Release v2.0.2

Fixed

• TOML config: remove unnecessary quotes around bare key in [mcp_servers] section (Codex CLI)
• /mcp/status endpoint: repository and documentation URLs now point to this repo
• selfTest error log filter: updated regex to match new addon ref
• Vector database filename: zotero-agent-vectors.sqlite

Release v2.0.1

Changed

• CLI commands in client config generator now use single-line format (Claude Code, Qwen Code)

Release v2.0.0

First public release. Relative to upstream v1.5.0 (27 tools), tools expanded from 27 → 42, with PSK authentication + security model + run_javascript as core additions.

Security & Authentication

• PSK Bearer auth: POST /mcp validates Authorization: Bearer <PSK>; PSK auto-generated on first launch and stored in prefs, copyable/regenerable from preferences panel; auth.enabled defaults to on
• Defense in depth: loopback-only binding → PSK → eval.enabled defaults to off → write.enabled defaults to off
• DNS rebinding defense: POST /mcp validates Origin header, returns 403 for disallowed origins (covers opaque origin / IP suffix / path suffix edge cases)
• Write tools are dry-run by default, confirm: true to execute; read-back verification after writes

Core Tools (15 new)

• run_javascript: AsyncFunction executing arbitrary JS in Zotero's privileged context, with Zotero/ZoteroPane/ztoolkit/console injected; structured return {result, logs, error}; guarded by eval.enabled pref (default off); timeout_ms + 100KB result cap
• import_by_identifier: DOI/arXiv/ISBN/PMID import with if_exists idempotency (case-insensitive DOI dedup including Extra field, adsBibcode dedup, post-import read-back verification)
• find_missing_pdfs: Library/collection PDF missing audit + Unpaywall OA auto-fill
• check_retractions: scite.ai retraction check (no API key required; truthfully reports unreachable on network errors)
• find_related_papers: OpenAlex citation graph expansion, marked with inLibrary; fetch default reduced to 5 to prevent client timeouts
• synthesize_annotations: Per-paper annotation summary bundles; unified scope resolution logic (DRY)
• find_duplicates / merge_duplicates: Reuse Zotero's native dedup engine, dry-run by default, merged leftovers go to trash; truncated result disclosure
• batch_update_tags: Batch add/remove/rename (rename uses Zotero.Tags.rename preserving item associations), dry-run previews impact
• extract_identifier_from_pdf: Extract identifiers from full-text cache — DOI frequency voting + arXiv ID extraction
• find_doi: CrossRef title-based DOI reverse lookup, title similarity fused scoring (with diacritic folding), dry-run by default
• enrich_item_metadata: Fill missing fields from doi.org CSL-JSON + OpenAlex, field-level merge rules, dry-run by default
• manage_pdf_resolvers: Register Sci-Hub / Anna's Archive etc. into Zotero's native findPDFs.resolvers pref
• reload_plugin / install_plugin_from_url: Deployment loop, guarded by eval gate; self_upgrade flag

Search Enhancements

• Zero-result fallback ladder, responses annotated with fallback
• Semantic search hybrid mode: RRF fusion + query-adaptive weights; keyword leg degradation noted, date field keeps year fallback

Grey-Literature PDF Download

• Sci-Hub / Anna's Archive support: Preferences panel toggle + mirror list management (11 built-in defaults, add/remove/reset), bidirectional sync with Zotero's native resolver pref
• Three entry points: preferences panel, MCP tool, native right-click "Find Available PDF"
• Relaxed PDF link selectors for mirror DOM variance
• Sci-Hub download proxy: PAC data-url mode proxying only grey-source domains (default port 7890), normal traffic direct

Developer Experience

• scripts/deploy-live.mjs: One-click deploy, xpi base64 via run_javascript write + install_plugin_from_url self-upgrade, ~5s reconnect to new version
• In-process selfTest harness (26 scenarios full-stack regression, driven via run_javascript)
• Pure-function unit tests (70 cases, test/*.test.cjs, Node.js direct run, no framework)
• Eval boundary tests: timeout clamp, UTF-16 surrogate truncation, circular reference returns
• GitHub Actions CI: tag push auto build + Release + update.json; release notes auto-extracted from CHANGELOG

Fixes

• Chinese request body mojibake: HTTP read layer changed to raw byte collection + single-pass decoding, dense CJK bodies no longer trigger -32700 parse errors
• Tool execution errors return result.isError instead of JSON-RPC -32603 (MCP spec compliance)
• Protocol version negotiation instead of hardcoded value
• When write.enabled is off, tools/list hides collection write tools (no more "listed but not callable")
• Basic fields correctly placed per item type (conference paper venue → proceedingsTitle); date zero-pad formatting
• Subtitle splitting heuristic tightened to ≥3 tokens
• buildResolver ignores explicit undefined keys
• selfTest uses privileged XHR (fetch silently drops forbidden headers)
• Protocol version string unified globally

Internationalization

• Full en / zh-CN / de / es / fr / ja six-language support

---

Upstream baseline: cookjohn/zotero-mcp v1.5.0 (27 tools, MCP server embedded in Zotero plugin, hand-written nsIServerSocket HTTP, port 23120). Upstream archive at refs/AI-plugins/zotero-mcp-cookjohn submodule.