Skip to content

BACKLOG.md

Latest commit

 

History

History
370 lines (278 loc) · 12.6 KB

BACKLOG.md

File metadata and controls

370 lines (278 loc) · 12.6 KB
Raw

Zouroboros Backlog

Active backlog for evaluation and future work. All P0/P1 items have been completed and moved to CHANGELOG.md.

Legend

Priority Meaning Timeline
🟢 P2 Nice to have - incremental Next 6 months
⚪ P3 Exploration - experimental Future

🟢 P2 — Nice to Have

Everything-Claude-Code Research (Continued)

[ECC-006] Manifest-Driven Selective Install

Priority: P2 Effort: Medium (1-2 weeks) Status: Proposed → Needs Evaluation Impact: Medium — enables selective deployment (e.g., memory-only, swarm-only)

Component manifest for installing only needed Zouroboros features:

# zo-ecosystem/manifest.yaml
components:
  memory: { skills: [zo-memory-system], agents: 1, tools: [memory.ts] }
  swarm: { skills: [zo-swarm-orchestrator], agents: 3, tools: [orchestrate-v4.ts] }
  vault: { skills: [zo-vault], agents: 1 }
  finance: { skills: [alpaca-trading-skill, backtesting-skill], agents: 1 }

Install: zo-ecosystem install --profile minimal or zo-ecosystem install memory swarm

Acceptance Criteria:

  • Manifest schema definition
  • Dependency resolution
  • Install/uninstall commands
  • State tracking of installed components
  • Incremental update support

[ECC-007] Multi-Harness Abstraction

Priority: P2 Effort: Medium-Large (2-3 weeks) Status: Proposed → Needs Evaluation Impact: Medium — extend Zouroboros to work with Claude Code, Gemini CLI, Codex, etc.

Make Zouroboros skills work equally well across harnesses:

  • Claude Code (.claude/)
  • Codex (.codex/)
  • Cursor (.cursor/)
  • Gemini CLI (via bridge scripts)
  • Hermes (native)

Acceptance Criteria:

  • Harness detection
  • Per-harness config overlays
  • Bridge script compatibility matrix
  • Harness-specific tool mapping
  • Unified skill format across harnesses

[ECC-008] Skill Evolution as Closed Loop

Priority: P2 Effort: Large (3-4 weeks) Status: Proposed → Needs Evaluation Impact: Medium — self-improving skill system

Closed-loop skill evolution:

Success → extract winning pattern → evolve skill version → test against eval fixture → promote
Failure → classify error type → flag skill for review → update instructions → retest

Acceptance Criteria:

  • Success pattern extraction
  • Skill version management
  • Automated test-on-evolve
  • Promotion/demotion workflows
  • Evolution history tracking

[ECC-009] Observer Loop Guard (5-Layer)

Priority: P2 Effort: Small (2-3 days) Status: Proposed → Needs Evaluation Impact: Medium — prevents recursive routing loops

Guard against self-referential routing loops:

  1. Request origin tracking
  2. Loop depth limit
  3. Cyclic call detection
  4. Timeout on recursive chains
  5. Circuit breaker for detected loops

Acceptance Criteria:

  • Origin header propagation
  • Loop detection algorithm
  • Automatic loop breaking
  • Alerting on loop incidents

[ECC-010] Memory Explosion Throttling

Priority: P2 Effort: Small (1-2 days) Status: Proposed → Needs Evaluation Impact: Medium — prevents vector DB bloat

Prevent excessive embedding generation:

  • Throttle: max N embeddings per conversation minute
  • Tail sampling: keep only last K captures when limit hit
  • Cooldown: don't re-embed same content within 5 minutes

Acceptance Criteria:

  • Rate limiting on embedding generation
  • Duplicate detection with cooldown
  • Sampling strategy for overflow
  • Metrics on throttled embeddings

Swarm Orchestrator

AgentKV/CortexDB Evaluation for Memory Backend

Status: Proposed → Spike Required Effort: Medium-Large (2-3 weeks for evaluation + decision) Impact: High — Potential replacement for SQLite+Ollama stack

Evaluate AgentKV (Python/C++) and CortexDB (Go) as potential backends for zo-memory-system. Both offer local-first, single-file graph+vector storage.

Criteria Current (SQLite+Ollama) AgentKV CortexDB
Language TypeScript/Bun Python/C++ Go
Storage SQLite file Single mmap'd file SQLite-backed
Vector Search Ollama (external) HNSW (embedded) HNSW (embedded)
Graph Custom adjacency table Property graph edges Knowledge graph

Acceptance Criteria:

  • Functional TypeScript bindings for CortexDB
  • Benchmark suite comparing current vs. CortexDB vs. AgentKV
  • Decision document with migration plan or rejection rationale
  • If accepted: migration path for existing users

Rationale: Current SQLite+Ollama stack works but has latency overhead. Embedded vector+graph could reduce search latency from ~4s to <100ms.

Memory System

Memory v4 Enhancements (Migrated from zouroboros-memory-system)

Status: Migrated 2026-04-01 Impact: High — 9 new capabilities in packages/memory/src/

Migrated scripts from zouroboros-memory-system into the monorepo:

ID Feature File
MEM-001 Context Budget Awareness context-budget.ts
MEM-002 Recursive Episode Summarization episode-summarizer.ts
MEM-003 Iterative Multi-Hop Retrieval multi-hop.ts
MEM-101 Memory System Metrics Dashboard metrics.ts
MEM-102 Import Pipeline (ChatGPT, Obsidian, Markdown) import-pipeline.ts
MEM-103 Memory Conflict Resolution conflict-resolver.ts
MEM-104 Cross-Persona Memory Sharing cross-persona.ts
MEM-105 Enhanced Knowledge Graph Traversal graph-traversal.ts
MEM-202 Embedding Model Benchmark embedding-benchmark.ts

[MEM-201] Vector Database Migration

Status: Not Started Effort: High (5-7 days) Impact: Medium — scale beyond 10k facts

Current: SQLite + in-memory vector similarity (brute force). Proposed: Add LanceDB or pgvector for >10k facts.

Acceptance Criteria:

  • LanceDB integration with SQLite fallback
  • Migration path for existing embeddings
  • Benchmark: query latency vs fact count
  • Hybrid search uses HNSW when available

Blocker: Current scale (~1k facts) doesn't justify complexity. Revisit at 10k+ facts.

Repository Consolidation ✅ Complete

All individual repos have been migrated to the monorepo and archived (2026-04-01):

Repo Monorepo Location Status
zouroboros-memory-system packages/memory/ ✅ Migrated + archived
zouroboros-swarm-orchestrator packages/swarm/ ✅ Archived
zouroboros-swarm-executors packages/swarm/src/executor/ ✅ Migrated + archived
zouroboros-persona-creator packages/persona-creator/ ✅ Archived
Projects/zouroboros-rag-expansion packages/rag/ ✅ Migrated (2026-04-01)

Skills migrated into monorepo (2026-04-02):

Skill Monorepo Location Status
spec-first-interview packages/workflow/docs/spec-first-interview/ ✅ Docs + refs migrated
three-stage-eval packages/workflow/docs/three-stage-eval/ ✅ Docs + refs migrated
autoloop packages/workflow/docs/autoloop/ + src/autoloop/standalone/ ✅ Docs + standalone scripts migrated
unstuck-lateral packages/workflow/docs/unstuck-lateral/ ✅ Docs + refs migrated
zouroboros-introspect packages/selfheal/docs/introspect/ + CLI + standalone ✅ Full migration
zouroboros-prescribe packages/selfheal/docs/prescribe/ + CLI + standalone ✅ Full migration
zouroboros-evolve packages/selfheal/docs/evolve/ + CLI + standalone ✅ Full migration
zo-swarm-orchestrator packages/swarm/docs/ + src/standalone/ (20 scripts) ✅ Full migration
zo-memory-system packages/memory/docs/ + src/standalone/ (41 scripts) ✅ Full migration

All 9 skills exportable via zouroboros skills install. Standalone copies in Skills/ deprecated with notices pointing to monorepo.

Remaining standalone repos (not part of monorepo):

  • zouroboros-seedkit — Standalone installer/bootstrap kit, stays independent
  • zo-ai-character-builder — Standalone skill, stays independent

Post-Archive Remaining Steps:

Phase Action Status
1 Archive individual repos with deprecation README ✅ Done
2 Subtree merge preserving full git history Deferred — current copy-based migration is sufficient
3 Export critical open issues to monorepo ✅ No open issues on archived repos
4 Update all package.json paths ✅ Dependencies point to monorepo
5 Pin monorepo to profile; update bio Pending

User Experience

  • Web dashboard — Browser-based UI (alternative to TUI)
  • VS Code extension — Integrated Zouroboros experience
  • Zo chat shortcuts — Natural language command shortcuts
  • Interactive tutorials — Built-in guided learning paths
  • Video documentation — Tutorial series for complex features

Platform Support

  • Windows support — Native Windows compatibility for all bridges
  • macOS optimizations — Apple Silicon optimizations
  • Cloud deployment — Kubernetes Helm charts
  • Serverless adapters — AWS Lambda, Cloud Functions support

Advanced Features

  • Multi-tenant support — Isolate data for multiple users
  • Federated memory — Share memory across Zo instances
  • A/B testing framework — Compare persona/skill variants
  • Cost tracking — Detailed per-feature cost attribution
  • Carbon footprint — Energy usage tracking for AI calls

Integrations

  • GitHub integration — PR automation, code review agents
  • Slack/Discord bots — Chat-based Zouroboros access
  • Notion/Linear sync — Bidirectional task/project sync
  • Obsidian plugin — Knowledge graph integration
  • Raycast/Alfred — Quick launcher integrations

⚪ P3 — Experimental

Research Directions

  • Learned routing — RL-based executor selection
  • Neural memory — Transformer-based memory retrieval
  • Speculative execution — Predict and pre-run likely tasks
  • Self-modifying code — Safe code generation and hot-reload
  • Multi-modal agents — Vision, audio, video processing

Architecture Evolution

  • Distributed swarm — Multi-node orchestration
  • Edge deployment — Run agents on edge devices
  • Blockchain anchoring — Immutable audit logs
  • Homomorphic encryption — Private computation on encrypted data
  • Quantum-ready crypto — Post-quantum security preparation

Swarm Orchestrator (Icebox)

Heartbeat Token Limit for Bridge Executors

Status: Icebox — Monitor triggers Effort: Medium Impact: Low (no active incidents)

Add bridge-level watchdog for hung bridges and per-task token budgets.

Trigger to Revive:

  • Swarm tasks exceed 50K tokens per task
  • Executor hangs >3 per week

Evidence: Analyzed 62 swarm runs — zero runaway output incidents, only 1 hung task.

Request Caching + Deduplication

Status: Icebox Effort: Small Impact: Medium (cost reduction)

Cache identical task requests to avoid redundant execution.

Trigger to Revive:

  • Observed duplicate task execution patterns
  • Cost concerns from redundant API calls

Visual Dashboard for Real-Time Monitoring

Status: Icebox Effort: Large Impact: Medium (UX improvement)

Web dashboard showing active swarm runs, task progress, circuit breaker states.

Trigger to Revive:

  • Multiple concurrent swarm campaigns become common

Metrics

Category P2 P3 Total
ECC Research 5 0 5
Swarm 1 3 4
Memory 1 0 1
Repo Consolidation 1 0 1
UX 5 0 5
Platform 4 0 4
Advanced 5 0 5
Integrations 5 0 5
Research 0 5 5
Architecture 0 5 5
Total 27 13 40

All P0 and P1 work: Completed → see CHANGELOG.md

Contributing

To add items to this backlog:

  1. Create an issue describing the feature/bug
  2. Add to this file under appropriate priority
  3. Tag with labels: enhancement, bug, docs, research
  4. Reference related issues in the item description

Template

- [ ] **Title** - Brief description
  - Owner: @username (optional)
  - Related: #123, #456
  - Notes: Additional context

Last updated: 2026-04-01 Next review: 2026-05-01