CHANGELOG.md

Changelog

Format (v0.95.0+): entries follow a lightweight Keep a Changelog convention going forward — within each version use ### Added / Changed / Deprecated / Removed / Fixed / Security where applicable, and flag breaking changes with a BREAKING: prefix. Per-release migration steps are mirrored in UPGRADE.md. Entries before v0.95.0 predate this convention and are kept verbatim (retrofitting 90+ historical entries was judged higher-risk than the consistency gain — G4 audit #3).

[Unreleased]

The UI 80/20 track (Phase B: ⌘K command palette + mobile breakpoints + _HOME wayfinding; Phase D: Zotero metadata correctness) is a SEPARATE future release — it was never folded into v1.1.0, which is the P2 "scale + provenance" release below. UI scope stays capped: the dashboard remains a thin status-mirror + palette + onboarding demo; no 3-pane / citation-graph rebuild (link out to the real tools instead).

[1.1.1] - 2026-06-14

Fixed

• vault gc dangling-link bug (found on a real vault during the v1.1.0 graph migration): find_orphan_mocs treated a hub/_moc/<name>.md page as orphan when no live cluster derived it — but did not check whether live paper notes still linked it. A note carrying a stale sub-MOC link after a rebind/merge (whose ## Hub block was never rewritten) would have had its _moc page deleted by vault gc --apply, stranding dangling wikilinks. run_gc now computes every wikilink target live (non-deleted) notes point at (note_linked_targets) and protects both the _moc pass AND the orphan-hub pass (a hub/<slug>/ linked by a cross-cluster [[slug/00_overview]] body reference is likewise protected). gc never makes the graph worse. Regression tests pin both guards. (The stale link itself is a separate, deferred cleanup.)

[1.1.0] - 2026-06-14

"Scale + provenance" — the roadmap P2 block (robustness sweep, PRISMA screening provenance, config-rot hygiene, vault gc), plus the v1.0.11 graph follow-ups that unblock the real-vault graph migration.

Added

• PRISMA screening provenance (P2-3): an append-only .research_hub/screening_log.jsonl records every relevance-gate decision (included / screened_out by reason, plus a unverified flag on papers the no-LLM gate kept without a clear verdict). New clusters prisma <slug> CLI + cluster_prisma MCP tool render the PRISMA flow counts. The no-LLM fit gate now emits these records and prints a cold-start nudge ("N papers ingested unverified — run an LLM fit-check to confirm").
• vault gc (P2-5d): garbage-collect aged raw/_deleted_<slug>/ residue (--older-than-days, default 30), orphan hub/<slug>/ dirs with no registry entry, and orphan hub/_moc/*.md pages no live cluster references; plus a paper-note ## Hub-block bare-parent strip (--no-strip-parents to skip). Dry-run by default; --apply to execute. PARENT_MOCS and query-derived sub-MOCs are protected from deletion.
• relevance_unverified is now persisted to paper-note frontmatter (it was computed by the no-LLM fit gate but discarded) — an unverified paper leaves an auditable trace.

Changed

• Windows FS-lock hardening (P2-1): the 6 remaining bare shutil.move sites (cluster archive / unarchive / cascade-delete soft-move; paper prune / unarchive / bulk-move) now route through fsops.robust_move (retry/backoff on transient PermissionError), completing the v1.0.7 sweep.
• Graph view label colors (P2-5a): vault graph-colors --refresh now emits label color groups ONLY for #label/* tags that actually exist in the vault, so a refresh no longer re-injects the 9 canonical label groups every time (a user who deleted the dead ones can now keep them deleted). The one-time init/bootstrap still pre-seeds the full palette.
• clusters._PARENT_MOCS is now frozenset(PARENT_MOCS) derived from the single source of truth in vault.hub_overview (was a hand-maintained literal mirror).
• The ## Hub-backfill path (vault hub-backlink-migrate) now drops the bare family parent on paper notes (for_paper_note=True), matching the live-ingest P1-4a policy (paper notes link the sub-MOC only).
• vault prune-footers now skips soft-deleted raw/_deleted_<slug>/ residue.

Removed

• Dead duplicate [tool.pytest.ini_options] block in pyproject.toml (pytest.ini is the authoritative pytest config; the pyproject block was ignored and silently drifting). Reconciled the stale "≥1-week RC bake" claim in docs/stable-api.md to the real release-gate cadence.

Security

• Typed MCP path-param validation contract (P2-5e): the MCP slug/identifier validation set is now an explicit module-level contract (_SLUG_PARAM_NAMES / _IDENTIFIER_PARAM_NAMES) and _validate_mcp_args fails closed on an unregistered slug-like kwarg. A new meta-test proves every slug-like parameter across the live tool registry is covered, closing the silent-bypass drift class (a new tool adding a path param under an unrecognized name).

Deferred (tracked for a follow-up)

• vault gc stale topic:-tag rewrite (the riskiest, content-mutating pass) and full PRISMA identified / deduped funnel emission (the screening log schema supports both stages; v1.1 auto-emits only the relevance-screen stage).

[1.0.11] - 2026-06-13

"Token economy + Obsidian graph" — make the vault graph a readable star-of-stars and stop the MCP surface from blowing the agent context window.

Changed

• Token-economy caps (P1-5). ask_cluster(mode='briefing') default cap lowered 100K→12K chars + a section= markdown-section projection. read_crystal at tldr/gist returns based_on_paper_count only — the full based_on_papers slug list + per-claim evidence + see_also ship only at level='full'. get_topic_digest paginates papers (default 10; limit=0 for all) and makes the markdown digest opt-in (include_markdown=True).
• Per-cluster MOC, capped footers (P1-4). derive_moc_links(for_paper_note=True) → paper notes link the per-cluster sub-MOC ONLY; the parent [[LLM-Agents]] / [[Water-Resources]] link now lives on sub-MOC + overview pages, not on every note — killing the O(n^2) super-hub clique. The backward footer-append re-ranks by tag overlap and caps at the same top-10 as the forward path (was unbounded). The v1.0.7 merged-cluster GC still protects the parent MOC (it uses the page variant of derive_moc_links).

Added

• research-hub vault prune-footers [--top N] [--apply] re-caps already-bloated Related-Papers footers (P1-4c). A doctor cluster/crystal_coverage check flags clusters that have papers but 0 crystals, and ask_cluster's digest fallback now returns an inline emit_crystal_prompt so an AI can crystalize on demand (P1-5d).

[1.0.10] - 2026-06-13

"Institutional access end-to-end" — make a paywalled-no-OA paper with a DOI localizable through the library's EZproxy by both the Zotero and auto-ingest paths, with fail-soft auth + a discoverable keepalive.

Added

• Institutional PDF localization (P1-1). find_pdf_url only ever returned open-access URLs, so EZproxy never fired for a paywalled-no-OA paper. New find_institutional_pdf_url resolves the publisher landing page (OpenAlex primary_location / DOI redirect), fetches it THROUGH the proxy, extracts the standard <meta name="citation_pdf_url">, and localizes the PDF — paper attach-pdfs and the NotebookLM auto-ingest path (fetch_paper_pdf(cfg=...)) both reach subscribed content now, reported as ezproxy-citation-pdf instead of a silent skip. The landing + PDF fetches reuse the v1.0.8 cookie-scoped _credentialed_get core, so the SSO session is never forwarded off-proxy.
• EZproxy fail-soft auth (P1-2). New ezproxy_probe (single-hop, cookie-safe) classifies the session live vs expired; doctor gains an ezproxy_session check with a real "last verified live: Nh ago" recency signal (from auth_state.json), and paper attach-pdfs surfaces a RequiresAuthRefresh one-liner (research-hub ezproxy login) when papers are paywalled and the session has lapsed — instead of a silent "no PDF".
• Keepalive discoverable + py3.14 --from-browser (P1-3). doctor nudges notebooklm keepalive --install-windows-task when the NLM session is live but the refresh task isn't registered; --from-browser gains a browser_cookie3 fallback (pure-Python, py3.14-compatible — rookiepy has no 3.14 wheel) behind the existing suffix-anchored cookie filter.

Changed

• _download_via_httpx_result refactored onto a shared _credentialed_get (the v1.0.8 bounded-redirect cookie-scoping core), now reused by the institutional landing-HTML fetch. browser-cookie3 added to the browser-auth extra.

[1.0.9] - 2026-06-13

CI-hardening release — no shipped-behavior change. Converts the test suite's remember-every-entry-point offline-stub denylist into a structural fence.

Added

• Structural network fence in the test suite (pytest-socket). An autouse fixture in tests/conftest.py restricts every test to LOOPBACK (127.0.0.1 / ::1 / localhost) + unix sockets by default; any connect to an external host fails loudly with SocketConnectBlockedError. The moment a new leak to a real API (arXiv, Crossref, Unpaywall, zotero.org, NotebookLM) is introduced — or an existing offline stub regresses — CI fails immediately instead of flaking on a network blip (the green-by-luck class behind several past CI saves). Tests that genuinely need a live external service opt in via @pytest.mark.network (plus the now-registered real_zotero / real_authenticity markers). The URL-quality active probe is also stubbed offline by default so the bundle tests make zero external connects. pytest-socket added to the dev extra; tests/test_v109_network_fence.py guards the fence (external connect blocked, loopback bind allowed).

[1.0.8] - 2026-06-13

Credential-safety release. Hardens the two highest-consequence trust paths the maturity audit flagged: the institutional-access credential path and the secret-at-rest key.

Security

• EZproxy session cookies are now scoped to the proxy host and never forwarded across an off-proxy redirect. The credentialed PDF fetch loaded cookies into a domain-less dict and followed redirects with follow_redirects=True, so a proxied publisher page that 302'd to any other host received the university SSO session — the hardest-to-renew credential in the system, leaked silently on the default path. The fetch now follows redirects manually (bounded), sending cookies ONLY to hosts within the proxy suffix and dropping them on every off-proxy hop.
• All credentialed / redirect-followed fetches are http(s)-only. A shared security.is_safe_fetch_url guard rejects file:// / ftp:// / data: / hostless URLs on the initial request AND every redirect hop, on both the httpx (Zotero PDF) and urllib (NotebookLM PDF) paths — closing the SSRF / local-file-read primitive a poisoned Unpaywall / OA / publisher record opened via urlopen.
• The secret-box encryption key is now OS-ACL-hardened. .secret_box.key was written with a bare os.chmod(0o600) — a no-op on Windows that left the key inheriting the parent directory ACL. It now routes through the real chmod_sensitive user-only Windows-ACL path (loud warning on failure), so the key gets the same at-rest protection as the secrets it guards.

Added

• research_hub.security.is_safe_fetch_url + host_in_suffix (shared URL and cookie-scope guards), a bounded manual-redirect loop in the Zotero PDF download path, and a _SafeRedirectHandler for the urllib fetch path. New tests/test_v108_credential_safety.py covers cookie non-forwarding, the scheme guard on both fetch paths, the redirect cap, and the key hardening.

[1.0.7] - 2026-06-12

Fixed

• Curation moves are now deep, retried, and fully index-synced (data-integrity keystone). operations.move_paper — the primitive under clusters merge / clusters split — was a shallow Path.replace with no retry and no index/link sync. On a transient Windows AV / Search-indexer lock it could strand a paper; worse, every move silently drifted the persisted dedup index (observed 126 stale entries vs 102 alive in the maintainer vault), left stale [[wikilink]] footers in the old cluster's siblings, and left the moved note's ## Hub block
  • topic: body tag pointing at the old cluster. It now: moves via the retry/backoff fsops.robust_move; invalidates the old path and registers the new one in dedup_index.json; rewires the source/target "Related Papers in This Cluster" footers; retargets the Hub block + topic: tag (backfilling topic_cluster frontmatter when a note was missing it); and returns a richer result dict carrying a sync_warnings list so degraded (non-fatal) syncs are observable instead of silent.
• merge / split are atomic. Each note move is verified (_verify_note_in_cluster); if any move fails the moves roll back and the source cluster stays intact (status unchanged) — never a half-merge or a stranded paper. A split whose moves fail also drops the just-created (empty) target cluster.
• Merge leaves a tombstone instead of dropping the cluster — fixes the duplicate-cluster bug. A merged source is kept as status: merged + merged_into: <target> and hidden from the active cluster list; match_by_query / resolve_merged redirect a re-ingest on the merged-away seed query to the target (cycle-guarded) instead of RE-CREATING the cluster and duplicating its papers. Orphaned hub/<source>/ dirs and source-only sub-MOCs are GC'd on merge (a shared sub-MOC referenced by any other active cluster always survives).

Added

• research_hub.fsops.robust_move — shared Windows-lock-safe move primitive (retry/backoff on transient PermissionError), promoted out of cluster_rebind (which now re-imports it). New Cluster.status / Cluster.merged_into fields; ClusterRegistry.resolve_merged and list(include_merged=False). New regression tests cover atomic-merge rollback, the dedup-index sync, the tombstone redirect, and a note missing topic_cluster frontmatter (no spurious rollback).

[1.0.6] - 2026-06-03

Fixed

• clusters rebind --auto-create-new no longer strands papers on a transient filesystem lock (Windows). The case-only folder rename (Foo → foo) did a two-step shutil.move (source → *.__rebind_tmp__ → target) and swallowed any exception into result.errors. A transient Windows AV / Search-indexer sharing-violation (WinError 5/32) on the second leg left the papers stranded in a *.__rebind_tmp__ folder with no cluster binding, surfacing as moved=0. Moves now retry transient PermissionError locks with backoff (_robust_move); a failed second leg rolls the temp folder back to its original name so papers are never stranded; and a partial cross-filesystem move is now reported instead of silently recording a short moved count. This was the root cause of the rare release-gate flake in test_v039_rebind_autocreate — it was not config-singleton pollution (disproven) or move_paper (not in this path).

[1.0.5] - 2026-06-02

Added

• ezproxy login --from-browser <browser> — import EZproxy session cookies from an already-logged-in real browser via rookiepy (no Playwright popup, no terminal interaction), mirroring notebooklm login --from-browser. Reuses the session you already have from browsing through your library proxy. Requires the browser-auth extra; rookiepy has no Python 3.14 wheel yet, so it degrades with a clear message (use the interactive ezproxy login on 3.14). New ezproxy.capture_cookies_from_browser.
• Auto-detected ezproxy_host_suffix — resolve_config now derives the hostname-rewrite suffix from captured cookie domains when it is not explicitly configured, so a one-time login enables hostname-rewrite without a manual config set ezproxy_host_suffix. New ezproxy.detect_host_suffix; an explicit config value still takes precedence.

Fixed

• ezproxy login now saves cookies reliably. It previously called storage_state() only AFTER the close-detection loop, but closing the window tears down the Playwright context first, so the save raised and nothing was persisted (0 cookies). Login now saves storage_state every second WHILE the window is open (plus a final save on close), so the session is captured even on an abrupt window close.

[1.0.4] - 2026-06-02

Added

• Hostname-rewriting EZproxy mode (ezproxy_host_suffix). Most modern institutions (e.g. Lehigh) run EZproxy in hostname-rewrite mode (www.nature.com → www-nature-com.ezproxy.lib.lehigh.edu) rather than the legacy /login?qurl= starting-point form — the latter returns an EZproxyCheckBack JavaScript interstitial that research-hub's non-browser httpx fetch cannot follow, so paywalled PDFs were never retrieved. Set research-hub config set ezproxy_host_suffix "<your-ezproxy-host>"; the publisher host is rewritten in place and paper attach-pdfs / auto fetch the proxied PDF directly. Hostname rewriting takes priority over ezproxy_url_template when both are set (the template remains a legacy fallback). New ezproxy.wrap_url_hostname; wrap_url gains a host_suffix arg; ezproxy status shows the suffix; RESEARCH_HUB_EZPROXY_HOST_SUFFIX env override. See docs/ezproxy.md.

[1.0.3] - 2026-06-02

Fixed

• auto now attaches Unpaywall open-access PDFs (previously silently skipped). auto's PDF-attach step (_run_pdf_attach_step) called plan_attach_for_items(items) without the unpaywall_email kwarg, so the Unpaywall lookup was skipped (the "Skipping Unpaywall" hint printed) and auto attached 0 OA PDFs even when unpaywall_email was configured. The standalone paper attach-pdfs (cli_paper) and the pipeline.run_pipeline path (oa_email) already passed it — only auto was missing it. Now threads cfg.unpaywall_email through, matching the other two call sites. Regression: tests/test_v087_pdf_attach_reporting.py::test_auto_pdf_attach_forwards_unpaywall_email.

[1.0.2] - 2026-06-01

Added

• Dashboard surfaces fit-check quarantined papers (FUNC-1 dashboard half). The Diagnostics tab gains a "Quarantined (N)" card listing papers the authenticity gate rejected (slug / cluster / layer / reason / date) -- completing FUNC-1 alongside the existing MCP list_quarantine and REST GET /clusters/{slug}/quarantine. Additive: new QuarantineRecord + DashboardData.quarantined, collected best-effort so a missing quarantine store never breaks dashboard render.
• mcp-tools.md coverage drift gate. docs/mcp-tools.md had drifted to 56 documented vs 79 live MCP tools; tests/test_mcp_tools_doc_coverage.py now enforces coverage (a new tool must be documented or consciously allowlisted; the gap allowlist may only shrink). The README already surfaces the live count via describe --filter mcp_tools.

Changed

• User-Agent centralized in research_hub._useragent (ARCH-1). The ~22 hardcoded research-hub/<version> UA literals across the search backends and verify / operations / importer / notebooklm now build from research_hub.__version__ via user_agent(). This fixes 3 STALE pins (0.4.1, 0.9.0, 0.43) that had drifted off the real version, and a version-scan gate test (test_useragent_version_sync.py) now fails CI if any hardcoded UA literal is reintroduced.
• CHANGELOG archived. Split the 8499-line CHANGELOG.md at the v1.0.0 boundary -- current + recent releases stay here (~1.4k lines); the v1.0.0 pre-release notes and all v0.95.x / earlier moved to CHANGELOG-archive/CHANGELOG-pre-1.0.md (linked at the bottom).

Fixed

• REST _optional_bool coerces string booleans. POST /api/v1/auto (and other bool flags) now accept case-insensitive "true"/"false" strings (REST clients commonly send string bools); other non-bool values still return 400.
• auto --force genuinely overwrites. force=true was previously a no-op beyond the FUNC-2 guard (indistinguishable from --append); it now clears the cluster's raw/<slug>/*.md notes before re-ingest (scoped, non-recursive, never on dry-run; Zotero / NotebookLM are not pruned). --force --dry-run advertises the clear step in the plan so the destructive action is previewable.
• zotero.fetch.get_notes no longer swallows all exceptions. Narrowed a bare except: to except Exception + a warning so KeyboardInterrupt / SystemExit propagate and real fetch errors surface (best-effort [] fallback retained).
• Added functional-test coverage for the restore_quarantine MCP tool.

[1.0.1] - 2026-06-01

Added

• tests/test_skill_structure.py — structural + trust invariants for the 11 SKILL.md files, closing the gap an ai-research-skills quality audit (2026-05-28) flagged: frontmatter shape and the installer mirror were already tested, but three guarantees were defended only by convention. Now asserted on every PR: (1) the ≤500-line progressive-disclosure ceiling (max today is research-hub at 244); (2) anti-fabrication safety strings survive every reword/sync (research-hub "Do not invent", paper-memory-builder status: gap, paper-summarize "hallucinated", notebooklm-brief-verifier "do NOT assume coverage"); (3) the sibling disambiguation arrows that keep auto-trigger from mis-firing (the paper-summarize / paper-memory-builder / literature-triage-matrix trio name each other; zotero-library-curator names zotero-skills; gap-to-topic names literature-triage-matrix
  • research-design-helper). Test-only — no skill content changed; all 21 parametrized cases pass against the current files.

Changed

• cli.py god-file split into cli_* domain modules (ARCH-2). The 8439-line src/research_hub/cli.py was broken into focused sibling modules (cli_common, cli_citations, cli_search, cli_notebooklm, cli_zotero, cli_summarize, cli_clusters, cli_vault, cli_pipeline, cli_paper, cli_maintenance); cli.py is now 3654 lines (−57%) and retains only the orchestrator core (build_parser, _main_dispatch, main, _sync_cli_dependencies). Pure move — byte-faithful handler bodies, no behaviour change; the CLI / MCP / REST surface is unchanged and the test files were not modified. Each domain module re-exports through cli.py and a _sync_cli_dependencies() forwarder keeps test monkeypatches on research_hub.cli.* reaching the relocated handlers. Also deduped the triplicated _load_zotero_if_configured into a single canonical cli_common definition and removed a dead, uncalled _read_doi_from_frontmatter copy.

• MCP tool docstrings rewritten for 12 worst-scoring tools (Glama TDQS lift) (mcp_server.py). Replaced terse one-line docstrings on the 12 tools that Glama's Tool Definition Quality Score flagged as weakest (cluster_rebind, ask_cluster, discover_continue, prune_cluster, discover_variants, examples_show, autofill_emit, autofill_apply, fit_check_audit, fit_check_drift, apply_fit_check_to_labels, propose_cluster_rebind). New docstrings hit all six TDQS dimensions — Behavior (verb-first opening), Conciseness, Completeness (full Returns dict-key list), Parameters (per-arg semantics + defaults + valid values), Purpose (relationship to neighbouring tools, deprecated-alias cross-refs), Usage Guidelines (When to use / When NOT to use bullets). The four canonical tools that LLM clients hit most (cluster_rebind, ask_cluster, discover_continue, prune_cluster) get the full multi-line template; the other eight get a compact 1-line-per-section variant. Zero behavioural change — function bodies / signatures / decorators byte-identical, only docstring text differs. Full pytest suite green (3101 passed). Expected TDQS impact: individual tool scores 1.8-2.4/5 lift toward 3-4/5; overall quality grade C should move to B-ish after Glama re-introspects on the next release.

• MCP deprecated aliases hidden from default surface (env-gated) (mcp_server.py, tests/_mcp_helpers.py). The 10 deprecated MCP tool aliases that were already scheduled for v2.0.0 removal are now hidden from the default @mcp.tool() registration. Glama's Tool Definition Quality Score (TDQS) penalises high tool count + poor disambiguation; this drops the LLM-facing surface from 86 → 76 tools without breaking the v1.x SemVer promise (the functions remain importable as Python module attributes; only the FastMCP registration is gated). Hidden by default — set RESEARCH_HUB_MCP_INCLUDE_DEPRECATED=1 to re-expose them during your migration window. Affected aliases (use the canonical replacement going forward):

  propose_cluster_rebind  → cluster_rebind(action='propose')
  apply_cluster_rebind    → cluster_rebind(action='apply')
  list_orphan_papers      → cluster_rebind(action='list_orphans')
  summarize_rebind_status → cluster_rebind(action='status')
  list_entities           → read_cluster_memory(kind='entities')
  list_claims             → read_cluster_memory(kind='claims')
  list_methods            → read_cluster_memory(kind='methods')
  read_briefing           → ask_cluster(source='notebooklm', mode='briefing')
  ask_cluster_notebooklm  → ask_cluster(source='notebooklm')
  brief_cluster           → ask_cluster(source='notebooklm', mode='brief')
  

  Test helper _get_mcp_tool gains an optional module= kwarg for callers that want to fall back to a Python attribute when an MCP registration is gated (back-compat: default None preserves the existing strict registry lookup).

• README.zh-TW.md full mirror of the new EN structure. Previously the zh-TW README was 152 lines / 6 sections (Real Screenshots, Why this exists, Start Here, First-Run Checklist, Connect your AI host, License only) — 10 sections behind the EN master + in the pre-restructure order. Now 430 lines / 18 sections matching the EN's 6-layer importance order: ## 快速開始 (Quick start) + ## 目錄 (Contents TOC) + the 16 content sections in the new order. 10 new sections translated from EN via gemini-delegate; existing 6 sections kept where the prose was operator-approved quality. All skill names, file paths, CLI commands, env var names, brand names, version refs, MCP terms, and code blocks preserved in English per locale convention.

• README.md reordered into 6 importance layers, with a new ## Quick start section and a 17-row ## Contents TOC inserted at the top. Previously install (## Start Here) lived at line 121 behind four selling sections (Real Screenshots / Why this exists / What it does / Is this for me?). A first-time visitor arriving from awesome-mcp-servers / SkillHub / MCP registry / promo posts had to scroll past ~100 lines of pitch before reaching a command they could run. New order: Quick start (line 26) → TOC → Real Screenshots → Is this for me? → full install (Start Here, First-Run Checklist, Credential Reference, Connect your AI host — moved up from line 243 to sit alongside install) → design rationale (Why this exists, What it does, Operator Modes — moved down from lines 51-73 into the design layer) → reference (Dashboard tour, Inside Zotero, Feature matrix) → troubleshooting + meta. No content removed; pure reorder + 2 new navigational sections (Quick start ~10 lines, Contents ~20 lines).

Fixed

• Pipeline ingest tests no longer make live Crossref / DOI-resolve network calls (CI flake fix). run_pipeline(dry_run=False) runs the authenticity gate, which corroborates DOIs over HTTP (_resolve_head_with_retry + CrossrefBackend). In the test suite that leaked real network calls which, on a CI network blip, hung until the 30s pytest-timeout killed them (the 2026-06-01 master CI failure on test_v041_pipeline_ingest_fixes / test_v062_note_enrich, while local + PR CI were green). A conftest.py autouse fixture now runs the gate OFFLINE for pipeline tests — stubbing only the two network entry points so the gate's real logic (local rejections such as L0 no_identifier) is preserved while DOI / Crossref checks take the designed fail-open path. Tests that drive the gate's network themselves are auto-excluded via source inspection of the gate internals. Suite stays 3144-green and deterministic.
• Subprocess reads now degrade gracefully on a stray byte instead of crashing the reader thread. Six diagnostic/system-tool subprocess calls used text=True with no errors= handler: doctor.py (the wmic / ps chrome-process lookup), security/__init__.py (the icacls ACL reset / grant / verify calls), dashboard/executor.py (the command runner's subprocess.run + Popen paths), and defuddle_extract.py (had encoding="utf-8" but no errors=). Under Python's UTF-8 mode (PYTHONUTF8=1), reading a Windows tool that emits cp950/Big5 raised UnicodeDecodeError: ... byte 0xa4 and killed the subprocess reader thread (10 tracebacks polluted doctor output; the affected lookup silently returned nothing). Added errors="replace" to all six — the codec is unchanged (these tools legitimately emit locale-encoded output, so forcing utf-8 would mangle them), only a bad byte now becomes a replacement char rather than a crash. The user-facing LLM-output reads (llm_cli.py, paper_summarize.py) were already protected with encoding="utf-8", errors="replace". Verified: doctor under PYTHONUTF8=1 goes from 10 UnicodeDecodeErrors to 0; the executor / doctor / security / defuddle suites stay green (141 passed). Surfaced during a live literature-search dogfood of the ai-research-skills audit follow-up.

[1.0.0] - 2026-05-26

First stable release. 129 commits since v0.91.1 (95 features + fixes; bumped from 0.91.x trail through 0.95.0 rc cycle to 1.0.0). Authenticity gate (v0.95+), MCP Server Registry preparation, and PDF attach reliability hardening are the headline themes.

Added

• mcp-name: io.github.WenyuChiou/research-hub line in README.md (HTML comment at the top, invisible in rendered Markdown). Required for ownership validation when publishing this package to the MCP Server Registry at registry.modelcontextprotocol.io. Without this line in the PyPI-rendered README, mcp-publisher publish fails with PyPI package 'research-hub-pipeline' ownership validation failed. The companion server.json manifest already declares the namespace io.github.WenyuChiou/research-hub; this line is the cross-verifier. No version bump or release flow is triggered by this change — the next PyPI release (currently staged as v1.0.0) will be the first PyPI artifact carrying the mcp-name line, at which point mcp-publisher publish can succeed.

Fixed

• PDF attach 0% regression from PR #108's requests → httpx port (zotero/pdf_attach.py). The requests → httpx migration in PR #108 used httpx.get() without overriding the default User-Agent (python-httpx/<ver>). MDPI, Frontiers, Springer-pdfdirect, IEEE, and other publishers' bot filters block that UA and return 403. Production E2E coverage crashed: a flood cluster got 0/15 OA PDFs attached, the Human-Nature cluster got 0/30. New _PDF_DOWNLOAD_HEADERS constant sends a real Chrome-on-Windows UA + Accept: application/pdf,... so the publishers' filters let the download through. Live verification: _download_via_httpx_result(<Springer pdfdirect URL>) now returns status 200 + a 1.6 MB PDF (was 403 / 406 bytes of HTML). Regression test in tests/test_v080_pdf_imported_file.py pins both that headers are passed AND that the UA looks like a real browser.

Added

• Explicit sub-MOC override via cluster.moc_links (vault/hub_overview.py, pipeline.py, vault/hub_backlink_migrate.py, cli.py). When a cluster's moc_links field contains a name with a family prefix (LLM-Agents-* or Water-Resources-*), the auto-derived slug-based sub-MOC for THAT family is now suppressed — the user-provided name wins instead of being appended alongside. Use case: slug generative-ai-large-language-models-coupled auto-derives LLM-Agents-Coupled, but the topic is really Human-Nature Systems; setting moc_links: [LLM-Agents-HumanNature] in clusters.yaml now yields [LLM-Agents-HumanNature, LLM-Agents] instead of also tacking on LLM-Agents-Coupled. Each family is independent: an LLM-Agents-* override does NOT suppress the Water-Resources auto sub-MOC, and vice-versa. Names that do NOT match a family prefix (e.g. MyCustomMOC) still pass through additively, as before. The override propagates through ALL derive_moc_links call sites — the cluster 00_overview.md, MOC pages, _HOME.md rendering, per-paper ## Hub section at ingest time (P1 fix: previously pipeline._render_obsidian_note did NOT pass cluster.moc_links, so the overview/MOC honoured the override while every paper wikilinked to the slug-derived sub-MOC), and hub-backlink-migrate backfills. Regression test test_explicit_override_propagates_to_paper_note_hub_section pins the pipeline call-site fix.
• Two-level hub-and-spoke MOC graph (per-cluster sub-MOC) (vault/hub_overview.py). Every LLM/water cluster now links to BOTH a parent MOC (e.g. LLM-Agents) AND a per-cluster sub-MOC derived from the slug's distinctive tail (e.g. LLM-Agents-Flood, LLM-Agents-ConsumerBehavior, Water-Resources-DataPipeline). Without the sub-MOC, every LLM cluster collapsed onto a single LLM-Agents node in Obsidian graph view; with it, each cluster has a distinct sub-hub between the parent and the paper notes — the cross-cluster centre stays, but each topic gets its own visible centre too. Two parser fixes along the way: hyphenated slug tokens like large-language-models-consumer-behavior now correctly match "large language model" substring (haystack normalised -/_ → space), and generic possessive prefixes (my, our, new, old) are stripped so my-cluster produces Cluster not MyCluster. Existing clusters need research-hub vault rebuild-overviews to backfill sub-MOC links into legacy paper-note frontmatter; new ingests get them automatically.
• EZproxy support for paywalled PDF downloads (ezproxy.py, cli.py, zotero/pdf_attach.py, pyproject.toml). Opt-in via cfg.ezproxy_url_template. New research-hub ezproxy login captures institutional SSO cookies via Playwright; paper attach-pdfs then wraps publisher URLs through the proxy and falls back to the direct URL on any proxy failure. Generic per-institution template — any university with an EZproxy gateway can configure it. Closes the IEEE/Wiley/Elsevier-via-paywall gap noted in the README's Known limitations.
• auto --year RANGE flag (cli.py, auto.py). The standalone search --year filter is now also exposed on the auto subcommand so users who want recent literature can write auto "LLM × X" --year 2024- directly instead of running a separate search step. Same syntax as search --year: 2024-2025 (closed range), 2024- (from 2024), -2024 (up to 2024). Wired end-to-end through _auto() → auto_pipeline(..., year_from, year_to) → _run_search() → search_papers(year_from, year_to). Unset = no year filter, behaviour unchanged.

Changed

• LLM-judge fit-check uses a stricter rubric when the cluster topic is LLM-narrowed (fit_check.py). For a cluster topic that explicitly mentions LLMs (any of LLM / large language model / ChatGPT / GPT-N / generative AI / LLM agent / AI agent / agentic AI), the LLM-judge prompt switches from the default 0-5 rubric ("squarely about / on-topic adjacent angle / tangentially related…") to a stricter one anchored on how central the LLM is to the paper's contribution. Under the new rubric an ML/DL-without-LLM paper in the same parent domain scores 2 (instead of 4 under the old rubric), so the default --fit-check-threshold 4 (PR #104) now actually filters those out. This addresses the empirical finding that running auto "LLM × flood" was still letting through ~12/15 pure ML-flood papers — they were getting "on-topic adjacent angle (4)" from the generic rubric. Detection scans BOTH the cluster definition and the slug, so freshly-created clusters (where the definition is the slugified topic fallback) trigger the strict rubric correctly. Five new tests pin every LLM-narrowing token variant.

• NLM keepalive: real refresh via SDK public API + minute-cadence default (notebooklm/keepalive.py, cli.py). The old rotate_and_persist_session used the SDK's private _rotate_cookies poke which returned success without verifying the session was still alive — every "ok" in the logs could have been a silent no-op against a revoked session, which is why the documented "持久一年" never actually held. New refresh_and_persist_session calls the SDK's public fetch_tokens_with_domains which GETs the NotebookLM homepage and extracts (csrf_token, session_id) as a side-effect of cookie rotation: if tokens come back, the cookies are observably good. The function returns a structured RefreshResult with before_metadata / after_metadata / changed (freshness-cookie expiry diff) so an operator can see PSIDTS actually moved forward. The old function name is retained as a thin bool shim for back-compat callers. keepalive_once was restructured to call refresh first (no pre-health gate — the Codex review of the old design flagged that as the source of permanently-stale sessions: a transient health-check error skipped the only refresh attempt). Default cadence flipped from hourly to 15 minutes (--interval 21600 → 900 s, floor 3600 → 600 s; --interval-minutes 6 → 15 with /SC MINUTE schtasks) — PSIDTS expires every ~3-4 hours, so hourly left only ~3 retries per expiry window and routinely lost races on flaky networks. --interval-hours is kept as a deprecated alias multiplied to minutes at dispatch. Existing tests in tests/test_v0950_nlm_keepalive_and_browser_login.py rewrote classes A (refresh) and B (keepalive_once) for the new contract; classes C/D/E (CLI / from-browser) unchanged. Note: "持久一年" is achievable only when (a) the scheduled task runs every 15 min without interruption, (b) Google does not revoke the account on a security event, and (c) the long-lived SID/PSID cookies (~1 year nominal) have not naturally expired — keepalive does not defeat any of those.

• auto --fit-check-threshold default raised 3 → 4 (cli.py). The LLM-judge fit-check now defaults to "clearly related" instead of "tangentially related and above". Rationale: at threshold 3 a "Large language models for X" cluster was consistently letting through pure ML/DL papers about X (no LLM at all) — the judge correctly scored them as 3 ("flood + AI, sort of related") and they passed. At threshold 4 the same papers are filtered out (still scored 3, now below the bar), cleaning the cluster to the actually-LLM-specific subset. Empirically on an "LLM × flood forecasting" run: kept dropped from 22/30 → 15/30, and the truly-LLM proportion rose from ~18% → ~20% (the LLM-judge is still generous on ML-flood papers; a sharper prompt is a follow-up). Use --fit-check-threshold 3 to get the old lax behaviour. Python API auto_pipeline(..., fit_check_threshold=3) deliberately stays at 3 for back-compat with programmatic callers and tests — the strict default is a CLI-UX choice, not an API contract change.

• auto --with-summary is now ON by default (cli.py). After ingest, the per-paper Key Findings / Methodology / Relevance sections are filled via the detected LLM CLI (claude / codex / gemini) on every run; previously the --with-summary flag had to be opted in explicitly. The flag uses argparse.BooleanOptionalAction, so --no-with-summary is the new opt-out. --full-auto no longer needs to re-set --with-summary and was simplified accordingly, which incidentally lets --full-auto --no-with-summary respect the opt-out (silently overridden before). Python API auto_pipeline(..., with_summary=False) deliberately stays opt-in (mirrors the --with-pdfs PR #90 split) so programmatic callers don't fire 20+ LLM CLI invocations per run silently.

• auto --with-pdfs is now ON by default (cli.py, auto.py). The auto subcommand attaches open-access PDFs from arXiv/OpenAlex/Unpaywall/Crossref to the ingested Zotero items as part of every run; previously --with-pdfs had to be opted in explicitly. The flag uses argparse.BooleanOptionalAction, so --no-with-pdfs is the new opt-out for runs where you only want the NotebookLM bundle. Rationale: the bundle ladder already downloaded the PDFs for NLM but they never made it into Zotero, leaving the cluster/pdf_coverage doctor check stuck at 0%. The Python API auto_pipeline(..., with_pdfs=False) deliberately stays opt-in so programmatic callers (tests, library users) don't fire the PDF-attach network round-trips silently — the CLI hands in an explicit True from argparse instead. --full-auto no longer needs to re-set --with-pdfs and was simplified accordingly, which incidentally lets --full-auto --no-with-pdfs respect the explicit opt-out (it was silently overridden before). The ingest and run subcommands stay opt-in (--with-pdfs only) — they are lower-level entry points where an explicit flag still makes sense.

Fixed

• Cluster overview LLM auto-fill now fires after populate_overview runs too (cluster_overview.py). PR #91 fixed _CHINESE_TEMPLATE_MARKER to match the Chinese scaffold (一到兩句話...) but in the real auto flow, populate_overview (vault/hub_overview.py) runs BEFORE apply_overview and overwrites the TL;DR with the cluster's topic-string fallback ("LLM for flood forecasting..."). The Chinese marker then no longer matches and apply_overview silently classified the topic string as "hand-curated", skipping LLM enrichment. Net effect: every auto-fresh cluster still ended with an empty TL;DR / Core Question / Scope. The scaffold check is now _is_scaffold_tldr(text, cluster_query, cluster_slug) which also recognises:
  • the English fallback "No cluster summary available yet." (from _render_tldr)
  • exact match against the cluster's first_query (the topic-string fallback _overview_tldr writes when there's no NLM brief yet)
  • exact match (case-insensitive) against the slug humanised (the last-resort fallback in _overview_tldr) Two regression tests pin the new branches.

Added

• brief_to_docx.js ships inside skills/research-design-helper/scripts/ (plugin 0.3.13 → 0.3.14). Sister script to skills/gap-to-topic/scripts/dossier_to_docx.js — same Markdown → Word converter (heading styles, bullet lists via numbering reference, tables with dual-width DXA, en / zh-TW font auto-select, Markdown separator-row skip), but default stem is design_brief instead of topic_dossier. Makes the Stage 3a artifact shareable with advisors / committee members who prefer Word over Markdown.

  • Not part of the contracted Stage 3a output. design_brief.md remains the canonical artifact; the .docx is an optional convenience for human consumption only. Downstream skills (Stage 3b research-context-compressor) read the .md frontmatter + section 1 directly, not the .docx.
  • Verdict-colour regex inherited verbatim but does NOT fire on design brief content — no "Do not pursue" / "Not assessed" / "不予推進" / "未評估" strings appear in a design brief. Keeping the regex unchanged means a single fix to the dossier generator lands in both scripts via parallel PRs.
  • SKILL.md adds ## Generate .docx (optional, plugin v0.3.14+) section between ## Output and ## Token-saving behavior, mirroring the gap-to-topic §4.5 Generate .docx pattern.
  • scripts/README.md ships alongside, documenting prereq (npm install -g docx), invocation examples (en + zh-TW), and the design-brief-specific deviation (the regex no-op).
  • New test test_design_helper_has_brief_to_docx_script_and_skill_md_section asserts scripts/brief_to_docx.js exists, the SKILL.md has the new ## Generate .docx (optional heading, and the script's default stem is design_brief not topic_dossier.
  • Mirrored to src/research_hub/skills_data/.

• Stage 2 → 3a → 3b handoff wiring (skills/research-design-helper/, skills/research-context-compressor/, tests/fixtures/topic_dossier_sample.gaps.yml, tests/test_handoff_gap_to_topic_design_helper.py, plugin 0.3.11 → 0.3.12). Two broken wires fixed as one coherent user-facing capability:

  • research-design-helper reads .research/topic_dossier.gaps.yml as new Input #2. New Workflow §0 preamble auto-pre-fills segment 1 (RQ) from the chosen gaps[].statement and segment 5 (risks) from open_questions[] + the specific concern hinted by gaps[].feasibility. Segments 2–4 (mechanism / identifiability / validation) are never pre-filled — the dossier doesn't carry that material and pre-filling with non-content corrupts the Socratic dialog. When the upstream .gaps.yml is absent, the skill behaves exactly as before (no regression for standalone users). Candidate selection is verdict-aware: filter gaps[] to verdict in {conditional-go, go}, then auto-pick if one, ask-the-user if 2+, halt with "nothing to frame" if zero.
  • design_brief.md frontmatter carries Stage 2 provenance. Two new optional fields: source: topic_dossier.gaps.yml#<gap-id> (URI-fragment pointer to the chosen gap) and gap_verdict: (frozen snapshot of verdict + first 60 chars of verdict_reason). The brief becomes self-contained — a future reader sees which dossier candidate this design was framed for. Provenance-protection: a refresh that would change the source gap-id triggers a confirm prompt instead of silent overwrite.
  • research-context-compressor reads .research/design_brief.md as new Input #2 under "For any project". Reads the frontmatter (project, source, gap_verdict) + section 1 (Research question) only. This implements the long-claimed pipeline.md contract; the brief is the authority on the sharpened RQ and the manifest mirrors it. If frontmatter source is set, copies that gap-id into the manifest's provenance.from_gap field (forward-compat with downstream tools).
  • First cross-skill handoff integration test ships at tests/test_handoff_gap_to_topic_design_helper.py with a frozen fixture at tests/fixtures/topic_dossier_sample.gaps.yml (copied from the v0.3.10 dogfood example). Asserts: fixture parses + has v0.3.10+ top-level keys + recall.screen sub-block + per-gap downstream-consumer fields + open_questions[].text; downstream SKILL.md prose lists the handoff input + has the §0 preamble + preserves the absent-.gaps.yml fallback; design_brief_template frontmatter has source + gap_verdict; research-context-compressor mentions design_brief.md; the schema reference covers every key the fixture contains (drift detector). Subsequent stage-to-stage wires (3a → 3b proper, 6 → 7, …) should follow this test shape.
  • Mirrored to src/research_hub/skills_data/. Pure additive change — no removed inputs, no breaking behaviour, no required new fields. Users without a .gaps.yml see identical UX to v0.3.11.

• dossier_to_docx.js ships inside skills/gap-to-topic/scripts/ (plugin 0.3.9 → 0.3.10). topic_dossier.docx is now a first-class contracted deliverable of the gap-to-topic skill. The generator converts topic_dossier.md to a styled Word document using docx 9.x: heading styles (H1/H2 navy, H3 dark grey), bullet lists via numbering reference, tables with dual-width DXA column sizing, bilingual verdict colour coding (light red "Do not pursue" / "不予推進", light yellow conditional "Worth pursuing … only if" / "值得推進…條件", light green unconditional, light grey "Not assessed" / "未評估"), Markdown separator-row skip, and an optional TOC + page break after the first table (--no-toc to suppress). Font auto-selects: filename matching .zh|zh-|zh_|-tw|-cn → Microsoft JhengHei; else Arial. Prerequisite: npm install -g docx. SKILL.md updated with §4.5 "Generate .docx" workflow step; front-matter description and "What it produces" updated to list .docx as a first-class artifact. Mirrored to src/research_hub/skills_data/gap-to-topic/scripts/.

Fixed

• paper-memory-builder doc enhancements — F-cross2 figures.yml sentinel + F-cross3 evidence-artifact scanning (skills/paper-memory-builder/SKILL.md, skills/paper-memory-builder/references/yaml-schemas.md, tests/test_handoff_gap_to_topic_design_helper.py, plugin 0.3.15 → 0.3.16). Two small doc improvements surfaced by the Stage 7-8 dogfood (~/.claude/audits/dogfood_runs/2026-05-23-paper-memory-academic-writing-stage-7-8/VERIFICATION.md):

  • F-cross2 — figures.yml file: sentinel values. The schema requires figures[].file (required field), but real-world Word-based research workflows often have figures embedded directly inside the .docx manuscript with no separable source file. Adds a ### file: sentinel values (v0.3.16+) section to references/yaml-schemas.md documenting three sentinels: embedded-in-manuscript, embedded-in-supporting-information, embedded-in-presentation. Downstream consumers (academic-writing-skills figure-text checks, future figure-archive tooling) treat these as "present in manuscript, no separable artifact to verify independently." Documented limitation, not a permanent solution — future versions may add a pre-processing step that extracts embedded figures.
  • F-cross3 — SKILL.md "Scanning the paper repo for evidence artifacts" sub-section. The previous Inputs section only described figures + manifest files but real research repos contain many non-figure evidence artifacts (simulation CSVs, analysis scripts, drawio sources, reviewer-response artifacts). The new sub-section documents typical artifact types and shows an example of populating claims[].evidence_artifacts with artifact PATHs for non-figure evidence (e.g. a chi-squared test claim might cite an outputs/llm-abm_decision_log.csv + the analysis script that ran it, in addition to the manuscript anchor). Makes the audit trail end-to-end traceable.
  • Tests: added 2 new test cases to test_handoff_gap_to_topic_design_helper.py: (1) test_paper_memory_yaml_schemas_documents_file_sentinel_values asserts all 3 sentinels appear in yaml-schemas.md; (2) test_paper_memory_skill_md_documents_evidence_artifact_scanning asserts the SKILL.md Inputs section has the new sub-section AND mentions at least 3 non-figure artifact types (Simulation, Analysis scripts, Drawio).
  • Pure additive prose change. Mirrored to src/research_hub/skills_data/paper-memory-builder/.

• Codex review tightenings — multi-eligible fixture + design_brief placeholder marker (tests/fixtures/topic_dossier_multi_eligible_sample.gaps.yml, skills/research-design-helper/references/design_brief_template.md, skills/research-design-helper/SKILL.md, tests/test_handoff_gap_to_topic_design_helper.py, plugin 0.3.14 → 0.3.15). Two follow-ups from the independent Codex evaluation of the v0.3.12 + v0.3.13 + v0.3.14 deliverables (.ai/codex_eval_report.md, verdict ship-with-fixes):

  • C2 — multi-eligible fixture. The §0 "2+ candidates, ask the user" branch was documented in SKILL.md and verified by the branch-coverage prose test, but never exercised against a real fixture with multiple verdict ∈ {conditional-go, go} entries. New synthetic fixture tests/fixtures/topic_dossier_multi_eligible_sample.gaps.yml ships 3 gaps (G1 conditional-go, G2 conditional-go, G3 no-go) so the filter produces 2 eligibles. Two new tests: (a) test_multi_eligible_fixture_parses_and_has_2plus_go_eligible asserts the fixture's filter result keeps the 2+ branch exercised (guards against fixture decay), and (b) a parametrized test_fixture_parses_and_drives_correct_section_0_branch runs across both fixtures asserting each drives its expected §0 branch (auto-pre-fill vs ask-the-user). A future third fixture for the zero-eligible halt branch can be added with one tuple.
  • C4 — design_brief placeholder marker. The v0.3.12 dogfood filled segments 2–4 with test-fit placeholder content (not from real Socratic dialog) and used an ad-hoc _TEST-FIT-PLACEHOLDER_ inline tag. Codified as a structured frontmatter field: placeholder_segments: [] (list of segment numbers). Example: placeholder_segments: [2, 3, 4] means segments 2–4 are placeholders, downstream tools should refuse to gate real research on a brief with non-empty list. SKILL.md adds a "Placeholder marker (v0.3.15+)" paragraph in the Output section explaining when to use it. New test test_design_brief_template_has_placeholder_segments_field asserts the frontmatter accepts the field.
  • Pure additive: no removed fields, no changed enum tokens, no behaviour change on absent/empty placeholder list. Mirrored to src/research_hub/skills_data/.

• research-context-compressor Output spec shows provenance.from_gap; research-design-helper §0 forbids in-file pre-fill annotations (skills/research-context-compressor/SKILL.md, skills/research-design-helper/SKILL.md, tests/test_handoff_gap_to_topic_design_helper.py, plugin 0.3.12 → 0.3.13). Two minor SKILL.md prose tightenings surfaced by the v0.3.12 Stage 3a + 3b dogfood (~/.claude/audits/dogfood_runs/2026-05-22-research-design-helper-llm-abm-socio-hydrology/):

  • F1 — compressor Output example now shows provenance.from_gap. The field was registered in docs/research-workspace-manifest.md (PR #95) and mentioned in the compressor ## Inputs section (PR #96), but the ## Outputs you must produce section had no example — a prose-driven skill is at risk of agents missing the wire. v0.3.13 adds an explicit example block under Outputs showing the provenance: { from_gap: ... } shape and codifies the absent-state rule (omit the block when there's no upstream; do NOT write provenance: {} or null). Provenance-protection on refresh now also explicitly applies to manifest.provenance.from_gap (mirrors the v0.3.12 rule in research-design-helper).
  • F2 — §0 forbids _PRE-FILL_-style annotations in design_brief.md content. The v0.3.12 §0 wrote gaps[].statement into segment 1 but didn't say HOW to format it. The dogfood agent added a _PRE-FILL (review/sharpen): ..._ annotation that wouldn't get cleaned up after segment 1 dialog sharpened the RQ. v0.3.13 clarifies: write the statement verbatim, no annotation in the file; the chat-level message flags it as pre-fill. Means segment 1 dialog can simply overwrite the statement with the sharpened RQ — no cleanup step needed.
  • Test coverage: added 2 new test cases to test_handoff_gap_to_topic_design_helper.py (14 → 16 total) asserting (1) the compressor Outputs section mentions provenance AND from_gap, (2) the §0 segment-1 pre-fill rule explicitly forbids _PRE-FILL_ annotations in file content (negation regex, not just word-presence — guards against rewrites that drop the prohibition).
  • Pure additive prose change (no removed inputs, no schema changes, no behaviour change on absent upstream). Mirrored to src/research_hub/skills_data/.

• gap-to-topic schema reference + research-hub discoverability (skills/gap-to-topic/references/dossier-template.md, docs/ai-research-skills.md, plugin 0.3.10 → 0.3.11). Two contract-vs-reality fixes ahead of the Stage 2 → 3a handoff wiring:

  • Schema reference refresh. The topic_dossier.gaps.yml schema in references/dossier-template.md was stale relative to what gap-to-topic actually emits since v0.3.6 (the --screen fit-check integration) and v0.3.9 (the 7-section reflow): missing top-level run_type / recall (with full screen sub-block) / pipeline; missing per-gap open_confidence / dead_end_evidence / borderline_reason / verdict / verdict_reason. The schema is now byte-aligned with the real output and adds a top-level downstream_consumer: research-design-helper key as a forward-compat hook recording the contract reader.
  • Discoverability. docs/ai-research-skills.md had zero mentions of gap-to-topic despite the skill being shipped at v0.3.10. Added a new Stage 2.5 row to the Stages table, a "Deciding whether a research gap is worth pursuing" entry to the When-to-use table, and a full ### gap-to-topic (v0.3.11) section in All packaged skills documenting Reads / Writes / handoff to Stage 3a / trigger phrases.
  • Pure additive changes (no removed fields, no enum tokens changed). Mirror at src/research_hub/skills_data/gap-to-topic/ updated in parity.

• gap-to-topic dossier reflowed as a 7-section research-grade decision memo (skills/gap-to-topic/, plugin 0.3.8 → 0.3.9). The v0.3.8 dossier rendered as "Markdown converted to Word" — wide scorecard with one-letter-wide columns, no decision-level visual cue, too many tables in the body. references/dossier-template.md is reorganised: §1 Executive Decision Summary (metadata + verdict cards + key uncertainty) → §2 Candidate Definitions (prose, no roster table) → §3 Decision Scorecards (per-candidate small tables, replacing the wide combined one) → §4 Evidence Base → §5 Gate-by-Gate Assessment (each gate uses a fixed Score / Evidence / Interpretation / Risk / Action-needed skeleton) → §6 Risks and Upgrade/Kill Tests (named risks; operational kill-test artifacts; salvage path) → §7 Recommended Next Steps (formal research-memo prose) → Appendix A Search and Screening Protocol (reproducibility log) → Appendix B Deliverable File List. SKILL.md "What it produces" updated to match. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• Cluster overview LLM auto-fill now actually fires on fresh clusters (cluster_overview.py, vault/hub_overview.py, tests/test_v071_cluster_overview.py, tests/test_hub_overview.py). The _CHINESE_TEMPLATE_MARKER constant in cluster_overview.py and the _SCAFFOLD_MARKERS tuple in vault/hub_overview.py both stored mojibake from a historical cp950 ↔ UTF-8 round-trip, so they never matched the actual Chinese placeholder phrases topic.py writes for a new cluster (一到兩句話說清楚... for TL;DR, 用一句話寫下... for 核心問題). Net effect: every brand-new cluster was immediately classified as "hand-curated", apply_overview silently refused to call the LLM, and the sibling populate_overview path likewise refused to refresh those sections — leaving the 00_overview.md stuck at the empty scaffolding (TL;DR / 核心問題 / 範圍定義 / 領域地圖 / 必讀論文 / 時間線 / 開放問題 all blank). The markers are now the real opening phrases (一到兩句話, 用一句話寫下); the legacy mojibake markers stay so vaults whose existing 00_overview.md still contains the corrupted bytes are still recognised and refreshed cleanly. The test fixture _template_text was likewise mojibaked and now mirrors topic.py verbatim, plus two dedicated regression tests (one per code path) pin the marker ↔ template round-trip so this can't drift silently again.

• gap-to-topic dossier — section-by-section rework for researcher value (skills/gap-to-topic/, plugin 0.3.7 → 0.3.8). A section-by-section review with the maintainer, plus a Codex evaluation, reworked references/dossier-template.md:

  • Decision scorecard is now a 5-point Likert per gate — each gate is a claim rated 1–5 (strongly agree … strongly disagree); the cell is N/5 label — description; a topic is worth pursuing only if every gate rates 3 or higher.
  • Gate 2 gains "What it would contribute" (one explicit claim) and a "Minimum validation sketch"; a caution paper is summarised by the kind of risk it raises, not offloaded to the reader.
  • Gate 3 gains "Design feasibility", a "Scale outline", and an explicit proposal-vs-dissertation feasibility split — it is no longer only a data-availability test.
  • The decision is yours is framed consistently as a conditional recommendation; the upgrade/kill test is written as concrete artifacts; a failed candidate gets a salvage-path line.
  • Appendix A is recast from a tool log into a method note (search scope, inclusion criteria, automated-vs-judgement).
  • What's in this deliverable gains a file tree. SKILL.md "What it produces" updated. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• gap-to-topic dossier reworked to read as a plain summary report (skills/gap-to-topic/, plugin 0.3.6 → 0.3.7). A review found the dossier still did not read cleanly in Word: the candidate roster showed an id column of G1/G2 codes; the Decision scorecard used ✓ ✗ ~ glyphs and the jargon verdicts "No-go / Conditional go / Go"; there was no index of the bundle's files. references/dossier-template.md now: drops the id column (the G1/G2 ids stay only in .gaps.yml) and states "why it could be a gap" in plain words; drops the scorecard glyphs for plain words; replaces the verdict jargon with plain phrasing ("Do not pursue — as stated" / "Worth pursuing — only if its open conditions hold" / "Worth pursuing"); adds a "What's in this deliverable" index section; gives the Bottom line one plain paragraph per candidate. The .gaps.yml schema moved to a trailing "not emitted" reference block. SKILL.md "What it produces" updated. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• ingest collapses in-batch duplicate papers (pipeline.py). Two search backends can return the SAME paper under different DOIs (a journal DOI vs a repository/preprint DOI). DOI-keyed search-merge dedup keeps both, and dedup.check() cannot catch them either — dedup.add() runs only in the note-writing loop, so in-batch siblings stay invisible during Zotero-item creation. Result: two Zotero items for one paper and two notes colliding on an identical filename slug (one silently overwriting the other). A new pre-ingest in-batch dedup pass now collapses candidates sharing a normalized DOI or normalized title (mirrors DedupIndex.add()'s >15-char title guard), keeping the first occurrence and logging a dup-in-batch manifest entry for each collapsed sibling.

• gap-to-topic §1 now applies the fit-check relevance gate (skills/gap-to-topic/, plugin 0.3.5 → 0.3.6). search --screen (PR #84) wired the BM25 relevance gate into the search command; §1 now uses it. SKILL.md §1: step 1 runs search --adversarial --screen --json and reads the new {screening_summary, results} shape (each paper carries a relevance field); steps 2-3 build the matrix and .bib from the on-topic results only (kept: true), excluding screened-out off-topic noise; step 4 reports the retrieved-vs-on-topic counts in the recall headline. references/dossier-template.md Appendix A pipeline + recall-mechanics rows updated. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• ingest dedup re-scoped to research-hub's own literature (pipeline.py). auto for a new topic produced a near-empty cluster — a search returned 25 papers, fit-check kept all 25, yet only ~1 Obsidian note was created. Cause: the dedup index was dominated by source:"zotero" entries mirroring the user's ENTIRE Zotero library (mostly never managed by research-hub) plus stale source:"obsidian" entries pointing at deleted note files. The dedup block treated any paper merely present in the Zotero library as a duplicate and continue-d past Obsidian-note creation, dropping it from the new cluster. Dedup now skips a paper only when it duplicates research-hub's OWN literature: (1) the obsidian_hit branch requires the note file to actually exist on disk — a stale index entry no longer blocks ingest; (2) the zotero_hit branch no longer skips — it still reuses the existing Zotero item (moves it into the cluster collection, adds hub tags + note, creates no duplicate item) but the paper is now ingested into research-hub with its Obsidian note created and bound to the cluster. New manifest action ingest-reuse-zotero records the reuse decision.

• fit-check no-LLM relevance gate rewritten (fit_check.py, auto.py). The old no_llm_fit_check gate split the topic into independent unigrams and kept any paper matching >= 0.1 of them — i.e. 1 of 5 words. A generic hydrology paper trivially matched water/model/resources and passed while the discriminating phrase "large language model" was destroyed; the llm-water-resources cluster ended up 38/43 off-topic. The gate is rebuilt as a pure-Python BM25 scorer over 1–3-gram topic terms — phrases like "large language model" survive intact, matching is plural-tolerant, and IDF is self-calibrated on the candidate batch. A paper is rejected only when the batch's sorted BM25 scores show a blatant bimodal split — a gap whose upper cluster out-scores the lower by ≥ 5× (the cross-field contamination signature: pure-hydrology papers score ~1 vs ~8 for genuine LLM papers). A focused, all-relevant search spreads only ~2× and is kept whole; cold-start (batch < 5, no topic terms, or no clear gap) defers — keeps all, flags relevance_unverified. Recall-biased: the no-LLM tier only catches blatant contamination; fine-grained relevance screening is the LLM-judge tier's job. New public API: extract_topic_terms, bm25_scores, screen_relevance. Grounded in a research sweep of ASReview (TF-IDF + Naive Bayes) and BM25 screening practice.

• gap-to-topic dossier — evidence-strength tags + an upgrade/kill test (skills/gap-to-topic/, plugin 0.3.4 → 0.3.5). A Codex evaluation of a real dossier found two gaps: (1) Gate 1 said a gap was "densely populated" but the reader had to open literature_matrix.md to see the occupancy signal was partly conference abstracts / artifacts, not primary studies; (2) a "conditional go" listed open conditions but no threshold — the advisor still had to supply the kill/upgrade logic. references/dossier-template.md now: Gate 1 tags each cited work by evidence type (primary study / review / close analogue / caution paper / conference abstract / preprint / data artifact) and ends with a one-sentence evidence-mix summary; "The decision is yours" carries an explicit upgrade / kill test per conditional-go candidate. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• gap-to-topic dossier — added tables for scannability (skills/gap-to-topic/, plugin 0.3.3 → 0.3.4). Follow-up to the reader-first redesign: the verdicts were still spread across prose bullets in each gate section. references/dossier-template.md now puts a Decision scorecard table in the Bottom line (candidates × the 3 gates

  • verdict, with a ✓/✗/~/— cell convention), a roster table in The candidates, and turns Appendix A (how it was produced) and Appendix B (companion files) into tables. Gate 1–3 bodies keep prose — the verdicts now live once, in the scorecard, so the gate sections carry only the evidence (no duplication). SKILL.md "What it produces" updated. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• gap-to-topic dossier was organised tool-first and code-first, not reader-first (skills/gap-to-topic/, plugin 0.3.2 → 0.3.3). A review of a real dossier found it opened with a metadata table of pipeline / API rows, wove search --adversarial / literature-triage-matrix / plugin versions through the body, and labelled candidates with bare [G1] / [G2] codes a researcher cannot decode. references/dossier-template.md is redesigned reader-first: a plain-language Bottom line leads; each candidate gets a readable name (the G1 id demoted to a .gaps.yml tag); each gate states its verdict in plain words before the evidence; all tool / pipeline / recall mechanics move to Appendix A, companion-file notes to Appendix B. SKILL.md "What it produces" + the §0 step updated to match; .gaps.yml schema gains a name: field. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• gap-to-topic §1 named literature-triage-matrix as the default prior-art tool but no step produced its matrix (skills/gap-to-topic/, plugin 0.3.1 → 0.3.2). The SKILL.md "orchestrates" paragraph + Inputs

  • the §2 component input-contracts all referenced .research/literature_matrix.md as an available input, yet the §1 numbered procedure only ran search — so an agent following §1 literally never built the matrix (a "phantom input"; surfaced by a 2026-05-21 workflow audit). §1 now has an explicit step 2 that feeds the search --adversarial --json results to literature-triage-matrix to produce the matrix; .bib and recall are renumbered to steps 3-4. references/dossier-template.md Pipeline line updated. Mirrored to src/research_hub/skills_data/gap-to-topic/.

• gap-to-topic §1 .bib instruction was unworkable (skills/gap-to-topic/, plugin 0.3.0 → 0.3.1). SKILL.md §1 step 2 and references/dossier-template.md told the skill to emit the §1 reference list via cite --format bibtex, but cite resolves identifiers only against an already-ingested Zotero library — at topic-selection time the candidate papers are not ingested, so cite returns "Could not resolve identifier" for every §1 DOI. Surfaced by the 2026-05-21 dogfood run. The .bib is now built from the search --adversarial --json metadata, which the skill already has in hand at §1. Mirrored to src/research_hub/skills_data/gap-to-topic/; plugin version bumped so the marketplace cache invalidates.

• notebooklm login --auto-detect now works (notebooklm/auth.py). The pre-fix implementation shelled out to the upstream notebooklm login subprocess and polled the patchright Chromium profile's on-disk Cookies SQLite for a notebooklm.google.com row. Chromium buffers cookies in memory and flushes them to that store only on a lazy timer, so a freshly-signed-in session stayed invisible on disk for minutes — the poll loop always timed out without ever firing the save. --auto-detect now drives the Chromium browser directly via Playwright (same stealth flags the SDK uses) and polls the live page.url: the moment the page settles on the NotebookLM host — and holds there for 3 consecutive polls, so a mid-redirect flash never triggers a premature save — the session is captured straight from the live browser context via storage_state. No terminal ENTER, no subprocess, no disk-cookie race. Fail-closed: a timeout, a browser-launch error, or a driver-start failure all return non-zero without saving and never raise into the CLI. Removed the dead disk-polling helpers (_patchright_cookies_db, _has_notebooklm_cookie, _cookies_db_modified_since).

• probe_cleared_failed_no_abstract URL quality signal now triggers text fallback in the NLM bundle builder (notebooklm/bundle.py). Springer / Wiley paywall skeleton pages return HTTP 200 with no body; the URL quality probe previously rated this ok, so the URL was uploaded to NotebookLM which also received an empty shell. The fix treats this reason code the same as likely_error_page — falling back to the abstract as a copied-text source, or skipping if no abstract is available.

• Zotero reparent on cluster reuse (auto.py). Clusters created before the parent-collection feature was added lived at the Zotero library root. On subsequent auto runs the pipeline found them by name and returned early, never calling ensure_parent_collection. New helper _maybe_reparent_collection PATCHes the collection's parentCollection to the configured parent (default "research-hub") when it is currently top-level (parentCollection=False). Idempotent and best-effort.

• PDF auto-attach step in auto_pipeline (auto.py). When --with-pdfs is passed, the pipeline now also calls the OA-PDF attachment chain (OpenAlex / Unpaywall / arXiv) for every Zotero item in the cluster after ingest, attaching the PDF as an imported_file child item. Previously --with-pdfs only fed PDFs to the NLM bundle, not Zotero.

• Summarize pending hint in auto_pipeline next-steps output. After a run, if any papers in the cluster have summarize_status: pending, the terminal output now prints a [HINT] line with the exact command to run (paper summarize --pending --cluster <slug>).

Added

• search --screen — fit-check BM25 relevance gate on the search command (cli.py). The BM25 relevance gate (screen_relevance / bm25_scores, from PRs #79/#81/#83) was only applied by the auto ingest pipeline; the standalone search command returned unscreened, potentially off-topic results. search --screen now runs that same gate over the retrieved papers: each result is tagged with a relevance score + a keep / screened-out verdict, and a screening summary (N retrieved → M kept, K screened out) is printed to stderr. It is recall-preserving — no paper is dropped from the output, so a caller can still audit the full retrieved count. --json switches to a {"screening_summary": {...}, "results": [...]} object (each result carrying a relevance block) when --screen is set; without --screen the output is byte-identical to before (bare array). Composable with --adversarial / --max-variants / --rank-by (orthogonal to ranking — --screen only annotates). Reuses the existing gate; no new scorer.

• gap-to-topic skill — a topic-decision dossier. New 11th packaged skill. Choosing a thesis/proposal topic is a go/no-go decision, not a literature review; gap-to-topic produces a .research/topic_dossier.md that runs a candidate breakthrough point through a 3-gate AND test — ① is the gap open (adversarial-recall search + a verifiable .bib), ② is it a contribution (dead-end history + problem-solving/incremental typing), ③ is it feasible (data/resource accessibility) — and hands the final "is it worth doing" verdict back to the researcher. skills/gap-to-topic/ (SKILL.md + 3 references + evals) mirrored to skills_data/; .claude-plugin/plugin.json 10→11 skills, version 0.2.0 → 0.3.0. The research-design-helper and literature-triage-matrix descriptions gain a one-line boundary clause routing topic-selection prompts to gap-to-topic.

• NLM session pre-flight in auto_pipeline. Before attempting any NotebookLM browser work, auto_pipeline now calls check_session_health on the stored state.json. When the session is missing or expired the pipeline skips NLM gracefully, sets nlm_deferred=True, and prints a [HINT] pointing at research-hub notebooklm login — instead of running a full headless browser session that would crash with an opaque Playwright error. The _print_next_steps footer now distinguishes "session expired → login first" from "transient error → retry". Health-check errors are caught so existing behaviour is preserved when the auth module is unavailable.

• Adversarial-recall search — research-hub search --adversarial. A single query phrasing systematically misses papers that use other vocabulary; in topic-scoping search a missed paper makes a research gap look open when it is not. --adversarial expands the query into several phrasings (LLM-generated when an LLM CLI is on PATH, deterministic fallback otherwise), searches each, unions the results by dedup_key, and prints a recall-confidence verdict (high/medium/low, derived from query saturation) to stderr. New search/query_expansion.py; adversarial_search() + RecallReport added to search/fallback.py; tests/test_search_adversarial.py (18 cases).

• Summary quality improvements (4 changes):

  • link_updater.find_related_in_cluster capped at 10 results (was unbounded) — prevents mega-hub nodes in the Obsidian graph for large clusters.
  • remove_paper_links(slug, raw_dir, cluster) — new function scrubs a deleted paper's slug from every sibling note's Related Papers section; removes the entire section header when the deleted slug was the last entry.
  • remove_paper() now cascades the cleanup automatically and returns links_cleaned in its result dict. dry_run=True is respected.
  • paper_summarize RELEVANCE prompt updated to require a specific dimension (method / empirical context / finding) rather than accepting a generic "this paper is relevant to [cluster]" sentence.
  • crystal emit prompt now exposes a first_finding field per paper (first Key Findings bullet, callout-format aware) to give the LLM more signal per paper.

• docs/literature-review-deliverable.md — format specification. Defines the consolidated document the skill pipeline (search → literature-triage-matrix → research-design-helper) produces end to end: the fixed 9-section contract, the .bib + .gaps.yml companion-file schemas, the per-paper summarization contract (and how it relates to the paper-summarize skill), honesty rules, and the bilingual / Markdown+Word bundle convention. A fully synthetic worked example ships in the ai-research-skills catalog repo.

• CI skill-version-guard job (.github/workflows/ci.yml). Blocks a PR that changes skill content (skills/ or src/research_hub/skills_data/) without also bumping .claude-plugin/plugin.json version. The marketplace plugin cache directory is keyed on the version string, so an un-bumped skill-content change ships to master but never reaches user installs. A 2026-05-21 dogfood verification (finding V1b) caught exactly this — a SKILL.md change merged without a version bump and silently failed to propagate. This guard makes that class of bug a hard CI failure at PR time.

• --peer-reviewed flag on search and auto. Drops preprint backends (arXiv/bioRxiv/chemRxiv/medRxiv), excludes gray doc types (preprint/posted-content/report/book-chapter/paratext/dataset), and floors corroboration. Closes the gap where auto ran search with zero filtering and the one-shot pipeline could not express "peer-reviewed only" (gray literature silently entered the vault).

• doctor check nlm_auth_paths. Reports which NotebookLM re-authentication paths actually work on this machine (interactive vs --from-browser/rookiepy) and the exact command to run.

• paper-memory-builder anti-leakage rule + E4 triad. New JSON Schema at skills/paper-memory-builder/references/claims.schema.json enforces: a claim with empty / absent evidence_artifacts MUST have status: gap and a non-empty gap_reason. Two new files alongside: scripts/check_claims_schema.py (validator with JSON-pointer-style error paths) and tests/test_check_claims_schema.py (14 cases — 1 meta, 4 positive, 5 negative, 4 schema-shape guardrails). The binding contract is restated in skills/paper-memory-builder/SKILL.md §"Anti-leakage rule" so it survives the SKILL.md-only marketplace install sync. Status enum extended: draft | supported | rejected | gap. Closes the long-standing Phase 2 Task B1 + E4 backlog item from the WenyuChiou/ai-research-skills plan.

• tests/test_skills_data_parity.py (21 cases) — guards the skills/ ↔ src/research_hub/skills_data/ byte-parity invariant. research_hub.skill_installer copies skills from skills_data/, so editing a SKILL.md in skills/ only ships the change to public-repo readers, not to research-hub install users. This test catches the divergence at PR time. Exception list (SHADOW_ONLY_IN_SKILLS_TREE) documents intentional shadows; one entry today (zotero-skills vendored copy, scheduled for removal in Phase 2 Wave C).

• Backup-first callout in zotero-library-curator. SKILL.md §"Output discipline" + references/report-template.md now lead the "Suggested follow-ups" section with a Zotero-RDF backup reminder before any apply/CRUD handoff to zotero-skills or research-hub zotero ... --apply. Closes the gap surfaced by the ai-research-skills Task #27 dogfood walk: read-only audit + apply step are different skills, and the callout was only on the marketplace README — never echoed to the user at handoff time.

Changed

• research-workspace plugin version bumped 0.1.0 → 0.2.0 (.claude-plugin/plugin.json). The marketplace plugin cache is keyed on this version; a dogfood test on 2026-05-20 confirmed fresh claude plugin install users were still receiving the pre-Phase-7 cached 0.1.0 skill bundle — i.e. the paper-memory-builder anti-leakage rule (Phase 7 Wave A) and the zotero-skills shadow removal (Wave C) had shipped to master but not to user installs. Bumping the plugin version forces a fresh cache directory so those changes actually propagate. No skill behavior changed by the bump itself.
• paper-summarize and literature-triage-matrix descriptions gain an "extract claims" disambiguation clause. The same dogfood test surfaced a fragile auto-trigger boundary: a user saying "extract claims from these papers" could land at paper-memory-builder (own draft), paper-summarize (per-cited-paper), or literature-triage-matrix (cross-paper matrix). Both descriptions now state the disambiguation explicitly so the router picks correctly.

Removed

• Vendored skills/zotero-skills/ shadow (Phase 7 Wave C). The 308-line vendored copy of the standalone zotero-skills plugin has been deleted; the canonical 60-line skill at WenyuChiou/zotero-skills is now the single source of truth. Coupled site updates landing in the same commit: skill_installer.py drops its zotero-skills exclusion (the dir no longer exists to filter), tests/test_skills_data_parity.py empties SHADOW_ONLY_IN_SKILLS_TREE, tests/test_v068_3_version_sync.py and tests/test_v066_skill_schema.py drop the now-stale exclusion comments / dead constants, and docs/interop-test-v068-2.md notes the historical snapshot vs the post-Wave-C state. Callers that resolve Skill(skill="zotero-skills") by bare name will now route to the canonical standalone plugin once that marketplace install exists in their environment. Users who relied on the vendored copy without separately installing WenyuChiou/zotero-skills will get a skill-not-found error until they run git clone https://github.com/WenyuChiou/zotero-skills ~/.claude/skills/zotero-skills (or claude plugin install zotero-skills@ai-research-skills --scope user). Closes Phase 2 backlog Item #3 from the WenyuChiou/ai-research-skills plan.
• .claude-plugin/plugin.json description count corrected 9 → 10 skills (paper-summarize was added in v0.69 but never reflected in the auto-discovery plugin manifest). Picked up the Wave B reviewer P9 follow-up.

Fixed

• probe_cleared_failed_no_abstract now routes to abstract-text fallback. When a URL's summarize_status is failed_no_abstract and the HTTP probe returns 200, the bundle was classifying the entry as url_quality=ok and uploading the URL to NotebookLM. This silently failed because publisher paywall pages (e.g. Springer) return HTTP 200 with a skeleton page that NLM also cannot read — the same paywall the summarizer hit at ingest time. probe_cleared_failed_no_abstract now resolves to quality=likely_error_page so the bundle falls back to abstract text when available, matching the same path already taken for cloudflare_block, tf_cookie_wall, and other confirmed paywall signals.
• Ubuntu CI OOM (issue #61). ubuntu-latest runners have ~7 GB RAM; 2800+ tests with lazy-loaded modules OOM the runner. test job split into test (windows/macOS — 14 GB, full suite) and test-ubuntu (4 pytest-split shards, ~700 tests each). Coverage moved to windows-latest (per-shard coverage would undercount). pytest-split>=0.8 added to dev deps.
• User-agent strings updated to research-hub/1.0.0 across all 14 search backends (biorxiv, chemrxiv, cinii, crossref, dblp, eric, kci, nasa_ads, openalex, pubmed, repec, websearch + the 3 already-fixed in the prior OSS-readiness commit). Mailto format preserved for APIs that recommend it (openalex, crossref, pubmed).
• cli.py build_parser() catches PackageNotFoundError (not bare Exception) when looking up the installed package version.
• pipeline_repair.py provenance tag bumped from pipeline-repair-v0.12.0 to pipeline-repair-v1.0.0.

Changed

• notebooklm login --help rewritten to the three real paths. Previously advertised --cdp / --from-chrome-profile / --use-system-chrome / --timeout / --keep-open as working modes; they were silently no-ops (the underlying aliases del'd their arguments). Help now states only: interactive default, --import-from, --from-browser.
• --from-browser failure on Python ≥3.14 is now actionable. Detects the missing-rookiepy + no-prebuilt-wheel case and points to the two paths that work (interactive in a terminal / --import-from) instead of a generic pip install hint that cannot succeed there.
• 5 SKILL.md descriptions tightened for keyword overlap with natural-language trigger prompts (Phase 7 Wave A polish): research-hub adds "organize them" (matches the verified trigger prompt "find papers and organize them"); paper-memory-builder adds "extract claims, supporting evidence, and figure key numbers" (matches the catalog phrase); research-design-helper adds "is my research question sharp enough to be falsifiable?" (matches the verified Phase 5.3.b trigger example); zotero-library-curator adds "bloated or under-used" (matches the SKILL body's own audit dimension) plus an explicit RDF-backup reminder line. research-hub-multi-ai reframes itself as the research-domain router vs agent-collab-workspace:agent-task-splitter (generic multi-agent decomposition), documenting the artifact asymmetry (.coord/multi_ai_plan.md vs .coord/plan.yml) so the two skills no longer silently shadow each other on the same prompt — the routing overlap the ai-research-skills Task #27 trigger verification surfaced.

Fixed

• auto / ingest now refresh _HOME.md, MOC bodies, and cluster overviews on success. The populate_all_overviews cascade (which writes _HOME.md, populates (populated by sync) MOC bodies with their cluster lists, and refreshes every hub/<slug>/00_overview.md) existed but was wired into vault rebuild-overviews only — auto never called it. Result: every research-session ingest left the vault-level navigation silently stale (empirically reproduced post-PR-D 4-leg E2E: no _HOME.md on disk, MOCs frozen at "(populated by sync)"). Now the cascade fires automatically when ingest wrote >0 papers; failures are logged to stderr and swallowed (the ingest itself already succeeded, navigation drift is non-fatal).
• --auto-detect cookies path hotfix. PR-D's _patchright_cookies_db hardcoded the LEGACY Default/Cookies path; modern Chromium (80+, including patchright's bundled chromium-1208) stores cookies under Default/Network/Cookies. Result: auto-detect polled the wrong file, never fired, user logged in successfully in the browser but the save never triggered. The path resolver now prefers Default/Network/Cookies, falls back to legacy Default/Cookies only if modern is missing AND legacy exists; if neither has been written yet (browser starting), returns the modern path so the next poll finds it the moment chromium writes it.

Added

• notebooklm login --auto-detect — fully automatic zero-touch login. Replaces the half-automatic --wait-file flow (which still required a manual file-touch after browser sign-in). With --auto-detect, research-hub polls the patchright Chromium profile's Cookies SQLite read-only for a notebooklm.google.com host_key. When you sign in and land on the NotebookLM homepage, the cookie appears, the script feeds both \n (any pending input() ENTER) and y\n (any pending click.confirm "Save anyway?" fallback) to the SDK subprocess, and the session is saved automatically. No terminal, no wait-file touch, no click.confirm response. Fail-closed on --wait-timeout (default 300 s): nothing is saved on timeout. Mutually exclusive with --wait-file, --import-from, --from-browser.

Fixed

• Ingest skips a paper missing one or more required core fields instead of aborting the whole batch. A real-world auto run (LLM reservoir management search) hit a CrossRef return with empty authors: [] and the pipeline fail-fast-aborted the entire ingest with "INPUT VALIDATION FAILED" even though 2 other valid papers were ready to write. The previous missing_doi_only skip branch was inert (doi is not in any required-fields list, so _validate_paper_input never emitted "missing required field 'doi'") and is replaced by a working _only_missing_required_field_errors predicate covering every field in REQUIRED_FIELDS_CORE (title / authors / year). When every error for a paper is a "missing required field 'X'" error, that paper is skipped from the batch with a logged "SKIPPED invalid input" entry, and the remaining valid papers still write to Zotero / Obsidian. Strict dry-run behaviour is preserved (every validation issue surfaces) so the operator sees the full picture before a real run.
• L1-deferred-but-L2-corroborated papers no longer fail-close at L1-deferred. When the DOI resolver HEAD is transient-blocked (anti-bot 401/403/406/418/451/etc. -- classified as *_check_unavailable by F7 / PR #51), the gate used to quarantine immediately at L1-deferred even when L2 corroboration (now augmented by PR #53's CrossRef-by-DOI metadata verify) would confirm it. The gate now lets such papers fall through to L2 / L3 / fit-check; if all pass, the paper is accepted with provenance.doi_recheck_pending = True and provenance.doi_recheck_details = {reason, status_code, url, resolved_via} so a future tool can re-verify the DOI when the publisher's anti-bot wall lifts. Definitive L1 failure (HTTP 404/410) still fail-closes as doi_unresolved / L1 -- the fabrication guarantee is unchanged; L2 corroboration + L3 metadata integrity remain the fabrication gate. The L1-deferred quarantine bucket is structurally empty post-fix; DEFERRED_LAYER survives as a public constant used by docs / reporting.
• L2 corroboration augments single-source DOIs with direct CrossRef metadata verify. A paper found by only one search backend (e.g. source: 'openalex') was quarantined L2 / uncorroborated even when its DOI was independently confirmable via CrossRef. The gate now does a direct https://api.crossref.org/works/{doi} metadata fetch for such papers; if CrossRef returns a record whose title/year/ authors match per the existing _records_agree predicate (fuzz.token_set_ratio >= 85, |year_delta| <= 1, surname intersection >= 1), CrossRef is recorded as a verified backend and the paper passes L2. Strictly augmentative: CrossRef is an authoritative metadata source, this only adds an existing-paper evidence check; the L2 corroboration bar itself is unchanged. Result cached in crossref_verify_cache.json (schema 1.0); failures are fail-quiet and not cached.
• zotero gc / zotero mark-kept --all-orphans: a real cluster collection whose Zotero key drifted from its cluster.zotero_collection_key binding (e.g. a Zotero-truncated date-prefixed name like 20260518-machine-learning-flood-forecas) is no longer mis-flagged as an orphan candidate. New conservative, suppression-only name-normalization match (_normalize_collection_name + bidirectional prefix match, min 12-char guard). No rebind/merge/delete-logic change; strictly fewer GC candidates. PR-A's non-empty hard-skip remains as an independent safety net.
• arXiv hits no longer leak past --exclude-type preprint. The arXiv backend left doc_type empty, so the type filter silently missed every raw arXiv result (bioRxiv/chemRxiv already set it). Now doc_type="preprint".
• Self-heal commands are environment-correct. New recommended_cli_invocation() picks the research-hub console script when on PATH, else the python -m research_hub module form. Wired into the NLM preflight, RequiresAuthRefresh, keepalive task resolution, and the auto-pipeline NLM-skip hint — these previously hardcoded a console-script command that fails on source checkouts.
• notebooklm keepalive no longer reports false-green. It now re-probes session health after rotating cookies; a rotation that does not preserve a usable session (server-side re-auth) returns non-zero instead of exit 0.
• F7: doi.org anti-bot 418 no longer quarantines every valid paper. The authenticity gate's DOI resolver sent the default python-requests User-Agent, which doi.org/Cloudflare answer with HTTP 418 — read as doi_unresolved → every peer-reviewed paper fail-closed-quarantined (reproduced: accepted: 0; quarantined: 2, DOIs accessible: 0; the same DOI returns 200 with a real UA). Fix: send a real User-Agent; classify 408/418/425/429/5xx + network errors as transient (bounded retry + backoff, not cached as a permanent miss, surfaced as *_check_unavailable); genuine 404/410 stay fail-closed doi_unresolved so the anti-fabrication guarantee is unchanged. A 0/malformed status also fails closed.
• F7 completion: access-blocked resolver statuses defer. DOI HEAD statuses other than resolved 2xx/3xx or definitive 404/410 non-registration (for example 401/403/406/451 anti-bot walls) now route to doi_check_unavailable / L1-deferred instead of permanent doi_unresolved / L1 quarantine.
• Semantic Scholar API keys are validated before use. A non-latin-1 SEMANTIC_SCHOLAR_API_KEY now emits a clear warning and is ignored so the backend queries anonymously instead of crashing while requests encodes the x-api-key header.
• DOI resolve cache auto-migrates pre-PR-#51 poisoning. The F7 completion (anti-bot HEAD defers, PR #51) was silently neutralised on any DOI cached under the old non-transient classification -- the cache-hit short-circuit returned the stale doi_unresolved outcome before the new logic ran. DoiResolveCache.load now performs a one-shot schema 1.0 -> 1.1 migration that prunes reason="doi_unresolved" entries whose status_code is not in {404, 410} (genuine non-existence) and rewrites the cache at the new schema version. Entries with status 404/410, success entries, and any other reason are preserved. Idempotent; logged at WARNING when any entry is pruned.

Removed

• Dead notebooklm login flags: --cdp, --chrome-binary, --use-system-chrome, --from-chrome-profile, --chrome-profile-path, --chrome-profile-name, --keep-open, --timeout, and the unused login_interactive / login_interactive_cdp aliases they delegated to.

Fixed (PR-A)

• zotero gc no longer presents non-empty real collections as deletion candidates. Any Zotero collection whose key was not a current cluster binding was flagged orphan-from-vault — including stale non-empty date-prefixed duplicate collections holding real items. delete_candidates already hard-skipped non-empty collections (so no data could actually be lost), but the output falsely implied a data-loss risk, eroding trust in the tool. Now: non-empty orphans get a distinct orphan-with-items(N) reason, are listed under a separate "NON-EMPTY ORPHANS — review only; gc cannot delete these" section, are never offered in the interactive prompt, and are excluded from --yes. Empty/test junk GC is unchanged.

Fixed (PR-B)

• F6: auto no longer prints [OK] ingest N papers when 0 were written. When every candidate was quarantined by the fail-closed authenticity gate the raw dir was never created, the exists() guard was skipped, and the tentative len(papers) count survived — the pipeline reported a clean ingest of N papers while the vault got nothing. The count is now authoritative (0 when nothing written); an all-quarantined ingest is reported as a failed step with an actionable quarantine list hint, not [OK]. A 0-written / 0-quarantined result (e.g. empty search) keeps the lenient path with an honest N written, M quarantined (of K candidates) message.
• F8: notebooklm upload no longer exits 0 when 0 sources were transferred. (Root cause later revised — the common cause is the URL-quality skip, not the upstream API; see the "Fixed (F8 real fix)" entry above.) When NotebookLM's source API drifts under the pinned notebooklm-py (e.g. Sources data ... is not a list (NoneType)) the notebook is created but no sources land; the old code returned 0 because fail_count == 0. A non-dry-run upload that transfers, caches, and prunes nothing is now a non-zero error naming the likely cause. (notebooklm-py was already pinned <0.5.0; this drift is server-side, so honest reporting is the only available remedy.)

Fixed (PR-C — deep F7)

• Transient DOI-resolution failures defer instead of being fail-closed-quarantined as fraud. PR-B's minimal F7 fix stopped the cache poisoning but the authenticity gate still quarantined ANY not ok at L1 — including *_check_unavailable (doi.org / Crossref rate-limit or network blip after the bounded retry). A sustained rate-limit window therefore still rejected valid papers, requiring manual quarantine restore. Now the L1 branch splits by reason family: transient → distinct L1-deferred layer (reported as "deferred, retryable"; recovers on a later run / quarantine restore); permanent (*_unresolved, 404/410, no-identifier, predatory/metadata/fit/uncorroborated) → unchanged L1 quarantine. The paper is still held out of ingest in both cases (fail-closed — the anti-fabrication L0–L5 guarantee is unchanged; a fabricated DOI returns 404 → permanent → quarantined). auto now reports N quarantined, D deferred distinctly, and an all-deferred run says the papers were not rejected and how to retry, instead of "quarantined". (An optional quarantine retry-deferred convenience CLI is deferred to a later follow-up — existing quarantine restore already recovers deferred entries.)

Added (PR-D)

• notebooklm login --wait-file PATH — non-interactive login (no terminal/ENTER). The upstream login blocks on input("press ENTER"), which can't be driven headless/scripted. With --wait-file you sign in in the browser then create PATH (touch PATH, or an automation wrapper does it); research-hub polls for it and feeds the newline that triggers the upstream session save. --wait-timeout (default 300s) fails closed — on timeout the subprocess is terminated and nothing is saved. Stale signal files are cleared first so a leftover from a previous run can't auto-trigger before sign-in. This formalizes (in Python, testable) the signal-pipe technique used to recover the maintainer's expired session.

Fixed (F8 real fix)

• auto can now actually upload publisher-URL clusters to NotebookLM. Diagnosed 2026-05-19: an all-URL cluster (DOIs → ScienceDirect/Elsevier) uploaded 0 sources because every entry was likely_error_page (our local probe can't read the anti-bot wall) and the conservative URL-quality gate skipped them all — and auto called upload_cluster positionally with no way to override, so there was no path to rescue it through the pipeline. The earlier F8 message also mis-blamed a NotebookLM API change (the SourcesAPI.list ... NoneType warning was a red herring — listing an empty new notebook). The conservative skip is intentional design and is unchanged; instead --include-suspect-urls is now exposed on auto and threaded auto → auto_pipeline → upload_cluster, and the 0-sources error message now lists the real likely cause first (URL sources skipped → re-run with --include-suspect-urls) instead of pointing at the upstream API.

Added (F8 first-principles fix — content-priority ladder)

• NotebookLM bundles now upload the abstract as a text source when no PDF is obtainable, instead of a paywalled URL. First-principles diagnosis: NLM needs content, not a URL — a paywalled publisher DOI carries none. Local PDF (rung 1) and Unpaywall/OA PDF (rung 2, via fetch_paper_pdf) already worked; the gap was rung 3. bundle.py now: no PDF + (no URL or a likely_error_page paywall URL) + a real ## Abstract in the note → emits action="text" carrying the abstract (title/DOI-prefixed); a good (non-suspect) URL is still preferred (full text). NotebookLMClient.upload_text → sources.add_text; upload.py dispatches action="text"; BundleReport.text_count. Net: an all-paywall-URL cluster that previously uploaded 0 sources now uploads the abstracts — real content NotebookLM can synthesise. The conservative URL-quality gate and --include-suspect-urls override (band-aid) are unchanged.
• socket.getfqdn CI flake — autouse stub in tests/conftest.py (master had been red on this for ≥3 consecutive runs: 26185448887, 26191738564, 26192402510). Multiple test files construct ThreadingHTTPServer(("127.0.0.1", 0), ...) whose server_bind() calls socket.getfqdn("127.0.0.1"). On macOS GitHub Actions runners (and Bonjour/mDNS-equipped environments generally) the reverse-DNS lookup hangs 30+ seconds before pytest-timeout fires — Python stdlib issue14914, 14-year-old bug. The first attempt of this fix patched only the artifact_server fixture in test_artifact_delete_endpoint.py, but the next CI re-run surfaced the same flake at test_dashboard_executor_e2e.py::test_e2e_sse_event_after_action. grep -rln 'ThreadingHTTPServer' found 7 test files with the same construction pattern, so the fix promoted to a conftest.py-level autouse fixture (_stub_socket_getfqdn) that stubs socket.getfqdn for every test in tests/. Production code path is unchanged — monkeypatch scope is per-test, reverting at teardown.

Removed

• Legacy root plugin.json deleted (Phase 7 Wave B from the WenyuChiou/ai-research-skills marketplace-maturity tracker — see research-hub issue #60). The file was an early Cowork-plugin placeholder with a self-described TODO ("Convert to full Cowork plugin manifest when plugin schema is finalized") that never got migrated. It declared 3 skill paths, one of which (skills/knowledge-base/SKILL.md) was the long-removed knowledge-base alias, while 11 actual skill dirs ship under skills/. The active manifest is .claude-plugin/plugin.json, which uses Claude Code marketplace auto-discovery and is unaffected. grep -rn 'plugin\.json' against tests/ + src/ + non-meta docs returns zero references to the root file — it is a pure orphan. Removing it removes a misleading source of truth without touching any install / marketplace / test path.

---

Older releases (the v1.0.0 pre-release planning notes, v0.95.x and earlier) are archived in CHANGELOG-archive/CHANGELOG-pre-1.0.md.