How this project generates a book from opencode's documentation, what has actually been run, and what has not.
| Document | What it covers |
|---|---|
| 01-build-process.md | The production pipeline: sync → build → assemble → publish |
| 02-poc-replay-process.md | The replay walkthrough used to prove the pipeline without a local model |
| 03-gaps-and-mitigations.md | Known gaps, ranked, with recommended fixes |
| 04-next-steps.md | Ordered next actions |
| 05-releases.md | Versioning, changelog, and release workflow (personal workflow — removable) |
It's worth being precise about this, because it's easy to remember it wrong.
There is one build process. The POC did not fork it or replace it. The pipeline code that ran during the walkthrough is the same code that will run in production — same corpus loader, same validator, same fingerprinting, same output writer.
What changed for the POC was the model backend. Every stage talks to an OpenAI-compatible HTTP endpoint, so the walkthrough pointed that endpoint at a replay server serving hand-authored responses instead of at a local model. From the pipeline's perspective nothing was different.
┌─ production: MLX / LM Studio / Ollama / hosted API
opencode-book build ──HTTP──────► │
└─ POC: tools/replay/server.py
Two changes were made permanently, and they are part of the production process now — they were motivated by POC economics but are not POC-only:
- One model for all four roles. The differentiated setup named three models totalling ~69 GB against a 64 GB ceiling, so they could never all stay resident. See 01 and 03.
- Stage-major builds. Every chapter through
extract, then every chapter throughdraft, rather than each chapter through the whole pipeline in turn. Cuts model loads from 4×chapters to 4 per run.
| Stage | State |
|---|---|
| Corpus sync | Working. 36 docs, 49,468 words, pinned at 040b8561 |
| Chapter specs | 8 written, all source refs verified to resolve |
| extract → validate → draft → verify → voice | Working. Proven end to end on chapter 05 |
| Publish (EPUB/PDF) | Not run. Documented approach only — see publish/README.md |
| Real local model | Never used. Every run so far has been mock or replay |
The single largest untested assumption is that a local model can produce prose you would actually publish. Nothing here has tested that yet.