Skip to content

CHANGELOG.md

Latest commit

 

History

History
665 lines (523 loc) · 40.4 KB

CHANGELOG.md

File metadata and controls

665 lines (523 loc) · 40.4 KB

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]

[0.5.2] - 2026-03-18

Fixed

  • VSX marketplace MCP deduplication: vscode-extension/src/extension.ts now detects when another installed VS Code extension already provides PubMed Search or Zotero Keeper MCP servers and skips both persistent tool installation and duplicate MCP registration, preventing doubled tool lists in Copilot Chat
  • VSX persistent tool upgrades: vscode-extension/src/uvManager.ts now runs uv tool upgrade for already-installed managed tools during activation, so older machine-level installs are refreshed on newer extension releases instead of staying pinned silently
  • VSX documentation alignment: Synced root README.md, README.zh-TW.md, and vscode-extension/README.md with current MCP counts, 11-phase Auto Paper wording, persistent-install behavior, and external VSX deduplication notes
  • Cross-platform setup reproducibility: scripts/setup.sh, scripts/setup.ps1, and vscode-extension/scripts/build.sh no longer auto-track remote submodule HEAD during install/build; they now use repository-pinned submodule commits, generate the full repo .vscode/mcp.json shape consistently, and verify current MedPaper/CGU startup paths
  • JSONC mcp.json support: scripts/migrate_mcp_json.py now strips // comments from .vscode/mcp.json before parsing, preventing failures when VS Code leaves JSONC-style comments in the config
  • setup.sh heredoc template corruption: Rewrote the mcp.json heredoc block with consistent 2-space indentation and all 6 canonical servers, fixing structural corruption from overlapping patches
  • CHANGELOG first-line corruption: Removed garbage prefix from CHANGELOG.md first line that would fail Prettier CI checks
  • Ruff lint/format compliance: Fixed E402 import ordering and formatting issues in new test files to pass CI checks

Added

  • MCP progress notifications for long-running tools: Audit/review pipeline tools now use FastMCP Context.report_progress() when the client provides a progress token, covering phase validation, pipeline heartbeat, review-round start/submit, quality audit, meta-learning, data artifact validation, writing hooks, and review hooks
  • Cross-platform migration script (scripts/migrate_mcp_json.py): Non-destructive mcp.json migration — detects missing servers, normalizes JSONC→JSON, creates timestamped backups; shared loader used by smoke_test.py
  • Installation smoke test (scripts/smoke_test.py): 13-check cross-platform verification covering environment, core imports, MCP server creation, CGU integration, mcp.json validation, git submodules, and migration dry-run

[0.5.1] - 2026-03-17

Added

  • Top 20 anesthesiology journal profiles: Added templates/journal-profiles/ with 20 ready-to-use journal YAML presets, _index.yaml, README, and scripts/generate_journal_profiles.py to regenerate the catalog from source metadata
  • Paper pre-commit hook (P-series): New scripts/hooks/paper_precommit.py registered in .pre-commit-config.yaml — runs P1/P2/P4/P5/P7 quality checks automatically on git commit, blocking on CRITICAL issues
  • Embedded post-write hooks: write_draft and patch_draft now auto-run A-series hooks (A1–A7, B9, B10, B15) after successful write — agent cannot skip, non-blocking advisory report appended to output
  • B2 protected content guard: patch_draft now blocks modification of 🔒-marked sections in concept.md — prevents agent from altering NOVELTY STATEMENT or KEY SELLING POINTS without user approval
  • Weak-model guardrails expansion: Added Code-Enforced B2/C2/P6 coverage, deprecated save_reference() warning + telemetry, and a dedicated weak-model regression suite in tests/test_weak_model_guardrails.py
  • Domain constraint expansion: Added shared anti-AI vocabulary core and broader paper-type coverage for meta-analysis, review-article, letter, and other, plus structural/temporal/evidential constraints such as methods-before-results, minimum references, and section-overlap detection

Fixed

  • Hook batch runner: run_post_manuscript_hooks() now calls C10 (full manuscript validation), C11 (citation distribution), C12 (citation decision audit), C13 (figure/table quality) — previously defined in mixin but not wired into the batch runner
  • MCP hook aliases: audit_hooks.py ALL set now includes A7 (reference sufficiency) and C7B (asset coverage); POST-WRITE +A7, POST-MANUSCRIPT +C7B; docstring corrected from 37→40 hooks
  • Hook effectiveness tracker: HOOK_CATEGORIES now tracks P (pre-commit) and G (git-hooks) event categories
  • Consistency checker accuracy: scripts/check_consistency.py now understands sub-hooks like A3b, counts 78 hooks correctly, and validates P/G hook categories instead of misreporting false failures
  • Documentation alignment: AGENTS.md, .github/copilot-instructions.md, vscode-extension/copilot-instructions.md, and vscode-extension/README.md now reflect 78 hooks / 55 Code-Enforced / 23 Agent-Driven and 88 MCP tools consistently
  • VSX bundled parity: Re-synced bundled Python sources and bundled skills with workspace source so local vitest sync checks and packaged extension content match the repository state

[0.5.0] - 2026-03-11

Added

  • Hook A3c: Voice Consistency Detector (Anti-AI 語體一致性偵測):

    • New Code-Enforced hook check_voice_consistency() in _post_write.py — detects paragraph-level style breaks via z-score outlier analysis
    • Per-paragraph metrics: avg_sent_len, avg_word_len, type_token_ratio, punct_complexity
    • Document baseline computation (mean + std per metric)
    • Outlier detection (z-score > 1.8) + vocabulary sophistication gap (max−min avg_word_len > 1.2)
    • Catches the #1 human-reviewer signal: ESL paragraphs suddenly switching to polished corporate-academic prose
    • Registered in WritingHooksEngine batch runner (run_post_write_hooks) and pre-commit P2c delegation
    • Integrated into R5 post-review anti-AI gate (now runs A3 + A3b + A3c)
    • MCP dispatch: run_writing_hooks(hooks="A3C") available in ALL and POST-WRITE hook sets
    • 18 new tests covering uniform text, mixed voice detection, sophistication gap, custom thresholds, markdown handling, and batch integration
    • Hook count: 77 → 78 (36 Code-Enforced / 42 Agent-Driven)

  • Dual-Subagent Anti-AI Cross-Audit Protocol (Stage C3):

    • SKILL.md Stage C3 rewritten: three-layer code scan (A3+A3B+A3C) → agent self-review (6 criteria incl. voice breaks) → dual-subagent cross-audit → verification
    • Concept-Challenger: new Anti-AI Surface Scanner role (blacklist + structural signals + GPTZero risk assessment)
    • Domain-Reviewer: new Voice Analyst role (per-paragraph AI probability scoring 1-5, ESL baseline analysis, TOP 5 suspicious sentences)
    • Cross-comparison protocol: 🔴 both flag → must rewrite, ⚠️ one flags → judge, ✅ both safe → pass

  • Asset Review Receipt Hard Gate for Figures/Tables:

    • New analysis tool review_asset_for_insertion() records auditable review receipts in .audit/data-artifacts.yaml
    • insert_figure() / insert_table() now BLOCK caption registration unless a matching review receipt exists first
    • Phase 5 gate now validates planned figure/table captions are backed by review receipts
    • Hook F / validate_data_artifacts() now report missing or incomplete asset review receipts as gate-blocking issues
    • 1 new targeted tool test file + expanded Phase 5 / Hook F coverage

  • Paper-Type-Aware Reference Minimum Enforcement (Hook A7 + Phase 2 Gate):

    • New Hook A7 check_reference_sufficiency() — Code-Enforced pre-write gate that blocks writing when reference library is insufficient
    • Phase 2 Gate now paper-type-aware: reads paper.type from journal-profile.yaml, resolves minimum via 3-tier chain
    • Per-type minimums: original-research (20), review-article (30), systematic-review (40), meta-analysis (40), case-report (8), letter (5), fallback (15)
    • journal-profile.template.yaml: new minimum_reference_limits field for per-project override
    • DomainConstraintEngine: B003 minimum_references BOUNDARY constraint added to all 3 paper types
    • Phase 3+ prerequisite checks also paper-type-aware (can't bypass by skipping Phase 2)
    • 8 new A7 hook tests + 7 new Phase 2 gate tests (839 tests total)
    • Hook count: 76 → 77 (35 Code-Enforced / 42 Agent-Driven)

  • Humanizer Anti-AI Enhancement:

    • ANTI_AI_PHRASES: 76 → 133 phrases across 12 semantic categories (overly_formal, unnecessary_hedging, ai_conclusions, filler_boosters, generic_linking, hollow_emphasis, ai_structuring, inflated_academic, ai_discourse, passive_deflectors, nominalised_verbs, hollow_intensifiers)
    • AI_TRANSITION_WORDS: 25 → 33 (added: Nevertheless, Conversely, Correspondingly, Notably, Importantly, Significantly, Fundamentally, Substantially)
    • 4 new A3b structural pattern checks: #6 negative parallelism, #7 copula avoidance, #8 em dash overuse, #9 false ranges (X to Y)
    • 12 new tests for anti-AI detection (826 Python tests total)

  • VS Code Copilot Lifecycle Hooks:

    • 7 hook scripts: session-init.sh, prompt-analyzer.sh, pre-tool-guard.sh, post-tool-check.sh, pre-compact-save.sh, subagent-init.sh, session-stop.sh
    • .github/hooks/mdpaper-lifecycle.json configuration
    • Design doc: docs/design/copilot-lifecycle-hooks.md
    • State communication via .github/hooks/_state/ (gitignored)
    • Graceful degradation when jq is not installed

Fixed

  • Asset Review Receipt — Caption Normalization & Auto-Review:
    • Caption comparison now uses _normalize_caption() — strips trailing punctuation, whitespace, and lowercases before matching (fixes brittle exact-match failures)
    • insert_table() with inline table_content now auto-records review receipt (agent has full content access), eliminating redundant review_asset_for_insertion() call
    • File-based tables (no table_content) still require explicit review receipt (hard gate preserved)
    • 5 new tests: TestCaptionNormalization (4 unit tests) + auto-review integration test
  • MCP Instructions: Removed ghost tool save_diagram_standalone (merged into save_diagram), added missing insert_figure/insert_table/list_assets to DATA ANALYSIS section, updated DIAGRAM WORKFLOW
  • Tool count drift: 85 → 86 tools (review/ 21 → 22), synced across 5 files via sync_repo_counts.py --fix

Changed

  • Tool count: 86 → 88 tools (analysis/ 9 → 10: review_asset_for_insertion, review/ 22 → 23: approve_concept_review)
  • Hook count: 76 → 78 (34 → 36 Code-Enforced): added A3c Voice Consistency + A7 Reference Sufficiency
  • VSX Extension: Synced copilot-instructions.md + auto-paper/SKILL.md, added 3 research skills (memory-checkpoint, memory-updater, project-init)

[0.4.6] - 2026-03-02

Added

  • VSX Zero-Config Marketplace Mode:
    • uvManager.ts — cross-platform uv auto-detection + headless installation (Windows PowerShell / Unix curl)
    • ensureUvReady() — auto-install uv on activation with VS Code progress notification
    • Marketplace mode uses uvx med-paper-assistant from PyPI (complete isolation, no PYTHONPATH contamination)
    • CGU marketplace mode uses uvx creativity-generation-unit
  • VSX Testability Refactor:
    • extensionHelpers.ts — 6 pure functions extracted from extension.ts (no vscode API dependency): shouldSkipMcpRegistration, isDevWorkspace, isMedPaperProject, determinePythonPath, countMissingBundledItems, buildDevPythonPath
    • extensionHelpers.test.ts — 30 tests covering all extracted helpers
    • packaging.test.ts — 21 tests for manifest schema, .vscodeignore, module structure, version consistency
    • uvManager.test.ts — 20 tests (expanded from 17) including async findUvPath tests

Fixed

  • Critical: shouldSkipMcpRegistration() now checks both "mdpaper" server name AND med_paper_assistant module path — prevents skipping when user has unrelated mcp.json entries
  • Critical: determinePythonPath() only returns 'uv' fallback for med-paper-assistant pyproject.toml — prevents wrongly treating any pyproject.toml as a valid Python path

Changed

  • VSX vitest: 52 → 106 passed (+54 new tests across 4 test files)
  • CGU bundled tools updated to latest version

[0.4.5] - 2026-03-02

Fixed

  • Critical: _strip_references_section() greedy regex (.* with re.DOTALL) deleted all content after ## References — any sections like ## Appendix or ## Supplementary Materials were silently lost during export. Fixed to non-greedy .*?(?=\n## |\Z) with lookahead.
  • 2 regression tests added for post-References content preservation

[0.4.4] - 2026-03-02

Added

  • Pre-export citation gate (C5 HARD GATE): export_docx() and export_pdf() now run Hook C5 (Wikilink Resolvable) as a blocking gate — export is refused if any [[wikilink]] citations cannot be resolved to saved references
  • Strict mode for prepare_for_pandoc(): New strict=True parameter raises ValueError on unresolved citation keys (used by export_docx/pdf internally as second defense layer)
  • Word export residual wikilink warning: save_document() scans the final Word document for any remaining [[...]] patterns and warns the user
  • 9 new tests covering all three defense layers

[0.4.3] - 2026-03-02

Fixed

  • Critical: MCP server now uses MEDPAPER_BASE_DIR env var (set by VSX extension to workspace folder) instead of CWD for projects/ and logs/ paths — fixes end-user issue where output went to extension install directory instead of workspace

[0.4.2] - 2026-03-02

Fixed

  • CI: Add # nosec B110 to checkpoint.py — bandit false positive on intentional try/except/pass causing CI failure

[0.4.1] - 2026-03-02

Fixed

  • Dashboard ThemeProvider: Replace setState-in-effect with useSyncExternalStore for hydration detection (ESLint react-hooks/set-state-in-effect)
  • Dashboard DrawioEditor: Fix ExportFormats type mismatch (xml → xmlsvg), remove unsupported style prop
  • Dashboard stats/route.ts: Next.js 16 params type changed to Promise<> with await
  • sync_repo_counts.py: Remove unused warnings variable (ruff F841)
  • Bundled Python sync: Fix rsync target to correct bundled/tool/ path

[0.4.0] - 2026-03-02

Added

  • VSX Extension Completeness:
    • agents/ directory with 9 bundled .agent.md files (concept-challenger, domain-reviewer, literature-searcher, meta-learner, methodology-reviewer, paper-reviewer, reference-analyzer, review-orchestrator, statistics-reviewer)
    • autoScaffoldIfNeeded() — auto-detect missing skills/agents/prompts and prompt user
    • build.sh step 2d — agent bundle copy
    • validate-build.sh V3c — agent sync validation
    • validateBundledAgents() utility function
    • mdpaper.audit bundled prompt
    • Agent sync tests (vitest)
  • macOS / VS Code Insiders Compatibility:
    • MCP environment inherits PATH, HOME, SHELL, LANG, USERPROFILE — fixes homebrew uv/uvx/git discovery on macOS
    • getPythonArgs supports versioned Python names (python3.12, python3.11) via regex
    • New test: versioned Python (homebrew/macOS) path matching
  • Dynamic Count Sync Script (scripts/sync_repo_counts.py): AST-based counting of all repo metrics, auto-sync into 7 docs, 43 stale counts auto-fixed
  • ReviewHooksEngine (R1-R6): Review report depth, author response completeness, EQUATOR compliance, review-fix traceability, post-review anti-AI, citation budget
  • EXPECTED_HOOKS: 40→58 (added B9-B16, C10-C13, R1-R6)
  • reset_review_loop MCP tool: Stuck review state recovery
  • citeproc-py: Moved to core dependencies (previously optional)

Fixed

  • Bug 1 (Critical): _compute_manuscript_hash() now hashes ALL .md files in drafts/ via sorted glob — resolves review pipeline deadlock when using multi-file drafts
  • Bug 2 (Critical): citeproc-py import failure crashes — added try/except lazy import with _CITEPROC_AVAILABLE flag at all 3 usage points
  • Bug 3 (Medium): start_document_session required template_name even when no templates available — now optional, creates blank document when omitted
  • Bug 4 (Medium): Writing Hooks false positives:
    • A1: Added _strip_frontmatter() to exclude YAML frontmatter from word counts
    • A6: Added statistical notation regex exclusion (F-tests, p-values, η²p, CI, OR/HR/RR, β, SD/SE) from n-gram overlap detection
  • Bug 5 (Minor): export_pdf citeproc crash chain — covered by Bug 2 lazy import guards
  • MCP server env: Child processes (uv/uvx/git) now inherit system PATH on macOS — prevents "command not found" errors
  • Hash mismatch error messages: Added diagnostic info showing current vs expected hashes

Changed

  • MCP tools: 85 (review/ 20→21 with reset_review_loop)
  • Hook architecture: 76 checks (34 Code-Enforced / 42 Agent-Driven)
  • Python tests: 698 → 730 passed (excl. external-dep tests)
  • VSX vitest: 34 → 35 passed
  • VSIX package: medpaper-assistant-0.4.0.vsix (615 KB, 224 files, 9 agents)
  • 4 outdated bundled skills synced from source

[0.3.12] - 2026-02-27

Added

  • B9-B16 Section-Specific Writing Quality Hooks (8 new Code-Enforced):
    • B9 Tense consistency — validates verb tense per section (past for Methods/Results, present for Discussion)
    • B10 Paragraph quality — minimum sentences, max length, transition word density
    • B11 Results objectivity — flags interpretive language in Results section
    • B12 Introduction structure — validates funnel structure (broad → specific → gap → aim)
    • B13 Discussion structure — validates interpretation → comparison → limitations → implications
    • B14 Ethical statements — checks for IRB/ethics approval and consent statements
    • B15 Hedging density — flags over-hedging (>15%) or under-hedging (<3%)
    • B16 Effect size reporting — checks for effect sizes alongside p-values
  • Self-Evolution Infrastructure (L2 complete):
    • ToolInvocationStore — tool telemetry persistence to .audit/tool-telemetry.yaml
    • PendingEvolutionStore — cross-conversation evolution persistence to .audit/pending-evolutions.yaml
    • guidance.py — build_guidance_hint + build_startup_guidance (auto-prompt pending evolutions on new conversations)
    • tool_health.py — diagnose_tool_health + flush health alerts to PendingEvolutionStore
  • Pipeline Flexibility (4 new MCP tools):
    • request_section_rewrite — Phase rollback with regression count guard
    • pause_pipeline / resume_pipeline — Pipeline suspend/resume with draft hash tracking
    • approve_section — Section-level approval gate (autopilot or manual)
  • Autonomous Audit Loop: autonomous_audit_loop.py — structured audit loop execution
  • CheckpointManager enhancements: Section approval tracking, pause/resume state, rollback history
  • WritingHooksEngine expansion: +1120 lines — Anti-AI pattern detection, data claim alignment, supplementary crossref, language consistency, overlap detection, data artifact validation
  • MetaLearningEngine: pending evolution flush (_flush_meta_learning_evolutions) for cross-conversation persistence
  • GitHub Actions: evolution-health.yml weekly health check workflow
  • Tool Logging: tool_logging.py — automatic tool invocation telemetry via decorator
  • New MCP tools: apply_pending_evolutions, diagnose_tool_health, run_autonomous_audit, get_tool_telemetry + 5 pipeline flexibility tools
  • CONSTITUTION v1.6.0: 新增第九章「核心哲學 — 逐步多輪演進」(§25-26)
  • Tests: +163 tests (test_writing_hooks +636 lines, test_meta_learning +191 lines, test_guidance, test_pending_evolution_store, test_pipeline_flexibility, test_tool_health, test_tool_invocation_store)

Fixed

  • C6 Word Count (ICMJE Convention): Body-only counting — excludes Abstract, References, Tables (markdown | | rows), Figure legends, Acknowledgments, Author info. New helpers: BODY_SECTIONS constant (12 sections), _strip_markdown_tables(), _extract_body_word_count(). Journal profile override via counts_toward_total flag.
  • count_words MCP tool: Rewritten to show Body? column (✅/—), ICMJE-labeled manuscript total
  • Deep review round 2: Defensive guards, assert removal, input validation
  • Code review: Input validation, regex robustness, safety guards across 6 files
  • Writing hooks: Word boundary regex for supplementary file matching, rglob for dir traversal, code block state tracking, wikilink stripping
  • Pipeline gate validator: Reference counting by PMID subdirs

Changed

  • MCP tools: 77 → 81 (pipeline flexibility + self-evolution tools)
  • Hook architecture: 23 Code-Enforced / 42 Agent-Driven (65 total)
  • Three-tier evolution: L1 Event-Driven Hooks ⚠️ / L2 Code-Level Enforcement ✅ / L3 Autonomous Self-Evolution ⚠️
  • Test count: 525 → 738 passed (42 test files)
  • AGENTS.md, copilot-instructions.md: complete rewrite reflecting actual implementation status

[0.3.11] - 2026-02-26

Added

  • Journal Support: medRxiv, JAMIA, JBI added to JOURNAL_REQUIREMENTS (total: 11 journals)
  • Pipeline Gate Validator: pipeline_gate_validator.py for phase-gate validation with reference counting fix (PMID subdirs)
  • Audit Hooks MCP Tool: audit_hooks.py for Hook D meta-learning integration

Fixed

  • check_formatting crash: 'Drafter' object has no attribute 'read_draft' — replaced with direct file reading via get_drafts_dir()
  • Pipeline gate: Added Path import, fixed _get_or_create_loop signature, removed unused variables
  • Pipeline gate: 29 API mismatches resolved in pipeline_gate.py
  • Gate validator: Count PMID subdirs instead of flat .md for reference counting

Changed

  • DRY refactor: Extracted 7 duplicate helper functions into _shared/project_context.py (get_project_path, get_drafts_dir, get_concept_path, validate_project_for_tool)
  • VSX bundled code: Full rsync from src/ — 38 files updated, 5 new files added
  • Auto-paper SKILL.md: Synced to VSX extension

[0.3.10] - 2026-02-25

Added

  • Author Info System: Author frozen dataclass value object with structured affiliations, ORCID, corresponding author flag
    • generate_author_block() produces markdown with superscript affiliations + corresponding author notation
    • update_authors MCP tool for managing project author metadata
    • create_project now accepts authors_json parameter for structured author input
    • get_current_project displays formatted author information
    • draft_section auto-injects author block for Title Page sections
    • journal-profile.template.yaml updated with authors: section template
  • Data Artifact Provenance: DataArtifactTracker persistence class for tracing data analysis artifacts
    • validate_data_artifacts MCP tool for cross-referencing data artifacts against drafts
    • All 4 analysis MCP tools (run_statistical_test, analyze_dataset, create_plot, generate_table_one) record provenance automatically
    • Phase 5/6 gate validators check data artifact completeness
    • CONSTITUTION §23.2 數據產出物溯源與交叉驗證 (v1.5.0)
  • Hook F1-F4: Data artifact validation hooks (溯源追蹤、manifest↔檔案一致、draft↔manifest 交叉引用、統計宣稱驗證)

Fixed

  • Full project ruff lint cleanup: 60 errors → 0 (unused imports, f-string placeholders, ambiguous variable names, import ordering, unused assignments, duplicate dict keys)
  • Pre-existing duplicate project_path key in get_project_info() dict literal (F601)

Changed

  • Hook count: 48 → 52 checks (added F1-F4 data-artifacts)

[0.3.9] - 2026-02-24

Added

  • Multi-Stage Review Architecture: 設計文件 docs/design/multi-stage-review-architecture.md，記錄 4 層審查機制（Prompt 15 + Skill 26 + Hook 42 + Agent Mode）
  • 42 Hooks 完整實作: Hook 系統從 38 → 42 checks
    • Hook C8 (Temporal Consistency Pass): post-manuscript 時間一致性檢查
    • Hook B5 (Methodology Audit): 方法學驗證
    • Hook B6 (Writing Order): 寫作順序強制
    • Hook B7 (Section Brief Compliance): Section Brief 合規
    • Hook D7 (Review Retrospective): 自我改進回顧
  • 11-Phase Pipeline (Phase 0-10): auto-paper 從 9-Phase 升級至 11-Phase
  • Auto-Paper Guide: docs/auto-paper-guide.md 完整操作手冊
  • VSX Template Bundling: journal-profile.template.yaml 打包進 VSX extension
    • setupWorkspace 自動複製 templates（含內容比對 auto-update）
    • mdpaper.autoPaper runtime safeguard（template 缺失時警告 + 一鍵修復）
    • validateBundledTemplates() 驗證函式 + 2 個新 unit tests
  • Citation Section: README.md + README.zh-TW.md 加入 BibTeX 引用區塊

Fixed

  • Bash Arithmetic Bug: build.sh / validate-build.sh 中 ((VAR++)) 在 set -e 下 VAR=0 時回傳 exit code 1 → 改用 VAR=$((VAR + 1))
  • V6 Vitest Detection: validate-build.sh 測試結果擷取修復（vitest v4 輸出格式變更）
  • Integration Test Async: 6 個 integration tests 修正為 async def + await（searcher.search() 已改為 async）
  • Integration Test CWD: test_drafter / test_insertion 改用 tmp_path fixture 避免 CWD 相依路徑問題
  • CI VSX Template Sync: ci.yml + release.yml 的 VSX sync step 加入 templates 同步
  • project_path resolution bug: get_project_info() 返回值缺少 project_path key

Changed

  • pytest default config: addopts = "-m 'not integration and not slow'" — 本地 pytest 預設只跑 unit tests（與 CI 行為一致）
  • VSX autopaper 描述: "9-Phase Pipeline + Hooks" → "11-Phase Pipeline + 42 Hooks"
  • Hook C7 (Temporal Consistency): 新增 post-manuscript hook 修正過時引用

[0.3.8] - 2026-02-20

Fixed

  • Lazy Import: matplotlib、seaborn、scipy 改為方法內 lazy import，修復 uvx 安裝時 ModuleNotFoundError: No module named 'matplotlib'（這些是可選依賴，核心 MCP server 啟動不應依賴）

[0.3.7] - 2026-02-20

Fixed

  • PyPI Entry Point: 加入 [project.scripts] 讓 uvx med-paper-assistant 正確啟動 MCP server（v0.3.6 缺少 CLI entry point 導致 Package does not provide any executables 錯誤）

[0.3.6] - 2026-02-20

Added

  • MedPaper: Setup Workspace 命令 ✅
    • 一鍵將 bundled skills (14)、prompts (12)、copilot-instructions.md 複製到 workspace
    • 只複製不存在的檔案（不覆寫已客製化的內容）
    • 完成後提示「重新載入」以啟用全部功能
    • Marketplace 安裝後執行一次即可獲得完整 Agent Mode 體驗

Fixed

  • CI Pipeline: test-vsx / publish-vsx 加入 skills/prompts/copilot-instructions.md 同步步驟
  • PyPI 重複發佈: publish-pypi 加入 skip-existing: true（v0.3.5 已上 PyPI）

Removed

  • Dead Code 清理: 移除 MDPAPER_INSTRUCTIONS、MDPAPER_EXTENSION_PATH 環境變數（Python MCP server 從未讀取）
  • Dead Code 清理: 移除 registerMcpServerProvider 中的 loadSkillsAsInstructions() 呼叫

0.3.5 - 2026-02-20

Added

  • Figure/Table Archive + Insert Tools ✅
    • Hook: _check_figure_table_archive 一致性驗證
    • 3 個新工具: insert_figure, insert_table, list_assets（54 → 57 tools）
  • GitHub Repo Metadata + Doc-Update Hook (G8) ✅
    • Repo description, 15 topics, 9 custom labels
    • scripts/check-doc-updates.py: 13 條規則映射檔案變更至文檔依賴
    • Pre-commit hook #15: doc-update-reminder (warn-only)
    • Hook 計數: 37 → 38 checks (G1-G8)
  • Prettier Markdown Formatter ✅
    • Pre-commit hook #14: mirrors-prettier v3.1.0
    • 格式化所有 121 個 .md 檔案
  • CI/CD Pipeline Upgrade ✅
    • CI: 2 → 5 jobs (python-lint, python-test, vsx, dashboard, markdown)
    • Release: 5-stage pipeline (validate → test → publish-pypi + publish-vsx → github-release)
    • Branch protection: 5 required CI checks, strict mode
    • 移除 Dependabot 配置

Fixed

  • README Submodule Links: 子模組相對路徑 404 → 改用 GitHub 絕對連結
  • VSX One-Click Install ✅
    • 移除 extensionDependencies 硬依賴（vscode-zotero-mcp 未上架會阻擋安裝）
    • Python fallback 改為 uvx（PyPI 已發布即可自動下載執行）
    • CGU MCP server 改為條件註冊（偵測到才啟用，避免錯誤訊息）

Documentation

  • README/README.zh-TW: 更新所有工具/Hook 計數（20 處：57 tools, ~107 total, 15 hooks）

Added

  • Placeholder Tools Implementation (Phase 8) ✅
    • 9 個佔位工具升級為完整實作（74→83 tools）
    • Analysis: analyze_dataset, detect_variable_types, list_data_files, create_plot, run_statistical_test, generate_table_one
    • Review: check_manuscript_consistency, create_reviewer_response, format_revision_changes
  • Tool Layer Architecture Audit (Phase 9) ✅
    • 7 個模板型工具（debate, critique, idea-validation）轉為 3 個 Skill 檔案
    • 新增 .claude/skills/academic-debate/SKILL.md
    • 新增 .claude/skills/idea-validation/SKILL.md
    • 新增 .claude/skills/manuscript-review/SKILL.md
    • 工具數量：83→76
  • Comprehensive Tool Consolidation (Phase 10) ✅
    • 6 大策略精簡 76→53 tools（-30%）
    • Strategy A: 移除無用工具 — close_other_project_files, export_word（legacy）
    • Strategy B: 簡單合併 — validate_for_section, get_project_paths, check_reference_exists 併入現有工具
    • Strategy C: 參數合併 — 6 組工具對合併（validate_concept +structure_only, get_current_project +include_files, update_project_settings +status/citation_style, save_diagram +output_dir, sync_workspace_state +clear, suggest_citations +claim_type/max_results, verify_document +limits_json）
    • Strategy D: 功能吸收 — consistency 檢查 + submission checklist 併入 check_formatting
    • Strategy E+F: Skill 轉換 — 7 個工具轉為 skill 知識（get_section_template, generate_cover_letter, list_supported_journals, generate_highlights, check_submission_checklist, create_reviewer_response, format_revision_changes）
    • 新增 .claude/skills/submission-preparation/SKILL.md（cover letter、highlights、journal requirements、reviewer response 模板）
    • 更新 draft-writing/SKILL.md、project-management/SKILL.md 反映工具變更
    • 測試驗證：35 passed / 21 pre-existing failures / 0 regressions
  • Citation-Aware Editing Tools (Layer 1+2) ✅
    • get_available_citations() — 列出所有可用 [[citation_key]]，含 PMID/作者/年份/標題表格
    • patch_draft(filename, old_text, new_text) — 部分編輯草稿，自動驗證所有 wikilinks
      • 唯一匹配檢查（防止模糊替換）
      • Wikilink 格式自動修復（[[12345678]] → [[author2024_12345678]]）
      • 引用存在驗證（拒絕 hallucinated citations）
    • 解決 Agent 使用 replace_string_in_file 繞過 MCP 驗證管線的核心問題
    • 14 個測試（3 test classes: GetAvailableCitations, PatchDraft, EditingIntegration）
    • SKILL.md 新增 Flow D: Citation-Aware 部分編輯
    • copilot-instructions.md 新增草稿編輯引用規則
  • Infrastructure & Quality Cleanup (Phase 3.5) ✅
    • Pre-commit hooks: 13 hooks（ruff, ruff-format, mypy, bandit, pytest, whitespace, yaml, json, toml, large files, merge conflicts, debug statements）全部通過
    • DDD Import 遷移：19 個測試檔從 core.* 遷移至 DDD 路徑
    • Test Isolation：所有測試改用 tmp_path fixture，不再污染專案根目錄
    • ARCHITECTURE.md 重寫：從 448 行過時文檔重寫為 ~240 行精確 DDD 架構文檔
    • Legacy Cleanup：刪除空的 core/ 目錄、多餘腳本、散落檔案
    • Copilot Hook 修復：AGENTS.md 補齊 7 skills + 8 prompts，修正 capability index
    • Coverage Baseline：32 passed / 1 skipped / 26 integration-deselected
    • 架構方向決策：選定 Direction C: Full VSX + Foam + Pandoc
  • Prompt Files 機制
    • 新增 .github/prompts/ 目錄，包含 9 個 prompt files
    • /mdpaper.project - 專案設置與切換
    • /mdpaper.concept - 研究概念發展（含 novelty 驗證）
    • /mdpaper.search - 智能文獻搜尋（情境 A/B 判斷）
    • /mdpaper.draft - 草稿撰寫（需先通過 concept 驗證）
    • /mdpaper.strategy - 搜尋策略配置
    • /mdpaper.analysis - 資料分析與 Table 1
    • /mdpaper.clarify - 內容改進與潤飾
    • /mdpaper.format - Word 匯出
    • /mdpaper.help - 指令說明
    • 參考 copilot-capability-manager 架構設計
  • 犀利回饋模式 (Sharp Reviewer Feedback)
    • concept_validator.py: 新增 _generate_novelty_feedback() 方法
    • 回饋格式：verdict / critical_issues / questions / actionable_fixes
    • CGU 創意工具整合建議
    • 像頂尖期刊 Reviewer 一樣審查：直指問題、用證據說話
  • 檔案保護模式 (File Protection)
    • .copilot-mode.json: 新增 protected_paths 設定
    • Normal/Research 模式下禁止修改開發檔案
    • 受保護路徑：.claude/, .github/, src/, tests/, integrations/
  • Session 檢視工具
    • scripts/view_session.py: 顯示 pubmed-search 搜尋紀錄
    • 可供人工驗證 Agent 確實執行了搜尋
  • 已知問題追蹤 (Known Issues)
    • ROADMAP.md 新增 4 個 Critical Issues
    • 新增 Phase 3.5: 學術品質保證系統

Changed

  • Novelty Check 改為 Advisory（不阻擋）
    • writing.py: _enforce_concept_validation() 改為建議性質
    • 用戶可選擇：直接寫 / 修正問題 / 用 CGU 發想
    • 仍然檢查結構完整性（NOVELTY STATEMENT, KEY SELLING POINTS）
  • concept-development SKILL 大幅更新
    • 新增犀利回饋原則和模板
    • 新增 CGU 工具使用指南
    • 新增危險信號處理流程
  • Pydantic V2 遷移
    • SearchCriteria: class Config → model_config = ConfigDict(frozen=True)
    • 消除 PydanticDeprecatedSince20 警告

Fixed

  • wikilink_validator.py: 移除未使用的 match.group(1) 呼叫
  • list_drafts / read_draft: 路徑解析改用 _get_drafts_dir() 取得專案路徑，與 patch_draft 一致

Documentation

  • AGENTS.md: 新增 Novelty Check 規則和 CGU 整合
  • copilot-instructions.md: 新增犀利回饋做法
  • pubmed-search-mcp ROADMAP.md: 新增 Phase 5.5 搜尋紀錄驗證機制

0.2.2 - 2025-12-18 (Previous)

Added

  • 完整靜態分析工具鏈
    • Ruff linter/formatter: 修復 1839 個錯誤
    • Mypy 類型檢查: 修復 49 個類型錯誤
    • Bandit 安全掃描: 7 個 Low severity 已加 # nosec 註解
    • Vulture 死代碼檢測: 0 個問題
  • 開發模式切換功能 (.copilot-mode.json)
    • development: 完整功能（所有 skills、Memory Bank 同步、靜態分析）
    • normal: 一般使用（僅研究技能）
    • research: 專注寫作（只同步專案 .memory/）
  • test-generator SKILL 擴展
    • 新增 Bandit/Vulture 工具文檔
    • 新增 # nosec 註解使用指南
    • 完整執行流程說明

Changed

  • 代碼品質改進
    • 所有 import * 改為明確導入
    • 所有 except: 改為 except Exception:
    • 統一使用 ruff format 風格
    • 修復所有 Optional type hints
  • pyproject.toml - 新增 dev 依賴: bandit>=1.9.2, vulture>=2.14

Fixed

  • 類型錯誤修復
    • concept_validator.py: 修正 result 變數衝突
    • project_context.py: 使用 get_project_info() 替代 get_current_project()
    • writing.py: 修正 Optional[str] 回傳類型
    • 多處 dict/list 變數加入明確類型註解

[0.2.1] - 2025-12-18 (靜態分析大掃除)

Added

  • MCP-to-MCP Direct Communication Architecture ✅ 已實作
    • pubmed-search 新增 HTTP API endpoints:
      • GET /api/cached_article/{pmid} - 取得單一文章
      • GET /api/cached_articles?pmids=... - 批量取得
      • GET /api/session/summary - Session 狀態
    • mdpaper 新增 PubMedAPIClient HTTP 客戶端
    • 新工具 save_reference_mcp(pmid, agent_notes):
      • Agent 只傳 PMID，無法修改書目資料
      • mdpaper 直接從 pubmed-search API 取得驗證資料
      • 防止 Agent 幻覺（hallucination）書目資訊
    • 分層信任 (Layered Trust) 參考檔案格式:
      • 🔒 VERIFIED: PubMed 資料（不可修改）
      • 🤖 AGENT: AI 筆記（AI 可更新）
      • ✏️ USER: 人類筆記（AI 絕不碰觸）
  • stdio + HTTP API 同時啟動
    • pubmed-search 在 stdio MCP 模式下自動啟動背景 HTTP API
    • start_http_api_background() 在 daemon thread 運行
    • 解決 VS Code MCP (stdio) 無法同時提供 HTTP API 的問題
  • Skill 文檔完整更新
    • literature-review/SKILL.md 完整重寫，含完整工具列表和 PICO 工作流
    • parallel-search/SKILL.md 新增工具表格和 Session 管理說明
    • concept-development/SKILL.md 擴展工具列表和 FAQ
    • 所有 skill 明確標示 save_reference_mcp 為 PRIMARY 方法

Changed

  • Reference 內容順序優化 - Abstract 移到 Citation Formats 之前
    • Foam hover preview 現在優先顯示 Abstract（更實用）
  • Foam settings 更新 - foam.files.ignore 改為 foam.files.exclude
  • sync_references Tool - Markdown 引用管理器
    • 掃描 [[wikilinks]] 自動生成 References 區塊
    • 可逆格式：[1]<!-- [[citation_key]] -->，支援重複同步
    • 按出現順序編號，支援 Vancouver/APA 等格式
  • Foam Project Isolation - 專案隔離功能
    • FoamSettingsManager 服務：動態更新 foam.files.ignore
    • switch_project() 整合：切換專案時自動排除其他專案
    • Whitelist 邏輯：只顯示當前專案的 references/
  • Reference Title Display - Foam 自動完成顯示文章標題
    • frontmatter 加入 title 欄位
    • foam.completion.label: "title" 設定
  • MCP Tool Logging System - 統一的工具日誌記錄
    • tool_logging.py: log_tool_call, log_tool_result, log_agent_misuse, log_tool_error
    • 日誌存放在專案目錄 logs/YYYYMMDD.log（跨平台支援）
    • 已整合至 draft/writing.py, project/crud.py, validation/concept.py, reference/manager.py
  • ReferenceConverter Domain Service - 支援多來源參考文獻
    • 支援 PubMed, Zotero, DOI 來源
    • ReferenceId Value Object 確保唯一識別符
    • Foam [[wikilink]] 整合
  • Reference Entity 更新 - 新增多來源識別符欄位
    • unique_id, citation_key, source 欄位
    • from_standardized() 類別方法

Changed

  • 授權變更 - 從 MIT 改為 Apache License 2.0
  • 日誌位置遷移 - 從系統 temp 目錄改為專案目錄 logs/
  • README.md - 新增 MCP 協調架構說明、更新工具列表
  • ARCHITECTURE.md - 新增 MCP Orchestration 架構圖
  • Prompts 更新 - /mdpaper.concept 和 /mdpaper.search 增加 MCP 協調流程說明
  • copilot-instructions.md - 簡化為參照 AGENTS.md，避免重複

Fixed

  • save_reference JSON 解析 - 處理 MCP 傳遞 JSON 字串的情況
    • 新增 Union[dict, str] 型別支援
    • 自動偵測並解析 JSON 字串輸入

Deprecated

  • save_reference_by_pmid - 改用 save_reference(article=metadata)

0.2.0 - 2025-12-17

Added

  • MCP 解耦架構：mdpaper 不再直接依賴 pubmed-search
  • 多 MCP 協調模式：Agent 協調 mdpaper + pubmed-search + drawio
  • 文獻探索工作區：start_exploration() / convert_exploration_to_project()
  • Concept 驗證系統：novelty scoring (3 rounds, 75+ threshold)
  • Paper type 支援：original-research, systematic-review, meta-analysis 等

Changed

  • Python 版本需求升級至 3.11+
  • ReferenceManager 重構：接受 article metadata dict 而非 PMID
  • 專案結構採用 DDD (Domain-Driven Design)

Removed

  • infrastructure/external/entrez/ - 文獻搜尋移至 pubmed-search MCP
  • infrastructure/external/pubmed/ - 同上
  • services/strategy_manager.py - 搜尋策略移至 pubmed-search MCP
  • tools/search/ - 搜尋工具改為 facade 委派

0.1.0 - 2025-12-01

Added

  • 初始版本
  • MCP Server 框架 (FastMCP)
  • 46 個 MCP 工具
  • Word 匯出功能
  • 參考文獻管理
  • 草稿撰寫流程