Skip to content

ROADMAP.md

Latest commit

 

History

History
83 lines (66 loc) · 5.19 KB

ROADMAP.md

File metadata and controls

83 lines (66 loc) · 5.19 KB
Raw

Roadmap

This roadmap reflects the actual state of the project. book-to-skill is already a mature generator (v1.6.1) — this is not a pre-1.0 plan. The version line is sourced from scripts/bookextract/__init__.py; shipped features are recorded in CHANGELOG.md, each tagged with its upgrade migration class.

Open work is tracked as GitHub issues; milestones below link to them where they exist.

Direction — Mycelia: book → OKF knowledge vault

The strategic default is shifting from one skill per book (doesn't scale past a few books — every skill description is injected into context each session) to an atomic, interlinked OKF knowledge vault that scales to hundreds of books. The new emitter is MYCELIA.md; the legacy SKILL.md skill flow stays supported.

Landed in this direction:

  • MYCELIA.md vault emitter — atomic notes (concept/framework/principle/entity/method/anti-pattern), OKF-native portable bundle, # Citations provenance, check-before-create cross-book reconciliation.
  • book-extract lint — stdlib OKF-bundle linter: zero dangling + type-required + reserved-file + citation-coverage + reciprocal-backlink checks.
  • Orchestrated build (P0) — book-extract build-plan + assemble: the LLM emits validated Note JSON, the code chunks, grounds (normalize-then-match, ligature/line-break tolerant), dedups, inserts reciprocal backlinks, computes folios, and gates with the linter. Removes hand-emission drift. See docs/mycelia-productization.md.
  • Deterministic reconciliation (P1) — reconcile.py: beyond slug/alias dedup, folds acronym ↔ expansion (bsc ↔ binary-symmetric-channel) and singular ↔ plural (encoder ↔ encoders) with zero dependencies. Conservative by design (basis/bases stay distinct); the folded slug becomes an alias so ## Related resolves through it.

Next for Mycelia (from the productization assessment):

  • Embedding reconciliation (P1+) — beyond the deterministic acronym/plural folding above: cosine-similarity + LLM arbitration for pure-semantic or cross-lingual synonyms that share no string ("Deutschland"/"Germany"). Deferred — would break the current zero-dependency stance.
  • Semantic QA pass — an independent verifier checking each note's body against its cited source.
  • Skill → vault converter — re-emit already-generated book-skills (chapters/*.md + manifest) as OKF bundles.
  • Obsidian vault → OKF converter — [[wikilink]] → bundle-relative links + frontmatter normalization.
  • Contradiction dashboard — aggregate contested: true notes into a reports/ view.
  • Headless autonomy — optional anthropic SDK / claude -p path if the in-session build proves too manual at scale.

v1.6.1 — current

Shipped (see CHANGELOG for the full list and migration classes):

  • Single-document → Claude Code skill generation across PDF / EPUB / DOCX / TXT / Markdown / reStructuredText / AsciiDoc / HTML / RTF / MOBI/AZW/AZW3.
  • Capability gating: study/reference always; code review, stack personalization and scaffold for code/buildable books; figures for technical extraction.
  • Grounded answers — chapter + verbatim quote (page folio for text PDFs), grep-verified.
  • Provenance manifest (.book-to-skill.json) + deterministic upgrade flow + .source/ archive.
  • Diagram/figure capture (#8); printed page folios (#11).
  • uv-first developer workflow — installable package, book-extract console script, optional-dependency extras, and GitHub Actions CI.
  • Public, copyright-safe demo under examples/ (CC0 input + generated sample skill, #14).
  • Strict quality gate (ruff · mypy · lizard · xenon) and Sphinx API docs.

v1.7.0 — next

Adoption & breadth (this release line):

  • Broader inputs beyond books — technical papers, internal wikis, video transcripts. (#7)
  • Propagate templates/ to existing technical skills generated before scaffolds existed. (#12)
  • A public, copyright-safe demo input + a published sample skill (planned). (#14)
  • Validation checks for generated skill folders (structure/manifest lint) (planned).
  • Pay down the cli.py complexity debt so lizard/xenon can become required gates. (#13)

v1.8.0+ — multi-document

  • Multi-book domain libraries — now realized through the Mycelia vault (above): N related books converge into one OKF bundle with shared canonical concept notes and cross-book citations, rather than one fused skill. (#6)
  • Cross-document references and a shared glossary across a domain library — native to the vault (# Citations + ## Related + reconciliation).

Items marked (planned) are not yet implemented. The dated, shipped record of truth is always CHANGELOG.md.