic AI harness · MIT License · Built with Deno 2.x + TypeScript