Research Tools:
- SearchTool: Multiple backends (Tavily, Serper, Brave, DuckDuckGo)
- WikipediaTool: Wikipedia lookup
- AcademicSearchTool: CrossRef, Semantic Scholar
- ResearchOrchestrator: Comprehensive multi-source research
ResearchAgent:
- NOT just fact-checking - actively discovers NEW information
- Identifies trends beyond training data cutoff
- Generates innovations from cross-referencing sources
- Deep research with subtopics
VerifiedFactChecker:
- Live claim verification against web sources
- Confidence scoring
- Citation needed detection
Dependencies added: tavily, wikipedia, arxiv, duckduckgo-search
- Merge llm.py + llm_sync.py into single unified client
- Remove llm_sync.py (now just llm.py with both sync/async)
- Add requests to dependencies
- Add Dockerfile for containerized deployment
- Add .dockerignore
All issues resolved!
- Web UI: novice-friendly interface at / and /ui
- Upload endpoint: /upload for file uploads
- S3 upload: /upload/s3 for uploading to S3/MinIO
- CLI: opus ui command to start web UI only
- Full HTML/CSS/JS interface with drag-drop, tabs, etc.
- LocalIngestor class for files and directories
- CLI: opus ingest-local PATH
- Generate from local: opus generate --local ./my-notes/
- Support for extensions, recursive scanning, summarize
- Pattern-based exclusion (.git, __pycache__, etc.)
- Complete architecture diagram
- All features documented with status
- Full CLI, Python, API client examples
- Configuration reference
- Project structure
- Handle dict result properly in /generate endpoint
- Remove use_autogen from API (not supported in run_opus)
- API client now works: tested via CLI with local server
Now CLI works in both local and client/server mode:
# Local mode (default)
opus generate --concept ...
# API client mode
opus --api-url http://localhost:8000 generate --concept ...
opus --api-url https://opus-api.example.com generate --repo owner/repo
OpusAPIClient class for programmatic API access.
CLI Commands:
- generate: Full manuscript generation with full GitHub content
- serve: Start FastAPI server with OpenAPI docs
- ingest: Standalone GitHub ingestion
- frameworks: List all story frameworks
- config: Show configuration
- docs: Show comprehensive docs (terminal/markdown/html)
- api: Export OpenAPI spec
Server:
- FastAPI with /docs, /redoc interactive docs
- /generate, /ingest, /frameworks, /health endpoints
- OpenAPI 3.0 specification
Documentation:
- Terminal, markdown, and HTML formats
- Full API reference
- Framework documentation
- Environment variables guide
- Project structure
Fix: Use full GitHub content as seed (not just 5000 chars)
- OpusPydanticAgent with schema validation
- StorySeed, CharacterProfile, ChapterOutline, ChapterDraft schemas
- CritiqueResult, StyleGuide schemas
- Factory functions for each agent type
- Test: successfully generated StyleGuide with validated output
Usage:
from opus_orchestrator import create_style_guide_agent
agent = create_style_guide_agent()
result = agent.run_sync('Create a style guide for...')
# result is a validated StyleGuide object
- Wire GitHubIngestor into the orchestrator
- Test: successfully ingested 251k chars from The-Last-Love-Story repo
- Also verified: run_opus generates 1k+ word stories with AutoGen critique
- GitHubIngestor class to fetch repo contents
- Support for .md, .txt, .notes, .draft files
- Method to ingest from GitHub directly into orchestrator
- Export GitHubIngestor in __init__.py
Usage:
orch = OpusOrchestrator(book_type='fiction', genre='memoir')
content = await orch.ingest_from_github('mrhavens/my-notes')
await orch.run()
- Add CritiqueCrew initialization in OpusGraph.__init__
- Add use_autogen flag to enable/disable
- Add critique_summary and critique_iterations to state
- Rewrite node_write_chapters with AutoGen critique loop
- Each chapter now gets multi-agent critique (LiteraryCritic, GenreExpert, StoryEditor)
- Iteration loop until approved or max iterations
- Full integration complete
Based on Gemini's analysis:
1. Nodes now return dicts instead of mutating state
2. run() uses stream_mode='values'
3. Falls back to get_state() from checkpointer
4. Uses model_copy() for Pydantic updates
Full pipeline runs end-to-end:
- All 7 pre-writing stages (seed → scene descriptions)
- Style guide generation
- Chapter writing (3 chapters, 2,833 words tested)
Fixed result extraction from graph.stream().
- Proper StateGraph with compiled nodes
- Conditional edges for writing iteration
- Checkpoint memory for resumability
- Structured state with Pydantic
- Node_validate for cross-stage validation
- Writing loop with iterate/next logic
- Full end-to-end workflow
This is the rigorous LangGraph integration.
- Structured OpusState with Pydantic models
- PreWriting schema with Character, ChapterPlan, PlotBeat
- Validation functions for each stage
- Checkpoint save/load to disk
- Workflow nodes for each stage
- Cross-validation between stages
- Structured parsing of LLM outputs
- Snowflake Method (already had)
- Three-Act Structure
- Save the Cat (Blake Snyder)
- Hero's Journey (Joseph Campbell)
- Story Circle (Dan Harmon)
- The 7-Point Plot (The Pantone)
- Fichtean Curve
Pass framework='hero-journey' or 'save-the-cat' etc. to use.
- Add .env to .gitignore (API keys stay local)
- Add LLM client with MiniMax and OpenAI support
- Update config to load from environment variables
- Wire up Architect agent to actually call the LLM
- Add MiniMax API key to local .env file
mrhavens
Mark Randall Havens △ The Empathic Technologist