Cross-Project Feature Map & Long-Term Plan for Gormes-Agent

Executive Summary

Analysis completed on 2026-04-30 of all 12 opensource projects in workspace-mineru:

• hermes-agent (Python upstream - 519K lines CLI, 702K lines agent runner)
• honcho (Python memory/session - Peer paradigm, 3-agent memory system)
• gbrain (TypeScript memory/runtime - Brain-first knowledge graph, minions queue)
• browser-harness (Python browser automation - CDP, Camofox, daemon lifecycle)
• go-browser-harness (Go browser automation - Chromedp, stateless actions)
• mercury-agent (TypeScript CLI/Telegram - Soul-driven, permission-hardened)
• space-agent (JavaScript browser-first - Agent reshapes interface, WebLLM)
• picoclaw (Go lightweight agent - 18+ channels, 30+ providers, MCP)
• go-agent-os references (8 donor repos - patterns for OAuth, retry, tools, state machines)

Gormes-agent current parity: ~20-30% of hermes-agent features

Master Feature Matrix

Category 1: Provider Adapters

Learnings from other projects:

• Picoclaw: 30+ providers with unified interface - could adopt pattern for provider registration
• Mercury: Provider fallback chain (DeepSeek → OpenAI → Anthropic → Grok → Ollama) - resilient routing
• GBrain: Multi-provider LLM abstraction with fallback models per provider

Category 2: Browser Automation

Learnings:

• Space-agent: Browser-first architecture with floating windows, DOM capture with typed refs ([link 12]), nested iframe support - completely different paradigm
• Hermes: 2,991-line browser_tool.py + 1,362-line browser_supervisor.py + 563-line browser_cdp_tool.py

Category 3: Tools Registry

Learnings from other projects:

• Mercury: 31 built-in tools with permission-hardened execution (shell blocklist, filesystem scoping)
• Space-agent: Skill system with metadata-driven placement (metadata.when, metadata.loaded, metadata.placement)
• Picoclaw: Cron, web search (DuckDuckGo, Baidu, Tavily, Brave, Perplexity, SearXNG), filesystem, shell, spawn
• GBrain: Cathedral II code navigation (call-graph edges, two-pass retrieval)

Category 4: Memory Systems

Learnings:

• Honcho: Three-agent memory system (Deriver/Dialectic/Dreamer) with observation levels (explicit/deductive/inductive/contradiction)
• GBrain: Brain-first lookup (5-step before external API), compiled truth + timeline pattern, tiered enrichment
• Mercury: Second Brain with 10 typed memories, hourly heartbeat consolidation, confidence/durability scoring

Category 5: Gateway/Channels

Learnings:

• Picoclaw: 18+ channels as donor repo for Go channel-edge work. Gateway Donor Map explicitly documents adaptation patterns.
• Mercury: Multi-user Telegram org model (admin/member roles, pairing codes)

Category 6: CLI/Operator Surface

Learnings:

• Mercury: Token budget system (daily enforcement, auto-concise at 70%), loop detection (5 types), permission manifest
• Picoclaw: Desktop (WebUI), Headless (TUI), Android APK, Docker, Termux deployment modes

Category 7: Security & Safety

Learnings:

• Mercury: Most mature permission system - shell blocklist, cwdOnly restriction, auto-approved patterns, YAML manifest
• Plandex (reference): Provider error classification with Retry-After parsing

Category 8: Personality/Soul Systems

Learnings:

• Mercury’s soul system: soul=heart, persona=face, taste=palate, heartbeat=breathing
• Space’s approach: single personality include file (simpler stepping stone)

Category 9: Scheduling & Background Tasks

Learnings:

• GBrain Minions: Postgres-native job queue, BullMQ-inspired, zero infra, stall detection, retry with backoff, supervisor auto-restart
• Mercury: Configurable heartbeat interval, episodic prune + second brain consolidate

Category 10: Unique Paradigms from Other Projects

Strategic Recommendations for Gormes

Immediate Actions (Next 30 Days)

1. Permission Hardening (from Mercury)

   • Port shell blocklist (36+ patterns)
   • Implement filesystem scoping (folder-level read/write)
   • Add permission approval UX (inline y/n/always)
   • Rationale: Gormes has zero permission hardening; this is foundational for safe agent operation

2. Provider Completion

   • Complete DeepSeek/Kimi reasoning isolation
   • Implement LM Studio provider
   • Add xAI/Grok provider
   • Rationale: Provider parity is critical path for Python-free normal agent turn

3. Browser Harness Gaps

   • Implement go-browser-harness doctor subcommand
   • Add daemon lifecycle (start/stop/ensure)
   • Rationale: Core browser tools work but lack production lifecycle

Short-Term (Next 90 Days)

4. Loop Detection (from Mercury)

   • Port 5-type loop detector (~200 lines of TypeScript)
   • Hard loop, failing loop, text repetition, no-action, same-tool
   • Rationale: Runaway loops are a real production problem

5. Structured Memory Enhancement (from Mercury + Honcho)

   • Extend Goncho with typed memory categories (start with 6: identity, preference, goal, habit, episode, reflection)
   • Add confidence/durability scoring
   • Implement conflict resolution (confidence wins, equal → newer)
   • Rationale: Current Goncho is session-based; structured memory is major UX improvement

6. Skill Metadata (from Space-agent)

   • Add metadata.when, metadata.loaded, metadata.placement to SKILL.md schema
   • Implement hierarchical routing skill pattern
   • Rationale: More granular skill activation control

7. Native Prompt Builder (Hermes parity)

   • Port agent/prompt_builder.py to Go
   • Context files, model-specific guidance, skills snapshots, memory/session-search assembly
   • Rationale: Critical path for Python-free normal agent turn

Medium-Term (Next 6 Months)

8. Context Compression Reconciliation (Hermes parity)

   • Reconcile with upstream 5006b220 changes
   • Tool-result pruning, protected head/tail invariants
   • Rationale: Compression behavior drift from upstream

9. GBrain Patterns (from GBrain)

   • Zero-LLM knowledge graph wiring (regex-based auto-links)
   • Brain-first lookup (5-step before external API)
   • Rationale: Significantly reduces LLM calls, compounds knowledge

10. Token Budget System (from Mercury)

    • Daily budget tracking with mutex-protected counter
    • Auto-concise at 70% threshold
    • Budget commands (/budget)
    • Rationale: Cost control for production deployments

11. Credential + OAuth (Hermes parity)

    • XDG-scoped token storage
    • Google OAuth flows
    • Rationale: Required for full provider parity

Long-Term (Next 12 Months)

12. Cathedral II Code Navigation (from GBrain)

    • Call-graph edges, two-pass retrieval
    • 5 commands: code-callers, code-callees, code-def, code-refs, query –near-symbol
    • Rationale: Code-aware agent capabilities

13. Minions Job Queue (from GBrain)

    • Postgres-native BullMQ-inspired queue
    • Parent-child DAGs, stall detection, retry with backoff
    • Rationale: Background task durability, subagent coordination

14. Soul/Personality System (Phase 6)

    • User-owned markdown personality files
    • soul.md (heart), persona.md (face), taste.md (palate), heartbeat.md (breathing)
    • Rationale: Planned for gormes Phase 6; Mercury’s implementation is reference

15. Web Dashboard (Hermes parity)

    • TypeScript/React web UI (Hermes has full web/ directory)
    • Session management, skills, chat, config, logs
    • Rationale: Hermes has 191K-line TUI gateway server

16. Multi-Memory Backends (from Honcho)

    • Turbopuffer, LanceDB vector store options
    • Redis caching layer
    • Rationale: Scale beyond single-node SQLite

Not Recommended for Port

• Space’s browser-first architecture (fundamentally different runtime model)
• Space’s WebLLM (requires browser runtime)
• Mercury’s Telegram org model (Gormes gateway channels are different abstraction)
• Picoclaw’s ultra-lightweight constraints (different target: $10 hardware vs server deployment)
• Full Kubernetes ACP controllers (from references) - use local state machine instead

Reference Implementation Priority

From references/go-agent-os/:

Success Metrics

• 30 days: Permission hardening shipped, provider parity >80%, browser harness doctor working
• 90 days: Loop detection, structured memory (6 types), skill metadata, native prompt builder
• 6 months: Context compression reconciled, GBrain patterns (auto-links, brain-first), token budget
• 12 months: Cathedral II, Minions queue, Soul system, Web dashboard, multi-memory backends

Current parity: ~20-30% → Target: 80%+ within 12 months