Skip to content

Milestone Tracker

⚠️ ARCHIVED — 2026-07-31

This tracker served as a running log through Phases 0–7. All milestones are completed.

Superseded by: tasks.md (structured task hierarchy with status per item).

Update this file whenever work is completed.

Milestone Summary

  • M0: Current-state audit completed
  • M0: Target architecture documented
  • M0: Feature roadmap documented
  • M0: Phased plan documented
  • M0: Initial DeepAgents scaffold added
  • M0: Phase 0 completed
  • M1: FastAPI route split completed
  • M1: Backend structural cleanup completed
  • M2: Real DeepAgents runtime build completed
  • M3: Skills and memory migration completed
  • M4: Human-in-the-loop approvals completed
  • M5: Frontend DeepAgents streaming integration completed
  • M6: Artifact quality and traceability completed
  • M7: Advanced capabilities completed

Completed Tasks

2026-07-28

  • Updated the docs to distinguish real agentic primitives from hardcoded, partial-agentic, and simulated paths.
  • Added upstream DeepAgents research notes to the architecture docs.
  • Reframed the roadmap around explicit retry/error states instead of silent fallback fabrication.
  • Replaced overstated completion claims with a realistic milestone tracker.
  • Documented the gap between workflow progress updates and a true live agent transcript.
  • Added implementation checklist items for typed runtime events, transcript separation, and explicit speaker labeling.
  • Added solution research checklist items for chat layout, transcript design, and patent/file adapter selection.
  • Added an explicit execution order for the complete agentic implementation path.
  • Captured phase-by-phase concrete solution paths in the phased plan.
  • Mapped each major issue family to a concrete agentic solution mechanism.
  • Removed sandboxing from the active roadmap and kept only safe validation notes.
  • Identified the next implementation slice: typed runtime events plus transcript-driven UI.
  • Implemented the first transcript/runtime slice: expanded event types, backend provenance metadata, and UI transcript compatibility.
  • Implemented transcript-backed sidebar loading so persisted runtime events appear in the live chat surface.
  • Added explicit transcript badges for interrupt, approval, retry, failed, and handover states.
  • Repaired transcript sidebar rendering so the live transcript remains stable with the new event badges.
  • Removed the remaining fabricated agent reply path from the chat route.
  • Added approval and rejection transcript events for HITL actions.
  • Added dedicated approval and interrupt transcript cards in the DeepAgents detail view.
  • Added a typed transcript event model plus transcript-backed chat history persistence.
  • Replaced the live chat sidebar's fake bootstrap conversation with real transcript event rendering.
  • Removed the remaining synthetic reply path from chat posting and streaming responses.
  • Added provenance and trust metadata to persisted transcript events.
  • Wired review-state transitions to create blocking approval interrupts.
  • Added provenance metadata to generated artifact sections.
  • Extended the historical timeline to render runtime transcript events with role-aware badges.

2026-07-29

  • Replaced the remaining scripted runtime narration with runtime-driven transcript event coercion.
  • Removed the last hardcoded idea-generation and heuristic scoring fallback paths.
  • Made pending HITL interrupts durable on disk instead of in-memory.
  • Replaced public patent scraping with a structured Google Patents XHR JSON adapter.
  • Revalidated the backend and frontend after the final adapter change.
  • Renamed generic backend modules to descriptive names (tools.py -> domain_tools.py / workflow_tools.py, ideas.py -> idea_workspace.py).
  • Consolidated dependency maintenance so the repo root requirements.txt now wraps backend/requirements.txt.

2026-07-30

  • Phase 10 M10 (Streaming Adapter) implemented: v3 streaming protocol with structured subagent/message/tool_call projections
  • Added message event type to transcript model — agent natural language output as first-class event
  • Rewrote execute_deep_agent_workflow_streaming to use astream_events(version="v3")
  • Queue-based concurrent consumption via asyncio.Queue + asyncio.gather
  • Graceful v2 fallback if v3 streaming is unavailable
  • Updated frontend StreamEventType with message type
  • Updated chat UI to render agent messages as chat bubbles with agent name header (not badges)
  • M11: Scoring moved into criteria.py — removed execute_llm_scoring() from subagent_executor.py
  • M11: Scoring reasoning emitted as SSE agent.progress events
  • M12: langchain-mcp-adapters installed, MCP_SERVERS env var config in runtime
  • M13: research-agent subagent defined with KB-scan instructions + save_research_note tool
  • M14: supervisor-agent subagent defined with full pipeline coordination instructions
  • M14.2: Procedural 13-state loop in run_full_pipeline() replaced with single supervisor invocation
  • M15: APScheduler disabled by default (WORKFLOW_SCHEDULER_ENABLED=false)
  • M16: Consecutive same-agent messages grouped into single chat bubble in UI
  • All 43 backend tests pass; frontend build succeeds
  • M13.3: Template generate_invention_ideas() replaced with agentic research agent invocation
  • M14.2: Procedural 13-state loop replaced with single supervisor execute_deep_agent_workflow() call
  • M16.2: Search bar added to chat sidebar — live filtering across all transcript messages
  • P9.3: Added search_processed_knowledge_base() research source — searches processed KB docs
  • All deferred items completed — no remaining PENDING tasks

Current reality

  • DeepAgents runtime is now the single source for the runtime entrypoint and event model.
  • Silent fallback behaviors have been removed from runtime code.
  • Human approval interrupts are now blocking while pending and resume on decision.
  • Real prior-art retrieval now has package-backed research adapters and source provenance.
  • Frontend error/retry trust signals are surfaced with explicit runtime badges.
  • Conversation threads now stay transcript-driven instead of using hardcoded agent/task narration.
  • The live chat surface now shows typed runtime transcript events instead of a fake bootstrap conversation.
  • Patent/filing research source adapters are selected and wired into the research path.
  • The UI distinguishes agent speakers from orchestration/status messages.
  • Agent roster and state transition ownership finalized.
  • The fully agentic runtime has replaced scripted chat narration in the transcript path.
  • The live transcript UI renders richer approval / retry / failed states.
  • The chat surface shows explicit cards/badges for interrupt, approval, retry, and failed events.
  • The live transcript UI includes dedicated event cards for approval / interrupt workflows.
  • Multimodal ingestion for PDFs and images is implemented.
  • LangSmith observability is configured through runtime env propagation.
  • Delete/archive requests now route through approval interrupts before destructive action.
  • Review analytics expose reviewer identity and pending-interrupt observability.
  • The streamed transcript is runtime-driven in backend/app/agent/runner.py.
  • Idea generation now uses runtime-driven prior-art and taxonomy inputs instead of hardcoded seed ideas.
  • Scoring now uses the LLM-backed JSON scoring path end to end.
  • Pending HITL interrupts are durable and survive restart.
  • Public patent search now uses a structured JSON adapter instead of HTML scraping.
  • Review timestamps now use durable real timestamps.
  • Repository cleanup renamed the most confusing generic backend modules to descriptive names.
  • Dependency entrypoint cleanup removed the duplicate root dependency list.

HITL blocking slice

  • Made HITL interrupts pause ideas until approval or rejection.
  • Fixed approval handling to use the correct workflow-state field and resume ideas after decisions.
  • Recorded interrupt/approval decisions in the transcript while preserving the pending interrupt queue.

Phase 7 hardening slice

  • Added reviewer identity normalization and approval analytics.
  • Recorded package/adaptation decision history in the living docs.
  • Documented the DB-migration planning decision to keep YAML until migration is warranted.
  • Completed the remaining Phase 7 capability checklist and closed M7.
  • Optional internal filing connector work remains a future enhancement, not a blocker.

Validation Notes

  • npm run build in frontend/ passed with 0 TypeScript/Vite compilation errors.
  • pytest backend/tests passed 43 out of 43 tests cleanly in the latest verified run.