Skip to content

CHANGELOG.md

Latest commit

 

History

History
281 lines (210 loc) · 34.9 KB

CHANGELOG.md

File metadata and controls

281 lines (210 loc) · 34.9 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[2.0.0-beta.1] — 2026-07-21

Product Direction

  • Functional convergence: Defined ScholarAIO as an All-in-One academic harness for agents. All-in-One now means one coherent workflow for evidence, persistent context, tools, outputs, and verification; it does not mean bundling every scientific package or expanding into a general autoresearch or multi-agent platform.
  • 2.x public contract: Documented the stable CLI, configuration, runtime-layout, skill-discovery, persistent-data, and published Python API surfaces, together with a deprecation policy for the 2.x line.
  • Bounded integration gate: New third-party integrations must demonstrate a core academic need, avoid duplicating agent-native or existing capabilities, isolate optional dependencies, pass a fixed-corpus or end-to-end smoke, and fail with an actionable fallback.

Added

  • Durable product strategy: Added the repository-root STRATEGY.md as the product-scope anchor for future planning and implementation work.
  • 2.0 upgrade guide: Added a no-migration path for 1.4/1.5 users, the explicit migration path for older runtimes, and replacements for removed or narrowed surfaces.
  • EndNote-style Library WebUI workflows (#114): Added canonical one-click BibTeX copy, separate inline and operating-system PDF actions, composable title/author/year/journal/type/DOI filters, and explicit Metadata/Keyword/Semantic/Unified search modes with stable ranked results and actionable index/vector diagnostics. Native PDF launch is stable-ID-only and anti-CSRF protected, bridges WSL PDFs into the Windows default application, automatically becomes a client download for remote deployments, and falls back to downloading if a native launch fails.
  • Persistent WSL PDF editing (#121): Added stable Windows-side edit mirrors with validated automatic reconciliation back to canonical WSL library PDFs and bounded recovery copies.
  • Repository knowledge map (#112): Added a published docs/DESIGN.md knowledge map and focused documentation indexes for agent navigation, while keeping maintenance plans, validation records, and audit notes out of the published MkDocs surface.
  • Automated cross-project agent setup (#111): Added scholaraio setup agent preview, apply, and check workflows for shell runtime wiring, Codex/OpenClaw skill discovery, Claude Code plugin instructions, and project-local wrappers for Qwen, Cursor, Cline, Windsurf, and GitHub Copilot.
  • Nature workflow bridge skill (#107): Added a ScholarAIO nature-workflow bridge skill that routes Nature Portfolio writing and figure workflows to the upstream nature-skills repository when installed, keeps ScholarAIO-native fallbacks explicit, documents the install and quick-start path, and includes deterministic plus product-demo fixtures that generate reviewable manuscript, figure, slide, and QA artifacts.

Changed

  • Optional MinerU cloud dependency: Moved mineru-open-api out of the mandatory base install into the dedicated mineru-cloud extra. The cloud parser remains supported, but users who only search existing libraries or use local PDF parsers no longer receive a provider-specific executable package. A clean-wheel CI smoke now verifies that the base package installs and starts without that external CLI.
  • Host-native skill routing (#127): Shared skills now request capabilities instead of hard-coding one agent host, prefer the active agent's native browsing and artifact abilities when appropriate, and keep local CLI fallbacks explicit. Oversized document instructions were split into focused references and the skill harness now enforces progressive disclosure.
  • Scientific-tool onboarding: Existing toolref workflows remain supported, but onboarding now starts with the 2.x integration gate and explicitly prefers external recipes or sidecars when a proposal does not belong in the core academic harness.
  • Paper2Any installation contract (#127): Setup and documentation now follow the current upstream base and paper requirement files, fail closed when the checkout shape is unsupported, and keep Paper2Any an optional sidecar.

Fixed

  • Semantic-filter independence (#116): Decoupled semantic-search metadata filtering from the FTS index so vector workflows do not require unrelated keyword-index state.
  • Cross-platform PDF delivery (#118, #119): Made local, WSL, headless, systemd, and remote WebUI PDF actions reflect what the host can actually do, including reliable Windows launcher discovery and client-download fallback.
  • Windows Unicode stdio fallback (#126; original report and patch #124): Prevented localized CLI output from crashing on legacy Windows console, file, or pipe encodings by relaxing only the affected stream error handlers, while preserving non-Windows behavior and explicit PYTHONIOENCODING settings.
  • Windows setup-check console compatibility (#113): Replaced non-cp1252 status symbols in the default English setup check output so Windows consoles no longer raise UnicodeEncodeError during final verification.
  • Webextract Markdown table-cell cleanup (#110): Sanitized malformed block-level code fences emitted inside qt-web-extractor table cells before HTTP/MCP extraction results reach webextract and ingest consumers, while preserving standalone fenced code blocks and pipe characters inside code-cell content.

Removed

  • Default external web-search surface: Removed the implicitly matched websearch skill, CLI command, default MCP registration, setup check, configuration template, and cross-skill routes. Live discovery now uses the host agent's native web search; webextract remains narrowly scoped to rendered content and ingest-link persistence.
  • Empty draw package extra (#127): Removed the dependency-free scholaraio[draw] extra and unowned package recommendations. Diagram source generation remains available, while Graphviz and Inkscape stay explicit system tools.

[1.5.0] — 2026-05-24

Added

  • Line-addressable evidence chunk search (#35): Added scholaraio index --chunks and scholaraio search --chunk so agents can build a paper-section chunk index from paper.md / meta.json["toc"] and retrieve source snippets with paper IDs, section titles, line ranges, and normal search filters such as --year, --journal, and --type.
  • Rights-respecting publisher PDF fetch (#98, #99): Added native scholaraio fetch-pdf support for DOI, landing-page URL, direct PDF URL, or title locators, plus --direct campus-network mode, safe temporary/single-file --ingest handoff, and selected/all-library canonical PDF refetch for already ingested papers.
  • Read-only local library WebUI: Added scholaraio gui, a packaged local WebUI for browsing the main paper library and proceedings child papers with live refresh, filters, metadata/quality inspection, Markdown-rendered abstracts/conclusions, local-only math rendering, and inline PDF preview/fullscreen controls.
  • Graphviz DOT/SVG diagram workflow: Added the Graphviz diagram guide, linked it from the draw skill and writing/CLI docs, and expanded setup check to report Graphviz dot and Inkscape with actionable install guidance for SVG rendering and Beamer insertion.
  • Paper2Any MCP sidecar integration: Added scholaraio paper2any setup, serve, status, and smoke commands, a lightweight MCP sidecar for Paper2Any conversion workflows, configuration and setup diagnostics for the extension checkout and backend API key, and user/agent documentation for running the sidecar from ScholarAIO.

Changed

  • English CLI and service messages: Migrated user-facing CLI help, status output, warnings, and service error messages from Chinese to English, and updated the corresponding regression tests so ScholarAIO no longer preserves Chinese compatibility for these message strings.

Fixed

  • Paper PDF preservation during ingest and repair: Kept original PDFs beside paper.md in the canonical paper directory, using the paper directory stem for the PDF filename, avoided overwriting an existing curated PDF when a duplicate DOI path only needs to restore missing Markdown, and made attach-pdf refuse to replace an existing canonical PDF unless --force is supplied.
  • Local WebUI robustness and privacy: Hardened the WebUI against malformed metadata, stale async list/detail responses, full-library re-audits on every poll, non-ASCII PDF filenames, large PDF buffering, stale PDF toolbar state, stale type filters, and remote runtime script loading.
  • Diagram CLI output noise: Stopped service-layer diagram generation logs from reusing the user-facing Generated: prefix, so diagram --from-text reports the generated artifact only once.
  • Local MinerU batch image assets: Saved images returned by the local MinerU API into per-PDF <pdf_stem>_images/ directories and rewrote Markdown references accordingly, preventing cmd_batch runs that share one output directory from overwriting generic image names such as image_1.png.

[1.4.0] — 2026-04-25

Added

  • Fresh-layout runtime and one-command upgrade: Standardized the current runtime layout under data/libraries/, data/spool/, data/state/, and workspace/_system/, and added scholaraio migrate upgrade --migration-id <id> --confirm as the release-grade path from supported old roots to the fresh layout.
  • Migration finalization safety gates: Hardened migrate finalize with journaled verify -> cleanup -> verify, workspace refs/papers.json migration, system-output migration into workspace/_system/, and archival cleanup for both populated and empty legacy roots.
  • Canonical package architecture: Completed the package split into core, providers, stores, projects, services, and interfaces, keeping scholaraio.cli as the published entrypoint while moving implementation code into canonical namespaces.
  • Webtools MCP support: Added generic MCP transport support for external webtools, including GUILessBingSearch (search_bing) and qt-web-extractor (fetch_url) while keeping HTTP transport available for hosts that need it.
  • Release validation evidence: Added release validation reports that cross-check current behavior against origin/main, v1.3.1, actual CLI canaries, migration rehearsal evidence, and docs/skill alignment.
  • Guided single-paper reading workflow: Added the paper-guided-reading skill plus the companion docs/writing-guide/paper-reading-framework.md so agents can start from fuzzy intent, confirm one target paper, perform structured deep reading, and persist reusable findings into notes.md
  • Incremental metadata scrub workflow: Added the scrub skill for post-enrich metadata cleanup of low-quality paper records, plus reusable .scrubbed marker helpers in papers.py and conservative scrub-suspect detection helpers in audit.py
  • Rsync backup workflow (#54): Added typed backup configuration, the scholaraio.backup module, scholaraio backup list/run, and the backup skill so ScholarAIO data can be synced to named remote targets through rsync instead of hand-written shell commands
  • macOS semantic-search smoke workflow: Added a dedicated GitHub Actions job on macos-14 to exercise the sentence-transformers -> faiss-cpu semantic-search stack and run targeted regression tests for vector-search paths
  • Academic writing router (#55): Added the academic-writing skill as a stable top-level entry point that routes users by deliverable and writing stage instead of forcing them to guess among multiple writing skills
  • Deliverable-first writing workflows (#55): Added lightweight poster and technical-report skills so conference posters, poster-style summaries, topic reports, and research briefings are first-class workflows rather than implicit combinations of lower-level skills
  • Writing workflow regression coverage: Added tests for skill frontmatter validity, router references, approximate host-style skill selection, and 11 rounds of documentation-alignment checks across docs, agent instructions, and marketplace metadata
  • Rendered web URL ingestion (#52): Added the native scholaraio ingest-link CLI, ingest-link skill, and sources.webtools connector flow so webpages and online PDFs can be ingested through an external qt-web-extractor daemon while preserving provenance fields such as source_url, source_type, extracted_at, and extraction_method
  • References-only metadata refresh: scholaraio refetch can now backfill structured references for DOI-bearing papers, with --references-only / --refs-only to target only papers whose references are still empty and a Semantic Scholar -> Crossref fallback chain when one source has no usable references

Fixed

  • Diagram IR rendering robustness: diagram --from-ir now accepts common source / target edge aliases in addition to the canonical from / to keys, and reports malformed edges as clean CLI errors instead of tracebacks.
  • Fresh-layout docs and agent entries: Slimmed agent entry docs into lightweight navigation surfaces, moved deeper runtime guidance into docs/guide/agent-reference.md, and aligned README, CLI docs, skills, and setup docs around explicit migration instead of implicit legacy runtime reads.
  • Skill metadata and routing docs: Normalized active project skill frontmatter to the cross-agent name + description shape, made descriptions trigger-focused, refreshed the skill harness validator, and fixed router wording that could steal explicit review-response requests.
  • Runtime-layout migration correctness: Fixed migration verification against real migrated libraries, empty legacy root cleanup, workspace output migration, and recovery/finalization edge cases found during repeated live CLI rehearsals.
  • CLI namespace refactor coverage: Moved command handlers and shared CLI helpers under scholaraio.interfaces.cli.*, keeping command behavior stable while removing implementation dependence on root-level facade modules.
  • Webtools runtime robustness: Hardened MCP/HTTP webtools configuration, command output, service error handling, and real local service canaries for websearch, webextract, and ingest-link.
  • Topic and search edge cases: Fixed offline topic CLI validation and search-result citation formatting for legacy scalar citation_count values.
  • Writing-skill discovery alignment: Synchronized academic-writing, docs/guide/writing.md, README.md, README_CN.md, docs/index.md, AGENTS.md, AGENTS_CN.md, CLAUDE.md, and clawhub.yaml so paper-guided-reading is discoverable consistently across router, docs, agent instructions, and marketplace metadata
  • Audit title matching and type-aware skips: audit now compares metadata titles against title-like candidates from the first 80 lines of paper.md, honors title_translated, keeps missing_doi / missing_journal warnings active when paper_type is blank, and skips front-matter-driven title_mismatch false positives for dissertation and document-like records
  • Backup runtime robustness: scholaraio backup run now reports missing rsync executables as controlled CLI errors, shell-quotes the displayed rsync command preview, defaults full-data backups to the safer default rsync mode instead of append-only behavior, forces SSH batch mode so runs fail fast instead of hanging on interactive authentication or host-key prompts, supports a config.local.yaml password fallback for password-only hosts, and prints concrete setup guidance when authentication or host trust is not ready yet
  • macOS semantic/unified search crash (#65): main-library and explore semantic search now embed and normalize the query before loading or searching FAISS indexes, avoiding a known faiss / sentence-transformers import-order segfault pattern on macOS while preserving existing ranking behavior
  • Academic writing docs alignment (#55): Synchronized docs/guide/writing.md, README.md, README_CN.md, docs/index.md, AGENTS.md, AGENTS_CN.md, CLAUDE.md, and clawhub.yaml around a router-first writing model so poster/report workflows and the academic-writing entry point are discoverable consistently across user and agent surfaces
  • ingest-link reliability and isolation (#52): URL ingest now preserves extractor PDF autodetect unless --pdf is explicitly requested, isolates both temporary inboxes from the real library, skips only failed URLs in multi-link batches, keeps warning-bearing extractions with usable text, retries transient extraction failures with exponential backoff, and avoids overlong fallback filenames for title-less URLs
  • Scrub workflow edge cases (#51): show now surfaces the stable UUID for partially corrupted records, direct-directory repair generates a UUID when recovering markdown-only papers, collision-suffixed directory names no longer get renumbered again by rename, and the scrub skill docs now distinguish invalid_metadata records from normal show-first review paths

Removed

  • Legacy runtime auto-detection as normal behavior: Fresh-layout accessors no longer auto-open old runtime roots such as data/papers/, data/explore/, data/proceedings/, or data/inbox*; those paths are migration inputs handled by scholaraio migrate upgrade.
  • Legacy root-level public facades: Removed obsolete public facade modules such as scholaraio.index, scholaraio.workspace, and scholaraio.translate; new code imports canonical namespaces directly.

[1.3.1] — 2026-04-14

Added

  • Qwen agent support: Added .qwen/skills symlink to .claude/skills/ so Qwen-based agents can discover ScholarAIO skills out of the box
  • Qwen project context: Added .qwen/QWEN.md so Qwen Code has a repository-native project context file instead of relying on AGENTS.md / CLAUDE.md
  • Cursor native project rules: Added .cursor/rules/scholaraio.mdc as the primary Cursor integration path, with AGENTS.md as the shared multi-agent instruction source and .cursorrules kept only as a legacy fallback
  • OpenAI-compatible embedding backend support: Added embed.provider config with local / openai-compat / none options; cloud API supports configurable api_base, api_key, api_timeout, batch_size, and max_retries; provider=none disables embeddings gracefully and falls back to keyword-only search

Fixed

  • Cursor compatibility CI coverage: Added regression coverage to keep the Cursor rule wrapper lightweight and explicitly MCP-free
  • Proceedings routing test stability: Tests no longer assume the first data/proceedings entry is a proceedings volume directory, avoiding CI-only failures when proceedings.db sorts before real volume folders
  • Unified search transparency: usearch and fsearch now print an explicit message when semantic retrieval is unavailable and the command degrades to keyword-only search
  • Zhipu GLM OpenAI-compatible chat routing (#60): call_llm() now maps open.bigmodel.cn/api/paas to the correct /v4/chat/completions endpoint instead of incorrectly forcing /v1/chat/completions

[1.3.0] — 2026-04-06

Added

  • AI-for-Science foundation: ScholarAIO v1.3.0 pushes the project beyond a paper-centric research terminal toward an AI-for-Science runtime. Added five lightweight scientific-computing domains for agents: Quantum ESPRESSO, LAMMPS, GROMACS, OpenFOAM, and bioinformatics
  • Versioned scientific tool docs via toolref: Added scholaraio toolref fetch/list/show/search/use plus the top-level scholaraio.toolref facade so agents can query exact official interfaces at runtime instead of guessing parameters from memory. Current indexed coverage includes Quantum ESPRESSO, LAMMPS, GROMACS, OpenFOAM, and curated bioinformatics tools
  • Extensible onboarding for new scientific software: Added a dedicated scientific-tool onboarding workflow so ScholarAIO can keep incorporating user-requested tools through official-doc ingestion, toolref integration, lightweight skill design, and end-to-end CLI verification, rather than being limited to the five tools already onboarded
  • Toolref-first scientific runtime design: Aligned tool-specific scientific skills around a clear separation of concerns: papers and notes hold scientific context, skills hold workflow and judgment, and toolref holds exact interface details. This keeps skills lightweight while letting agents stay grounded in both literature and tool docs
  • Semantic Scholar API key support: Configure ingest.s2_api_key (or env var S2_API_KEY) to authenticate Semantic Scholar requests, increasing rate limits from 100 req/5min (public) to 1 req/s (authenticated); polite delay automatically reduced from 3s to 1s when key is present
  • PDF parser benchmark harness: Added scholaraio/ingest/parser_matrix_benchmark.py plus tests for comparing Docling / MinerU / PyMuPDF parser runs and configuration matrices
  • Parser-aware setup guidance: scholaraio setup and the setup skill now explain MinerU vs Docling selection, provide official deployment links, note that MinerU tokens for mineru-open-api are free to apply for, and warn agent users about sandbox/network mis-detection
  • Insights analytics module coverage: scholaraio.insights now owns reusable behavior-analysis helpers, with dedicated tests plus CLI smoke coverage for scholaraio insights

Fixed

  • PDF parser fallback flow: Batch conversion and attach-pdf now follow the same MinerU → fallback behavior as the main ingest path; fallback assets are preserved; unsupported parser options from the previous broader design were removed so the active chain matches the current MinerU / Docling / PyMuPDF strategy
  • MinerU cloud backend + chunking limits: All MinerU cloud ingest entrypoints now use the mineru-open-api / ModelScope-backed path instead of the old raw API flow, and cloud chunk planning now respects both the 600-page and 200MB single-file limits with size-aware chunk estimation
  • Proceedings ingest routing: Regular data/inbox/ items no longer auto-route into data/proceedings/; proceedings now enter that workflow only through the dedicated data/inbox-proceedings/ inbox, and misclassified real-library proceedings shells were cleaned back into normal paper ingest
  • Setup robustness for agents: setup / setup check no longer fail hard when metrics.db is locked, parser recommendations honor an already-configured MinerU token before network probing, and interactive prompts treat EOF as empty input so agent-driven stdin does not crash the wizard
  • Docs consistency: README, README_CN, AGENTS, and CLAUDE now describe the current parser stack and setup behavior consistently
  • arXiv ingest edge cases: scholaraio.sources.arxiv no longer makes bs4 a transitive hard dependency for normal metadata flows, and old-style arXiv IDs like hep-th/9901001 now create parent directories correctly during PDF download
  • Scientific runtime docs compatibility: toolref runtime behavior, scientific skills, and published setup/docs metadata now match the refactored toolref facade and current public CLI/package surface
  • Optional dependency guidance: missing-dependency messages and setup check point Python dependencies to scholaraio[import], scholaraio[pdf], and scholaraio[office]; Graphviz and Inkscape remain explicit system-tool checks instead of a dependency-free draw extra
  • Translate / enrich CLI feedback and recovery: translate now reports chunk-level progress, persists per-chunk state in .translate_{lang}/, resumes unfinished work safely, and avoids writing fake success output when every chunk fails; enrich-toc now reports start/success/failure with extracted TOC counts for single-paper runs
  • Workspace removal and refetch status accuracy: ws remove now falls back to exact workspace dir_name matching when registry lookup misses, and refetch no longer reports spurious updates when API enrichment returns no authoritative data

Removed

  • MCP server: Removed scholaraio/mcp_server.py (1585 lines, 32 tools) and the scholaraio-mcp entry point. All agent interactions now go through CLI + skills, which are agent-agnostic and supported across Claude Code, Codex, Cursor, Windsurf, Cline, and GitHub Copilot. The [mcp] optional dependency group has also been removed.

[1.2.0] — 2026-03-26

Added

  • Agent analysis notes (T2): Per-paper notes.md for persistent cross-session analysis notes; show now auto-displays existing notes, show --append-notes appends new notes, and loader.load_notes() / loader.append_notes() expose the workflow in Python
  • Context management guidance: Workspace skill and 4 academic writing skills updated with notes.md read/write workflow and large-content delegation guidance for subagent-heavy analysis

Fixed

  • Zotero LaTeX filename too long (#32): Titles containing LaTeX math (e.g. $\mathrm{La}{\mathrm{BH}}_8$) or HTML/MathML entities now get properly cleaned before directory naming; added 255-byte filename length limit as safety net

[1.1.0] — 2026-03-24

Added

  • Patent literature management: New data/inbox-patent/ inbox for patent documents; automatic publication number extraction (CN/US/EP/WO/JP/KR/DE/FR/GB/TW/IN/AU + more formats); deduplication by publication number; paper_type: patent auto-tagging; publication_number field in PaperMetadata and papers_registry
  • Paper translation (translate CLI + skill): LLM-based markdown translation preserving LaTeX formulas, code blocks, and images; language detection heuristic; configurable defaults (config.yaml translate section) with per-call --lang/--force override; single paper and batch modes; show --lang to view translated versions; pipeline --steps translate for batch processing
  • Federated search (fsearch CLI + federated_search MCP tool): search across main library, explore silos (explore:NAME / explore:*), and arXiv in a single command; arXiv results annotated with "已入库" when DOI matches the main library
  • arXiv source module (sources/arxiv.py): shared Atom API client using defusedxml for safe XML parsing
  • Insights analytics (scholaraio insights): behavior dashboard showing top search keywords, most-read papers, weekly reading trend, semantic neighbor recommendations, and active workspaces with paper counts
  • Metrics recording for search/read: search, usearch, vsearch, and show commands now record events to metrics.db for behavior analysis
  • MetricsStore.query_distinct_names(): efficient distinct-name query with supporting (category, name) index, used by insights recommendations
  • Skill YAML front matter: all 26 skills now carry standardized version/author/license/tags metadata; new insights and document skills added
  • clawhub.yaml: marketplace manifest listing all available skills for discovery
  • explore fetch --limit: cap the number of papers fetched from OpenAlex (useful for quick sampling)
  • attach-pdf --dry-run: preview what attach-pdf will do without actually running MinerU conversion
  • document inspect (scholaraio document inspect <file>): inspect Office documents (DOCX/PPTX/XLSX) showing structure, layout, content preview, and overflow warnings; new document.py module with inspect_pptx/inspect_docx/inspect_xlsx functions
  • Office format ingest: inbox-doc/ now accepts .docx, .xlsx, .pptx files; new step_office_convert pipeline step converts them to Markdown via MarkItDown before ingestion
  • RIS export: export ris outputs RIS format compatible with Zotero, Endnote, and Mendeley (zero dependencies)
  • Markdown reference list export: export markdown generates formatted reference lists with configurable citation styles (APA, Vancouver, Chicago, MLA); supports ordered/unordered lists
  • DOCX export: export docx converts any Markdown content to a Word .docx file, supporting headings, paragraphs, tables, lists, code blocks, and bold/italic text
  • Citation styles module (citation_styles.py): manages built-in (APA/Vancouver/Chicago/MLA) and custom citation formats; custom styles loaded from data/citation_styles/*.py with path-traversal protection
  • draw skill (.claude/skills/draw/): generate diagrams (Mermaid flowcharts, sequence diagrams, ER diagrams, Gantt charts, mind maps) and vector graphics (cli-anything-inkscape); outputs to workspace/figures/
  • [office] optional dependency group: markitdown[docx,pptx,xlsx] + python-docx

Fixed

  • Chicago citation format: empty authors list no longer causes IndexError; condition reordered to check not authors first (consistent with APA/Vancouver)

  • Federated search DOI annotation: WHERE doi IN (...) replaced with WHERE LOWER(doi) IN (...) in cli.py, preventing false negatives when stored DOIs have different casing

  • insights --days validation: replaced args.days or 30 with explicit days <= 0 check; --days 0 or negative values now produce a clear error instead of silently defaulting to 30

  • CLI error messages and output text unified to Chinese

  • citation_styles: show_style(), list_styles(), get_formatter() error messages Chinese-ified; Google-style docstrings added

  • Translation same-language skip: language detection now recognizes common German/French/Spanish inputs, avoiding unnecessary same-language translation calls for supported targets

[1.0.0] — 2026-03-14

Added

  • Workspace batch add: ws add now supports --search "<query>", --topic <id>, and --all flags for bulk paper addition, with --limit/--year/--journal/--type filter support (--top remains a compatibility alias)
  • PDF optional dependency: pymupdf declared in pyproject.toml as [pdf] extra (included in [full]), fixing undeclared dependency for long PDF splitting
  • Subagent information tiers: T1/T2/T3 architecture documented in CLAUDE.md and AGENTS.md for structured context management

Fixed

  • MCP build_topics: nr_topics=0 now correctly maps to "auto" (automatic topic merging/reduction) instead of None (no reduction); added -1 as explicit "no reduction" value

[0.1.0] — 2026-03-13

Knowledge Base

  • PDF ingestion via MinerU (local API / mineru-open-api cloud CLI), with auto-splitting for long PDFs (>100 pages)
  • Three inboxes: regular papers (inbox/), theses (inbox-thesis/), general documents (inbox-doc/)
  • DOI-based deduplication; unresolved papers held in pending/ for manual review
  • Metadata extraction with 4 modes: regex, auto (regex + LLM fallback), robust (regex + LLM cross-check), llm
  • API-based metadata enrichment (Crossref, Semantic Scholar, OpenAlex)
  • L1–L4 layered content loading (metadata → abstract → conclusion → full text)
  • FTS5 full-text search index
  • FAISS semantic search with Qwen3-Embedding-0.6B, GPU-adaptive batch profiling
  • Unified search with Reciprocal Rank Fusion (RRF) combining keyword + semantic results
  • Author search and top-cited paper ranking
  • BibTeX export with year/journal filtering
  • Data quality audit with structured issue reports and LLM-assisted repair
  • BERTopic topic modeling with 6 HTML visualizations (hierarchy, 2D map, barchart, heatmap, term rank, topics over time)
  • Citation graph queries (references, citing papers, shared references)
  • Citation count fetching from Semantic Scholar / OpenAlex APIs
  • Workspace management for organizing paper subsets (search, export within workspace)

Content Enrichment

  • Table of contents (TOC) extraction via LLM
  • Conclusion (L3) extraction via LLM, with skip logic for non-article types (thesis, book, document, etc.)
  • Abstract backfill via LLM for papers missing abstracts
  • Concurrent LLM calls for batch enrichment (configurable worker count)

Literature Exploration

  • Multi-dimensional OpenAlex exploration (ISSN, concept, topic, author, institution, source type, year range, min citations)
  • Isolated explore datasets (data/explore/<name>/) with independent FTS5 + FAISS + BERTopic
  • Explore-specific unified/semantic/keyword search

Import & Export

  • Endnote import (XML and RIS formats)
  • Zotero import (Web API and local SQLite)
  • PDF attachment to existing papers
  • BibTeX export with filtering by year, journal, or paper IDs

LLM & Embedding

  • Multi-LLM backend support: OpenAI-compatible (DeepSeek/OpenAI/vLLM/Ollama), Anthropic (Claude), Google (Gemini)
  • API key resolution: config → environment variable → vendor-specific env vars
  • LLM token usage and API call timing via MetricsStore
  • GPU-adaptive batch embedding with automatic profiling and OOM fallback

AI Agent Integration

  • 22 Claude Code skills following AgentSkills.io open standard
  • MCP server with 31 tools
  • CLI with 29 subcommands (scholaraio --help)
  • Multi-agent compatibility: AGENTS.md, .cursorrules, .windsurfrules, .clinerules, .github/copilot-instructions.md
  • Claude Code plugin packaging (.claude-plugin/plugin.json, marketplace.json)
  • SessionStart hook for auto-installing dependencies in plugin mode
  • Global config fallback (~/.scholaraio/) for plugin usage outside the project repo

Project Infrastructure

  • Bilingual setup wizard (EN/ZH) with environment diagnostics
  • Code quality toolchain: ruff linter/formatter, mypy type checking, pre-commit hooks
  • CI workflow: lint, typecheck, test matrix (Python 3.10–3.12)
  • Contract-level test suite (36 tests across 6 modules)
  • Community governance: CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md
  • GitHub issue/PR templates (bug report, feature request)
  • CITATION.cff for academic citation
  • MkDocs documentation site with API reference (mkdocstrings)
  • Release workflow for PyPI publishing (trusted OIDC)