Features¶

Last updated: 2026-07-29

Legend¶

0. Authentication and Authorization¶

• ✅ Google-only Firebase sign-in with persistent, automatically refreshed sessions
• ✅ Adaptive shadcn/Tailwind sign-in page with desktop image panel
• ✅ Protected frontend routes, destination restoration, user navigation, and sign-out
• ✅ FastAPI Bearer-token verification for REST and streaming APIs
• ✅ Public /api/health and /api/ready; all other APIs authenticated
• ✅ Atomic Firestore users/{uid} profile bootstrap from verified claims
• ✅ Owner-only Firestore Rules with immutable identity/audit fields
• ✅ Sanitized success, failure, and session-expiry toasts; no raw auth errors
• 📋 Roles, tenant isolation, App Check, and MFA are future authorization milestones

1. Agent Runtime (DeepAgents)¶

1.1 Runtime Factory¶

• ✅ get_deep_agent_runtime() — Creates compiled DeepAgents graph
• ✅ create_deep_agent with model, system prompt, backend, permissions, subagents
• ✅ SqliteSaver checkpointer for state persistence
• ✅ interrupt_on configuration for write/edit/delete operations
• ✅ System prompt loaded from instructions/global-agent-instructions.md

1.2 Backend Configuration¶

• ✅ CompositeBackend with route-based filesystem backends
• ✅ StateBackend as default for thread-scoped scratch state
• ✅ /workspace/ — FilesystemBackend (virtual mode)
• ✅ /kb/ — FilesystemBackend (virtual mode)
• ✅ /instructions/ — FilesystemBackend (virtual mode)
• ✅ /memories/ — FilesystemBackend (virtual mode)
• ✅ /skills/ — FilesystemBackend (virtual mode)

1.3 Permissions Model¶

• ✅ Path-based FilesystemPermission rules
• ✅ Read/write on /workspace/** and /memories/**
• ✅ Read-only on /kb/**, /instructions/**, /skills/**
• ✅ Interrupt on /workspace/submissions/** and /workspace/final/**
• ✅ Deny write on /kb/**, /instructions/**, /skills/**
• ✅ Deny all external paths

1.4 Middleware Stack¶

• ✅ TodoListMiddleware — Task planning and progress tracking
• ✅ SkillsMiddleware — Skill file loading
• ✅ FilesystemMiddleware — With permissions applied
• ✅ SubAgentMiddleware — Subagent delegation
• ✅ SummarizationMiddleware — Context window management
• ✅ PatchToolCallsMiddleware — Tool call patching
• ✅ MemoryMiddleware — Long-term memory
• ✅ HumanInTheLoopMiddleware — Approval interrupts

1.5 Subagent Definitions¶

• ✅ 11 subagents from workflow roles
• ✅ knowledge-curator — Document ingestion, signal extraction
• ✅ idea-discoverer — Signal to structured idea transformation
• ✅ problem-framer — Problem statement refinement
• ✅ novelty-analyst — Novelty claims articulation
• ✅ prior-art-researcher — Prior art search and analysis
• ✅ detectability-analyst — Infringement detectability assessment
• ✅ business-value-analyst — Siemens business value quantification
• ✅ siemens-alignment — Strategic domain alignment validation
• ✅ checklist-validator — Gate checklist enforcement
• ✅ reviewer-summarizer — Review packet creation
• ✅ patent-drafter — IdeaScope document drafting

1.6 Domain Tools¶

• ✅ generate_invention_ideas() — Structured idea generation
• ✅ query_prior_art_taxonomy() — KB taxonomy queries
• ✅ draft_patent_section() — Patent disclosure drafting
• ✅ evaluate_patentability() — Scoring engine integration
• ✅ record_approval_decision() — Review decision recording

1.7 Credential Management¶

• ✅ .env file loading via pydantic-settings
• ✅ Credential propagation to os.environ for LangChain compatibility
• ✅ deepagents_model auto-derivation from openai_model_name

2. Thread System¶

• ✅ Thread metadata persistence in SQLite
• ✅ thread_manager.py creates, updates, deletes, and lists threads
• ✅ updated_at sort order for active thread navigation
• ✅ idea_id, tags, and agent_names metadata fields
• ✅ LangGraph checkpoint-backed message history
• ✅ GET /api/threads/{id}/messages reads from the saver state
• ✅ POST /api/threads/{id}/stream streams events immediately as they arrive
• ✅ Thread touch/update on every send for active sorting
• ✅ Frontend thread experience
• ✅ Right sidebar thread list
• ✅ New thread creation on first message
• ✅ Thread switching reloads persisted history
• ✅ Streaming UI renders event-bound updates

2. Workflow Engine¶

2.1 State Machine¶

• ✅ 18-state PatentWorkflowMachine using transitions library
• ✅ Discovery phase (3 states): raw_signal → idea_discovery → idea_clarification
• ✅ Research phase (3 states): novelty → prior_art → detectability
• ✅ Analysis phase (2 states): business_value → siemens_alignment
• ✅ Drafting phase (2 states): ideascope_draft → filing_check
• ✅ Review phase (3 states): manager → ip_review → counsel_validation
• ✅ Done phase (5 states): ready → submitted → feedback → revision → accepted

2.2 State Transitions¶

• ✅ Linear workflow with sequential advancement
• ✅ advance_to_next() — Auto-advance to next state
• ✅ advance_to(target) — Advance to specific next state
• ✅ Gate validation before transition
• ✅ Lifecycle hooks: on_entry, validate, on_exit
• ✅ State history recording

2.3 Gate Checklists¶

• ✅ 12 gate checklists defined in config/checklist-config.yaml
• ✅ discovery → clarification (3 items)
• ✅ clarification → novelty (3 items)
• ✅ novelty → prior_art (3 items)
• ✅ prior_art → detectability (3 items)
• ✅ detectability → business_value (3 items)
• ✅ business_value → alignment (3 items)
• ✅ alignment → drafting (5 items)
• ✅ drafting → filing_check (7 items)
• ✅ filing_check → manager (4 items)
• ✅ manager → ip_review (3 items)
• ✅ ip_review → counsel (4 items)
• ✅ counsel → ready (4 items)
• ✅ Evidence checking logic in state/gates.py

2.4 Workflow Orchestration¶

• ✅ run_generation_cycle() — Scheduler-driven cycle
• ✅ run_full_pipeline() — End-to-end pipeline execution
• ✅ seed_ideas() — Autonomous idea seeding
• ✅ Idea prioritization and focus selection
• ✅ Active processing state management
• ✅ Pause/resume idea processing

3. Scoring Engine¶

3.1 Criteria¶

• ✅ 7 weighted criteria (0-100 each)
• ✅ novelty (25%) — Prior art gap analysis
• ✅ siemens_alignment (15%) — Strategic domain match
• ✅ technical_feasibility (15%) — Implementation viability
• ✅ detectability (10%) — Infringement detection
• ✅ business_value (15%) — Siemens market impact
• ✅ originality (10%) — Non-obviousness
• ✅ completeness (10%) — Documentation quality

3.2 Scoring Methods¶

• ✅ LLM-powered scoring (primary)
• ✅ Composite score calculation (weighted sum)
• ✅ Strength rating assignment (Very Strong / Strong / Moderate / Weak / Reject)
• ✅ Filing threshold validation (composite ≥ 70, no critical < 50%)
• ✅ Score history tracking in scores.yaml

3.3 Scoring Models¶

• ✅ ScoreBreakdown — Per-criterion scores
• ✅ ScoreRecord — Full score record with timestamp
• ✅ CriterionDetail — Score + reasoning + confidence
• ✅ ScoringEngine — Orchestrates scoring
• ✅ CriterionEvaluator — LLM-powered evaluation

4. Human-in-the-Loop¶

4.1 Interrupt Configuration¶

• ✅ Manager review interrupt
• ✅ IP review interrupt
• ✅ Counsel validation interrupt
• ✅ Delete/archive approval interrupt
• ✅ Durable pending interrupt persistence (disk-backed)

4.2 Approval Flow¶

• ✅ Pending interrupt listing endpoint
• ✅ Approval endpoint with reviewer identity
• ✅ Rejection endpoint with reason
• ✅ Resume workflow after decision
• ✅ Transcript event recording for decisions

4.3 Review Analytics¶

• ✅ Reviewer identity normalization
• ✅ Approval/rejection counts by role
• ✅ Pending interrupt observability

5. Transcript & Events¶

5.1 Event Types¶

• ✅ TranscriptEventType enum with 14 event types
• ✅ thinking, tool_call, tool_result, subagent, handover
• ✅ interrupt, approval, retry, failed, completion
• ✅ done, token, tasks_update, transition, user_message

5.2 Event Roles¶

• ✅ TranscriptRole enum with 7 roles
• ✅ user, orchestrator, subagent, reviewer, tool, workflow, system

5.3 Event Persistence¶

• ✅ Typed TranscriptEvent model (Pydantic)
• ✅ normalize_transcript_event() — Metadata enrichment
• ✅ Transcript event storage in transcript.yaml
• ✅ Provenance tracking per event
• ✅ Trust level assignment per event type

5.4 Event Streaming¶

• ✅ SSE-based real-time event streaming
• ✅ Runtime event coercion from DeepAgents messages
• ✅ Task update events with progress tracking

6. API Layer¶

6.1 REST Endpoints¶

• ✅ Health check (GET /api/health)
• ✅ SSE stream (GET /api/sse)
• ✅ Idea CRUD (GET/POST /api/ideas, GET /api/ideas/{id})
• ✅ Workflow advancement (POST /api/ideas/{id}/advance)
• ✅ Scoring (POST /api/ideas/{id}/score)
• ✅ Gate validation (POST /api/ideas/{id}/validate-gate)
• ✅ Idea update (POST /api/ideas/{id}/update)
• ✅ Evidence addition (POST /api/ideas/{id}/evidence)
• ✅ Idea delete with interrupt (DELETE /api/ideas/{id})
• ✅ Idea archive with interrupt (POST /api/ideas/{id}/archive)
• ✅ Pause/resume (POST /api/ideas/{id}/pause, /resume)
• ✅ Comments (POST /api/ideas/{id}/comment)
• ✅ Chat (GET/POST /api/chat, /api/ideas/{id}/chat)
• ✅ Streaming chat (POST /api/chat/stream, /api/ideas/{id}/chat/stream)
• ✅ Agent tasks (GET /api/agent-tasks)
• ✅ Workflow interrupts (GET /api/workflow/interrupts)
• ✅ Approval/rejection (POST /api/workflow/{id}/approve, /reject)
• ✅ Review analytics (GET /api/workflow/analytics)
• ✅ Workflow cycle/seed (POST /api/workflow/cycle, /seed)
• ✅ Pipeline execution (POST /api/submit-pipeline, /auto-pipeline)
• ✅ System stats (GET /api/stats)
• ✅ Knowledge base (GET /api/knowledge-base)
• ✅ Siemens domains (GET /api/config/siemens-domains)
• ✅ Idea files (GET /api/ideas/{id}/files)
• ✅ Artifact revisions (GET /api/ideas/{id}/revisions)
• ✅ Artifact diff (GET /api/ideas/{id}/artifacts/{name}/diff)

7. Agentic Organization Platform¶

• ✅ Thread persistence is the source of truth for conversations
• ✅ Active thread sorting uses updated_at DESC
• ✅ Threads can be associated with an idea_id for idea-scoped conversations
• ❌ @mention routing is intentionally deferred for this iteration
• ❌ Siemens-specific workflow generalization is pending the idea/thread cleanup

6.2 Route Organization¶

• ✅ Routes split into modular files
• ✅ Each route file < 150 lines
• ✅ Clear separation of concerns

7. Frontend¶

7.1 Pages¶

• ✅ Dashboard — Stats, idea grid, action buttons
• ✅ IdeaDetail — Full idea view with tabs
• ✅ KnowledgeBase — Document browser
• ✅ SiemensControls — Domain alignment view

7.2 Components¶

• ✅ shadcn/ui components (Button, Card, Dialog, Input, Select, Tabs, Badge, Sidebar, Sheet, Tooltip, Separator, ScrollArea)
• ✅ DeepAgents components
• ✅ AgentTodoPanel — Task/progress panel
• ✅ SubagentActivityCard — Subagent activity display
• ✅ ToolCallTimeline — Tool call inspection
• ✅ InterruptInbox — Approval interrupt inbox
• ✅ ArtifactDiffPanel — Artifact diff viewer
• ✅ Idea components
• ✅ IdeaCard — Idea summary card
• ✅ ScoreRadar — Score radar chart
• ✅ WorkflowTimeline — Workflow state timeline
• ✅ Layout components
• ✅ AppSidebar — Left navigation
• ✅ SiteHeader — Top header
• ✅ RightChatSidebar — Live transcript sidebar
• ✅ IdeaHistoryTimeline — Historical timeline

7.3 API Integration¶

• ✅ Centralized API client (api/client.ts)
• ✅ SSE streaming hook (hooks/useDeepAgentStream.ts)
• ✅ Interrupt polling hook (hooks/useInterrupts.ts)

7.4 Design System¶

• ✅ shadcn/ui configured via components.json
• ✅ Tailwind CSS with Zinc base color
• ✅ Radix UI primitives for accessibility
• ✅ Lucide React icons

8. Storage & Persistence¶

8.1 Filesystem Storage¶

• ✅ YAML/Markdown persistence in workspace/ideas/
• ✅ Per-idea folder structure
• ✅ idea.yaml — Main record
• ✅ state.yaml — State history
• ✅ scores.yaml — Score history
• ✅ transcript.yaml — Transcript events
• ✅ ideascope-draft.md — IdeaScope document
• ✅ submission-summary.md — Submission packet
• ✅ handovers/ — Per-transition handover packets
• ✅ revisions/changelog.md — Audit trail

8.2 Storage Modules¶

• ✅ base.py — Read/write YAML and Markdown
• ✅ idea_workspace.py — Idea folder CRUD
• ✅ registry.py — Idea registry management
• ✅ knowledge_base.py — KB document loading
• ✅ artifacts.py — Artifact revision tracking
• ✅ recovery.py — Filesystem recovery

8.3 Artifact Management¶

• ✅ Artifact revision tracking with versioning
• ✅ Artifact diff generation
• ✅ Provenance metadata per artifact
• ✅ Evidence reference tracking

9. Research & Knowledge¶

9.1 Prior Art Search¶

• ✅ Structured Google Patents XHR JSON adapter
• ✅ Source provenance tracking
• ✅ Configurable search limits

9.2 Knowledge Base¶

• ✅ Raw document storage (knowledge-base/raw/)
• ✅ Processed document storage (knowledge-base/processed/)
• ✅ Siemens tech domains (knowledge-base/siemens/)
• ✅ Prior art taxonomy (knowledge-base/prior_art_taxonomy.json)
• ✅ Multimodal ingestion (PDFs and images)

9.3 Duplicate Detection¶

• ✅ Lexical similarity matching (SequenceMatcher)
• ✅ Token overlap scoring
• ✅ Configurable threshold
• ✅ Provenance tracking for matches

10. Observability¶

10.1 LangSmith Tracing¶

• ✅ Environment variable configuration
• ✅ Tracing enable/disable toggle
• ✅ Project and endpoint configuration

10.2 System Monitoring¶

• ✅ Health check endpoint
• ✅ System statistics endpoint
• ✅ Review analytics endpoint
• ✅ Active processing state tracking

11. Configuration¶

11.1 System Configuration (config/system-config.yaml)¶

• ✅ Workflow interval, retries, timeout
• ✅ Scoring weights and thresholds
• ✅ Strength rating boundaries
• ✅ State definitions

11.2 Gate Checklists (config/checklist-config.yaml)¶

• ✅ 12 gate definitions with checklist items
• ✅ Per-item descriptions

11.3 Environment Configuration¶

• ✅ .env file loading
• ✅ Pydantic Settings model
• ✅ Docker environment variable passthrough

12. Deployment¶

12.1 Docker¶

• ✅ Backend Dockerfile (FastAPI)
• ✅ Frontend Dockerfile (Nginx + Vite)
• ✅ Docker Compose configuration
• ✅ Volume mounts for config, workspace, KB

12.2 Environment¶

• ✅ Configurable host/port
• ✅ Configurable model and API settings
• ✅ APP_ROOT_DIR for Docker path resolution

Feature Linkages¶

Dependency Map¶

Agent Runtime (1)
├── Workflow Engine (2) — uses runtime for state execution
│   ├── Scoring Engine (3) — called after state transitions
│   ├── HITL (4) — interrupts at review states
│   └── Transcript & Events (5) — records all transitions
├── API Layer (6) — exposes runtime capabilities
│   └── Frontend (7) — consumes API
├── Storage & Persistence (8) — backend for all data
├── Research & Knowledge (9) — feeds idea generation
└── Observability (10) — monitors all components

Cross-Cutting Concerns¶

Related Documents¶

• Architecture — System architecture
• PRD — Product requirements
• Coding Guidelines — Development standards
• Architecture Decisions — ADR log
• Tasks — Implementation task hierarchy