Feature Roadmap¶

⚠️ ARCHIVED — 2026-07-31

This roadmap was the planning document for Phases 0–7. All features listed are now implemented: - Agent runtime: DeepAgents with middleware, subagents, permissions - Governance: HITL interrupts for manager/IP/counsel; approval endpoints - Artifact quality: versioning, diffs, evidence traceability, duplicate detection - Frontend: subagent cards, todo panel, tool-call inspection, approval inbox, diff views - Research: Google Patents adapter, source provenance, multimodal ingestion - Operations: LangSmith tracing, RBAC, review analytics

Superseded by: features.md (complete feature tree), tasks.md (current status).

---

Current Features¶

• idea creation from manual signal text
• autonomous idea seeding from knowledge-base content
• custom workflow state progression
• heuristic and LLM-assisted scoring, with some paths still requiring cleanup before the app can be called fully agentic
• gate validation from checklist config
• filesystem artifact storage per idea
• comments on ideas
• workflow status and progress views
• SSE-based live dashboard updates
• knowledge-base listing

Trust Constraints¶

• No silent fallback to fabricated agent output
• No simulated human approval for production-trust stages
• Explicit retry/error states are preferred over hidden heuristics

Features To Build Next¶

Agent runtime and orchestration¶

• real DeepAgents runtime using create_deep_agent
• structured subagent delegation
• task/todo state surfaced to UI
• event-stream based tool and subagent updates
• real DeepAgents memory and skills loading

The existing upstream DeepAgents package already supports these primitives, so the roadmap should adapt the package rather than reimplementing them.

Governance and review¶

• human-in-the-loop approvals for manager, IP, and counsel stages
• approval and rejection endpoints
• reviewer assignment and queues
• audit log of approval decisions
• protected final artifact writes

If the review cannot be completed by a real human, the workflow should pause instead of auto-approving.

Workflow and artifact quality¶

• split research, drafting, and review skills into maintainable packs
• better structured artifact generation
• artifact versioning and revision comparisons
• evidence traceability per draft section
• duplicate idea detection against existing ideas

Frontend¶

• subagent activity cards
• live todo list
• tool-call inspection
• approval inbox
• artifact diff views
• richer workflow event timeline

Later Features¶

Knowledge and research¶

• real prior-art search integrations
• evidence normalization and citation extraction
• cross-idea clustering and deduplication
• multimodal input support for PDFs, slides, and images

Operations¶

• LangSmith tracing and run analytics
• role-based access control
• reviewer identity and audit export
• background job and queue processing
• database-backed repositories

Quality systems¶

• rubric-driven iterative drafting
• domain-specific evaluation suites
• acceptance/rejection analytics
• memory consolidation and policy updates

Explicitly Deferred For Now¶

• sandbox execution
• shell command execution by the agent
• remote code runner integration