Gormes

Must-Have Features — Long-Term Plan

Comprehensive feature catalogue derived from 12+ open-source AI agent projects. Every feature a production Go agent runtime must ship.

Must-Have Features — Long-Term Plan

Source scope: 12+ open-source projects studied April 2026. Projects: hermes-agent (Python upstream), honcho (Python memory), gbrain (TypeScript memory/runtime), browser-harness / go-browser-harness, mercury-agent (TypeScript), space-agent (JS browser-first), picoclaw (Go lightweight agent), MLT-OSS/hermes-agent-go (Go port), trpc-agent-go, AgenticGoKit, claw-prime, langchaingo, 8 go-agent-os reference repos. Audience: gormes-agent planner, builder, and reviewer skills. Not end-user marketing. Relationship to progress.json: This document names the features gormes must eventually have. The canonical execution queue is progress.json. This document is the why and the what. progress.json is the when and the how.


Feature Architecture

Every feature is classified by three axes:

AxisValuesMeaning
Criticalityfoundational / production / differentiator / commodityHow essential the feature is to gormes’s existence
Maturityshipped / in-flight / planned / gapCurrent implementation status
Complexitytrivial / moderate / large / subsystemImplementation effort required

A feature is must-have if it scores foundational on criticality OR production on criticality with gap maturity. Differentiators are must-have when they define gormes’s unique value. Commodity features are must-have only when they block user adoption.


1. Provider Adapters — The Model Surface

Donor projects: hermes-agent, picoclaw, mercury-agent, gbrain, trpc-agent-go, langchaingo Current parity: ~70% (7/10 major providers) Criticality: foundational

Every provider gormes cannot speak is a provider the operator cannot use without running Python Hermes alongside. This is not a nice-to-have. It is the single biggest block to a Python-free agent turn.

1.1 Must-Have Provider Completion

ProviderStatusRationale
Anthropic Messages APIShippedPrimary provider for Gormes development
OpenAI Chat CompletionsShippedLargest model ecosystem
Codex Responses APIPartialRequired for Codex CLI parity
AWS Bedrock (Converse + Anthropic)In-flightEnterprise deployment vector
Gemini / Google Cloud CodePartialGoogle ecosystem
Google Code AssistPartialGoogle developer tooling
OpenRouterPartialMulti-model routing; 250+ models
DeepSeek / Kimi reasoningPartialReasoning-model isolation needed
Moonshot / Kimi tool schemaPartialChinese model ecosystem
xAI / GrokGapGrowing provider, no Go adapter
LM Studio / local inferenceGapOffline-capable agent runtime
Azure Foundry (OpenAI + Anthropic)In-flightEnterprise Azure deployment

1.2 Provider Infrastructure (Must-Have)

FeatureSourceStatus
Provider-neutral stream contracthermes-agent, gbrainShipped
Tool-call normalization + continuationhermes-agentShipped
Provider registry + alias manifesthermes-agent, picoclawShipped
Retry-After header parsing + backoffplandex (go-agent-os)Shipped
Provider error taxonomy (rate-limit vs auth vs capacity)plandex, hermes-agentShipped
Prompt-cache capability guardhermes-agentShipped
Provider rate guard + degraded-state evidencehermes-agentShipped
Provider timeout config fail-closedhermes-agentShipped
Cross-provider reasoning-tag sanitizationhermes-agentShipped
Tool-call argument repair + schema sanitizerhermes-agentShipped
Model metadata registry (context limits, pricing, capabilities)hermes-agent, gbrainIn-flight
Smart model routing with fallback chainmercury-agent (DeepSeek→OpenAI→Anthropic→Grok→Ollama)In-flight
Per-turn model selection + reasoning effort propagationhermes-agentShipped
Provider account usage read model (/usage)hermes-agentShipped

2. Native Prompt Assembly — The Agent’s Brain

Donor projects: hermes-agent (prompt_builder.py), mercury-agent (SOUL/persona system), space-agent (prompt-items.js) Current parity: ~40% Criticality: foundational

Without native prompt assembly, the Go runtime cannot construct a correct Hermes-compatible system prompt. Every turn depends on Python’s prompt builder. This is the critical path for Python-free operation.

2.1 System Prompt Block Assembly

BlockSourceStatus
SOUL.md identity + AGENTS.md project contexthermes-agent, mercury-agent (soul.md)Shipped
USER.md + MEMORY.md durable user contexthermes-agent, space-agent (plain markdown includes)Shipped
Tool/memory guidance constants (byte-equivalent port)hermes-agent (prompt_builder.py)Shipped
Skills guidance block (toolset-aware)hermes-agent, space-agent (metadata-driven placement)In-flight
Model-specific role + tool-use enforcementhermes-agentShipped
Timestamp + model/provider/session metadatahermes-agent (run_agent.py:3770-3779)Shipped
Platform/session context blockhermes-agentShipped
Ephemeral memory/plugin recall (injected into user turn)hermes-agentGap
WSL environment hinthermes-agentShipped
Developer-role model routinghermes-agentShipped

2.2 Prompt Budgeting (Must-Have from space-agent design)

FeatureSourceStatus
Prompt contributors as keyed, budgeted itemsspace-agent (prompt-items.js)Gap
Cached token counts per contributorspace-agentGap
Trim policy with runtime read-back seamspace-agentGap
Long-context trimming without silent text dropspace-agentGap

3. Context Compression — The Long Session Engine

Donor projects: hermes-agent (ContextEngine), gbrain (tiered enrichment) Current parity: ~60% Criticality: production

Real agent sessions run long. Without compression, the context window fills and the agent degrades. Hermes’s ContextEngine is a well-defined contract. Gormes must port it completely.

3.1 Compression Engine

FeatureSourceStatus
ContextEngine interface + status contracthermes-agentShipped
Token-budget trigger + summary sizinghermes-agentShipped
Protected head/tail message invariantshermes-agentShipped
Tool call/result pairs kept togetherhermes-agentShipped
Compression lineage recordhermes-agentShipped
Auxiliary compression headroom for system + tool schemashermes-agentShipped
Provider-aware context cap resolutionhermes-agentShipped
Tool-result pruninghermes-agentShipped
Image-token budget chargehermes-agentShipped
Compression-boundary callback (notify kernel)hermes-agentIn-flight
Manual compression feedback + context referenceshermes-agentGap
Kernel compression-boundary callback bindinghermes-agentGap
Recalculate budgets on model-switchhermes-agent (2026-04-25 sync)Gap

4. Gateway + Channels — The Operator Surface

Donor projects: hermes-agent (gateway/run.py), picoclaw (18+ channels), mercury-agent (permission-hardened Telegram) Current parity: ~60% (5/19 channels shipped; shared chassis in place)

4.1 Shared Gateway Infrastructure (Must-Have)

FeatureSourceStatus
Shared gateway chassis (multi-channel)hermes-agent, gormesShipped
Slash command registry + active-turn policyhermes-agentShipped
Session store + SessionSource parityhermes-agentShipped
Session metadata refresh (created, activity, platform, title)hermes-agentShipped
Gateway auto-title generationhermes-agentShipped
Gateway /status Hermes-format closeouthermes-agentShipped
Gateway /title manual title commandhermes-agentShipped
Gateway session token accountinghermes-agentShipped
Gateway stream/tool trace formattinghermes-agentShipped
Gateway typing-action wiring during streamhermes-agentShipped
Placeholder edit-failure fallback hardeninghermes-agentShipped
Gateway fresh-final send/delete fallbackhermes-agentShipped
Gateway message deduplicatorhermes-agentShipped
Gateway inbound dedup manager bindinghermes-agentShipped
Notify-to delivery routinghermes-agentGap
Channel directory refresh + stale-target invalidationhermes-agentGap
Mirror + sticker cache surfaceshermes-agentGap

4.2 Channel Adapters (Must-Have by Priority)

ChannelStatusPriority Rationale
TelegramShippedPrimary dogfood channel
DiscordShippedDeveloper community channel
SlackShippedEnterprise channel
WhatsAppShippedGlobal messaging
WeChat (WeCom + WeiXin)ShippedChinese enterprise
BlueBubbles (iMessage)ShippedApple ecosystem
Home AssistantShippedSmart home integration
SignalPartialPrivacy-focused users
Email + SMSShippedFallback delivery
MatrixGapOpen federated protocol
MattermostGapSelf-hosted enterprise
Feishu / LarkPartialChinese enterprise
DingTalkPartialChinese enterprise
QQPartialChinese consumer
LINEGapJapanese market
VKGapRussian market
IRCGapLegacy community
YuanbaoPartialChinese AI platform
Webhook (generic)ShippedUniversal integration

4.3 Gateway Active-Turn Policy (Must-Have from Hermes Design)

FeatureSourceStatus
/stop interrupthermes-agentShipped
/queue waithermes-agentShipped
/steer mid-turn injectionhermes-agentShipped
Busy command guard for compression + long CLI actionshermes-agentShipped
Gateway-handled slash commands bypass active-session guardhermes-agentShipped
Mid-run steer injection between tool callshermes-agentShipped
Hardline command blocks (yolo/approval-off/cron)hermes-agent (2026-04-26 sync)Partial
Approvals bypass interrupt semanticshermes-agentGap

5. Tool Registry — The Agent’s Hands

Donor projects: hermes-agent (40+ tools), mercury-agent (31 tools, permission-hardened), picoclaw (Go native tools), gbrain (operation catalog) Current parity: ~25% (15/61 tools) Criticality: foundational

5.1 Tool Infrastructure (Must-Have Before Tool Breadth)

FeatureSourceStatus
Typed Tool interface + in-process registryhermes-agent, gormesShipped
Tool descriptor layer (toolset, availability, mutating flag, trust class, timeout, result budget, audit kind, prompt-visible flag)gbrain (OperationSpec), hermes-agentGap
Streamed tool_calls accumulationhermes-agentShipped
Kernel tool loop (90-turn budget, summary grace turn)hermes-agentShipped
Doctor verification (--offline)hermes-agent, gbrainShipped
Tool registry inventory + schema parity harnesshermes-agentShipped
Executor-enforced trust class + timeout before handler runsgbrain (trust-class enforcement)Gap

5.2 Core Task Tools (Must-Have for Agent Utility)

ToolSourceStatus
read_filehermes-agentShipped
search_fileshermes-agentShipped
write_filehermes-agentShipped
patchhermes-agentShipped
terminal (foreground)hermes-agentShipped
Terminal (background, PTY)hermes-agentPartial
todo (session-bound)hermes-agentShipped
session_search (session-bound)hermes-agentShipped
clarifyhermes-agentShipped
Checkpoint manager (atomic snapshot, shadow-repo)hermes-agentIn-flight
Code execution (sandboxed)hermes-agentShipped
Process registryhermes-agentGap
Fuzzy patch/lint/checkpoint restorehermes-agentGap

5.3 Web + Browser Tools

ToolSourceStatus
Browser navigate/snapshot/click/type/scrollhermes-agent, go-browser-harnessShipped
Browser console/vision/CDPhermes-agent, go-browser-harnessShipped
Browser dialog/get-imageshermes-agent, go-browser-harnessShipped
Web search (DuckDuckGo, Tavily, Brave, Perplexity, SearXNG)hermes-agent, picoclawShipped
Web extract + crawlhermes-agentShipped
Browser daemon lifecycle (start/stop/ensure)hermes-agentGap
Browser profile managementhermes-agentGap
Browser tab enumeration/switchhermes-agentGap
Browser doctor/diagnosticshermes-agentGap
Browser event drainhermes-agentGap
Browser Use cloud + Go harness bridgehermes-agent, go-browser-harnessShipped
HTTP GET (no-browser)hermes-agentGap
Iframe targetinghermes-agentGap

5.4 Media + Vision Tools

ToolSourceStatus
Image generationhermes-agentShipped
Image input mode router + native content partshermes-agentShipped
Vision analysishermes-agentGap
TTS synthesis + voice-mode statehermes-agentGap
Transcription toolhermes-agentShipped

5.5 Delegation + Subagents

FeatureSourceStatus
Deterministic subagent runtime (goroutines)hermes-agent, gormesShipped
Max-depth guard + bounded batch executionhermes-agent, gormesShipped
Timeout + cancellation scopes (context.Context)hermes-agent, gormesShipped
Typed result envelope + append-only run loghermes-agent, gormesShipped
Runner-enforced tool allowlists + blocked-tool policyhermes-agent, gormesShipped
Interrupt propagation to concurrent-tool workershermes-agent, gormesShipped
Child-agent message/tool-call replay from ledgergbrain (subagent_messages)Shipped
Durable subagent/job ledger (SQLite)gbrain (Minions), gormesShipped
Background review toolset restrictionhermes-agentShipped

5.6 Scheduling + Background Tasks

FeatureSourceStatus
robfig/cron scheduler + bbolt/SQLite job storehermes-agentShipped
Heartbeat [SYSTEM:] + [SILENT] delivery contracthermes-agentShipped
Durable job queue (claim, renew, complete, fail, retry, cancel)gbrain (Minions), gormesShipped
Durable job backpressure + timeout auditgbrain, gormesShipped
Durable worker supervisor status seamgbrain, gormesShipped
Durable pause/resume intent contractgbrain, gormesShipped
Durable replay and inbox message contractgbrain, gormesShipped
Cronjob tool API + schedule parser parityhermes-agentGap
Parent-child DAGs with completion policiesgbrain (Minions)Gap
Per-job token accountinggbrainGap
Cron multi-target deliveryhermes-agentShipped
Cron context_from output chaininghermes-agentShipped

5.7 MCP (Model Context Protocol)

FeatureSourceStatus
MCP client (stdio + HTTP transport)hermes-agent, gormesShipped
MCP tool discovery + schema normalizationhermes-agent, gormesShipped
MCP OAuth state store + token refreshhermes-agent, gormesShipped
Goncho MCP tool cataloggormesShipped
Gormes-native MCP host runtime boundarygormesShipped

6. Memory System — Goncho / Honcho Parity

Donor projects: honcho (3-agent memory, peer paradigm), gbrain (knowledge graph, brain-first lookup), mercury-agent (Second Brain, 10 typed memories), hermes-agent (MemoryProvider interface) Current parity: Shipped (Goncho Phase 3 complete) Criticality: production (core is shipped; enhancements are differentiators)

6.1 Shipped Memory Foundation

FeatureStatus
SQLite + FTS5 lattice + schema migrationsShipped
Ontological graph + LLM extractor + dead-letter queueShipped
Neural recall (2-layer seed selection, CTE traversal)Shipped
Semantic fusion (Ollama embeddings, cosine similarity, hybrid)Shipped
Memory mirror (USER.md async export)Shipped
Session index mirror (bbolt→index.yaml)Shipped
Tool execution audit log (JSONL)Shipped
Transcript export (gormes session export)Shipped
Extraction state visibility (gormes memory status)Shipped
Insights audit log (daily usage.jsonl)Shipped
Memory decay (relationships.last_seen + weight attenuation)Shipped
Cross-chat synthesis (user_id concept, opt-in cross-chat recall)Shipped
Session lineage + cross-source searchShipped
Honcho-compatible scope/source tool schemasShipped
Honcho host + SillyTavern integration fixturesShipped
Goncho configuration namespace + dreaming schedulerShipped
Goncho streaming chat persistenceShipped
Hermes memory tool over Goncho/local storeShipped
Goncho drop-in compatibility (keys, webhooks, HTTP, CLI, SDK)Shipped

6.2 Memory Enhancements (Must-Have Differentiators)

FeatureSourceStatus
Typed memory categories (identity, preference, goal, habit, episode, reflection)mercury-agent (10 types), honchoGap
Confidence + durability scoring per observationmercury-agent, honchoGap
Conflict resolution (confidence wins; equal→newer)mercury-agent, honchoGap
Auto-pruning (21-day stale memories)mercury-agentGap
Zero-LLM knowledge graph wiring (regex-based auto-links)gbrainGap
Brain-first lookup (5-step before external API)gbrainGap
Retrieval eval harness (precision@k, recall@k, cross-chat negative tests)gbrainGap
Source-aware retrieval ranking (curated/reviewed boost, bulk-source damping)gbrainGap
Compiled truth + timeline pattern (per-entity page)gbrainGap
Directional peer cards (users AND agents as peers)honchoGap
Multi-memory backends (Turbopuffer, LanceDB, Redis cache layer)honchoGap
Three-agent memory loop (Deriver/Dialectic/Dreamer)honchoGap

7. CLI / Operator Surface — The Control Panel

Donor projects: hermes-agent (49-file CLI tree), mercury-agent (token budget, loop detection) Current parity: ~50% Criticality: production

7.1 CLI Commands (Must-Have by Function)

CommandSourceStatus
gormes (TUI, oneshot, gateway modes)hermes-agentShipped
gormes doctorhermes-agentShipped
gormes gateway status/start/stophermes-agentShipped
gormes session exporthermes-agentShipped
gormes memory statusgormesShipped
gormes config (check, edit, migrate)hermes-agentShipped
gormes auth (login, logout, status, add)hermes-agentIn-flight
gormes model (interactive picker)hermes-agentShipped
gormes setup (wizard)hermes-agentShipped
gormes profilehermes-agentShipped
gormes agent-template-resethermes-agentShipped
gormes mcp loginhermes-agentShipped
gormes backup / gormes restorehermes-agentGap
gormes logshermes-agentGap
gormes web-serverhermes-agentGap
gormes uninstallhermes-agentGap
gormes cron management CLIhermes-agentGap
gormes plugins managementhermes-agentGap
Diagnostics, backup, and status CLI closeouthermes-agentGap

7.2 API Server (Must-Have)

FeatureSourceStatus
OpenAI-compatible chat-completions APIhermes-agentShipped
Responses API store + run event streamhermes-agentShipped
API server disconnect snapshot persistencehermes-agentShipped
Gateway proxy mode forwardinghermes-agentShipped
Dashboard API client contracthermes-agentShipped
API server detailed health endpointhermes-agentShipped
API server cron admin endpointshermes-agentShipped
SSE streaming to Bubble Tea TUIgormesShipped

7.3 TUI (Terminal UI)

FeatureSourceStatus
Bubble Tea shell + 16ms coalescing mailboxgormesShipped
SSE reconnectgormesShipped
Hermes skin token rendererhermes-agent, gormesShipped
Hermes status bar + bottom-pinned chrome layouthermes-agent, gormesShipped
Native TUI slash-command dispatchgormesShipped
Native TUI /save, /branch, conversation viewportgormesShipped
Web dashboard (TypeScript/React parity with Hermes UI)hermes-agentGap

8. Security + Safety — The Trust Boundary

Donor projects: mercury-agent (permission system), hermes-agent (approval, URL safety, Tirith), gbrain (trust-class enforcement) Current parity: ~30% Criticality: foundational

Gormes has zero permission hardening. This is the most critical gap in the entire project. An agent that can execute arbitrary shell commands, write to any filesystem path, and make arbitrary network requests is not safe to operate.

8.1 Must-Have Security Features

FeatureSourceStatus
Shell blocklist (36+ dangerous patterns)mercury-agentGap
Filesystem scoping (folder-level read/write restrictions)mercury-agentGap
Permission approval UX (inline y/n/always)mercury-agent, hermes-agentGap
Hardline command pattern detection (non-recoverable host destruction)hermes-agent, gormesIn-flight
Recoverable dangerous patterns + blocked-result schemahermes-agent, gormesIn-flight
Approval mode config normalizationhermes-agent, gormesIn-flight
Gateway hook auto-accept strict parserhermes-agent, gormesShipped
Subagent dangerous-command non-interactive policygormesShipped
URL safety / SSRF guardhermes-agentPartial
Path security (traversal prevention, workspace confinement)hermes-agentGap
Credential redaction in logs and telemetryhermes-agentPartial
Website policy (per-domain allow/block)hermes-agentGap
Cron dangerous-command approval modehermes-agentGap
Tirith security integrationhermes-agentGap
OSV vulnerability checkhermes-agentGap
Trust-class enforcement in shared tool executorgbrain, mercury-agentGap
WhatsApp identity safety (ASCII-only, anti-traversal)hermes-agent (2026-04-27 sync)Shipped

8.2 Operator Safety (Must-Have from Mercury)

FeatureSourceStatus
Loop detection (5 types: hard/failing/text/no-action/same-tool)mercury-agentGap
Token budget system (daily + per-session, auto-concise at 70%)mercury-agentGap
Permission manifest (YAML-based, per-agent allow/deny)mercury-agentGap
Filesystem operation preflight (batch validation before writes)space-agentGap

9. Skills System — The Agent’s Knowledge

Donor projects: hermes-agent (skills_hub.py, 118K lines), gbrain (fat skills, thin harness), space-agent (metadata-driven skills), mercury-agent (skill preprocessing) Current parity: ~40% Criticality: differentiator

9.1 Skills Infrastructure

FeatureSourceStatus
SKILL.md parsing + active storehermes-agent, gormesShipped
Deterministic selection + prompt injection blockhermes-agent, gormesShipped
Kernel injection + usage loghermes-agent, gormesShipped
Inactive candidate drafting + explicit promotionhermes-agent, gormesShipped
SKILL.md frontmatter validation guardgbrain, gormesShipped
Hermes creative skill metadata compatibilityhermes-agent, gormesShipped
Skill preprocessing + dynamic slash commandshermes-agent, gormesShipped
Skills hub search (in-memory registry)hermes-agent, gormesShipped
Skills list (enabled/disabled status + filter)gormesShipped
Metadata-driven skill placement (when, loaded, placement, context tags)space-agentGap
Skill conformance checks + routing evalsgbrainGap
Skill conflict detection (overlapping triggers)gbrainGap
Bundled productivity skills (Airtable, Spotify, TouchDesigner, Google Meet)hermes-agentIn-flight
Skill registries (external repo fetching)hermes-agentGap
Skill confidence/feedback scoringhermes-agentGap
Skill versioning + review historygbrainGap

10. Sandboxing + Execution Environments

Donor projects: hermes-agent (Docker, Modal, Daytona, Singularity), picoclaw (7 terminal backends) Current parity: ~30% Criticality: production (Docker is must-have; others are differentiators)

BackendSourceStatusPriority
Dockerhermes-agentIn-flightMust-have
Environment interface + file sync contracthermes-agent, gormesShippedMust-have
Modalhermes-agentGapDifferentiator
Daytonahermes-agentGapDifferentiator
Singularityhermes-agentGapDifferentiator
Local (SSH, direct exec)hermes-agent, picoclawShippedShipped
Code execution mode policy (strict/project/default resolver)hermes-agentGapMust-have

11. Packaging + Distribution

Donor projects: hermes-agent, picoclaw (Desktop/Headless/Android/Docker/Termux), claw-prime Current parity: ~60% Criticality: production

ArtifactStatusPriority
Single Go binary (Linux, macOS, Windows)ShippedFoundational
OCI / Docker imageShippedMust-have
Homebrew formulaShippedMust-have
Unix installer (install.sh) source-backed update flowGapMust-have
Windows installer (install.ps1 + install.cmd)GapMust-have
Installer site asset/route coverageGapMust-have
Debian/RPM packagesGapCommodity
Android APK / TermuxGapCommodity

12. Soul / Personality System — Phase 6

Donor projects: mercury-agent (soul.md, persona.md, taste.md, heartbeat.md), space-agent (personality.system.include.md) Current parity: Not started (Phase 6) Criticality: differentiator

FeatureSourceStatus
SOUL.md identity (heart)mercury-agentPlanned
persona.md behavior (face)mercury-agentPlanned
taste.md preferences (palate)mercury-agentPlanned
heartbeat.md lifecycle (breathing)mercury-agentPlanned
User-owned markdown files (operator-editable)mercury-agent, space-agentPlanned
Plain-text memory includes as first-classspace-agentPlanned
Git-backed personality time travel (rollback/revert)space-agentPlanned

13. Learning Loop — Phase 6

Donor projects: gbrain (skill extraction, maintenance cycles), hermes-agent (skill generation) Current parity: Not started (Phase 6) Criticality: differentiator

FeatureSourceStatus
Complexity detector (heuristic/LLM signal)hermes-agent, gormesGap
Skill extractor (LLM-assisted pattern distillation)hermes-agent, gormesGap
Hybrid lexical + semantic skill lookupgbrain, gormesGap
Skill effectiveness scoring + feedback loopgbrain, gormesGap
TUI + Telegram skill browsinggormesIn-flight
Delta-bounded skill/memory maintenance (changed-source IDs only)gbrain (e2961c0)In-flight
Source-aware retrieval damping fixturesgbrainGap
Code Cathedral II (call-graph edges, two-pass retrieval)gbrainGap
Incremental extract + embed-stale filtersgbrain (v0.22.1)Gap
Per-source sync anchor (source-scoped last-commit)gbrain (v0.22.5)Gap
Fail-improve loop (logs regex failures, generates better patterns)gbrainGap

14. Observability + Telemetry

Donor projects: hermes-agent (logs, status, tips), gbrain (doctor, degraded mode), mercury-agent (heartbeat) Current parity: ~40% Criticality: production

FeatureSourceStatus
gormes doctor --offlinehermes-agentShipped
gormes gateway status (read-only)hermes-agent, gormesShipped
gormes memory status (extractor queue, dead-letter)gormesShipped
Tool execution audit log (JSONL)gormesShipped
Insights audit log (daily usage.jsonl)gormesShipped
CLI log redactor (secret shapes)hermes-agentShipped
CLI log snapshot readerhermes-agentShipped
CLI dump support-summary helperhermes-agentShipped
Visible degraded mode (every fallback reported in status/doctor/audit)gbrain, mercury-agentGap
Semantic recall disabled warning (no embedding model)gbrainGap
Extractor queue depth + dead-letter countgbrain, gormesShipped
Graph extraction agegbrainGap
Skill resolver warningsgbrainGap
Durable job stalled countgbrain, gormesShipped
Child-agent replay not available warninggbrainGap
Tips system (deterministic, seen-state)hermes-agent, gormesShipped
Web dashboard (session, skills, chat, config, logs)hermes-agentGap

15. Plugin Architecture

Donor projects: hermes-agent (plugin SDK, hooks, tools, slash commands), gbrain (isolated MCP plugins) Current parity: ~40% Criticality: commodity (must-have infrastructure; plugins themselves are differentiators)

FeatureSourceStatus
Plugin SDKhermes-agentShipped
Dashboard theme/plugin slot inventorygormesShipped
First-party Spotify plugin fixturehermes-agent, gormesShipped
First-party Google Meet plugin fixturehermes-agent, gormesShipped
Third-party extension systemhermes-agentGap
Plugin isolation (subprocess/WASM for untrusted)gbrain, hermes-agentGap
Plugin manifest + capability declarationgbrainGap
Operator enablement required before activationgbrainGap
Hindsight memory setuphermes-agent, gormesGap

16. Mixture of Agents

Donor projects: hermes-agent (multi-model coordination), gbrain (subagent orchestration) Current parity: Gap Criticality: differentiator

FeatureSourceStatus
Multi-model coordinationhermes-agentGap
Agent ensemble with voting/consensushermes-agentGap
Model-specific task routinghermes-agentGap

Cross-Cutting Must-Have Patterns (from All Projects)

These are not features but architectural patterns that every subsystem must adopt:

PatternSourceApplied To
Contract-first operations (one descriptor → CLI, MCP, tool schemas, doctor)gbrainTools, commands, gateways
Trust-class enforcement in shared executor (operator/gateway/child/system)gbrain, hermes-agentAll tool execution
Visible degraded mode (every fallback reported)gbrain, mercury-agentAll subsystems
Provider-neutral events (adapters own quirks; kernel sees one contract)hermes-agentProvider layer
Thin harness, fat skills (intelligence in skills, not runtime)gbrainSkills system
Hermetic fixtures over live tests (replayable, no live credentials)gbrain, hermes-agentAll tests
Metadata-driven discovery (not hardcoded branches)space-agentSkills, tools, commands
Prompt contributors as keyed, budgeted itemsspace-agentPrompt assembly
Goroutines + channels (not asyncio threads)All Go projectsConcurrency
Context.Context propagation (cancellation, deadlines)All Go projectsAll I/O
SQLite + FTS5 first (Postgres optional behind interface)gbrain, gormesMemory, jobs, state
Single binary, zero runtime depsclaw-prime, gormesDistribution
Durable job ledger (restartable work, not ephemeral goroutines)gbrainCron, subagents, long work
Provenance-rich memory (source, extractor, confidence, freshness)gbrain, honchoGoncho relationships
Skills as reviewed code (frontmatter, conformance, routing evals)gbrain, hermes-agentSkills store

Feature Gap Summary by Criticality

Foundational Gaps (Block Python-Free Agent Turn)

#FeatureSource ProjectEstimated Effort
F1Provider completion (xAI/Grok, LM Studio, DeepSeek, Moonshot)hermes-agent, picoclaw4-6 weeks
F2Native prompt builder assembly (skills snapshot, memory/search guidance)hermes-agent2-3 weeks
F3Tool descriptor layer (toolset, trust class, timeout, audit, budget)gbrain2-3 weeks
F4Shell blocklist + filesystem scopingmercury-agent2-3 weeks
F5Permission approval UX (inline y/n/always)mercury-agent, hermes-agent2-3 weeks
F6Trust-class enforcement in shared executorgbrain1-2 weeks

Production Gaps (Block Real-World Operation)

#FeatureSource ProjectEstimated Effort
P1Context compression complete (manual feedback, model-switch recalculation)hermes-agent2-3 weeks
P2Loop detection (5 types)mercury-agent1-2 weeks
P3Token budget system + auto-concisemercury-agent2-3 weeks
P4URL safety / SSRF guard completehermes-agent1-2 weeks
P5Browser daemon lifecycle + doctorhermes-agent, go-browser-harness2-3 weeks
P6Docker sandbox backendhermes-agent1-2 weeks
P7Code execution mode policy (strict/project/default)hermes-agent2-3 weeks
P8CLI closeout (backup, logs, web-server, uninstall, diagnostics)hermes-agent3-4 weeks
P9Packaging closeout (install.sh, install.ps1, install.cmd)hermes-agent2-3 weeks
P10Visible degraded mode in all subsystemsgbrain, mercury-agentOngoing
P11Channel adapters (Matrix, Mattermost, LINE, IRC, VK, Feishu/DingTalk closeout)hermes-agent, picoclaw4-6 weeks

Differentiator Gaps (Define Gormes’s Unique Value)

#FeatureSource ProjectEstimated Effort
D1Typed memory categories + confidence scoringmercury-agent, honcho3-4 weeks
D2Zero-LLM knowledge graph wiring (regex auto-links)gbrain2-3 weeks
D3Brain-first lookup (5-step before external API)gbrain2-3 weeks
D4Retrieval eval harness (precision@k, recall@k)gbrain2-3 weeks
D5Metadata-driven skill placement (when, loaded, context tags)space-agent2-3 weeks
D6Minions DAG job dependencies (parent-child with policies)gbrain3-4 weeks
D7Soul/personality system (soul.md, persona.md, taste.md, heartbeat.md)mercury-agent2-3 weeks
D8Learning loop (complexity detector, skill extractor, feedback)hermes-agent, gbrain6-8 weeks
D9Code Cathedral II (call-graph edges, two-pass retrieval)gbrain4-6 weeks
D10Web dashboard (TypeScript/React parity with Hermes UI)hermes-agent6-8 weeks
D11Multi-memory backends (Turbopuffer, LanceDB, Redis)honcho3-4 weeks
D12Three-agent memory loop (Deriver/Dialectic/Dreamer)honcho4-6 weeks
D13Mixture of agents (multi-model coordination)hermes-agent6-8 weeks

Immediate (Next 30 Days) — Close the Safety Gap + Provider Gap

Week 1-2: Provider completion (xAI/Grok, LM Studio, DeepSeek/Kimi)
Week 1-2: Tool descriptor layer (OperationSpec with trust classes)
Week 2-3: Prompt builder assembly final closeout
Week 3-4: Shell blocklist + filesystem scoping + permission approval
Week 3-4: Trust-class enforcement in tool executor

Short-Term (Next 90 Days) — Production Hardening

Week 5-6: Context compression complete (manual feedback, recalculation)
Week 5-6: Loop detection (5 types)
Week 7-8: Token budget system + auto-concise
Week 7-8: Docker sandbox backend
Week 9-10: Browser daemon lifecycle + doctor
Week 9-10: Code execution mode policy
Week 11-12: CLI closeout (backup, logs, diagnostics)
Week 11-12: Packaging closeout (installers)

Medium-Term (Next 6 Months) — Differentiators

Month 3: Typed memory categories + confidence scoring
Month 3: Zero-LLM knowledge graph wiring
Month 4: Brain-first lookup + retrieval eval harness
Month 4: Minions DAG job dependencies
Month 5: Metadata-driven skill placement
Month 5: Soul/personality system
Month 6: Channel adapters (Matrix, Mattermost, LINE, IRC)

Long-Term (Next 12 Months) — Capstone Features

Month 7-8: Learning loop (skill extraction, feedback)
Month 8-9: Web dashboard
Month 9-10: Code Cathedral II
Month 10-11: Multi-memory backends
Month 11-12: Three-agent memory loop + Mixture of agents

Features studied but explicitly excluded from the must-have list:

FeatureSourceReason
Browser-first runtime (agent lives in browser)space-agentFundamentally different runtime model
WebLLM / HuggingFace browser-side inferencespace-agentRequires browser runtime, not Go
Mercury’s Telegram org model (admin/member roles)mercury-agentGormes gateway channels use different abstraction
Picoclaw’s ultra-lightweight constraints ($10 hardware)picoclawDifferent deployment target
Full Kubernetes ACP controllersgo-agent-os referencesUse local state machine instead
TypeScript indexer / tree-sitter WASM in runtimegbrainKeep runtime small; optional code-context evidence only
Hermes’s dynamic Python import-time tool registrationhermes-agentExplicit registration is safer for Go
Hermes’s sync/async/thread complexityhermes-agentGo goroutines + context.Context are sufficient
Hermes’s Postgres-backed cronhermes-agentSQLite-backed is correct for single-binary promise

References


Generated: April 30, 2026 Source: 12+ open-source projects, 7-phase gormes roadmap, cross-project feature map, upstream lessons