Skip to content

ROADMAP.md

Latest commit

 

History

History
139 lines (111 loc) · 7.35 KB

ROADMAP.md

File metadata and controls

139 lines (111 loc) · 7.35 KB

Ctx AI Roadmap

Last updated: 2026-03-19

✅ Completed

🔐 Security

Item Detail
Safe eval safe_eval.py — primitive-only values, dunder-key rejection, fail-closed
Context limits MAX_CONTEXTS=64, LRU eviction of non-running contexts
Concurrency limits MAX_CONCURRENT_TASKS=16 semaphore on agent execution
Plugin auth All plugin API handlers enforce auth+CSRF+API-key unconditionally
Login rate limit 10 attempts / 5 min window on /login
Trace correlation trace_id on AgentContext and LogItem output

🏗️ Architecture (Layered Separation)

Layer Module Responsibility
Agent Orchestrator helpers/agent_orchestrator.py Monologue loop, process chain, lifecycle hooks
Reasoning Engine helpers/reasoning_engine.py Prompt assembly, LLM calls, tool resolution & execution
Event Bus helpers/event_bus.py Typed pub/sub for 15+ lifecycle events
Task Router helpers/task_router.py Priority queue, skill-based dispatch, load tracking
Orchestrator Agent helpers/orchestrator_agent.py Worker pool, per-task timeouts, retry with backoff

Agent class reduced 27% (993→727 lines). All @extension.extensible methods preserved as thin delegates.

🤖 Agent System

  • AgentMemory — observations (transient, TTL), facts (persistent), rolling tool results window
  • execute_tool_with_retry() — configurable retries, exponential backoff, fallback tool support
  • AgentContext.timeout — per-context execution timeout enforced via asyncio.wait_for

🧠 Memory & Context

Component Detail
MemoryManager remember/retrieve/forget API, importance scoring, TTL, short/long memory types
Backends InMemoryBackend (default), QdrantBackend stub with fallback
Compression Message.importance field — high-importance messages skipped during eviction
Budget eviction forget(policy="budget|age|importance|expired")

🏷️ CI/CD & Tooling

  • .github/labeler.yml — path-based PR labels (17 area/type labels)
  • .github/workflows/labeler.yml — auto-label PRs on open/sync
  • .github/workflows/pre-commit.yml — pre-commit hooks in CI
  • scripts/ — all .sh files set executable (755)

🔄 In Progress

Memory Backends

  • Qdrant — embedding computation, vector search, multi-tenant collections
  • Chroma backend
  • Redis backend for high-concurrency short-term memory

Observability

  • Structured metrics export (Prometheus/OpenTelemetry)
  • Dashboard — agent context pool, task queue depth, memory stats

Testing

  • Unit tests for History.compress() and Topic.summarize()
  • Integration tests for agent + tool loops with known inputs
  • Load tests for concurrent contexts

📋 Planned

Phase 2 — Stability

  • Streaming tool responses — async iterator for progressive output
  • Agent skill profiles in agent.yaml — auto-register with TaskRouter
  • Context-aware agent selection — map tasks to skill agents
  • Graceful degradation — fallback to simpler model on overload
  • Typed ToolRequest dataclass and centralized parser
  • Standardized error handling — exceptions → user-friendly error with retry policy

Phase 3 — Scaling

  • Distributed context store (Redis/DB + workers)
  • Persistent vector DB (Qdrant/Chroma optional plugin)
  • Multi-tenant isolation — per-user context pools and memory namespaces
  • Cluster orchestration for multi-host, with leader election
  • Backpressure and queue limits for new contexts

Phase 4 — Advanced AI

  • Self-improving learning loop — post-task evaluation + autoprompt refinement
  • Multi-agent collaboration graph — capability-based routing
  • Conversation branching — snapshot/restore context state
  • Knowledge graph integration — structured knowledge beyond vector similarity
  • Plugin sandboxing — restrict filesystem and network access
  • Dynamic tool-function schema generation and safe-call enforcement
  • Agent personality — persistent persona traits learned from interactions
  • Webhook triggers — external events drive agent tasks (email, git push, cron)

Refactoring Priorities

Code quality

  • Break large functions (Agent.monologue, History.compress, Agent.process_tools)
  • Move history compression policy to configurable strategy classes
  • Remove dead/commented code in handle_exception

Dependencies

  • Split core runtime vs optional features (browser, speech, document, MCP)
  • Regularly run pip-audit and safety scans
  • CI step for dependency drift

Architecture

┌───────────────────────────────────────────────────────────┐
│                  API / WebUI / WebSocket                  │
├───────────────────────────────────────────────────────────┤
│                   Agent Orchestrator                      │
│  ┌───────────────┐  ┌──────────────┐  ┌───────────────┐  │
│  │ Process Chain  │  │ Monologue    │  │ Intervention  │  │
│  └───────────────┘  └──────────────┘  └───────────────┘  │
├───────────────────────────────────────────────────────────┤
│                   Reasoning Engine                        │
│  ┌───────────────┐  ┌──────────────┐  ┌───────────────┐  │
│  │ Prompt Build   │  │ LLM Call     │  │ Tool Execute  │  │
│  └───────────────┘  └──────────────┘  └───────────────┘  │
├───────────────────────────────────────────────────────────┤
│  ┌───────────────┐  ┌──────────────┐  ┌───────────────┐  │
│  │ History/       │  │ MemoryMgr    │  │ Vector DB     │  │
│  │ Compression    │  │ short/long   │  │ FAISS         │  │
│  └───────────────┘  └──────────────┘  └───────────────┘  │
├───────────────────────────────────────────────────────────┤
│       EventBus  ·  Extensions  ·  Plugins                │
└───────────────────────────────────────────────────────────┘

See AGENTS.md for development patterns.