Skip to content

ROADMAP.md

Latest commit

 

History

History
90 lines (69 loc) · 4.52 KB

ROADMAP.md

File metadata and controls

90 lines (69 loc) · 4.52 KB
Raw

Roadmap to a stable public release

Current version: v1.0.0-rc1.

The frozen content/generalization review is recorded in v1.0.0-rc1-frozen-validation.md. The RC work is frozen for public feedback; the note template and semantic workflow stay unchanged unless a release-blocking fidelity, security, privacy or data-protection defect is found.

The next milestone is not a larger feature set. It is a stable, understandable and recoverable install-plus-read experience for users running a local-capable Agent on Windows:

install LitAnchor from its GitHub repository
→ deep-read one local PDF into Markdown
or deep-read one Zotero paper into a named Obsidian Vault

The example wording is not a fixed command contract. Routing must recognize equivalent installation and single-paper reading requests.

Stable-release priorities

P0 — isolated-environment installation

  • Run install, doctor, setup, upgrade, rollback and uninstall in at least three isolated Windows application profiles. When independent users or machines are unavailable, label same-account profiles accurately and use public RC feedback as the independent-usability evidence.
  • Cover a normal English path, a Chinese user/Vault path and an interrupted-install recovery path.
  • Require no manual pip, JSON editing, Skill copying or intermediate-ledger repair from the user.
  • Keep Python at >=3.10 and Zotero at >=7; probe newer untested versions instead of imposing arbitrary maximum versions.

P0 — end-to-end reliability

  • Resolve one Zotero item and one local PDF without silently accepting ambiguous matches.
  • Preserve PyMuPDF physical pages as the evidence authority.
  • Automatically invoke MinerU for eligible files only under stored consent, and fall back safely on timeout, rate limit, empty output or network failure.
  • Prove non-overwriting, path-contained Obsidian export and rollback behavior.

P0 — frozen generalization checks

  • Freeze the Skill, schemas, prompts and rubric before running new holdouts.
  • Use at least three papers that were not used to shape the current rules.
  • Do not expose reference notes or manually prefill Evidence/Claim records.
  • If a holdout causes a workflow change, move it to development and replace it before making a final generalization claim.

P1 — public usability

  • Keep the repository homepage task-oriented and bilingual.
  • Maintain issue forms that prevent accidental uploads of copyrighted PDFs, private notes or credentials.
  • Turn recurring public failures into reproducible, paper-independent tests and controlled EvolutionCandidates.
  • Publish a copyright-safe demonstration artifact when one is available.

RC qualification evidence

v1.0.0-rc1 was qualified against the following gates:

  • isolated install/read lifecycle passes in three Windows application profiles, including one real core-dependency installation;
  • repository tests, Skill validation, compilation, Markdown links, privacy and anti-leak audits pass in CI;
  • severe unsupported claims, data overwrite and privacy leakage are zero;
  • physical-page accuracy is at least 99% and core numeric errors are zero;
  • supported-path end-to-end success is at least 90%;
  • MinerU failure falls back without corrupting authoritative evidence;
  • upgrade, rollback and uninstall affect only receipt-owned LitAnchor files;
  • documentation clearly distinguishes supported, warning, partial and blocked outcomes.

The final unseen paper exposed captions separated from their figure plates. The generic cropper now supports an explicitly verified caption-page and plate-page binding, while keeping crop provenance, page identity and the same quality gate. Ambiguous bindings remain blocked instead of being silently accepted.

During the RC period, accept only release-blocking fidelity, security, privacy, data-protection and supported-path installation fixes. Before stable v1.0.0, freeze the exact candidate and run one final source-closed smoke test without changing the rules. A clean result permits the stable release. Independent users are valuable but not a hard dependency; isolated profiles, frozen holdouts and public RC issue feedback provide the fallback validation path.

Deferred beyond v1.0

  • Zotero write-back or bidirectional synchronization;
  • batch processing and multi-paper literature reviews;
  • knowledge graphs and automatic Vault-wide linking;
  • paid MinerU APIs;
  • unattended modification or publication of the formal Skill;
  • support claims for browser-only agents without local filesystem and Shell access.