Releases
v0.2.0
v0.2.0 — Plan A: Orchestration + Memory Subsystem
Compare
Sorry, something went wrong.
No results found
v0.2.0 — Plan A: Orchestration + Memory Subsystem
Orchestration Engine
Task Graph — DAG-based task decomposition with dependency resolution, cycle detection, retry logic
Orchestrator — State-machine tick loop (idle → classify → plan → review → execute → complete/failed)
Planner — LLM-powered goal decomposition with custom roles, cost estimation, dependency validation
Plan Mode — Execution state controller with phase transitions, plan persistence, replan triggers
Health Monitor — Agent health checks (heartbeat, stuck tasks, error loops, credit depletion) with auto-heal
Messaging — Typed inter-agent messaging with priority routing, retry backoff, pluggable transport
Local Worker Pool — In-process multi-agent execution for local dev (no Conway sandbox required)
Attention Pattern — Manus-style todo.md injection for persistent goal tracking every turn
Memory Subsystem
Context Manager — Model-aware context assembly with token-budget enforcement
Compression Engine — 5-stage progressive compression cascade (compact → compress → summarize → checkpoint → truncate)
Event Stream — Append-only event log with compaction and pruning
Knowledge Store — Cross-agent knowledge base with category-based retrieval
Enhanced Retriever — Metadata-based relevance scoring, dynamic budget, feedback tracking, query enhancement
Agent Context Aggregator — Triage-based child update aggregation
Inference
Unified Inference Client — Multi-provider client with tier routing, retry/failover, circuit breaker, survival mode
Provider Registry — Model configuration with tier resolution, Conway Compute API fallback
Agent Improvements
Comprehensive system prompt — Structured orchestration context with capabilities, constraints, state machine, error handling, anti-patterns, circuit breakers, and turn protocol
7 orchestration tools — create_goal, list_goals, cancel_goal, get_plan, complete_task, orchestrator_status + existing spawn/fund tools
Agent spawning — Conway sandbox → local worker → self-assignment fallback chain
Goal dedup — Max 1 active goal, duplicate detection, BLOCKED message tells agent to sleep
Identity registration — Automaton registers with Conway on startup (merged PR #216 )
Bug Fixes
Fixed duplicate normalizeTaskResult (shared export from task-graph.ts)
Renamed FundingProtocol.addr → childAddress per spec
Added role column to children table (V9 ALTER migration)
Wired handlePlanReviewPhase to call reviewPlan()
Fixed SimpleFundingProtocol.getBalance returning parent's balance instead of child's
Fixed orchestrator stuck in failed phase (now resets to idle)
Fixed planner empty-task fallback (single generalist task instead of failing)
Fixed spawn ENOENT (mkdir before writeFile in sandbox)
Skip funding/messaging for local workers and self-assigned tasks
Bridge config API keys to env vars for provider registry
Conway Compute API as inference fallback when no OpenAI key
Tests
190+ new tests across 10 test files (unit + integration)
Orchestrator, planner, simple-tracker, health-monitor, agent-context-aggregator
Plan-execute flow, multi-agent coordination, compression cascade, inference failover, memory retrieval
You can’t perform that action at this time.