architecture

Architecture

Overview

SaiLoR is a single-codebase React app that runs as both an Electron desktop application and a static web SPA. The key architectural seam is the PlatformAdapter interface, which abstracts all file I/O and PDF loading so the React UI is identical in both runtimes.

Platform Adapter Pattern

The entire file-system and PDF-loading layer is abstracted behind a single interface:

src/platform/adapter.ts  →  PlatformAdapter interface
src/platform/index.ts    →  getPlatform() singleton factory
src/platform/electron.ts →  ElectronAdapter
src/platform/browser.ts   →  BrowserAdapter

PlatformAdapter (src/platform/adapter.ts) defines these operations (the list below is a guided tour, not an exhaustive count, which would just go stale as the interface grows):

• getRecents() — return the list of recently opened projects (RecentEntry[] with id + name)
• openRecent(id) — re-open a project by its recent-entry id (path on Electron, IndexedDB handle key on browser)
• openProject() — show an open dialog/picker, return JSON text + a SaveHandle
• saveProject(text, handle) — write back to the handle's location
• rebasePdfPaths(pdfPaths, from, to) — re-express PDF paths that were relative to from's directory as relative to to's. "Save as" depends on this: a paper's pdf is stored relative to the project file, so writing the old paths to a new location left every PDF pointing at nothing. store.saveAs() therefore picks the destination first (pickProjectLocation), rebases, and only then serializes and writes. Electron does the real path math via a paths:rebase IPC (relative(dirname(to), resolve(dirname(from), rel))); the browser has no paths and returns the input unchanged.
• getPdfSource(pdfPath, projectHandle) — resolve a paper's relative PDF path into a URL react-pdf can load. On Electron it re-asserts the main process's project directory from projectHandle first, so PDFs always resolve against the project actually being rendered (the project editor repoints that directory when picking a location).

Three more exist for the project editor (see below):

• pickProjectLocation(suggestedName) — ask where the project JSON should live; writes nothing. Returns a ProjectLocation (handle, name, and — Electron only — an absolute path).
• pickPdfs() — pick PDFs to reference; returns PickedPdf[] (name, plus an absolute path on Electron).
• pickPdfFolder() — pick a folder and return every PDF inside it, recursively, as PickedPdf[]. Electron walks the filesystem via IPC (pdf:pickFolder); the browser reuses the same webkitdirectory <input> the PDF-viewer's one-time folder grant uses (see pickPdfFolderViaInput below), filtered to .pdf names.
• pickReferenceFile() — pick a single .bib/.ris/.json reference-manager export; returns { text, name } or null if cancelled. Parsed by src/model/references.ts.
• relativePdfPaths(pdfs, location) — the pdf values to store, relative to the JSON's directory. Electron computes real relative paths via IPC; the browser returns bare file names (the File System Access API exposes no paths).
• absolutePdfPaths(pdfPaths, from) — the inverse: absolute paths for values relative to from's directory. Added for startFromScreening/importFromScreening (see "Screening" below) — a paper carried in from a screening project needs a real sourcePath, not just the relative path the source file stored, or changeLocation cannot re-derive it later. The browser returns undefined per entry (no paths to compute from), leaving those rows exactly where an edited project's rows already sit.
• siblingProjectLocation(source, fileName) — the location fileName would have if it sat next to source's directory; writes and prompts nothing. What makes "save the new annotation project next to the screening JSON" the default rather than a dialog suggestion. Returns null in the browser (no paths), and callers fall back to pickProjectLocation.

One more is its own capability object rather than a flat method: getGit(): GitPlatform | null — git operations against the user's own git installation, or null where the runtime cannot reach one (the browser). See "Git" below.

Project title. A project JSON may set a top-level title; the app shows it wherever it would otherwise show the file name (toolbar, recents list, Open menu), falling back to the file name when absent. It is a first-class key on Project (not swallowed into extra, which would duplicate it on save) and is only written when non-empty. The project editor exposes it as a Project title field next to the JSON location.

Recents are re-read from disk, not trusted. The title shown for a recent is stored on the entry, but that copy goes stale the moment the project is renamed elsewhere — most obviously by changing the Project title in the editor, where the old name would otherwise still be sitting in the list after closing it. So checkRecents() re-reads each project's current title from the file: on Electron a project:peek IPC returns { exists, title } per path in one round trip (parse failures are contained per file); in the browser it reads through the retained handle, but only when read permission is already granted — startup must never throw a permission prompt at the user just to refresh a label, so without permission the stored title is kept. The refreshed list is written back (replaceRecents, order preserved) so the fresh titles survive a restart, and editorStore.save() triggers a refresh so a rename shows up as soon as the editor is closed. A title removed from a file correctly reverts the entry to showing its file name.

A recent whose file has gone is kept, not forgotten. checkRecents() (run on startup via refreshRecents()) re-tests each entry — fs:exists on Electron, "do we still hold the IndexedDB handle" in the browser — and sets a runtime-only available flag (never persisted). A missing project stays in the list, faded, badged not found, and unselectable, because the drive may come back; the user can still dismiss it with the ×. Opening one that has since vanished marks it unavailable rather than pruning it. Note the flag is deliberately tri-state: undefined means not checked yet, and only an explicit false greys an entry out, so nothing flickers on first paint.

Closing a project (requestCloseProject) returns to the start screen, and when there are unsaved changes it first asks via ClosePrompt — the same three choices and wording as Electron's native quit dialog (Save / Don't Save / Cancel). A failed save keeps the project open rather than closing and losing the work.

Recents entries carry path and title beyond id/name. The path is what lets a user tell two projects called review.json apart: the start screen shows it under the title (truncated from the left, since the tail is what differs — direction: rtl + unicode-bidi: plaintext, the latter so the leading / isn't reordered to the end), and the Open menu shows it as a hover tooltip. On the start screen each entry also carries a ✎ pen (useEditorStore().startEditRecent(id)) that opens the project straight into the schema editor instead of the annotation view — the same edit path as Edit annotation JSON…, but for a known recent rather than a file picker. Both places offer an × to drop an entry (forgetRecent), which in the menu deliberately does not close it, so several can be cleared in a row. The title is only known once the JSON is parsed, so loadFromText re-pushes the entry via rememberProject() — pushRecent dedupes by id, so this enriches the existing entry rather than duplicating it.

Recent projects are managed by src/platform/recents.ts — a platform-opaque module that stores up to 5 entries in localStorage (separate keys for Electron and browser). On Electron, the entry id is the absolute file path; on browser it is a key into the IndexedDB handle store.

getPlatform() (src/platform/index.ts) returns a singleton: ElectronAdapter if window.slr exists (preload bridge), otherwise BrowserAdapter. Detection uses isElectron() which checks for the preload-bridged window.slr object.

ElectronAdapter (src/platform/electron.ts)

Delegates to window.slr (the preload bridge). File operations use IPC to the main process. PDFs are served via the custom slr-file://project/<encoded-path> protocol — the main process resolves paths relative to the project directory. setProjectDir is called on open/save-as so the protocol knows the base directory. On open and save-as, the adapter pushes an entry to the recents list (slr.recents.electron localStorage key). openRecent(id) calls bridge().openPath(id) to read a file by absolute path; if the file no longer exists the entry is pruned from recents.

BrowserAdapter (src/platform/browser.ts)

Has three tiers of capability:

Capability	Chromium (FSAPI)	Other browsers	Server mode
Open	showOpenFilePicker — in-place handle retained	Hidden <input type=file>	fetch(url) via ?project=
Save	createWritable on retained handle	Download blob	Download blob
Save as	showSaveFilePicker	Download blob	Download blob
PDF loading	One-time folder grant via showDirectoryPicker, blob URL	One-time folder grant via a webkitdirectory <input>, blob URL	fetch relative to project URL

setServerBase(url) records the URL a project was fetched from so sibling PDFs resolve correctly in server mode. The adapter stores FileSystemFileHandle references in an internal map keyed by generated IDs.

Every locally opened project — FSAPI handle or <input> fallback alike — resolves its PDFs through a folder the reviewer grants access to once, never a fetch. A local project has no URL for its PDFs to live at, so getPdfSource branches purely on whether serverBase is set: unset means "resolve locally", full stop, regardless of how the JSON itself was opened. resolveLocalPdf calls the shared ensureLocalPdfGrant() (prompts only if neither pdfDir nor pdfFileMap is set yet — a no-op once one is), then dispatches to whichever grant it produced: resolveViaDir walks getDirectoryHandle/getFileHandle per path segment on the FSAPI directory handle; resolveViaFileMap looks the path up in the map ensureLocalPdfGrant built from a folder-picking <input>'s webkitdirectory attribute (non-standard name, but implemented everywhere that matters — Firefox, Safari, Chromium), which reads an entire folder tree in one go and keys each file by its webkitRelativePath with the picked folder's own name stripped (that name isn't part of the project-relative paths stored in the JSON). Either mechanism prompts once per session.

needsPdfFolderGrant() / grantPdfFolderAccess() are PlatformAdapter methods that expose that same grant explicitly, so a caller can ask for it from a real click instead of letting getPdfSource pop the native picker unannounced. PdfViewer.tsx checks needsPdfFolderGrant() before ever calling getPdfSource for the first PDF of a locally opened project; if true, it renders an in-app explanation with a Choose folder… button instead, and only calls getPdfSource after the button's click has driven grantPdfFolderAccess() (which just calls the same ensureLocalPdfGrant()) to completion. getPdfSource itself still auto-prompts via ensureLocalPdfGrant() if called directly without this check — a caller like aiStore.ts's run(), which reads a paper's PDF bytes for the AI-annotation flow and doesn't go through PdfViewer's gate, still gets a working (if unannounced) prompt rather than an error. ElectronAdapter implements both methods as false/no-op — there is no such prompt on the desktop build, PDFs are read straight off disk via slr-file://.

pickPdfFolderViaInput's "did the user cancel" detection uses the cancel event, not a focus-return guess — and this isn't a stylistic choice. A focus-based guess (the pattern the older pickFileViaInput/pickFilesViaInput still use, for a plain — non-directory — file pick) is unsafe here specifically: Firefox inserts its own "Upload N files from this folder?" confirmation after the OS folder dialog closes, and that OS dialog closing already returns window focus — well before the user has answered Firefox's prompt. A short focus-based timeout reads that in-between moment as a cancel and resolves empty while the real selection is still on its way, which is indistinguishable from a genuine cancel to the caller. The dedicated cancel event (Chrome 113+, Firefox 106+, Safari 16.4+) fires only when the picker was actually dismissed with nothing chosen, so it doesn't have this problem; a focus-based fallback still exists for an engine with neither event, but at a much longer delay (2s, up from the old 300ms) specifically to give a pending confirmation step like Firefox's room to resolve first.

This used to only run for an FSAPI handle — every other local-open path (Firefox/Safari, or Chromium without the FSAPI grant) fell straight through to a fetch against the app's own page URL, which obviously never contains the reviewer's PDFs. That fetch didn't even fail cleanly: on a dev server or any SPA-style static host it 200s with index.html instead of 404ing, so the failure surfaced three layers away as pdf.js's opaque InvalidPDFException: Invalid PDF structure — indistinguishable from an actually-corrupt file. The byte-level check described below catches that case when it's genuinely a server-mode misconfiguration; this local-resolution fix is what stops it from happening at all for the (very common) case of a reviewer just opening their own project file.

Every promise in src/platform/idb.ts settles on abort and blocked, not just complete/error. A transaction can abort with no request having failed — site data cleared, a devtools deleteDatabase, origin eviction under storage pressure — and that fires onabort alone, so a promise waiting only on oncomplete/onerror never settled and its finally never ran. That is not an abstract leak: openProject awaits rememberHandle inline, so the picker returned a file and the app simply never showed it — no error, no timeout, busy stuck on forever, and the caller's own try/catch powerless because there was no rejection to catch. tx.error is null until a transaction aborts, so the error paths carry an explicit fallback rather than rejecting with null, and an open that was reported blocked but later succeeds closes the connection it can no longer hand back.

Recents identify a file, not a file name. Two reviews both saved as review.json used to collapse into one recents entry and one IndexedDB key, so the second open overwrote the first's handle and the surviving entry opened the wrong project. Electron sidesteps this with the absolute path; the File System Access API exposes none, so rememberHandle mints an opaque id — after asking the existing entries, via the API's own isSameEntry, whether they already hold this file, so that reopening one reuses its entry instead of piling up duplicates. Entries written before this keep their name-based ids and are adopted rather than duplicated.

Write permission is requested at open time, not at save time. createWritable() needs readwrite, and a handle restored from IndexedDB is back at prompt after a reload — but requesting permission needs transient user activation, and by the time a save has serialized the project the click that started it may no longer count. openRecent therefore asks for readwrite instead of read: one prompt, at the moment the reviewer expects one, on the reasoning that opening a project in an editor implies intent to save it. writeFsApi keeps a check as a fallback. Worth knowing when reading this: the failure it guards against was reasoned from the API contract, then partly verified against a real Chromium File System Access implementation: queryPermission/requestPermission exist on handles, queryPermission with readwrite short-circuits when access is already granted (measured: requestPermission called zero times), a handle survives an IndexedDB structured-clone round trip and still answers isSameEntry, and createWritable round-trips content. Still unverified is whether a picker-derived handle restored after a page reload prompts on createWritable() by itself — that needs a native file dialog and a human permission decision no automation can drive. ensurePermission also returns true when a handle exposes no permission API at all (optional chaining yields undefined, which is not 'granted'), so a browser that implements handles without the permission methods can still save rather than being blocked outright.

When the File System Access API is available, the BrowserAdapter also persists handles in IndexedDB (src/platform/idb.ts) so they survive page reloads. openProject() and saveProject() (via the FSAPI Save As flow) call rememberHandle() which stores the handle under a recent:<id> key (opaque id, not file name — see above) and pushes an entry to slr.recents.browser. openRecent(id) retrieves the handle from IndexedDB, re-requests readwrite permission (via ensureWritePermission), and re-reads the file. If the handle is missing the entry is kept and marked unavailable; if permission is denied the entry is also kept (so the user can retry) and the error is thrown.

Opening any new project — local or server-mode — clears whatever the previous project's PDFs resolved through, via clearLocalPdfGrants() (drops pdfDir/pdfFileMap) alongside resetting serverBase, from all three project-load paths plus setServerBase() itself. Without this, switching from Project A (with a granted folder or a set server base) to Project B would silently keep resolving PDFs through A's leftover grant — at best a confusing "not found", at worst the wrong PDF's bytes for a path that happens to collide between the two projects.

getPdfSource's server-mode fetch branch validates the response bytes, not just the HTTP status. A dev server's SPA fallback, a static host's catch-all rewrite, or a reverse-proxy login page can all answer 200 with HTML for a PDF path that doesn't actually exist — res.ok alone can't tell that apart from a real PDF. Handing pdf.js those bytes produces its own genuine InvalidPDFException: Invalid PDF structure, which reads as a corrupt file when the real problem is upstream (wrong base, missing file, wrong deploy config). So before trusting a fetched body, hasPdfMagic() checks it actually starts with PDF's own magic number (%PDF-) — read from res.arrayBuffer(), not Content-Type (plenty of static hosts serve everything as application/octet-stream, so a header check alone would reject real PDFs) — and throws a specific "the server answered, but not with a PDF" error naming the URL it tried, instead of letting the failure surface three layers away in pdf.js.

State Management

The entire app state lives in a single Zustand store with immer middleware:

src/state/store.ts → useStore

State Shape

Field	Type	Purpose
project	Project | null	The loaded, normalized project (schema + papers)
currentPaperId	string | null	Currently selected paper
saveHandle	SaveHandle | null	Where to write back (path, FSAPI handle id, or download)
projectName	string	Display name for title bar
dirty	boolean	Unsaved changes flag; gates beforeunload guard
loadError	LoadError | null	Error overlay data
busy	boolean	Disables toolbar buttons during async operations
sidebarCollapsed	boolean	Paper list visibility. Driven by SidebarToggle, which renders inside the paper list's own header while it is open and in the toolbar once collapsed — otherwise the button would disappear with the pane it reopens.
pdfSelection	string	Latest text selected in the PDF viewer (for "grab from PDF")
theme	Theme ('light' | 'dark')	Current app theme (persisted in localStorage via src/state/settings.ts)
fontScale	number	Current font scale factor (0.7–2.0, persisted in localStorage)
pdfZoom	number	PDF zoom multiplier (0.4–3.0, session-only, default 1)
recents	RecentEntry[]	Recently opened projects (max 5, from platform.getRecents())
helpOpen	boolean	Help dialog visibility
past / future	HistoryEntry[]	Undo/redo stacks for annotation edits (session-only, capped at 100). Each entry is { project, paperId }; thanks to immer's structural sharing, snapshots are cheap references
aiMarks	Record<string, true>	Fields the AI filled and the reviewer has not looked at yet, keyed `${paperId}::${canonicalPath}` for a single-reviewer project, or `${paperId}::${reviewer}::${canonicalPath}` for a multi-reviewer one (see "AI marks" below). Session-only: it lives beside the project, so serializeProject cannot see it
currentReviewer	string | null	Which reviewer's tree is shown/edited — "1".."N", "consolidation", or null. Always null for a single-reviewer project; also starts null for a multi-reviewer one until picked (see "Multiple reviewers & Consolidation" below). Persisted per project in localStorage; not an undo step, does not set dirty
consolidationTarget	{ path, name, index } | null	The field the Consolidation "compare" popup (ConsolidationDialog) is showing, or null when closed. Session-only
consolidationOverviewOpen	boolean	Whether the project-wide ConsolidationOverview modal is open. Session-only
deferredConsolidations	Record<string, true>	Fields where the consolidator chose "Enter a different value" — waiting for a manually entered value. Keyed by deferredConsolidationKey(paperId, canonicalPath). Session-only; cleared on project close/load
screeningFilter	'all' | 'included' | 'excluded' | 'undecided'	Which decisions the screening paper list shows. Screening projects only; session-only, resets on closeProject/loadFromText
screeningShowPdf	boolean	Whether the middle pane shows the PDF instead of ScreeningRecord's title+abstract. Session-only, see "Screening" above
screeningSummaryOpen	boolean	Whether ScreeningSummary (the PRISMA-style counts modal) is open. Session-only

Key Actions

• openProject() — delegates to platform.openProject(), then loadFromText(), refreshes recents
• openRecent(id) — delegates to platform.openRecent(id); on success → loadFromText + refreshes recents; on null → prunes recents and sets loadError
• loadFromUrl(url) — fetch the project JSON, set serverBase on browser adapter, then loadFromText()
• loadFromText(text, handle, name) — calls loadProject(text) from the model layer, sets state, selects first paper
• save() / saveAs() — serializeProject(project) → delegate to platform, clears dirty, refreshes recents
• selectPaper(id) — switches paper, clears pdfSelection
• setFieldValue(path, name, index, value) — navigates the annotation tree via containerAt(), sets inst.value, marks dirty
• addInstance(path, def) / removeInstance(path, name, index) — manages repeatable annotation instances, respects max/min
• toggleTheme() / setTheme(theme) — flips or sets the app theme, applies via applyTheme() (sets data-theme attribute on <html>)
• increaseFont() / decreaseFont() / resetFont() — adjusts fontScale by ±0.1 (clamped to 0.7–2.0), applies via applyFontScale() (sets --app-font-scale CSS variable)
• zoomInPdf() / zoomOutPdf() / resetPdfZoom() — adjusts pdfZoom by ±0.2 (clamped to 0.4–3.0, rounded to 2 decimals) or resets to 1; session-only, not persisted
• applyAiSuggestions(suggestions) — writes the reviewer-approved AI proposals into the current paper as one undo step, and marks every field it wrote (see "AI-assisted annotation" below)
• confirmAiMark(paperId, canonicalPath) — drops one AI mark; the reviewer clicked into that field (or its label)
• undo() / redo() — swap the current project snapshot with one from the past/future stack (and switch to the affected paper). The mutating actions push a snapshot before applying; consecutive edits to the same field coalesce into one undo step (a module-level lastFieldKey tracks this), while add/remove/paper-switch reset it. History is cleared on project load.
• setHelpOpen(open) — shows/hides the help dialog
• selectReviewer(reviewer) — switches currentReviewer and persists the choice per project. A view switch: no undo step, no dirty
• openConsolidation(path, name, index, returnToDisagreements?) / closeConsolidation() — open/close the compare popup for one field; when returnToDisagreements is set, closeConsolidation reopens the per-paper disagreement list
• resolveConsolidationValue(path, name, index, value) — writes a chosen reviewer value, marks the field equal, and clears any deferral in one undo step
• deferConsolidationValue(path, name, index) — marks a field as deferred (waiting for manual entry); setFieldValue auto-clears the deferral when a non-empty value is entered
• setConsolidationOverviewOpen(open) — toggle the project-wide ConsolidationOverview modal
• openAgreementFromOverview() / closeAgreement() — closes overview, opens agreement, restores overview on close
• openDisagreementsFromOverview(paperId) / closeDisagreements() — selects paper, closes overview, opens per-paper disagreement list, restores overview on close
• alignConsolidationNode(paperId, nodeName, coalesce) — match the reviewers' repeated entries under one node and write the result in (reorder + grow); see "Matching the reviewers' repeated entries" below
• adoptUnanimousValues(paperId, coalesce) — fill the consolidated fields every reviewer answered the same way, marking each via aiMarks; runs after the matching for a paper
• setScreeningDecision(decision, reason?) / setScreeningReason(reason) — screening-only field writes, routed through currentTree like every other write; see "Screening" below for the auto-advance and reason-clearing rules
• adoptAllUnanimousScreening() — adoptUnanimousValues for every paper in one undo step; safe unscheduled (unlike the per-paper alignment scheduler) because a screening schema has nothing for align.ts to line up — see "Screening" below
• adoptAllUnanimousAnnotations() — the ordinary-schema counterpart: aligns, then adopts, paper by paper, in one undo step, skipping any paper the consolidator has already partly answered; see "Batch-adopting across the whole project" above

The containerAt(root, path) helper walks the annotation tree following PathSeg[] (name + index pairs) to reach the container for a given path. currentTree(project, currentReviewer, paper, create?) decides which tree that walk starts from — see "Multiple reviewers & Consolidation" below; setFieldValue, addInstance, removeInstance, and applyAiSuggestions all call it before touching anything.

Component Tree

App (src/App.tsx)
├── Toolbar (src/components/Toolbar.tsx)
│     Open ▾ dropdown (Open file… + recent projects) + Save ▾ dropdown (Save / Save as…)
│     Font controls (A− A A+), theme toggle (☾/☀), help (?)
│     Reviewer switch (multi-reviewer projects only), centered on the toolbar — Reviewer 1..N + Consolidation, hidden entirely for a single-reviewer project; pills at ≤5 reviewers, a dropdown above that; see "Multiple reviewers & Consolidation" below
├── [if project loaded: workspace — a CSS grid whose column widths come from resizable panes]
│   ├── PaperList (src/components/PaperList.tsx)
│   │     List of papers with search box (META / TAGS modes, see below); a dot showing the active reviewer's completeness — a conic-gradient partial fill, not just touched/untouched, see "Completeness dot" below (screening and Consolidation keep their own tri-state/binary markers); click to select
│   ├── Splitter (src/components/Splitter.tsx) ×2  — drag handles between the panes
│   ├── [screening project + PDF pane not toggled on: ScreeningRecord (src/components/ScreeningRecord.tsx)]
│   │     Title/authors/DOI header + the abstract (or "No abstract recorded"); a "Read the PDF" button swaps to PdfViewer when paper.pdf !== ''
│   ├── [otherwise, including the screening PDF toggle: PdfViewer (src/components/PdfViewer.tsx)]
│   │     react-pdf Document+Page; ResizeObserver for width; zoom controls; multi-page navigation; jump history (back/forward); in-PDF search (Ctrl+F); text selection capture; empty state for paper.pdf === '' (screening only)
│   ├── [screening project: ScreeningPanel (src/components/ScreeningPanel.tsx)]
│   │     Include/Exclude decision buttons + a Reason ComboBox (disabled unless Exclude) + progress line; ◧ Summary always, ⚖ Agreement / ⚠ Disagreements in the Consolidation seat — see "Screening" below
│   └── [otherwise: AnnotationPanel (src/components/AnnotationPanel.tsx)]
│         ✦ AI button in the column header (opens AiDialog; disabled while busy, when the paper has no PDF, when the project forbids it, when no reviewer is picked yet, or — by default — always, until the hidden unlock; see "AI-assisted annotation" below)
│         renders the tree `currentTree()` routes to for the active reviewer; prompts to pick a reviewer instead of the form when a multi-reviewer project has none selected yet
│         Consolidation seat only: "☰ Overview" button (opens `ConsolidationOverview`) + "⚠ Disagreements" button (opens per-paper `DisagreementOverview`); builds a `ConsolidationVerdictsContext` map (per-field agree/disagree status) shared to all child `Field` components via React Context
│         └── AnnotationNode (src/components/AnnotationNode.tsx) [recursive]
│               └── Field (src/components/Field.tsx)
│                     Input control (text/number/checkbox/enum ComboBox) + ⧉ grab-from-PDF button + (Consolidation mode only) ⇄ compare button → opens ConsolidationDialog for that field; reads `useConsolidationFieldStatus(canonical)` from context to add `consolidation-agree`/`consolidation-disagree` CSS classes; checks `deferredConsolidations` for `consolidation-pending` visual state
├── [if no project: welcome screen with "Open project…" button, "New from screening…", and (if any) recent projects list]
├── AiDialog (src/components/AiDialog.tsx)
│     Modal driven by useAiStore's phase: pick a target → Start → review table → Apply
│     └── LlmSettingsDialog (src/components/LlmSettingsDialog.tsx) — manage LLM targets (stacks on top)
├── HelpDialog (src/components/HelpDialog.tsx)
│     Modal overlay with app intro + keyboard shortcuts table
├── ValidationDialog (src/components/ValidationDialog.tsx)
│     Modal overlay showing the results of "Validate", scoped to the active reviewer's own tree (or the consolidated one, for Consolidation) — see below
├── ConsolidationDialog (src/components/ConsolidationDialog.tsx)
│     Modal overlay reachable only from Consolidation mode's ⇄ button: every reviewer's answer for one field, side by side; picking one calls `resolveConsolidationValue` (writes value + marks equal + clears deferral), or "Enter a different value" defers for manual entry
├── ConsolidationOverview (src/components/ConsolidationOverview.tsx)
│     Project-wide modal for Consolidation's batch actions: lists all papers with ≥1 disagreement, houses "Adopt all unanimous" (with run progress), and opens Agreement / per-paper DisagreementOverview via return-to-flag navigation
├── ScreeningSummary (src/components/ScreeningSummary.tsx)
│     Modal: progress + PRISMA-style include/exclude/reason counts for a screening project
├── ScreeningImportDialog (src/components/ScreeningImportDialog.tsx)
│     Modal: the pre-commit summary for "New from screening…" / "Import from screening…", including "New from screening…"'s annotation/screening target-kind radio — see "Screening" below
├── DuplicateReviewDialog (src/components/DuplicateReviewDialog.tsx)
│     Modal: reviewing the *probable* duplicates a reference import found — one row per pair, a merge/separate choice defaulting to undecided — see "Duplicate detection at import" below
├── GitCloneDialog (src/components/GitCloneDialog.tsx)
│     Import-from-git modal, driven by useGitStore's clone.phase: setup (URL + destination) → cloning (spinner + elapsed seconds) → error (git's exact text, back to setup) → done (pick the project JSON, opened inside the clone) — Electron only, see "Git" above
├── GitDialog (src/components/GitDialog.tsx)
│     Modal for the open project's own repository: changes + diff, a commit message, Commit, Pull, Push — shown only when the toolbar's Git button is present. When the open project's own file is a tracked, field-diffable modification, its changes are reviewed field by field (Use/Ignore/Discard per row) instead of as one whole-file checkbox — see "Field-level commit review" below. When every row is marked Discard, the primary button relabels to "Discard all" (danger-red) and reverts marked rows directly via `runDiscard` without committing. When a Discard row is mixed in among Use rows, `mixedDiscardConfirmMessage()` warns that the Discard field's change will be lost on commit before proceeding
├── GitMergeDialog (src/components/GitMergeDialog.tsx)
│     The pull's conflict-resolution list, grouped by paper (one collapsible section per paper, auto-collapsing when its last conflict is decided): your value left, the remote's right, an editable final value in the middle, with full-text wrapping instead of one-line clipping. "Use all mine"/"Use all remote" exclude conflicts in another reviewer's own tree (`isForeignReview()`), badged "another reviewer", leaving them for individual ◀/▶ resolution. No Escape, no backdrop-click, no × — see "Git" above for why
└── ErrorPanel (src/components/ErrorPanel.tsx)
      Modal overlay for load/save errors

Dropdown component

Dropdown (src/components/Dropdown.tsx) is a reusable click-to-open menu. Items are a union of item (label, optional shortcut, disabled, onSelect), separator, or header. It closes on outside-mousedown, Escape, or item selection. The Toolbar uses two Dropdown instances: Open ▾ (with recent projects list) and Save ▾ (with shortcut labels).

AnnotationNode (recursive renderer)

AnnotationNode is the core recursive component. Given a ResolvedDef and a container (AnnotationValueTree), it:

1. For a single non-repeatable leaf — renders one Field on a single row (the common case).
2. For repeatable or group nodes — renders a header with "+ Add" button (if repeatable), then iterates instances. Each instance shows a remove ("×") control (if repeatable), an optional field, and recursively renders children with an extended path.

The path (PathSeg[]) is extended at each nesting level: [...path, { name: def.name, index: i }]. This path is passed to store actions to navigate the tree.

Field component

Field renders the appropriate input based on def.type:

• boolean → checkbox
• number → <input type=number>
• string with options (enum) → ComboBox (src/components/ComboBox.tsx) — a filterable dropdown of allowed values; no free-text grab button
• string without options → auto-expanding <textarea> (single line when idle, grows on focus up to 240px, max 500 chars)

The grab-from-PDF button (⧉) reads useStore.getState().pdfSelection and inserts it. It is shown for string (non-enum) and number fields. For number fields, it extracts the first numeric token via parseNumber() (handles comma decimals).

Annotation names and descriptions

src/components/NodeName.tsx renders schema node names. When a definition has a description, the UI adds an ⓘ marker, shows the description as a hover/focus tooltip, and renders that tooltip in a portal so it is not clipped by the annotation panel scroll container. The wrapper also includes an aria-label that combines the name and description for assistive technology.

Update check

The start screen shows the running version at the bottom (SaiLoR v0.1.0) and, when a newer release exists, a notice linking to it. src/model/version.ts fetches api.github.com/repos/Gram21/SaiLoR/releases/latest and compares the tag with __APP_VERSION__ — injected from package.json by vite.config.ts, so package.json stays the single source of truth for the version.

This only works while the repository is public. GitHub answers 404 to an unauthenticated request for a private repo's releases, and the app is distributed — embedding a token to get around that would ship a credential to every user, so we don't. The check is therefore written to fail silently: a 404 (private), 403 (rate-limited), network error, draft release, or unparseable version all yield null, and the app simply shows no notice. It starts working the moment the repo is made public, with no code change.

When a newer release exists, the notice offers a direct download of the installer for this machine rather than dumping the user on the releases page: pickInstaller matches the release's assets against the OS/arch reported by getOsInfo() (Electron exposes process.platform/process.arch through the preload bridge; the browser returns null, since a web deployment has no installer and updates by redeploying). Matching keys off the OS/arch we bake into the artifact names (SaiLoR-0.2.0-macos-arm64.dmg), so it stays correct as long as package.json's artifactName patterns do. When nothing matches it offers nothing rather than the wrong binary, and the release-notes link is always there as a fallback.

isNewerVersion compares numeric components, not strings (0.10.0 > 0.9.0, which a string compare gets wrong), strips a leading v, and sorts a pre-release below its release (1.0.0-beta < 1.0.0). It never claims an update from a version it cannot parse.

Pre-release tags are compared the same way, by comparePre, following semver §11.4: dot-separated identifiers left to right, numeric ones compared as numbers and sorting below alphanumeric ones, and a longer tag beating its own prefix. Comparing the raw strings agreed with this exactly until the tenth pre-release and then inverted — "rc.10" < "rc.2" lexicographically — so a user on rc.2 was never offered rc.10, and a user on rc.10 was offered rc.2 as an update, with a live download button. The project ships pre-release versions, so this was live rather than theoretical.

The result — including a null — is cached in localStorage (slr.updateCheck) for 24 h, so a private repo or an offline launch doesn't re-request on every startup, and the 60-requests-per-hour unauthenticated rate limit is never a concern.

PaperList search modes

The paper-list search box (src/components/PaperList.tsx) has two modes, toggled by a trigger sitting inside the input's right-hand edge, labelled META (title + authors + DOI + PDF file name + paper ID, the default) and TAGS (the annotation values recorded in the active reviewer's tree; see currentTree above). The trigger is given a fixed width rather than relying on its two labels being the same length: the active state is bold, and bold vs. regular text of equal character count still measures a few px apart, which would reintroduce exactly the reflow the fixed width exists to prevent. It replaced a 🔎/🏷 emoji pair whose differing glyph widths visibly resized the control on every toggle. Both modes share one ranking: filter to papers where every query word matches, then sort by distinct words matched, then total matched characters, then original order — only the haystack differs per mode. The META haystack (paperMetadataHaystack) includes paper.pdf (the file name) and paper.id so a reviewer can find a paper by the PDF they remember or by its identifier, not just by bibliographic metadata.

Both haystacks are precomputed once per paper in a useMemo keyed on [papers, schema], not re-walked on every keystroke. papers is a safe key for annotation content too: the store's immer set produces a new paper object — and therefore a new papers array — on every field edit, so the memo is invalidated exactly when annotation content actually changes, never stale.

The annotation haystack comes from annotationText(schema, tree) (src/model/annotations.ts), which mirrors hasAnnotations's recursive walk: it collects every field value that is a non-empty string or a number, joined and lowercased. Booleans are skipped — every paper has one per boolean field (defaulting to false, never absent), so including "true"/"false" would make almost every paper match. Like the rest of that module, it never throws on a tree that doesn't match the schema's shape.

Completeness dot

The paper-list dot used to be binary: touched or not, via hasAnnotations. src/model/completeness.ts replaces that with a real fraction, rendered as a conic-gradient partial fill (.status-dot.partial { background: conic-gradient(var(--ok) var(--fill), transparent 0) }, --fill set inline per row) instead of a flat colour.

• completeness(defs, tree) walks the schema the same way hasAnnotations does and returns { filled, total }. hasRequiredFields(defs) decides the denominator: if the schema declares any required field anywhere (including nested), only required fields are counted; otherwise every field counts. This is the fix completeness.test.ts calls out by name as "the headline bug this module fixes" — a schema with one required field out of thirty used to make finishing that one field look like 3% progress.
• Booleans are excluded from the count entirely, for the same reason the Validation section below documents isEmptyValue/hasAnnotations disagreeing on them: an untouched checkbox reads false, indistinguishable from a deliberate "no", so there is no way to count it as "filled" or "unfilled" that isn't a guess. Completeness is a third function with its own opinion here — it doesn't count the field at all.
• Repeatable instances are counted per present instance, not once per node, so adding a second Finding grows the denominator along with the numerator.
• completenessPercent(c) returns null — the signal to fall back to the old binary dot — whenever c.total === 0 (an empty schema). It otherwise clamps to [5, 99] for any non-zero, non-total count, so a paper at 1/200 fields doesn't visually read as empty and one at 199/200 doesn't read as done; the literal old status-dot/status-dot done classes are reused verbatim at exactly 0% and 100%, rather than rendering a conic gradient that would look the same as those two classes anyway.
• Screening projects and the Consolidation seat opt out, keeping their existing tri-state/binary markers: paperCompleteness() (PaperList.tsx) returns null for them, the same "nothing to compute" signal total === 0 produces, so both routes land on the same old-dot fallback with no separate branch.

PaperRow was pulled out of PaperList's .map() and wrapped in React.memo. Its props (paper, active, onSelect, dotClassName, dotLabel, dotFill) are deliberately primitives — dotFill is passed as a bare number rather than the Completeness object, specifically so memo's shallow comparison is cheap and correct. This pays off because of how the store's immer set() already works (see "State Management" above): editing one field produces a new object for exactly that one paper, leaving the other 1999 array entries referencing their old objects — confirmed directly by PaperList.perf.test.ts, which asserts exactly 1 of 2000 paper objects changes identity per edit. So a single edit re-renders exactly one row's PaperRow, not all 2000. Measured cost (from the feature's own commit message): recomputing completeness (and the search haystack) over 2000 papers costs on the order of 3–4ms, and the row re-render this memoization buys back drops to single-digit milliseconds — comfortably fast enough that no windowing/virtualization was added for this.

Validation

src/model/validate.ts checks a reviewer's annotations against the schema; the Validate button in the toolbar runs validateProject(project) and ValidationDialog shows the result grouped by paper (click a paper to jump to it). Four issue kinds: required (a field marked required is empty), type (the stored value doesn't match the field's type — the JSON is hand-editable), enum (a value outside the field's options), and cardinality (an instance count outside [min, max]).

Only papers with at least one annotation are actually validated. validateProject returns { issues, unannotated }, not a bare array: a paper nobody has touched yet fails every required field for the single reason that it hasn't been started, which would drown the results in noise that says nothing a reviewer doesn't already know from the paper list's own "not annotated yet" dot. Such a paper is skipped from issues entirely and instead added to unannotated — the check is hasAnnotations(schema, paper.annotations) (src/model/annotations.ts, the same primitive that drives that paper-list dot, see below), so "annotated" means exactly the same thing in both places: on the first field genuinely filled in, not merely present in the tree. This matters for booleans specifically, since hasAnnotations disagrees with isEmptyValue on purpose — a boolean left at its untouched false does not count as an answer here (only an explicit true does), whereas isEmptyValue treats a boolean as never empty for the required check. The two functions are answering different questions ("has this paper been started at all" vs. "is this one required field answered") and are not meant to agree. ValidationDialog renders unannotated as a separate, plain "not annotated yet" checklist below the issues — clickable like everything else, but with no kind badge or message, since there is nothing to report beyond "this one hasn't been started."

Emptiness is the load-bearing definition, and booleans are the special case: a boolean field is never empty. An unticked box is a real answer (false), and a missing/null boolean reads as false — so a required boolean can never raise a required issue. 0 is likewise a real number, and ''/whitespace is empty only for strings. A type mismatch suppresses the required/enum checks for that field, so one broken value yields one issue rather than a cascade. Everything is defensive: a malformed tree produces issues rather than throwing.

Fields are marked required by required: true in the schema (ResolvedDef.required, defaulting to false; rejected on a group, which holds no value). The schema editor exposes it as a Required checkbox on non-group rows, and the annotation form marks such fields with a red *.

Note that loadProject normalizes the tree to each node's min/max, so in practice cardinality issues only arise from a project that bypasses the loader.

On a multi-reviewer project, Validate checks the active reviewer's own tree, not the consolidated one — unless the active reviewer is Consolidation, in which case it checks the tree that actually ships. runValidation (store.ts) builds this by mapping each paper's annotations through currentTree() before calling validateProject, so validate.ts itself needs no reviewer awareness. The Validate button is disabled until a reviewer is picked, for the same reason the annotation form is withheld then: there is no "the reviewer" to validate yet.

Project editor

A second screen (src/components/ProjectEditor.tsx, shown instead of the workspace while useEditorStore().open) lets users create or edit a project JSON — its annotation schema and the PDFs it references — without hand-writing JSON. It is entered from the welcome screen's New annotation JSON… / Edit annotation JSON… buttons, or from the ✎ pen on a recent project (startEditRecent, which loads that recent by id rather than prompting a file picker; startEdit and startEditRecent share editorStateFromOpened).

The help dialog is mode-aware. HelpDialog derives a mode from useEditorStore().open and whether a project is loaded, then renders one of three guides, each with only the shortcuts that actually do something there, plus a badge in the title naming the mode:

• Getting started (start screen, nothing open) — what an SLR project JSON is, and what the three buttons do (open / new / edit).
• Annotating (a project is open) — pick a paper, read the PDF, fill the fields, save.
• Editing the annotation JSON (the editor is open) — schema building, drag-to-nest, adding PDFs, the two save buttons.

Shared sections (appearance, license) render in all three.

Two ways out of the editor: Save JSON writes the file and stays put (so you can keep building), while Save JSON & Begin Annotating writes it and hands it to the annotation view (loadFromText) — that split is save() vs saveAndAnnotate(). Both validate first, so an invalid draft neither writes nor closes. The editor has its own undo/redo history (past/future snapshots, same shape as the annotation store, with consecutive keystrokes in one input coalesced into a single step), and useKeybindings / useElectronCloseGuard route Ctrl+S / Ctrl+Z / Ctrl+Shift+Z / Ctrl+Y to it whenever it is open.

• src/state/editorStore.ts — a separate Zustand+immer store holding the draft. It deliberately works on the raw JSON shape, not the loaded Project: each paper's annotations object is carried through verbatim while the schema is edited, so editing the schema never prunes existing annotation data (it is normalized against the new schema the next time the project is opened for annotating). Key pieces: EditorNode (a schema node with a client-side uid, where kind: 'group' means "no type" — a name-only sub-tree), EditorPaper (which also keeps the PDF's absolute sourcePath), toAnnotationDefs/fromAnnotationDefs (conversion to/from the compact on-disk AnnotationDef), moveNodeIn (tree move that refuses to drop a node into itself or its own subtree), buildProjectJson, and validateDraft — which runs the real projectSchema + resolveSchema validators, so the editor cannot produce a file the loader would reject. On save it writes the JSON and hands it straight to the main store via loadFromText.

• SchemaTreeEditor.tsx — recursive tree exposing the schema's full expressiveness: name, kind (Group / Text / Number / Yes-no), min, max (with an ∞ checkbox for max: null = unbounded repeats), description, enum options (string fields only), nesting, add/remove. Native HTML5 drag-and-drop reorders rows and builds nesting: the drop position comes from the pointer's Y within the target row — top 25% → before, bottom 25% → after, middle → inside (nest as a child).

  Renaming or removing a field warns first when papers still record answers under it. Answers are keyed by field name and nothing migrates them: normalizeTree builds its output by iterating the schema's defs and drops any key the schema no longer has, so the next load quietly prunes an orphaned answer and the next save makes that permanent. countPapersUsingField (src/model/fieldUsage.ts) is the counterpart of reasonUsage.ts below, walking the consolidated tree and every reviewer's own; an unticked boolean and a blank string are not answers, so neither makes a rename look destructive when it is not.

  It matches the field's path from the schema root, not its name anywhere in the tree. Matching the bare name over-warned during the most ordinary editor sequence there is — add a field, type a name another field already uses, change your mind, delete it — and a guard that cries wolf over a node holding nothing is one people learn to click through, which disarms it exactly when it matters. The check runs on blur (a committed rename), not per keystroke, and the × button checks both the live name and any uncommitted one, because clicking a <button> does not move focus on macOS/Chromium and the input's blur would otherwise never fire.

  Dragging carries the same guard, since it is the same loss reached by a different gesture: moving a field into or out of a group changes the path its answers live under. Only a change of parent asks — reordering among siblings leaves the path alone, because answers are keyed by name at each level and never by position, so warning there would be the crying-wolf failure again. nodePathNames and parentUidOf (src/state/editorStore.ts) are what tell the two apart; parentUidOf returns null for a root-level node and undefined for a uid that is not in the tree, and the guard branches on that difference rather than treating an unknown node as safe.

  One deliberate gap remains: a group renamed but not yet saved makes its children's paths miss the old answers, so removing a child then asks nothing. That is intentional — renaming the group is itself guarded, so reaching that state means the reviewer has already been told those answers will be discarded and agreed.

• PapersEditor.tsx — add PDFs one at a time, a whole folder at once, import a reference-manager export, or (outside a screening project) import from a screening project's results (four buttons next to each other in the header/empty state), edit each paper's id/title/authors/DOI/abstract and its pdf path, reorder by drag, remove.

• ScreeningReasonsEditor.tsx — replaces SchemaTreeEditor in the editor body whenever useEditorStore().screening is set: an ordered, editable list of exclusion reasons (add / remove / reorder with plain ↑/↓ buttons rather than drag — the list is short enough that drag-and-drop's extra affordance isn't worth it), writing through setScreeningReasons. On blur after renaming a reason, the editor checks countPapersUsingReason (src/screening/reasonUsage.ts) across both consolidated and per-reviewer trees; if papers still record the old label, it offers to migrate those decisions to the new label (or warns when there is no new label to migrate to), so a rename never silently orphans a decision. See "Screening" below.

• ProtocolEditor.tsx — a collapsible Review protocol section in the project editor for Project.protocol (research questions, search strings, databases, search date, notes). See Data Model's "The review protocol" section for the field's shape and merge behavior.

Adding PDFs (addPdfs) does two things beyond appending rows:

• Duplicate rejection. A PDF already referenced is skipped rather than added twice, and the skipped names are reported in a dismissible notice. Identity is pdfKeys(): the absolute path when known (Electron) and the stored relative path — so re-picking the same file, picking it twice in one dialog, or picking one already listed in an opened project all collapse to one entry. Same-named PDFs in different folders stay distinct.
• Title/author auto-fill. src/model/pdfMeta.ts reads each added PDF (PickedPdf.read() — an IPC in Electron, File.arrayBuffer() in the browser) and pre-fills the fields. It tries the PDF's embedded Title/Author metadata first, validating it (isPlausibleTitle rejects artefacts like "Microsoft Word - paper_final_v3.doc"), then falls back to a layout heuristic on page 1: the largest text near the top is the title (joined across wrapped lines), and the lines under it are the authors. A line is not a flat string but a list of segments, split wherever two runs on the same baseline sit more than COLUMN_GAP_RATIO font-sizes apart — a word space is a fraction of the font size even in justified text, a column gutter is several times it. This matters because the common two-column author block puts each author on the same baseline with nothing between them but the gutter: joined first, Jan Keim and Angelika Kaplan read as the single name "Jan KeimAngelika Kaplan", and no amount of later parsing can recover the boundary, since there is no punctuation at it. So authors are parsed per segment. A run whose width pdf.js does not report never starts a segment: without it there is no way to know where a run ends, and guessing could cut mid-phrase. parseAuthorList strips affiliation superscripts, footnote daggers, emails and an "Authors:" label; in strict mode (used only for the heuristic, which is a guess) it also requires each entry to look like a person's name, so a body sentence can't become an author list.

A long author list wraps, and the break lands inside a name — a real seven-author paper (samples/pdfs/A1-37.pdf) ends one line … Niklas Ewald, Tobias and opens the next Thirolf, and Anne Koziolek. Parsing each line on its own loses two authors and cannot get them back: Tobias is a lone token strict mode correctly rejects, and Thirolf never recovers its first name. So namesFromAuthorBlock joins the lines before parsing — per column, matching segments by x (COLUMN_X_TOLERANCE), so a two-column block's wraps stay inside their own column instead of inventing a name across the gutter. Superscript affiliation keys, which sit on their own raised baseline and so arrive as a line of their own ("1 1") between the halves of a wrapped list, are stepped over rather than stopped at (SUPERSCRIPT_SIZE_RATIO).

Growing that block is guarded, and the guard is what makes it safe. The line under the authors is far more often an affiliation or an email row than the rest of the list, and absorbing one does not merely add noise — it destroys names, since the last author fuses with it into a single entry ("John Smith Karlsruhe Institute of Technology") that parseAuthorList then drops wholesale as an affiliation. So a candidate line is kept only when the block including it yields strictly more names than the block without it. That tests the join on its own evidence, which is both simpler and harder to fool than trying to recognise an affiliation line up front — a real continuation adds names by construction, and every kind of line that shouldn't be absorbed takes them away. Everything is best-effort and only ever pre-fills — rows appear immediately with a name-derived placeholder, extraction patches them in the background, and it never overwrites a value the user has already typed. The pdf.js worker is configured once in src/platform/pdfjs.ts, shared by the viewer and the extractor. The same background pass also tries an abstract (abstractFromLines) when the row has none — see the Screening section's "A missing abstract is extracted from the PDF, and flagged durably" for why that one gets a persisted abstractFromPdf disclosure the title/author guesses above do not need.

Adding a whole folder of PDFs (addPdfFolder) is addPdfs with a different picker: both funnel into a shared addPickedPdfs(picked) closure in editorStore.ts that does the duplicate rejection, row creation, and background title/author extraction described above. Only pickPdfs() vs pickPdfFolder() differs. pickPdfFolder() returns every .pdf found recursively under the chosen folder — Electron via a plain recursive readdir walk over the real filesystem, the browser via the same webkitdirectory <input> mechanism ensureLocalPdfGrant() uses for the PDF-viewer's one-time folder grant (see pickPdfFolderViaInput above) — filtered to file names ending in .pdf. Reusing that input, rather than writing a second one, keeps the one careful piece of cross-browser logic (the cancel-event-not-focus-guess cancellation detection, needed because Firefox's own "Upload N files?" confirmation appears after the OS dialog already returns focus) in one place.

Importing references (importReferences) reads a BibTeX/RIS/CSL-JSON export from a reference manager (Zotero, Mendeley, JabRef, EndNote) and turns each entry into a paper row, without requiring a PDF to already be attached. src/model/references.ts (parseReferences(text, filename)) is a pure, defensive parser — the format is picked from the extension, content-sniffed as a fallback — that never throws: a malformed entry (unbalanced braces, a missing field, an entry with no title) is skipped rather than failing the whole file. For each parsed RefEntry, editorStore.ts dedupes against existing rows via classifyImport (see "Duplicate detection at import" below): a certain match fills in that row's empty fields only (never overwriting something the reviewer already typed) and counts as "updated" or "already complete" in the summary notice; a probable match is held back for the reviewer to decide (DuplicateReviewDialog); a new entry adds a fresh row with pdf: '' (or the imported pdfHint's file name as a placeholder, if the reference file named one) for the reviewer to attach a PDF to afterward. The whole import (including any duplicate-review decisions) is one undo step. Because pdf: '' is something addPdfs/addPdfFolder never produce, validateDraft reports it as a named per-paper issue ("Paper N has no PDF attached") rather than letting it reach buildProjectJson/save(), which still requires the on-disk pdf to be non-empty (paperSchema in src/model/schema.ts is unchanged — this tolerance is draft-only).

Duplicate detection at import (src/model/duplicates.ts)

Before this existed, importReferences' dedup was two flat tiers — an exact DOI match, else an exact normalized-title match — and anything short of that silently became a second row for a paper already in the project. classifyImport(existing: DupRecord[], incoming: DupRecord[]): DupVerdict[] replaces both tiers with a graded one: every incoming record is classified new, certain, or probable, checked both against the papers already in the project and against earlier records in the same incoming batch (so importing the same paper twice in one file is caught too), in this priority:

1. Exact DOI (case-insensitive) — certain, and the only tier the year check below never touches.
2. Exact normalized title (lowercased, whitespace-collapsed, punctuation-stripped) — certain.
3. Fuzzy full-title match — stringSimilarity (src/consolidate/similarity.ts) ≥ 0.90 — probable.
4. Subtitle-stripped base title (everything before the first :) at the same ≥ 0.90 threshold, gated on an author-surname Dice coefficient ≥ 0.50 — probable. This is what catches "same paper, different subtitle" without also catching two unrelated papers that happen to share a generic title fragment.

A year gap of YEAR_GAP_VETO (2) or more demotes an otherwise-matching title pair straight to new — two papers four years apart are not the same paper no matter how alike their titles read. The veto is title-only: a DOI match is definitive on its own and is never second-guessed by a year gap.

classifyImport reuses stringSimilarity rather than reimplementing fuzzy title matching — the same one-shared-implementation rule comparable() follows for consolidation. Before this feature stringSimilarity was internal to similarity.ts (called only by its own valueSimilarity wrapper, which align.ts uses for reviewer-entry matching); duplicates.ts is its first outside caller. A cheap pre-check (a token-Dice pass, then a length-bound, then a character-histogram bound) skips the real, more expensive stringSimilarity call for pairs that are obviously nothing alike.

A certain verdict merges silently, exactly as the old exact-match tier did (fill empty fields only). A probable verdict is held back for DuplicateReviewDialog (src/components/DuplicateReviewDialog.tsx, styled by src/styles/duplicates.css): one row per pair, a merge / separate choice per row that starts undecided — Import stays disabled until every row has one, so a probable duplicate is never silently merged or silently doubled. editorStore.ts wires this as duplicateReview state plus commitImport (the shared writer both a duplicate-free import and a resolved review call into), resolveDuplicateReview, setDuplicateDecision, and setAllDuplicateDecisions.

findMatchingPaper — the function the rest of the editor already called to ask "is this paper already in the project" — is now a thin adapter over classifyImport: it returns a match only for a certain verdict, so every other caller's behavior is unchanged by the grading in between.

Highlighting newly added papers. Every row added by addPdfs, addPdfFolder, or importReferences gets a light-blue border so the reviewer notices the addition and checks the (possibly guessed) metadata — the same idea, and the same --ai-mark CSS variable, as the AI-annotation "unconfirmed" marks below. editorStore.ts tracks this the same way aiMarks does in the main store: a session-only justAdded: Record<uid, true>, not part of EditorSnapshot/undo (an add already has its own undo step) and never written to the file. It is cleared wholesale on save, on opening a project into the editor, and on closing the editor (openEditorSession, close, save). PapersEditor.tsx clears one row's mark (confirmAdded) the moment the reviewer focuses any of its fields, mirroring how Field.tsx clears an AI mark on focus/click.

Relative PDF paths. The JSON's location is chosen up front (and changeable any time via Change…), because a paper's pdf is stored relative to the JSON file. Each picked PDF keeps its absolute sourcePath, so when the location changes changeLocation() re-derives every pdf against the new directory (/reviews/x.json + /reviews/pdfs/a.pdf → pdfs/a.pdf; move the JSON up a level and it becomes reviews/pdfs/a.pdf). This only works in Electron, where real filesystem paths exist: the platform methods pickProjectLocation / pickPdfs / relativePdfPaths (see below) compute it via a paths:relative IPC (path.relative(dirname(json), pdf), POSIX-separated). In the browser the File System Access API exposes no paths, so a picked PDF is stored as its bare file name and the user places it next to the JSON or edits the relative path by hand — the papers editor says so. A row added by importReferences with no PDF yet (pdf: '') has no path to rederive either way, and is left alone until the reviewer attaches one.

Reference-file parsing (src/model/references.ts)

parseReferences(text, filename) turns a BibTeX/RIS/CSL-JSON export from a reference manager (Zotero, Mendeley, JabRef, EndNote) into RefEntry[] (title, authors, doi?, year?, pdfHint?). The format is picked from the file extension, content-sniffed as a fallback (@ → BibTeX, a TY - line → RIS, leading [/{ → CSL-JSON). It is a pure, defensive parser — a malformed entry is skipped rather than failing the whole file, and parseReferences itself never throws.

LaTeX escapes. BibTeX (and RIS, when it originated from a .bib round-tripped through a converter) routinely spells non-ASCII letters as LaTeX escapes rather than UTF-8 — an accent command on a base letter (\"o, \'e, \c{c}) or a handful of standalone letters that aren't "letter + accent" at all (\ss, \o, \ae, \i, ...). unescapeLatex resolves both via lookup tables (LATEX_ACCENTS, LATEX_LETTERS) rather than a chain of per-letter replacements, and handles all three shapes a command can appear in — bare (\"o), its own braced argument (\"{o}), or wrapped in an extra capitalization-protecting brace pair ({\"o}) — as one case: it only ever matches backslash-led sequences, so surrounding braces are inert to it either way. This is why cleanBibValue runs it before stripping braces — {\"o}'s braces need to still be there for the bare-letter branch to see past them, and are gone by the time the generic {} strip (still needed for plain capitalization protection like {DNA}) runs afterward. A bare control word also consumes the single space that terminates it — S\o ren is "Søren", because TeX ends a control word at that space and does not treat it as text; getting this wrong mangles most Nordic names. That is safe only because the author list is split on " and " before any unescaping (parseBibEntry), which is the order BibTeX itself works in: the separator belongs to the field, not to any one name. Unescape first and a name-final control word (Hans Wei\ss and Sven) would swallow the separator's space and glue the two names together. An unrecognized escape degrades gracefully: the backslash is dropped and the rest is left as-is, never a crash or a stray backslash. This unescaping does not apply to the BibTeX file field or RIS L1/UR (cleanBibPathSegment, kept deliberately narrower than cleanBibValue) — a Windows path like C:\Users\name\file.pdf is mostly backslashes, and the graceful-degradation fallback would otherwise mangle it. CSL-JSON is real JSON, already UTF-8, so none of this runs there.

Merged author names. Some exports lose the and separator between BibTeX authors entirely ("Jan KeimAngelika Kaplan") or partially ("Jan Keimand Angelika Kaplan", "Jan Keim andAngelika Kaplan"). splitAuthorList repairs the unambiguous half-loss (andX with no space and a capital right after — never legitimate text) unconditionally, then repairMergedAuthorNames looks for a lowercase→uppercase seam either inside a bare "and"-suffixed token or inside a single fused token, and only commits to a split when both halves come out as plausible multi-token "First Last" names. This is a deliberately conservative heuristic — a wrong split silently corrupts a real name, which is worse than leaving two names glued together — so a Mc/Mac/De/Di/La/Van/Du prefix allowlist protects McDonald-style surnames, an apostrophe or hyphen at the capital (O'Brien, Smith-Jones) is never even seen as a seam (the check requires direct lowercase→uppercase adjacency), and the both-sides-multi-token requirement is what keeps genuine names ending in "and" (Roland, Armand, Ferdinand) from being torn apart when nothing follows them worth calling a second name. RIS and CSL-JSON don't get this treatment: their authors are already structurally one-per-entry (AU/A1 lines, author[] array), so the separator-loss problem this heuristic exists for cannot occur there.

PdfViewer

Uses react-pdf's Document + Page components. The pdf.js worker is loaded from the bundled dependency URL. A ResizeObserver tracks container width so pages scale to fit; the final render width is the fit-to-width size multiplied by the store-level pdfZoom factor. The PDF header shows the paper title, authors, and DOI, plus zoom controls (−, percentage, +) wired to zoomOutPdf / resetPdfZoom / zoomInPdf. For multi-page PDFs, the header also shows page navigation (prev/next buttons, a page-number input, and a total page count). The current page is tracked from scroll position via onScroll — the last page whose top has scrolled past 30% of the viewport height — and typing a page number jumps to that page. The PDF text and annotation layers are both rendered. Pages are align-items: safe center so horizontal scrolling remains reachable when zoomed wider than the pane. Text selection is captured via onMouseUp/onKeyUp → window.getSelection() → setPdfSelection().

External links. The <Document> is given externalLinkTarget="_blank" and externalLinkRel="noopener noreferrer", so link annotations pointing at a website render as target="_blank" anchors and open in a new browser tab rather than navigating the SPA away. Internal (in-PDF) destinations are unaffected — pdf.js renders them without a target and its LinkService scrolls to them. In Electron, main.ts intercepts the resulting window.open with webContents.setWindowOpenHandler, hands the URL to shell.openExternal (the user's default browser), and returns { action: 'deny' } so no Electron window is created. A will-navigate listener is a safety net that prevents anything from navigating the app window away, routing off-app URLs to the browser too. Both paths go through openExternalUrl, which only opens http:/https:/mailto: — never file: or other schemes, which the OS could use to launch programs.

Jump history (back/forward). Clicking an internal PDF link (e.g. a reference) lets the pdf.js LinkService scroll to the destination. An onClickCapture on the scroll container notices clicks on <a> elements and, after polling briefly, records the pre-jump scrollTop on a back stack only if the view actually moved (so external links, which don't scroll, are ignored). Two header buttons (↩ / ↪, shown once history exists) then move between positions like a browser: jumpBack pops the back stack, pushes the live position onto the forward stack, and scrolls there instantly (scrollTo with default behavior — reliable under prefers-reduced-motion); jumpForward is symmetric. The stacks are refs (with canJumpBack/canJumpForward state mirroring their lengths) and are cleared when the paper changes.

In-PDF search. A 🔍 button in the header (and Ctrl/Cmd+F) toggles a find bar below the header; opening it focuses the input so the user can type immediately (via a searchOpen effect, since the input isn't mounted on the open transition). findMatches walks the text nodes of each rendered text layer (.react-pdf__Page__textContent), concatenating them per layer so a query can span multiple spans, and returns DOM Ranges. Matches are painted with the CSS Custom Highlight API (CSS.highlights + ::highlight(slr-pdf-search) / ::highlight(slr-pdf-search-active)) — this tints the transparent text-layer glyphs without mutating react-pdf's DOM, and degrades gracefully where the API is unavailable. The active match is centered in the scroll container; Enter / Shift+Enter (and the ‹ › buttons) cycle matches. Crucially, the <Page> elements are memoized (useMemo on [numPages, renderWidth, onTextLayerRendered]) with a stable onRenderTextLayerSuccess callback, so typing in the search box reuses the same element references and React skips re-rendering the pages — otherwise every keystroke would tear down and re-render the text layers (a "TextLayer task cancelled" flood) and matches would never resolve.

The page count is capped at MAX_PDF_PAGES (5 000). There is no virtualization here: every page becomes a React element with its own canvas, text layer and annotation layer, plus an entry in pageRefs. pdf.js correctly ignores a lying /Count, but it does not dedupe a page tree that is a DAG — a 2.4 KB file whose /Pages nodes each list the same child twice, 24 levels deep, reports 16 777 216 pages. Building the element array alone measures 3.6 s and 3.6 GB at that count, which is a certain renderer crash from clicking a paper. When a document exceeds the cap the header shows the count with a + and the real total in its tooltip, so a truncated document says so rather than quietly showing fewer pages. extractPdfText has the matching DEFAULT_MAX_PAGES (2 000): its maxPages option previously defaulted to the document's own page count, which made the guard dead code, and there is no cancellation on that path.

PDF source resolution is async: getPlatform().getPdfSource(paper.pdf, saveHandle) returns a { url, revoke? }. The effect cleans up (revokes blob URLs) on paper/handle change or unmount.

Granting folder access carries the same cancellation guard as the load effect. In the browser build, "Choose folder…" leaves the page interactive while the directory walk and the PDF read are still in flight, so the reviewer can select another paper meanwhile — and the load effect then starts its own fetch, since the grant it was waiting for has arrived. Without the guard whichever settled last won, which could leave paper B selected in the list and the annotation panel with paper A's PDF on screen: annotating one paper from another's text, with nothing to show anything was wrong. The grant path checks the live paper id before calling setUrl and revokes the loser's blob URL.

Multiple reviewers & Consolidation

An SLR is normally annotated by ≥2 reviewers independently, then reconciled into one final answer. config.reviewers (2–10) turns this on for a project; see data-model.md for the file format and Paper.reviews shape. This section covers the store/UI wiring.

currentTree(project, currentReviewer, paper, create?) (src/state/store.ts) is the single routing decision every mutating action and every reader goes through: single-reviewer → paper.annotations (unchanged from before this feature); Consolidation → paper.annotations (the tree that ships); a numbered reviewer → paper.reviews[N]; nobody picked yet on a multi-reviewer project → null. create (default false) controls whether a numbered reviewer's missing tree is lazily initialised and normalized in place (only safe inside an immer set() producer — the mutating store actions pass true) or a fresh throwaway default is returned for display/validation without touching the project (selectors and read-only computations pass nothing). setFieldValue, addInstance, removeInstance, and applyAiSuggestions all resolve their target tree this way before doing anything else, and bail out — no write, no undo entry — if a multi-reviewer project has no reviewer selected yet.

Matching the reviewers' repeated entries (src/consolidate/)

Two reviewers who both record three Findings need not record them in the same order: Reviewer 1's Finding #1 may be Reviewer 2's Finding #3. Comparing them position by position would then report disagreement that isn't there. Whenever Consolidation is the active seat, useConsolidationAlignment (src/hooks/) works out which of each reviewer's entries are the same entry and lines them up.

Module	Does
similarity.ts	How alike two answers are, as {score, weight}. weight is what makes "agrees on five fields" outrank "agrees on one" — averaging scores alone cannot tell those apart, as both average to 1.0. Weight 0 means the pair said nothing (a field only one reviewer filled abstains rather than voting against). Text is max(levenshtein ratio, token Dice); enums compare as labels, never as characters ("High"/"Low" overlap and mean the opposite); a false boolean carries no evidence in the matching context; a year field is scored as an identity, not a magnitude — 1999 vs. 2999 scores 0 exactly like 1999 vs. 2000, because two different publication years are two different papers, not a near-match the way two head-counts might be (valueSimilarity's dedicated 'year' branch, checked before the number branch's relative-closeness scoring)
assign.ts	Hungarian max-weight assignment. Greedy is not merely worse but wrong here: one locally good pair can force two later entries into a much worse one, and greedy cannot trade the first against the second
align.ts	The recursion. alignNode returns slots per repeatable node; alignableNodes lists what is worth doing
apply.ts	Writes an alignment into the data
unanimous.ts	Finds the fields every reviewer answered identically, for adoptUnanimousValues to fill. Owns comparable() — the one rule for "did they say the same answer", shared with disagreements.ts and the compare popup so the three cannot drift into different verdicts
disagreements.ts	The per-field cross-reviewer verdict (FieldVerdict): who answered, which category their answer falls in, whether that is agreement. Boolean fields use a different answeredBy test (!== undefined && !== null instead of !isUnanswered) so a present false counts as a real answer. What both the overview and the statistics read
metrics.ts	Cohen's κ, Fleiss' κ, Krippendorff's α over abstract units × raters, each with an applicability check that explains refusal in a sentence. Knows nothing about papers or schemas, so it can be checked against published worked examples — and has been, including α with missing data
agreement.ts	Turns projectVerdicts into a MetricInput. Booleans are first-class answers (see below). Also produces perField: FieldAgreement[] — units bucketed by canonical field path, each with its own MetricInput for the per-field breakdown table
ConsolidationVerdicts (src/components/ConsolidationVerdicts.ts)	consolidationFieldStatus(answeredCount, reviewerCount, agree) → 'agree' / 'disagree' / undefined. Computed once by AnnotationPanel via useMemo, shared to all Field components via ConsolidationVerdictsContext (a ReadonlyMap keyed by canonical field path); consumed by Field.tsx via useConsolidationFieldStatus(canonical)

Matching cannot cross, which is a requirement of the feature: a group's sub-entries are only ever matched inside an already-matched pair of parents, because the recursion never offers a candidate from another group. It is structural, not a rule applied afterwards.

The mapping is stored as the ordering itself — there is no mapping field in the file. Every reviewer's entries are reordered so position N means the same entry for everyone, and the consolidated tree is grown to one entry per slot (the feature's "add the maximum number automatically" rule). This is why pruneTree keeps interior gaps and drops only trailing empties: a reviewer with no entry for slot 2 holds an empty one there, and closing that gap would slide every later entry down a slot and silently re-point the alignment on the next load.

A slot is scored per member, not summed over them. As reviewers are matched onto slots in turn, each candidate is compared against everyone already in the slot — and combine averages the score but adds the weights up, while the assignment maximises score × weight. Summed, a slot two reviewers had already landed in scored about twice one holding a single reviewer, and could outbid it on headcount alone. simAgainstSlot therefore divides by the member count, which makes slots comparable however full they are.

This needs three or more reviewers and one of them recording fewer entries than the anchor: with two reviewers every slot holds exactly one member and the bias cancels exactly, which is why it went unnoticed. The symptom was the one thing the feature exists to prevent — an exact match pulled into a crowded slot while the anchor's identical entry sat alone reporting agreement 0, showing two reviewers as disagreeing about an answer they had both given the same way. Note that a project consolidated under the old scoring may re-permute its reviewer entries (and go dirty) the first time it is opened after the fix; the new alignment is the correct one, but the change is not announced.

It runs a node at a time, off the paint path. Matching is not cheap — a large paper measures in the hundreds of milliseconds — so the hook yields to the browser between nodes rather than freezing the window as it opens. Whatever the reviewer opens the ⇄ compare popup on jumps the queue. alignConsolidationNode(paperId, nodeName, coalesce) is the store action; coalesce folds later nodes into the undo entry the run's first node pushed, so lining a paper up is one undo press.

A node the consolidator has already answered is never re-matched. Their entry N means a particular thing to them; re-matching could move a different entry into slot N, leaving their recorded answer describing something it was never about. The cost is that entries added after consolidation began are not auto-matched for that node — the safe side of the trade, since a stale match is visible and a silently re-pointed answer is not.

Unanimous answers are adopted (adoptUnanimousValues)

Once the matching for a paper is done, the scheduler runs adoptUnanimousValues, which fills every still-unanswered consolidated field on which all reviewers gave the same answer, and marks each one via aiMarks — the same light-blue border an AI fill gets, meaning "the app did this, you have not looked at it yet". There is nothing to reconcile where everyone already agrees, and copying those across by hand is the kind of task done on autopilot, which is when the real disagreement two rows down gets missed.

It runs after matching, necessarily: it reads every reviewer at a fixed index, which only means anything once their entries line up. The rules:

• Comparison folds case and whitespace only ("Controlled experiment" == "controlled experiment "). Not punctuation, and no fuzzy matching — the matcher's job is to pair entries, where near-enough is right; writing a value into the shipping tree unasked is a higher bar. The lowest-numbered reviewer's wording is kept, trimmed, so the choice is deterministic.
• Every reviewer must have answered. Two agreeing and a third blank is not unanimous — silence is not assent. This is also what keeps booleans honest: isUnanswered (src/llm/fields.ts) counts a boolean as answered only once ticked, so untouched checkboxes — which all read false — do not count as a unanimous "no" and mark every checkbox in the project.
• A field the consolidator already answered is left alone, per field (unlike the matching guard, which is per node).

Batch-adopting across the whole project (adoptAllUnanimousAnnotations)

adoptUnanimousValues/alignConsolidationNode above are per-paper, driven by whichever paper is open. adoptAllUnanimousAnnotations() (store.ts) is the multi-paper version for an ordinary (non-screening) multi-reviewer project's Consolidation seat: one button, one pass over every paper.

For each paper it aligns every alignable node, then adopts unanimous values — in that order, and the order is load-bearing. Adopting at a fixed index without aligning first would read reviewer 1's Finding #2 against reviewer 2's Finding #3 whenever the two recorded their findings in a different order, and report that as agreement neither reviewer actually gave. This is exactly why the existing adoptAllUnanimousScreening (see "Screening" below) never needed an alignment step: screening's Decision/Reason are both plain, non-repeatable leaf fields, so alignableNodes already returns [] for it and reading at a fixed index was always safe there. An ordinary schema's repeatable groups are exactly the case that safety doesn't extend to, which is what this feature adds the missing align step for.

• A paper the consolidator has already answered under any alignable node is skipped whole — no alignment, no adoption, nothing touched — via the same consolidatorHasAnswered predicate alignConsolidationNode's own per-node guard uses, now factored out into src/consolidate/readiness.ts alongside readyToConsolidate (previously inlined in store.ts's alignConsolidationNode). Re-matching a node the consolidator has already recorded an answer under could move a different entry into the slot their answer describes; skipping the whole paper is the safe side of that trade.
• One undo step for the whole batch, not one per paper — the same coalescing shape alignConsolidationNode's own coalesce parameter already provides per paper, threaded across the outer loop here too.
• It yields to the browser between papers (await a setTimeout(resolve, 0)), the same off-the-paint-path shape useConsolidationAlignment already uses between nodes — a paper is the smallest unit that can safely yield here, since every one of its nodes must finish aligning before any value can be read across it.
• Progress is reported as it runs: unanimousRun ({ done, total, filled, skipped, running }) drives the button's own label while busy and a dismissible result banner once it finishes. The button lives in ConsolidationOverview.tsx's project-wide overview, alongside ⚖ Agreement.

A known, currently unfixed limitation, stated plainly rather than glossed over: the interactive per-paper path (useConsolidationAlignment, driving an ordinary Consolidation-seat paper view) does not have this batch's paper-level guard. It still calls adoptUnanimousValues unconditionally the moment its alignment queue drains, with no check for whether any node in that queue was itself guard-refused along the way (a node the consolidator had already answered, per the per-node guard above). So a paper the consolidator has partially answered — one node done, a different node still open, perhaps because a reviewer added a new entry after consolidation began — can still have adoptUnanimousValues read that still-unaligned node at a stale index and fabricate an agreement nobody gave, through the ordinary act of just opening the paper. The batch path above avoids this not by fixing the interactive path, but by being strictly more conservative: it refuses to touch a paper at all once any alignable node on it is already answered, which the interactive path has no equivalent of before its own final adoption call. This is not covered by a regression test; fixing it is future work, not something this feature did.

Picking a seat (ReviewerPrompt)

A multi-reviewer project opens with currentReviewer === null — never Reviewer 1, since an unattributed edit is worse than a prompt. ReviewerPrompt renders over the app in exactly that state: it explains what multi-review means (independent annotation, one reconciling pass) and makes the reviewer choose. There is no dismiss — the form is withheld without a seat anyway, so a dismiss button would only offer a state in which nothing can be done — but it yields to Help (helpOpen), which is otherwise unreachable behind it.

Because the selection is persisted per project, "prompt shows" means "first open of this file". The exception is a project with no stable path — a ?project= URL, or a browser download-only save — where reviewerStorageKey deliberately persists nothing rather than risk restoring a seat under the wrong project. Those ask on every load. That is honest (the app genuinely cannot tell who you are) but it is the one case where the prompt is not a once-per-project event.

Seat selection is a local view switch, not project data: it is persisted per project in localStorage only, never written to the project file, and never compared against anyone else's claim — so the copy explains that the selection is remembered only on this machine for this project.

Readiness: what Consolidation can act on (readiness.ts)

readyToConsolidate(schema, paper, reviewerCount) — every numbered reviewer has recorded something on this paper, by hasAnnotations (so a ticked box counts as work; an unticked one never does, since every boolean reads false whether or not anyone looked).

consolidatorHasAnswered(def, tree) lives here too — the predicate behind alignConsolidationNode's per-node "never re-match a node the consolidator already answered" guard, and the same predicate the batch adoption above uses per paper. One shared function rather than two independent checks is what keeps the per-node and per-paper guards from silently drifting apart on what "already answered" means.

The Consolidation seat itself is not gated — a consolidator may legitimately start on the papers that are ready while the rest are still being reviewed. The gate is per paper instead, and one rule drives both places it shows, so they cannot disagree:

• the paper list's dot (paperIsMarkedDone) reads "ready to consolidate" in that seat;
• a field's ⇄ compare button is disabled on a paper not every reviewer has annotated. That reviewer's column would otherwise render empty, which reads as "they found nothing here" when the truth is "they have not looked yet" — inviting a decision on evidence that does not exist.

Alignment and unanimous adoption already decline such papers on their own (alignConsolidationNode needs two reviewer trees; adoptUnanimousValues needs every reviewer to have answered), so no extra guard is needed there.

Agreement and the disagreement overview

The Consolidation seat's UI uses a layered navigation of modals:

ConsolidationOverview (project-wide, modal)
  ├─ "Adopt all unanimous" batch action + run-progress display  (moved here from AnnotationPanel)
  ├─ "Agreement" → AgreementDialog (overview restored on close)
  └─ Paper row click → DisagreementOverview (per-current-paper, modal)
       ├─ "Overview" button → back to ConsolidationOverview
       └─ Field click → ConsolidationDialog (field comparison popup)
            ├─ reviewer answers to pick from
            ├─ "Enter a different value" (defer)
            └─ on close → returns to DisagreementOverview (via returnToDisagreements flag)

AnnotationPanel now shows two buttons in the Consolidation seat: ☰ Overview (opens ConsolidationOverview) and ⚠ Disagreements (opens the per-paper DisagreementOverview). Navigation between modals uses return-to flags in the store (agreementReturnToOverview, disagreementsReturnToOverview, returnToDisagreements) so closing one modal restores the one that opened it.

ConsolidationOverview (src/components/ConsolidationOverview.tsx) lists all papers that have ≥1 disagreement (count per paper), filterable by text search. It houses the "Adopt all unanimous" batch action with its run-progress display and post-run summary notice. Its "Agreement" button opens AgreementDialog; clicking a paper row opens the per-paper DisagreementOverview.

⚖ Agreement (AgreementDialog) computes the coefficients the reviewer ticks. A unit is one annotation field on one paper; agreement.ts includes only the fields at least two reviewers answered, since a field with one answer carries no agreement information (disagreements.ts says so, and warns that callers must gate on answeredBy.length >= 2 rather than trust agree, which is vacuously true for a single answer). The dialog states the unit count and how many were skipped, so the reader can see what was measured rather than guess.

A metric that cannot honestly be computed is disabled, at half opacity, with its Applicability.reason verbatim on hover — those strings are written as complete user-facing sentences for exactly that ("Cohen's κ compares exactly two reviewers; this project has 3"). Cohen's needs exactly 2 raters; Fleiss' needs every reviewer to have rated every unit; Krippendorff's α survives both, which is why it is worth having all three.

The dialog also renders a per-field breakdown table (PerFieldTable) — each metric's coefficient bucketed by schema field, so incomparable category spaces (a Year field and free-text Claim do not share categories) are no longer pooled into one misleading number. A "Copy as TSV" button exports the table. An alignment warning banner appears when ≥2 reviewers recorded entries in a repeatable group that Consolidation hasn't aligned yet, with a "Line them up now" button that triggers adoptAllUnanimousAnnotations().

⚠ Disagreements (DisagreementOverview) lists every field where the answering reviewers gave different categories, filtered to the current paper only (it was project-wide before; the project-wide view now lives in ConsolidationOverview). A row jumps to it (selectPaper → openConsolidation), which is the point — finding a disagreement is useless if you then have to hunt for it.

Per-field verdict coloring (ConsolidationVerdicts)

src/components/ConsolidationVerdicts.ts provides a React context that gives every Field component its per-field consolidation status at a glance. AnnotationPanel computes a Map<string, ConsolidationFieldStatus> once per render (via consolidationFieldStatus()) and shares it through ConsolidationVerdictsContext. Field.tsx reads useConsolidationFieldStatus(canonical) and adds a consolidation-agree (green) or consolidation-disagree (red) CSS class — green when all reviewers answered and agree, red when ≥2 answered and disagree.

What the agreement numbers do and do not cover

The three coefficients themselves are verified against published worked examples, including Krippendorff's α with missing data. Two things about the input are worth knowing before quoting a number from this dialog.

Yes/no fields are now first-class answers. Previously boolean fields were excluded entirely from agreement statistics — the rationale was that an unticked box (false) is indistinguishable from "never looked at" — but this meant the only boolean units that survived the answeredBy.length >= 2 gate were guaranteed agreements (both ticked true), and every real boolean disagreement was silently discarded. The coefficient came out higher than the truth (κ 0.500 measured where the honest value was 0.000). Booleans are now treated as real answers: disagreements.ts' walk() uses a different answeredBy test for booleans — values[r] !== undefined && values[r] !== null instead of !isUnanswered(def, values[r]) — so a present false counts as a real answer, and a true/false split between two reviewers registers as a disagreement.

Repeatable-group entries are compared in stored order, not aligned order. This is the known limitation described under disagreements.ts, and the metrics inherit it: two reviewers who recorded the same findings in a different order score as disagreeing. Since this dialog computes over the whole project while alignment runs only on papers someone has opened in Consolidation, a coefficient from an un-consolidated project can be arbitrarily pessimistic — κ = −0.333 measured on a two-paper case whose true value is 1.0. Align the papers first. needsAlignment()/needsAlignmentCount() (readiness.ts) detect papers where ≥2 reviewers recorded entries in a repeatable group that Consolidation hasn't reviewed yet; AgreementDialog shows a warning banner with a "Line them up now" button that triggers adoptAllUnanimousAnnotations() to align + adopt.

All fields share one pooled category universe for the headline coefficient. agreementInput pools every field of every paper into one unit list, so p_e is computed over a merged marginal distribution mixing enum labels and free-text answers. Free text contributes near-unique categories, which drives p_e toward 0 and κ toward raw percent agreement: an enum field with honest κ 0.0 and a free-text field with honest κ 1.0 pool to 0.706. Screening escapes this — decisionOnly narrows to the Decision enum, and that path is sound. A per-field breakdown table (perField) mitigates this for the reader: each field gets its own coefficient (bucketed by canonical path joined with /), rendered in AgreementDialog's PerFieldTable, with a "Copy as TSV" button (perFieldTsv()). Cells show — when a metric isn't applicable, and blank when the coefficient is null (e.g. perfect agreement on one unit).

Semantic equality (Paper.equal)

Reviewers write the same thing differently ("RCT" / "randomized controlled trial"). Nothing captured that, so every statistic understated agreement. The compare popup carries a tick — "these answers mean the same thing" — persisted as Paper.equal, a list of canonical field paths. disagreements.ts then gives those reviewers one shared category, so the field reads as agreement everywhere: the badge, the overview, and the coefficients. Measured on a two-reviewer demo, marking one such pair moved Cohen's κ from 0.533 to 0.682.

The box only appears where there is something to declare — when the answers already match after comparable(), there is nothing to add.

The mark alone is not a resolution, and the popup will not let it pass for one. It settles that the reviewers agreed, which is enough to drop the field out of the disagreement list and count it as agreement — but it says nothing about what they agreed on, so the consolidated value stays blank. Marked-but-blank is the worst state available: resolved everywhere, holding no answer, and never surfaced again. closingWouldStrand catches it on every exit (×, Escape, backdrop, and taking a reviewer's blank answer, which records nothing and leaves the same hole). Leaving anyway un-marks the field, so it returns to the disagreement list rather than disappearing quietly. Escape backs out of that warning rather than through it — discarding the mark should take a deliberate click.

A boolean can never strand: isEmptyValue says a boolean is never empty, and rightly — an unticked box is a real false, not a gap, so there is no value left to pick and the warning would be one the reviewer could not clear.

Known limit, documented at the field: it is one boolean for the whole field, i.e. "all the answering reviewers here are equivalent". Exact for two reviewers; with three or more it cannot say "these two agree but that one does not".

AI is not available in Consolidation

AnnotationPanel does not render the ✦ AI button in this seat at all (not disabled, not the transparent treatment the locked state uses — absent), and applyAiSuggestions refuses when currentReviewer === 'consolidation', so opening the dialog as a reviewer and then switching seats cannot get round it. Consolidation decides between what the reviewers actually said; a model's answer would be a fresh opinion invented after the fact, written into the tree that ships and dressed as a reconciliation of the others.

Matching is lexical only. Bundling a local embedding model (nomic-embed-text and similar) was evaluated and rejected: ~185 MB installed at best and plausibly several hundred MB of RAM, against peer-reviewed evidence (VLDB 2023 entity-resolution, DeepMatcher, fuzzylink) that on short, structured strings — which annotation values are — good lexical methods match or beat embeddings, which win on long dirty text, paraphrase and multilingual instead. onnxruntime-node has also dropped Intel-Mac support. If semantic matching is ever wanted, the shape is an opt-in call to a local Ollama (/api/embed) layered on top of lexical scoring, not replacing it.

currentReviewer is a view selection, not project data: switching it is not an undo step and does not set dirty. It defaults to null (unselected) whenever a multi-reviewer project is opened — never silently to Reviewer 1, since an unattributed edit is worse than a prompt — unless a previous selection for this project was persisted. Persistence is per project, keyed by the save handle's path (slr.currentReviewer.<path> via safeGet/safeSet/safeRemove in src/state/settings.ts); a project with no stable path (server mode, or a browser download-only save) simply doesn't persist a selection. It resets to null on closeProject/loadFromText, same as validation/aiMarks.

Toolbar (src/components/Toolbar.tsx) renders a reviewer switch — buttons for Reviewer 1..N plus a visually distinct Consolidation pill — only when the open project is multi-reviewer; hidden entirely otherwise. The active seat is highlighted; an unselected state gets a warning border and a "Pick a reviewer:" prompt. Validate is additionally disabled until a reviewer is picked (see "Validation" above). Note this shares the toolbar with the hidden AI-unlock click gesture on the app title — the two are unrelated and don't interact.

The switch sits in a dedicated center track of the toolbar (.toolbar-left / .toolbar-center / .toolbar-right in index.css, a 3-column grid with 1fr auto 1fr) so it's centered on the toolbar itself rather than merely between its flanking groups — a plain margin: auto would drift as the project title (.toolbar-status, on the right) changes length. The two outer tracks are bare 1fr (not minmax(0, 1fr)), so each floors at its own min-content width and can never be squeezed into overlapping the center; only .toolbar-right is allowed to shrink further, because .project-name beneath it already truncates with an ellipsis. Under real space pressure the row simply asks for more width than the window has (or the title truncates harder), never overlapping text.

Above REVIEWER_DROPDOWN_THRESHOLD (5) reviewers, the pill row would crowd the toolbar, so the same choice renders as a Dropdown (reusing src/components/Dropdown.tsx, extended with an optional className for the warning/Consolidation styling hooks) instead of one button per reviewer. The closed trigger always names the active seat ("Reviewer 3", "Consolidation", or "Pick a reviewer" while unselected) rather than a bare caret — the whole point of the switch is that the active seat reads at a glance without opening anything. The open menu marks the current selection with a checkmark and keeps Consolidation visually distinct (its own label styling, matching the pill form's colors) rather than listing it as "reviewer N+1". At 5 or fewer reviewers the pill row is unchanged.

AnnotationPanel withholds the annotation form (showing a prompt instead) whenever a multi-reviewer project has no reviewer selected, and otherwise renders the tree currentTree() routes to for the active reviewer — falling back to a schema-shaped empty tree (not creating anything) when that reviewer hasn't written on this paper yet. A small badge next to the paper title echoes which seat is active, redundant with the toolbar switch on purpose: this is the one piece of state that must never be ambiguous, since it decides which tree every edit lands in.

Consolidation compare popup. When currentReviewer === 'consolidation', Field.tsx shows an extra ⇄ button next to the existing ⧉ grab-from-PDF button (and, for a boolean field, next to the checkbox) that opens ConsolidationDialog for that exact field path. The dialog lists every reviewer's raw value for the path (via the same peekValue/fieldPath machinery store.ts already uses, and resolvePath from src/llm/paths.ts to resolve the ResolvedDef for display), including reviewers who left it empty, and flags whether the answered reviewers agree. Picking a reviewer's value calls resolveConsolidationValue — which writes the value, marks the field equal, and clears any deferral in one undo step. Taking a reviewer's blank answer calls deferConsolidationValue instead of writing an empty value — the field stays marked as "pending a different value" rather than silently leaving a hole. An "Enter a different value" button defers and closes the dialog, letting the consolidator type a custom answer directly in the annotation panel; a deferred field gets a consolidation-pending CSS class, and entering a non-empty value via setFieldValue auto-clears the deferral and adds the field to paper.equal. openConsolidation accepts a returnToDisagreements flag so closeConsolidation reopens the per-paper disagreement list when the dialog was opened from it. Closing without picking changes nothing.

AI marks and AI-assisted annotation are reviewer-scoped too — see "AI marks" below and unansweredFields's call site in aiStore.ts's openDialog(), which now proposes values for the active reviewer's empty fields via currentTree(), not unconditionally paper.annotations.

PaperList follows the same rule on both of its reads: the "annotated" dot and the annotation search mode each go through currentTree(), so as Reviewer 2 the sidebar tracks your progress and answers "which papers did I record this in" — the same tree the form and validation show. When the project is multi-reviewer and nobody has picked a seat yet, currentTree() returns null and both degrade to "nothing annotated / nothing to match" rather than falling back to the consolidated tree, which would attribute someone else's work to the unselected reviewer.

Screening

A screening project is an ordinary project whose schema is derived rather than authored. config.screening: { reasons: [...] }'s presence is the opt-in marker and the single source of truth for the one authorable thing (the exclusion-reason list); config.schema is a projection of it — src/screening/schema.ts's screeningSchemaDefs — re-derived on every loadProject and rewritten on every serializeProject, never read back for a screening project. That one decision is what makes almost everything else in the codebase need zero changes: to resolveSchema, normalizeTree, pruneTree, validate.ts, align.ts, unanimous.ts, disagreements.ts, metrics.ts, readiness.ts, currentTree, and undo/redo, a screening project is simply a project with a two-node schema.

Rejected: config.schema authoritative, with a sync check against config.screening.reasons. Two sources of truth for one thing, and hand-editing reasons would silently desync the dropdown from the actual protocol — a sync check only catches the drift after the fact, it doesn't prevent it. Rejected: omitting schema from the in-memory Project for a screening project. Every downstream module reads project.schema; a screening project having a real one is exactly what lets them all work unmodified. Drift between file and in-memory state is structurally impossible here because the derived value always wins on load and is what gets written on save.

Why Decision is a two-option enum, not a boolean

The obvious reading of "a boolean field if the paper should be excluded" is an Exclude checkbox. This codebase cannot represent an unanswered boolean, and says so in three separate places: isEmptyValue in src/model/validate.ts ("booleans are NEVER empty" — an unticked box is a real, answered false); isUnanswered in src/llm/fields.ts (same rule, for the AI layer); and hasAnnotations in src/model/annotations.ts, which only counts a boolean once it is true. An Exclude checkbox therefore cannot tell "I decided to include this" apart from "I have not looked at this yet" — both are false. Screening is the one phase where that distinction is the output: the progress count, the PRISMA include/exclude/pending numbers, and — above all — which papers survive an import (startFromScreening) all depend on it.

A checkbox would also have broken every piece of machinery this feature exists to reuse, each in a different way: hasAnnotations would call an included paper unannotated (so its progress dot would never light); readyToConsolidate would say a paper both reviewers included is not ready to consolidate; unanimousFills would refuse to adopt a unanimous "include" (isUnanswered treating false as unanswered-or-not is exactly the ambiguity that breaks); and κ would only ever see the excluded papers, since an included paper's category would be indistinguishable from an untouched one. Each of those would need a screening-specific special case — precisely the parallel machinery this design set out to avoid. Spelling the field as Decision: "Include" | "Exclude" (schema in src/screening/schema.ts) gets the tri-state (null until chosen) for free, and every one of those modules is then correct with zero code changes — see src/screening/status.ts's screeningStatus, the one place that reads the field.

The polarity the user's brief asked for is kept where it matters: Reason belongs to the exclusion side, and only an explicit Decision === "Exclude" ever drops a paper on import — see below.

align.ts is a structural no-op here, and needs no guard

alignableNodes(schema) (src/consolidate/align.ts) returns [] for the screening schema: its hasAnythingToMatch predicate is isRepeatable(def) || def.children.some(...), and both Decision and Reason are plain max: 1 leaf fields with no children. useConsolidationAlignment's queue is therefore always empty for a screening project, and it goes straight to adoptUnanimousValues — no special-casing added, none needed. This falls out of the schema shape, not of any screening-aware code; if a future version ever made Reason repeatable this reasoning would need re-checking, but nothing here defends against that on purpose.

Reused verbatim, and the one thing that had to be scoped

readiness.ts, align.ts, unanimous.ts, disagreements.ts, metrics.ts, apply.ts, assign.ts, similarity.ts — zero changes. Two reviewers screening independently with a reconciliation pass is the standard SLR screening protocol, and κ over the include/exclude decision is exactly the statistic a screening phase reports; building anything parallel to the existing multi-reviewer machinery for that would have been indefensible.

The one scoped change is agreement.ts's agreementInput: for a screening project it filters projectVerdicts down to Decision only, before the answeredBy.length < 2 skip accounting. Without it, Reason — a field only even defined on the subset of papers both reviewers excluded — would be folded into the same coefficient as a different question, producing a number that answers neither honestly. Reason verdicts are filtered out entirely rather than counted as "skipped" (which means "too few reviewers answered" — not true here).

ConsolidationDialog gets one guard too: the "these answers mean the same thing" checkbox is suppressed on the screening Decision field (isScreeningDecision, checked against SCREENING_DECISION and an empty container path). "Include and Exclude mean the same thing" is not a claim anyone can make; without the guard a consolidator could tick it and make a real disagreement vanish from the overview while inflating agreement. The box stays available on Reason, where two overlapping reasons genuinely can be equivalent — the same reasoning closingWouldStrand already applies to ordinary fields.

AI is not offered for screening, on the same precedent as Consolidation

ScreeningPanel renders no AI button at all (mirroring AnnotationPanel's Consolidation-seat treatment — absent, not disabled), and applyAiSuggestions (store.ts) refuses whenever project.screening !== null, so opening the dialog on a different project and switching cannot get round it either. Screening decides the review's corpus; a model's include/exclude pass would be the difference between a systematic review and a generated one. It is also mechanically moot most of the time: the ordinary AI path needs a PDF (aiDisabled = … || !paper.pdf), and screening papers usually have none (pdf: "" — see below).

The middle pane swaps, it does not nest

App.tsx renders ScreeningRecord (title + abstract) in place of PdfViewer by default for a screening project, toggled to the actual PDF via a session-only screeningShowPdf flag (toggleScreeningPdf). Rejected: nesting PdfViewer inside a record component. PdfViewer owns its own panel, header, ResizeObserver, and in-PDF search state — fiddly to embed for no gain when a flat swap does the same job. Rejected: two conditional layouts keyed off paper.pdf === ''. A swap reuses PdfViewer whole (including its own now-necessary empty state for pdf === '' — trap below) and degrades cleanly whether or not a PDF is attached, without a second code path to keep in sync.

PdfViewer's missing empty state was a latent gap this feature makes reachable. Before screening, paperSchema.pdf was z.string().min(1), so pdf: '' could never survive loadProject and PdfViewer never needed to render anything for it — it only guarded !paperId. Screening relaxes pdf to z.string().default(''), scoped by a superRefine that still requires a non-empty pdf for every non-screening paper (src/model/schema.ts), so the relaxation cannot leak into ordinary projects. PdfViewer now also guards !pdfPath with an explicit "This paper has no PDF attached" panel, right next to the pre-existing !paperId guard.

A missing abstract is extracted from the PDF, and flagged durably

Screening is normally decided from abstract alone, but a paper added straight from a PDF (rather than through a reference-manager export, which usually carries one already — see references.ts's RefEntry.abstract) may have none. pdfMeta.ts's abstractFromLines fills that gap with the same kind of best-effort layout heuristic the module already used for title/authors: find the "Abstract" heading and follow the column it sits in to that column's next section heading (see the function's own comment for why column-following, not line-reading, is the whole problem). Two call sites, one function:

• editorStore.ts's addPickedPdfs — the existing background title/author fill for a directly added PDF now also tries the abstract, pre-filling the draft row the same way (never clobbering something the reviewer already typed while the read was in flight).
• store.ts's extractScreeningAbstract(paperId), fired by selectPaper — and by loadFromText for the paper a project opens on, which nothing else would ever fire for. Selection is the trigger because the abstract is the screening surface: ScreeningRecord must simply have one to show, without the reviewer opening the PDF to make it appear. An earlier version hung this off toggleScreeningPdf instead, which inverted the feature — it only produced an abstract once you had already opened the PDF you were trying to avoid opening. Auto-advance (setScreeningDecision) routes through selectPaper, so the read-decide-read rhythm gets this for free with no second call site.

Why this writes abstract immediately, with no per-paper confirmation step, unlike AI suggestions. The AI-annotation flow (applyAiSuggestions, below) always gates a machine-produced value behind an explicit reviewer "Apply" click on a reviewed table. Screening's whole design optimizes for the opposite: hundreds of papers, seconds each, one keystroke per decision (I/E/1-9, see "Auto-advance" above) — a confirmation dialog per PDF-opened paper would defeat exactly the throughput this feature exists for. The chosen substitute for that review gate is a permanent, unmissable warning rather than a one-time confirmation: Paper.abstractFromPdf is written to the file (project.ts, alongside abstract), not held in a session-only structure like aiMarks — a co-reviewer opening the same file later, or the same reviewer in a future session, must see the same "this is a guess, verify against the PDF" caution the extracting session did, which a per-session mark cannot promise. ScreeningRecord.tsx renders that warning inline, above the abstract text, whenever the flag is set — text, not just a colored border, since a claim this consequential to a systematic review's corpus needs to survive a glance. This mirrors the precedent Paper.aiUsage already set: a durable disclosure, not a UI hint, for exactly this reason (see "AI-assisted annotation" below).

The flag is meaningless without a value to describe it, so it never survives without one. loadProject drops abstractFromPdf whenever abstract itself is empty (p.abstract && p.abstractFromPdf === true, project.ts) — a hand-edited file that deletes the abstract but leaves the flag behind gets the defensive treatment every other structurally-validated field in this loader gets, not trust.

A heuristic abstract is explicitly lower-confidence than one actually recorded somewhere. Every other field fillFromRef (editorStore.ts) considers is "fill only if currently blank, never overwrite" — the reviewer's own input always wins. abstract is the one exception: a later reference-manager import is allowed to replace an abstractFromPdf-flagged value with the entry's real one, clearing the flag in the same write. This is a real answer to a real ordering problem — a PDF added first, a reference file imported second — not an oversight in the general rule.

Staleness here is about the project being gone, not about anything having changed — and loadFromText's get().project === project reference check is the wrong tool for that, despite being right for its own narrower job. immer hands back a new project object on every edit, so a reference comparison reads one screening decision as "the project was replaced". During screening that is not an edge case, it is the loop: decide a paper every second or two, while a PDF read is in flight. Guarding on identity threw away a perfectly good abstract because someone pressed I — and then marked the PDF as having none, so it never retried. projectGeneration (a module-level counter, bumped only by loadFromText and closeProject) answers the question actually being asked. The write additionally re-checks, inside the producer, that the paper still has no abstract, so a hand-typed one that landed first is never clobbered. What it deliberately does not check is the selection: the abstract belongs to paperId, so a late result is still written for the paper it was read from rather than discarded because the reviewer scrolled on.

It pushes no undo step, unlike adoptUnanimousValues — and the difference is who asked. Adopting unanimous values is an action a consolidator invokes; this is a passive background fill triggered merely by looking at a paper. On the undo stack it would mean Ctrl+Z after a decision silently removes an abstract instead of undoing that decision, and with auto-advance firing an extraction per paper the stack would fill with them. It follows the editor's own background title/author fill (addPickedPdfs), which likewise patches rows with no undo entry of its own. dirty is still set, so the ordinary unsaved-changes path persists it. A corollary the screeningAbstractReads bookkeeping has to respect: an undo can still restore a snapshot taken before an abstract landed, so a successful read deliberately leaves no marker behind — re-selecting that paper simply extracts again, rather than the loss being permanent for the session.

The two-column layout is the whole difficulty, and hand-built test fixtures could not see it. pdf.js reports a two-column paper's left-column "Abstract" heading and its right-column "1 Introduction" on the same baseline, so toLines yields one Line reading "Abstract 1Introduction", and every body line below is likewise one Line holding a strip of each column. The first implementation read line.text and stopped at the first multi-segment line — which on a real paper is the line immediately after the heading — so it extracted nothing at all from exactly the documents the feature exists for, while every synthetic unit test passed, because those fixtures were built from the same wrong mental model as the code. Following the column requires each segment's x, which toLines already computed and then discarded; Segment now carries it. The lesson is pinned in place by pdfMeta.test.ts's real-PDF integration tests (against samples/pdfs/, including a genuine two-column ICSE paper) and by editorStore.realPdf.test.ts for the composition — a class of bug no amount of hand-built Line[] could have caught.

Rejected: treating the extracted abstract as session-only until confirmed, never writing it to the file. This was the first design considered, closer to aiMarks' "shown but not yet real" treatment. It fails the actual requirement: a co-reviewer must see the same paper's abstract (and the same caution about it) the extracting reviewer saw, which a value that lives only in one person's browser session cannot deliver on a shared, git-tracked project file — and re-extracting on every single open, for every reviewer, of every paper, is wasted work abstractFromPdf avoids for free once the first extraction lands.

The paper-list dot's meaning changes in Consolidation, and the mitigation

A screening project's paper list gets a tri-state marker (undecided / included / excluded, via paperScreeningStatus in PaperList.tsx) instead of the plain done/not-done dot — a single boolean dot is nearly meaningless once every paper's whole state is one field. The rule is uniform across every seat: it reads whatever currentTree routes to, same as paperIsMarkedDone always has. In the Consolidation seat, currentTree is paper.annotations, so the dot reads as "the final decision so far" — which is the right thing for that seat to show.

The cost, stated plainly: in Consolidation this dot stops meaning readyToConsolidate — the signal paperIsMarkedDone computes there for an ordinary project (every numbered reviewer has recorded something). Losing that signal outright would be a regression, since it is what tells the consolidator which papers are actually workable. The mitigation is that it has not actually been lost, only moved: readyToConsolidate still runs and is reported in the marker's title tooltip, and — this is the part that matters — the ⇄ compare button's own readiness gate in Field.tsx is completely untouched. That gate is the thing that actually protects a consolidator from deciding based on an absent reviewer's empty column; the paper-list dot was always a secondary, at-a-glance cue on top of it, never the enforcement mechanism itself.

Auto-advance: one rule, forward-only

Deciding a paper (setScreeningDecision in store.ts) advances the active seat to the next undecided paper only when this seat's own decision went from undecided to decided — computed by reading screeningStatus before the mutation, not after. Re-deciding an already-decided paper, or un-deciding one (setScreeningDecision(null)), never advances. This is deliberately the only rule, applying identically to the keyboard shortcuts and the panel's own Include/Exclude buttons (both call the same store action), so the two cannot drift apart on it: a reviewer moving forward through a stack gets a read-decide-read rhythm, and a reviewer who jumps back to fix an earlier call is never yanked away from the paper they just navigated to.

Importing from a screening project

startFromScreening / importFromScreening (editorStore.ts) read a screening project through the real loadProject — not a raw JSON parse — specifically so that paper.annotations (the consolidated tree in the multi-reviewer case) is what "included" is read against, never an individual reviewer's own opinion in paper.reviews. partitionScreeningPapers buckets every paper by screeningStatus(paper.annotations); only 'excluded' is ever dropped. 'undecided' covers both "genuinely never touched" and "a hand-edited file holds an unrecognised decision string" — screeningStatus treats both identically, on purpose (see src/screening/status.ts): carrying a paper you meant to drop is recoverable (it shows up, flagged, in the new project); dropping a paper you meant to keep is invisible and silently corrupts the review. The import dialog (ScreeningImportDialog) states the three counts and lets the reviewer choose to leave the undecided ones out instead, but never drops them without that explicit choice.

"New from screening…" can start a second screening pass, not just an annotation project. A radio in ScreeningImportDialog (defaulting to annotation, the original behavior) sets startKind: 'annotation' | 'screening' on the import draft (setScreeningImportKind); resolveScreeningImport reads it back when the reviewer confirms. Choosing screening used to be refused outright — a screening-to-screening import had nowhere sensible to go, and a PDF-less screening paper carried into an annotation project simply couldn't be saved, since only a screening project's paperSchema tolerates pdf: ''. Building a second screening project fixes both: PRISMA's title/abstract pass and full-text pass are two rounds of the same protocol, so the destination is a screening project again, seeded with the source project's own screening.reasons — never DEFAULT_SCREENING_REASONS — because a full-text pass needs to report against the same reason vocabulary the first pass used, not a fresh generic one. It also inherits the source's reviewers count and suggests a -fulltext.json filename, where the annotation target keeps reviewers: 1 and suggests -annotation.json. Either way, every carried paper's annotations (and, for the screening target, reviews) starts empty — a first-pass title/abstract decision must not silently anchor the second pass's full-text one, whether that second pass is an annotation schema or another screening round.

The import records where the new project came from, and the project editor shows it. Project.provenance: ProjectProvenance | null (src/model/project.ts) —

interface ProjectProvenance {
  kind: 'screening-import'
  source: { title?: string; file: string }
  importedAt: string  // ISO 8601
  counts: { included: number; undecided: number; excluded: number; carried: number }
}

It is a first-class field on Project, not tucked under config or extra: config's zod schema has no .passthrough(), so an unrecognized key nested under it is silently stripped before project.ts ever sees it, and extra is reset to {} by a "New from screening…" import in any case — neither is a place a value can be trusted to survive. parseProvenance reads it defensively (malformed input becomes null, never a thrown error), and both serializeProject and the editor's own buildProjectJson write it, so a draft never in-memory-only holds provenance the eventual save would drop. Because a nested record like this has no FieldConflict shape, mergeProjects refuses the whole merge if both sides changed it differently (see "Merging two copies of a project" in data-model.md), and detectFieldChanges treats any provenance difference as structural, falling back to the plain whole-file commit checkbox — the same treatment config.schema and friends already get, for the same reason: there is no field-level answer to "which import history is right."

Provenance is now surfaced in the project editor as a read-only "Imported from" line (ProvenanceNote in ProjectEditor.tsx), showing the source file/title, the import date, and the carried/dropped counts — previously written to the file but displayed nowhere.

resolveScreeningImport also has to solve a real path-integrity problem, not just a metadata carry-over: a paper's pdf in the screening file is relative to that file's directory, so a carried row needs a real absolute sourcePath, or the moment the reviewer later uses Change… on the new project, changeLocation (which only re-derives pdf for rows that have a sourcePath — see the project-editor section above) would silently leave every PDF pointing at nothing. PlatformAdapter.absolutePdfPaths exists for exactly this (the inverse of relativePdfPaths); siblingProjectLocation is the platform op that makes "save the new project next to the screening JSON" the default rather than a suggestion in a dialog — a sibling location is what keeps every relative pdf path resolving with zero rewriting at all.

For target:'import' (merging carried papers into an already-open session rather than starting a new project), the path-integrity problem is sharper: the open session can live in any directory, almost certainly not the screening project's own. resolveScreeningImport now rebases each carried paper's pdf against the open project's location via relativePdfPaths — the same pattern changeLocation uses for "Save as" — rather than copying the verbatim relative path from the source, which would point at nothing the moment the reviewer looked. (target:'start' doesn't need this because it defaults to a sibling of the source.)

pendingUnanimous (src/screening/counts.ts) exists because adoptUnanimousValues only ever runs for the paper currently open in the Consolidation seat (useConsolidationAlignment is driven by currentPaperId) — so a multi-reviewer screening project whose consolidator never happened to open paper X has no consolidated Decision for X even though every reviewer agreed on it. The import deliberately does not auto-consolidate to paper over that gap — writing decisions nobody actually recorded, from an import dialog, would be worse than surfacing the gap. Instead the import dialog reports the count and points at adoptAllUnanimousScreening (one click, one undo step, Consolidation seat only) as the fix.

There are two counts, deliberately, because there are two questions. pendingUnanimous counts papers with any pending unanimous fill — a decision, a reason, or both — and drives the notice sitting beside the Adopt all button in ScreeningPanel, so the notice and the button it offers cannot disagree. Counting only decisions there meant two reviewers who both excluded a paper for the same reason, where the consolidator set the decision by hand and left the reason blank, produced no notice and nothing offering to adopt the reason — and the paper booked as excluded-without-a-reason permanently.

pendingUnanimousDecisions counts only papers that are still undecided, and is what the screening-import dialog reads. That dialog speaks specifically about "the not-yet-screened papers" and about the project having no final decision for them, so a paper that is already decided and merely lacks a unanimous reason must not be counted there: the sentence would promise that adopting changes an inclusion count which is in fact already settled.

AI-assisted annotation (src/llm)

A ✦ AI button in the annotation column's header asks an LLM to read the current paper and propose values for the fields that are still empty. The reviewer gets a table — field, proposed value, the supporting quote from the paper, the model's confidence, and a checkbox per row — and nothing is written until they press Apply.

Availability is gated twice, and defaults to off

The button being clickable requires both of two independent things to be true — Project.aiEnabled && useStore().aiUnlocked — and either one being false disables it identically:

1. Project.aiEnabled (config.ai in the file, default false for new projects) — the project's say. A provider of a project file can set config.ai: false to forbid AI use on that file outright; see docs/annotation-schema.md. New projects created in the editor now default to aiEnabled: false (writing config.ai: false into the file), and the AI-annotation checkbox has been removed from the project editor — with no reachable entry point for the feature itself (aiUnlocked below), a control that configures it promises something the app doesn't currently deliver. An existing file's own config.ai setting is still read and preserved on save.
2. useStore().aiUnlocked (default false, in-memory only, never persisted) — the app's say, for now: AI-assisted annotation ships in the app but is off by default regardless of what config.ai says. config.ai: true (or omitting it) is necessary but no longer sufficient. The only way to flip this for the running session is the hidden gesture in Toolbar.tsx: UNLOCK_CLICK_COUNT (12) clicks on the "SaiLoR" title within UNLOCK_CLICK_WINDOW_MS (2500ms) of each other, counted by the pure nextTitleClickState() (deliberately kept out of the component and unit-tested in Toolbar.test.ts, since the "N clicks within a window, else the run resets" rule is exactly the kind of off-by-one/timing logic worth pinning down). The title itself carries no title attribute, no pointer cursor, and no other affordance hinting that it does anything — nothing changes even mid-run, since the click count lives in a useRef specifically so counting does not trigger a re-render. aiStore.openDialog() re-checks both flags as a second line of defense in case the dialog is ever reached another way; the disabled button is the primary gate.

Whenever the button is disabled — for any reason (busy, no PDF, aiEnabled false, or not aiUnlocked) — its tooltip is the uninformative "Coming soon" rather than a reason, so the button reads as an ordinary disabled control rather than one hinting it can be unlocked. Reaching for a specific one of these reasons in a new codepath is a sign the tooltip logic needs revisiting, not extending — see AnnotationPanel.tsx.

Not-yet-unlocked (!aiUnlocked) goes one step further than the other disabled reasons: the button is fully invisible, not merely dimmed. AnnotationPanel.tsx computes aiHidden = !aiUnlocked separately from aiDisabled, and applies an ai-btn-hidden class (opacity: 0 in index.css, overriding the ordinary .ai-btn:disabled dimming via the higher-specificity .ai-btn:disabled.ai-btn-hidden selector) plus aria-hidden="true" and a title of undefined — no visual presence, no accessibility-tree presence, no tooltip, nothing for a reviewer who hasn't found the click gesture to notice. The button is still in the DOM at opacity: 0 (not display: none), so unlocking mid-session doesn't shift the header's layout; it is also already disabled in this state (aiHidden implies aiDisabled), so it was never actually clickable regardless of visibility. Once unlocked, the other disabled reasons — busy, no PDF, and in particular a project that explicitly sets config.ai: false — go back to the ordinary dimmed-but-visible treatment: a reviewer who already knows the feature exists benefits from seeing that this specific project has turned it off, which hiding the button again would only obscure.

If a later product decision makes AI available by default again, aiUnlocked and the whole gesture can simply be deleted and Project.aiEnabled alone regains control — the two checks are additive (both must hold), not layered logic that needs untangling.

The modules

Module	Job
src/llm/types.ts	The shared shapes: LlmConfig (one configured target — provider, base URL, model, attach, reasoningEffort?), ModelInfo/ReasoningProfile/ModelsPage (the model-listing shapes, see below), LlmHttpRequest/LlmHttpResponse, Suggestion, SkippedField, RejectedSuggestion, and the API_KEY_SENTINEL ({{apiKey}}).
src/llm/fields.ts	unansweredFields(schema, tree) — every field the AI will be asked about, in schema order. Its isUnanswered() is validate.ts's notion of empty for strings/numbers; a boolean is offered unless it is already ticked, since the data model cannot represent an unanswered boolean and the archetypal AI field ("Relevant") is one.
src/llm/prompt.ts	The system prompt: the schema format description (SCHEMA_FORMAT_DOC), the path syntax, the schema itself (dehydrateSchema, i.e. exactly as it appears on disk), one line per field to fill, the rules, and the output shape. Plus buildUserText / buildUserPdfCaption for the user turn.
src/llm/paths.ts	The path language of the LLM contract — "Findings[1]/Evidence[0]/Metric". parsePath / formatPath / displayPath, and resolvePath(schema, raw), which checks a path against the schema (not the current data), so the model may name the next free index of a repeatable node to record a further entry.
src/llm/providers.ts	Everything that differs per vendor for the annotation call: PROVIDERS (base URL, whether the URL is editable, whether the provider can take a PDF, which output-length parameter it wants — see below), buildRequest (Anthropic /v1/messages, Google /v1beta/models/{model}:generateContent, or OpenAI-style /v1/chat/completions for the rest — plus, when cfg.reasoningEffort is set, each provider's own reasoning-effort field, see below), extractText and extractError. join() tolerates a user-typed base URL that already ends in /v1 or the full chat path. Also exports baseOf/join for models.ts to reuse, and googleThinkingMechanism/GOOGLE_BUDGET_BY_LEVEL for Gemini's level-vs-budget split.
src/llm/models.ts	The model-listing call: buildModelsRequest(cfg, cursor?) (a GET, per-provider endpoint/auth/pagination) and parseModelsResponse(provider, json) (→ ModelsPage, defensive against every shape above). Also where reasoning-effort support is detected per model — see below.
src/llm/parse.ts	parseAnswer(schema, raw) — the trust boundary. Digs the JSON object out of whatever the model sent (fences, prose, stray braces), then validates every proposal against its ResolvedDef.
src/model/pdfText.ts	extractPdfText(bytes) — the paper as plain text, one [page N] block per page, using the same reading-order heuristic as pdfMeta.ts. Reports empty: true when a document yields almost no characters (a scan).

src/state/aiStore.ts (a separate Zustand+immer store, kept out of the main store for the same reason the project editor is) owns the flow; src/components/AiDialog.tsx and LlmSettingsDialog.tsx are views over it.

The flow

1. openDialog() computes targets = unansweredFields(schema, paper.annotations) and loads the configured targets from the platform. The dialog names how many empty fields will be proposed and — before anything is sent — states what leaves the machine, and to whom.
2. run() fetches the paper's bytes from the same URL the viewer renders (getPdfSource, so slr-file:// in Electron and blob:/http in the browser), then:
   • attach: 'text' (the default, and the only option for an openai-compatible target) → extractPdfText. If the extraction is empty the run stops with an error rather than sending a title and inviting the model to invent a paper from it.
   • attach: 'pdf' → the PDF is base64'd and sent as an attachment (Anthropic / OpenAI / Google / OpenRouter only — the others have no way to take one inline in a single request). A target set to pdf against a provider that cannot take one silently falls back to text, and the dialog's consent line mirrors that fallback rather than promising the PDF.
3. buildSystemPrompt(schema, targets, delivery) + buildRequest(config, …) → an LlmHttpRequest. The delivery matters: with extracted text the model is told the extraction is lossy, or it will confidently reconstruct a mangled table into numbers.
4. platform.callLlm(request, signal) sends it (see below) and returns the raw body.
5. parseAnswer validates it; the surviving Suggestion[] become review rows, all pre-ticked — the reviewer's job is to remove what is wrong.
6. apply() hands the ticked suggestions to useStore.applyAiSuggestions.

The store keeps an AbortController outside the state (it is not serializable), so Cancel aborts a call in flight; an elapsed-seconds ticker drives the progress line.

Model listing and reasoning effort

The model field in LlmSettingsDialog.tsx is not free-typed against nothing: aiStore.fetchModels(config, apiKey?, opts?) asks the provider itself what models exist, via the same platform.callLlm transport the annotation call uses — buildModelsRequest (src/llm/models.ts) builds a GET (the one place LlmHttpRequest.method is ever 'GET'; everything else is 'POST'), and parseModelsResponse turns whatever comes back into ModelInfo[]. Results are cached in aiStore's models/modelsLoading/modelsError/modelsFetchedAt, keyed by LlmConfig.id, for MODELS_TTL_MS (1h) unless opts.force bypasses it. Fetching still requires a stored key — same requirement, and same "save first" pattern, as verifyConfig.

ModelPicker.tsx renders the fetched list as a searchable dropdown, but — unlike ComboBox.tsx (used for closed enum fields elsewhere) — it never reverts what the reviewer typed. A provider's catalog can be incomplete (a brand-new model, a private fine-tune) or simply not fetched yet, so the model field always accepts free text; once the field has been left with text that doesn't match any fetched model, ModelPicker marks itself invalid (red border, a tooltip naming the provider) rather than silently discarding it. An empty/unfetched list is never "invalid" — only "unknown".

Reasoning-effort detection is per-model, not per-provider, because it depends on which model, not which vendor: models.ts attaches a ReasoningProfile | null ({ levels, defaultLevel }) to each ModelInfo, from two different sources depending on the provider:

• Read from the response itself, where the provider says so: Anthropic's capabilities.effort.{level}.supported (/v1/models — the exact levels a model takes, no guessing), Google's thinking: boolean flag, OpenRouter's per-model reasoning.supported_efforts. These need no maintenance as new models ship.
• A model-ID pattern, where the provider's list endpoint says nothing about it (OpenAI: o[0-9]/gpt-5* minus -chat variants; Groq: openai/gpt-oss* only; xAI: grok-4.5 only; Mistral: mistral-(small|medium)*). These are best-effort allowlists, verified against each provider's docs at the time they were written, and will go stale as new models ship — revisit them the same way tokenParam in providers.ts already has to be revisited (see Change Guidance below).
• Deliberately null for every model on two providers: DeepSeek (its reasoning_effort contract for the newer V4 models could not be confirmed against primary docs) and openai-compatible (no single agreed-upon shape across llama.cpp/vLLM/LM Studio — llama.cpp is actively removing its per-request toggle in favor of a server-startup flag). Offering a control that silently does nothing on some servers is worse than not offering one.

LlmSettingsDialog.tsx shows the reasoning-effort <select> only when the currently selected model has a profile, defaults it to defaultLevel ("medium" when the model offers it, else the level in the middle of its range — "if reasoning effort is available, default to medium or the nearest equivalent") the moment such a model is picked, and clears LlmConfig.reasoningEffort the moment a non-reasoning model is picked instead — a stale effort level must never outlive the model it was chosen for. buildRequest (providers.ts) then injects it in whatever shape that provider's chat-completions call actually wants: Anthropic gets thinking: {type:'adaptive'}, output_config: {effort}; Google gets generationConfig.thinkingConfig.thinkingLevel (Gemini 3.x) or a token-count .thinkingBudget translated via GOOGLE_BUDGET_BY_LEVEL (Gemini 2.5.x — the two shapes are mutually exclusive on one request, so a model only ever gets one); OpenRouter gets a nested reasoning: {effort}; everyone else (OpenAI, Groq, Mistral, xAI) gets a flat reasoning_effort field.

Switching provider mid-edit clears both the typed model and the cached model list for that draft (clearModels(id)) — the list answered "what does the old provider have", which is not an answer to "what does the new one have", even though the draft's id (and so the cache key) stays the same across the switch.

Why the call goes through the Electron main process

A request to an LLM API is a POST with Content-Type: application/json and an auth header — which makes it a preflighted cross-origin request, sent from the renderer's origin (file:// in a packaged app). It is the same CORS wall that forced corsEnabled on the slr-file:// protocol (documented under Electron Main Process below): Chromium rejects the request before it ever reaches the network, and the failure surfaces as an opaque TypeError with no detail. So on the desktop the renderer never sends the request itself — ElectronAdapter.callLlm hands it to llm:call, and the main process sends it with net.fetch, where no origin and no CORS check apply.

The second reason is the one the whole layer is built around: the API key never enters the renderer.

• LlmConfig has no apiKey field — only hasKey: boolean. A stored key can never be read back, not even by the settings dialog that wrote it (so leaving the key box blank on an edit keeps the stored key; that is the only way to edit a target without retyping it).
• The renderer builds the entire request, but can only put API_KEY_SENTINEL ({{apiKey}}) where the key goes. electron/main.ts substitutes the real key into the headers immediately before sending.
• Before it does, it checks the origin: new URL(request.url).origin must equal new URL(config.baseUrl).origin, or it refuses. The renderer names the URL, so a compromised renderer must not be able to post the key to a host of its choosing.
• Keys are stored in userData/llm-config.json, encrypted with safeStorage (mode 0600). If safeStorage.isEncryptionAvailable() is false the app refuses to save rather than silently writing the key in the clear.

llm:call also takes a requestId and keeps the AbortController in an inFlight map, because an AbortSignal cannot cross IPC — Cancel sends a separate llm:abort message that main matches against the id.

The same channel carries the list-models requests from models.ts: LlmHttpRequest.method ('GET' or 'POST', defaulting to 'POST') tells both electron/main.ts and src/platform/browser.ts whether to send a body at all — a GET carries none, and passing one is a fetch error on some runtimes. The origin check and sentinel substitution apply identically either way.

The browser build cannot make those promises, and says so instead of pretending otherwise (src/platform/browser.ts): the key lives unencrypted in localStorage (slr.llm.configs), the request goes out from the page, so the provider must be willing to answer a cross-origin call (Anthropic needs the explicit anthropic-dangerous-direct-browser-access header, which the adapter adds; a self-hosted OpenAI-compatible endpoint usually sends no CORS headers at all and simply fails), and a cross-origin block is caught and re-thrown with an error that names the likely cause rather than reading as "the provider is down". The settings dialog shows a red notice in that runtime and points at the desktop app.

Why a misbehaving model cannot corrupt a project

Two gates, and both are unconditional:

• parse.ts validates every proposal against the schema. resolvePath rejects unknown names, group paths (a group holds no value), non-final segments that have no children, and any index at or beyond a node's max — plus, for the LLM callers only, any index at or beyond MAX_UNBOUNDED_INDEX (10 000) on a node declared max: null, since applyAiSuggestions materializes every instance up to the index and a reply of Findings[9007199254740990] would otherwise be an out-of-memory kill. That ceiling is opt-in (ResolveOptions.maxUnboundedIndex) precisely because the same function also resolves paths that already exist in a project: applied unconditionally it made git/merge.ts's applyOne skip a conflict the reviewer had explicitly resolved by hand, keeping "ours" with no error. Only llm/parse.ts and applyAiSuggestions ask for it; then the value must typecheck against its ResolvedDef. It bends only where models misbehave in a way with exactly one honest reading ("2021" → 2021, "True" → true, a case-off enum value snapped onto its option). Everything else — "about 20", a value outside the enum, a duplicate answer for the same field — is rejected, never guessed at, and rejections are shown to the reviewer, because a silently dropped answer looks like the model never said anything. parseAnswer never throws: it sits on a network response, where garbage is a normal outcome.
• A reply is only ever applied to what it was asked about. The run records the paper, the seat and the target that answered (AiState.runFor), and applyAiSuggestions refuses if any of them has changed. This is not defensive padding: the dialog stays open and the paper list and seat picker stay usable while a call is in flight, so selecting another paper and pressing Apply used to write the reply about paper A onto paper B — fabricated content on a paper nobody read, with an aiUsage record vouching for it — and switching seats did the same across reviewers, corrupting the inter-rater data. It refuses rather than quietly retargeting: writing to the run's paper while the reviewer looks at a different one would be correct attribution but invisible work, so they would see "applied" and no change. The recorded target is also what aiUsage discloses, rather than whichever target the picker happens to show at Apply time.
• applyAiSuggestions (src/state/store.ts) is one undo step. It decides what to write before touching anything — a suggestion is dropped if its path no longer resolves, or if the field has been answered since the model was asked, so the reviewer's own work is never overwritten — and if nothing survives, it leaves no empty entry on the undo stack. It then snapshots once and mutates, creating any instances of repeatable nodes the model addressed but that did not exist yet. lastFieldKey is reset so the reviewer's next keystroke is not coalesced into the AI's step. Ctrl/Cmd+Z therefore undoes the whole fill in one go. It returns an AiApplyResult (filled / skipped) for the summary the dialog shows.

AI marks

A field the model filled gets a light-blue border (--ai-mark, .ai-marked in src/styles/index.css) so the reviewer can see at a glance which values are not their own. Clicking into the control — or on its label (NodeName) — clears it: that click is the confirmation. Field and AnnotationNode read the flag through the useAiMark(path, name, index) hook and clear it via confirmAiMark.

Two properties make the marks safe:

• They are session-only, by construction. aiMarks is a Record<string, true> in the store, not a field of Project, so serializeProject has nothing to write even by accident — a saved file is byte-identical to one saved without the feature (pinned by src/state/store.aimarks.test.ts). Loading or closing a project clears them.
• They only ever point at real AI values. applyAiSuggestions marks the suggestions it wrote, never the ones it skipped. undo/redo clear all marks: undoing an AI run empties exactly the fields the marks point at, and a blue border on a now-empty field would be a lie. History restores values, not marks.

The key is `${paperId}::${formatPath([...path, { name, index }])}` — paper-scoped because every paper shares the same paths, and canonical (src/llm/paths.ts) so a mark set from a model suggestion and one looked up by the UI meet on the same string. Index 0 stays implicit, which is what keeps Findings[1]/Claim a different field from Findings/Claim.

On a multi-reviewer project, the key also folds in the active reviewer (`${paperId}::${reviewer}::${canonicalPath}`, via aiMarkKey's third argument) — otherwise Reviewer 1 running the AI would leave marks that appear to belong to Reviewer 2's identical field paths. A single-reviewer project's keys are exactly the two-part form above (the reviewer argument defaults to null, which reproduces it), so nothing about single-reviewer behavior changes.

AI usage disclosure (Paper.aiUsage)

Deliberately the opposite design from AI marks: where a mark is session-only and never reaches the file, Paper.aiUsage (src/model/project.ts) is a permanent, append-only log — { provider, model, appliedAt } — meant to survive into the saved file so a co-reviewer (or the reviewer themself, later) can see that, and how, AI was used on a paper.

• Written by applyAiSuggestions (src/state/store.ts), inside the same set() mutation — and therefore the same undo step — as the field writes themselves, only when the run actually filled something (filled > 0). This is a deliberate coupling: undoing the AI fill that produced a disclosure entry undoes the entry with it, consistent with how AI marks are already tied to the data they describe. There is no cheap way to make a field "survive" an undo of the action that created it under this app's whole-project-tree undo/redo (see State management above), so rather than fight that, the entry rides along with the fill it discloses.
• Parsed defensively (parseAiUsage in project.ts): the file is hand-editable, so a malformed entry — wrong types, a non-array aiUsage, missing fields — is dropped, never thrown over (aiUsage: z.unknown().optional() in schema.ts intentionally does not typecheck the shape at the zod layer, for the same reason annotations doesn't: "loosely typed here, validated/normalized structurally in project.ts").
• Written only when non-empty (serializeProject): a paper AI was never used on carries no aiUsage key at all, so a normal, hand-annotated project's file stays exactly as it always was.
• Array order is the ordering guarantee ("the order of use should be apparent"): entries are pushed oldest-first, and appliedAt (an ISO 8601 timestamp) makes that order explicit even if the array is ever hand-edited or reordered.

Git

Why Electron-only

Electron: yes. The main process is Node, so child_process.execFile('git', […]) spawns the user's own git binary, which reads their real ~/.gitconfig, their credential helper (osxkeychain, manager-core, …), their ~/.ssh/config, and their SSH agent. No npm dependency is needed and none was added — the runtime dependency list stays at 6 (immer, react, react-dom, react-pdf, zod, zustand).

Browser: no, and there is nothing to fall back to. src/platform/browser.ts has no process, no fs, and no path (rebasePdfPaths returns its input unchanged with the comment "the browser has no paths"; getOsInfo() returns null; PDFs need an explicit user folder grant). A web page cannot spawn a process, cannot read ~/.gitconfig, and cannot reach an SSH agent — this is the sandbox, not a missing feature, and no flag or permission changes it. The feature request's own fallback clause — "if this is not possible, it should still try to use the local git configurations" — has no referent in the browser: there is no local git installation reachable, so there is no local git configuration to try.

Rejected: a pure-JS reimplementation (isomorphic-git or similar). It would answer a different question than the one asked. It is not "the local git installation" — it has its own credential handling, does not read the user's ~/.gitconfig, and does not use their SSH agent or credential helper. Shipping it and calling it "git support" would be dishonest about what actually ran.

The conclusion: git support is Electron-only. This mirrors how the codebase already handles other capability gaps — getOsInfo(): OsInfo | null, needsPdfFolderGrant() returning false on Electron (there is no such prompt there), the pathless rebasePdfPaths — and it is expressed in the type system, not left as a runtime convention: PlatformAdapter.getGit(): GitPlatform | null returns null in the browser, so a caller cannot invoke a git operation without first proving the runtime has one. A flat GitPlatform capability object rather than fifteen individual methods on PlatformAdapter is deliberate too: fifteen flat methods would mean fifteen browser stubs, each of which either throws at runtime or silently no-ops — "unavailable" would be something a caller discovers by calling it, not something the type checker catches for them. getOsInfo(): OsInfo | null already established the pattern this follows.

The browser build shows every git entry point disabled, not absent. Toolbar.tsx checks getGit() the same way any other caller does, but where the annotation UI usually treats "no capability" as "no control" (there is no PDF-grant button on Electron, because needsPdfFolderGrant() is false there and the concept doesn't apply), the toolbar's Git button and the Open menu's Import from git… item stay in the layout in the browser build too — dimmed, with GIT_BROWSER_DISABLED_HINT (Toolbar.tsx) as the tooltip, telling a reviewer why it doesn't work and that the desktop app is where it does, rather than a control that quietly isn't there for them to notice missing. This is a genuinely different choice from needsPdfFolderGrant()'s: that gap has no button anywhere to hide, while this one is choosing to surface a capability boundary explicitly instead of omitting the entry point the way the type-level GitPlatform | null design would most simply suggest.

On Electron the Git button is now always rendered, disabled with an honest tooltip whenever it can't be used — including the start screen (no project open) and when the open project's file isn't in a git repository. The policy is unified through gitButtonState() (Toolbar.tsx), a pure function mirroring the "Import from git…" menu item's precedence: hidden only in the browser (where getGit() is null and the concept doesn't apply), disabled-with-reason everywhere else the capability exists but can't currently be used, and enabled only when a project is open in a real git repo and no other operation is busy. Showing the button disabled on the start screen — rather than hiding it — keeps the layout stable so a reviewer who opens a project doesn't see the toolbar jump.

A third, distinct case: git is not installed. GitPlatform.probe() runs git --version. On failure the app says so honestly, with git's own error text, and the git entry points are shown disabled with that reason rather than hidden entirely — a control this machine could offer if git were installed is worth showing dimmed; a control the runtime can never offer (the browser) is noise and is hidden outright. This is the same asymmetry AnnotationPanel.tsx already applies to the ✦ AI button: invisible when the runtime/session gate (aiUnlocked) is off, dimmed-but-visible when the project turned it off (config.ai: false).

The renderer never names an argv

Every git operation the renderer can ask for is one of fifteen enumerated IPC handlers in electron/main.ts (git:probe, git:pickCloneDir, git:clone, git:pickProjectIn, git:info, git:status, git:headContent, git:workingContent, git:commitPartial, git:commit, git:push, git:pullBegin, git:pullFinish, git:pullAbort, git:writeWorking) — never a general git <args> channel. Git has --exec-path, aliases, and the ext:: remote-helper transport; a channel that let the renderer choose the argv would be handing it arbitrary code execution wearing a "just run git" label. Main decides what git is actually asked to do; the renderer only supplies data (a URL, a path, a commit message, a resolved text).

Every call uses execFile with an argument array, never a shell string, and a -- terminator before any user-supplied path or URL — a repository URL or a destination path is user input reaching a spawned process, and without both, a URL like --upload-pack=/bin/sh would be read as an option rather than an argument. src/git/url.ts (validateGitUrl/validateClonePath) is the actual gate, and electron/main.ts imports it — the first, and (with src/git/output.ts's gitErrorText) one of only two, imports of src/ into electron/. That is deliberate and load-bearing: a security gate must not exist twice, the same reason comparable() in src/consolidate/unanimous.ts is one shared function rather than three copies that could drift. Both modules import nothing themselves, so they typecheck identically under tsconfig.node.json (types: ["node"]) and tsconfig.app.json — a file appearing in two TypeScript programs is fine as long as neither is composite and both are noEmit, which is already true here.

validateGitUrl is an allowlist of transports (https://, http://, ssh://, git://, git+ssh://, file://, the user@host:path scp shorthand, or an absolute local path), not a blocklist of characters — because the dangerous case is not a stray shell metacharacter (execFile

• the argument array already close that door) but git's own ext:: remote-helper syntax, which makes git run a program named by the URL. The :: check requires two consecutive colons, so an ordinary https:// (which has none between the scheme and its slashes) can never trip it — src/git/url.test.ts pins this explicitly, because it is exactly the kind of regex a later "simplification" could quietly break.

Not weakening git, only removing what cannot work here

gitEnv() (electron/main.ts) sets GIT_TERMINAL_PROMPT=0, GIT_EDITOR=true, and GIT_SEQUENCE_EDITOR=true, and strips any inherited GIT_DIR/GIT_WORK_TREE/GIT_CONFIG-family variable. None of this weakens anything a user configured:

• The spawned process has no terminal. A git that decided to prompt for a username, or open an editor for a commit message, would otherwise block forever on a tty that does not exist, and the app would look frozen with no way out. Turning both off makes git fail immediately with its own message instead of hanging — the opposite of weakening, since a hang is a worse failure mode than an honest error.
• Credential helpers, askpass programs, SSH agents and host-key checking are never touched. None of them is a terminal prompt, which is exactly the point: the user's configured way of authenticating still works unchanged, and only the "type it at the console" path — which cannot work in a spawned child with no tty — is turned off.
• The GIT_DIR-family variables are stripped because SaiLoR may have been launched from a shell that happens to be sitting inside some other git repository; an inherited GIT_DIR would silently point every git call below at that repository instead of the project's own.

The repository's own config is not trusted (GIT_SAFE_CONFIG)

The environment is only half the story. A repository's .git/config is read before git does anything, and several of its keys name commands git runs — and this app's documented workflow is receiving a project folder from a collaborator. A folder that arrives by zip, USB, or shared drive brings its .git/ with it, so those keys are attacker-controlled input, not user configuration.

The concrete path: a core.fsmonitor of printf PWNED > /tmp/proof; false executes on git status, which the Git button reaches in one click, and git:info runs automatically on project open. .git/hooks/pre-commit gives the same on commit. Git's safe.directory guard does not apply, because the copied folder belongs to the reviewer who copied it.

Every runGit call therefore prepends a fixed list of -c overrides. -c outranks every config file, so this is a hard override rather than a request:

core.fsmonitor=false, core.hooksPath=<nonexistent>, core.pager=cat, core.editor=false, core.alternateRefsCommand=, uploadpack.packObjectsHook=, protocol.ext.allow=never.

Two deliberate exclusions, and both are the interesting part:

• core.sshCommand, credential.helper and gpg.program are left alone. Because -c outranks the global config too, overriding them would break the ordinary setups the section above is careful not to touch. They also only run on an explicit network action the user asked for — never on merely opening a folder, which is the boundary that actually matters here.
• diff.external is not in the list, and this is not an oversight. Setting it empty does not mean "no external diff": git tries to exec the empty string and the diff dies with cannot run :. Swapping an attacker's differ for a guaranteed failure is not a fix — it silently emptied the Git panel's diff for every user while porcelain kept working, so nothing looked wrong. --no-ext-diff, passed on the diff invocation itself, is the flag that actually means "use your own", and it is what the diff call uses.

--no-textconv on the diff is not optional either. diff.<driver>.textconv is selected by an in-tree .gitattributes, so -c cannot pre-empt it, and --no-ext-diff does not cover it — they are separate mechanisms. Verified: a received folder carrying * diff=evil plus a diff.evil.textconv in its own config executes that command on git status, which is one click from opening the project, exactly like the core.fsmonitor case.

filter.* clean/smudge drivers and custom merge.<driver>.drivers cannot be disabled by -c at all and are the known residuals. Both run only on an explicit commit or pull, not on opening a folder — which is the boundary this section is about.

Writes never follow a symlink

writeFile follows a symlink and writes the target, and one save path is never confirmed by a dialog: the sibling <name>-fulltext.json that "Start full-text screening" derives from the project's own location. Shipping that name as a symlink to ~/.zshrc turned one click into an overwrite of a shell startup file — with substantially attacker-chosen content, since serializeProject round-trips unknown keys verbatim. Every write in electron/main.ts calls assertNotSymlink first, which uses lstat (reporting the link itself rather than what it points at). A parent directory that is a symlink is fine and stays working — that is ordinary on macOS, where /tmp is one.

Checking the leaf is not enough. assertNotSymlink inspects only the final component, which leaves the directory case open: a repository carrying sub -> /elsewhere accepts a relative path of sub/project.json — no .., and the leaf really is an ordinary file — and the write lands outside the repository. assertInsideRoot resolves the parent with realpath, following every link in the chain, so containment is checked against where the write actually goes rather than where the path string claims it goes. slr-file:// had the read-side version of the same hole and now realpaths both sides before serving: path.resolve collapses .. but follows no links, so a pdfs/paper.pdf symlinked to /etc/passwd used to resolve inside the project and be served.

The relative-path rule itself lives in src/git/relpath.ts, not in electron/main.ts, for the same reason validateGitUrl does: electron/ is outside vitest's include, and a security gate no test can reach is one nobody can change safely. It splits on both separators (p.split('/') left ..\..\Users\victim\.bashrc as a single opaque segment on POSIX while path.win32.join honours it), rejects Windows absolute forms explicitly, and rejects any .git segment — a valid relative path, but never project data, and a write-to-hooks primitive. That comparison strips trailing dots and spaces and lowercases, because Win32 strips them from path components itself, so .git.\config, .git/config and .GIT/config all reach the same directory.

The module layout

Module	Purpose
src/git/types.ts	Shared shapes crossing the platform seam: GitRun, GitProbe, GitFileChange, GitStatus, GitRepoInfo, CloneOutcome, PullStart, and the GitPlatform interface itself.
src/git/url.ts	Pure. validateGitUrl, validateClonePath, repoNameFromUrl — the security gate, imported by electron/main.ts (see above).
src/git/output.ts	Pure. parsePorcelain (turns git status --porcelain=v1 -z into GitFileChange[]), capDiff (caps a diff for the DOM), diffLines (splits a unified diff into per-line add/remove/context for the coloured view — see below), gitErrorText (what to show when a git command failed) — also imported by electron/main.ts, so the "what does a failed run's message say" logic exists once.
src/git/merge.ts	Pure. The field-level three-way merge — see below. Knows nothing about git or the DOM, the same shape src/consolidate/ follows.
src/git/changes.ts	Pure. Field-level local change detection and composition for the commit panel — see "Field-level commit review" below. Reuses merge.ts's conflictId/MergeTree for row identity, but not its three-way merge3 rule (only one side, the working tree, has changed here).
src/state/gitStore.ts	The clone flow and the commit/pull/push panel; owns the pull orchestration and the field-review state.
src/components/GitCloneDialog.tsx, GitDialog.tsx, GitMergeDialog.tsx	Views over gitStore.

Each dialog's width class is .modal.git-*-dialog, not bare .git-*-dialog — a single-class selector has the same specificity as index.css's own .modal (which also sets width), so which one wins is decided by stylesheet order rather than anything about the rules, and that order is not even the same between dev mode (Vite injects each imported stylesheet as its own <style>, in module-evaluation order) and the production bundle (Rollup concatenates them, and did so the other way round). Concretely: with the bare selector, every git dialog's own width was silently losing to .modal's generic one in the shipped build — all three stuck at the same 680px regardless of what git.css said. Qualifying with .modal too makes the outcome deterministic instead of an accident of build order.

The clone dialog's URL field needed its own width for an unrelated reason. .field-input's width comes from flex: 1 1 auto, which does nothing outside a flex row — every other place it is used is one, but the URL field is the sole control in the clone dialog's body, so it still rendered at the browser's default input width no matter how wide the dialog around it was. .git-clone-url-input gives it width: 100% directly, which is what actually makes it "big" — the dialog itself is a bounded min(600px, 94vw), the same shape as the other two dialogs, not scaled to window width. An earlier version made the dialog 92vw to get a wide field, which worked but took everything else in the dialog (the destination path, the buttons) along with it, past a comfortable reading width on a large monitor for no reason those needed to grow at all. Fixing the input's own width instead of the dialog's gets the same big field at whatever size the dialog actually is.

The diff is coloured per line, not as one block. GitDialog.tsx renders each line diffLines (src/git/output.ts) returns as its own <span className="git-diff-line git-diff-{kind}">, since a unified diff interleaves added, removed, and unchanged context lines — anything coarser would colour context text too. Classifying a line as add/remove by a bare +/- prefix check would misread two things at once: a genuinely added line whose own content starts with +/- (++counter;), and — worse — the +++ b/path/--- a/path file-header pair the moment a file's first real change happens to start with those same three characters. diffLines instead tracks whether a hunk has started for the current file (reset on every diff --git, set on that file's first @@): the header pair, which by git's own format only ever appears once per file and always immediately before that file's first hunk, is context regardless of what it looks like, and every +/- line once inside a hunk is real content regardless of what it looks like.

State management: gitStore

useGitStore (src/state/gitStore.ts) is a separate Zustand+immer store, kept out of the main store for the same reason aiStore and the project editor are: a self-contained flow with its own lifecycle that the ordinary annotation path never needs to know exists. The dependency direction is one-way — gitStore reads and drives the main store via useStore.getState(), but store.ts never imports gitStore (it does not import aiStore either, for the same reason). Refreshing repo (where the open project sits git-wise) whenever the open project's saveHandle changes is therefore an effect in App.tsx, not a call made from inside store.ts — the same shape the AI store's own wiring already has.

The pull command sequence

git:pullBegin (electron/main.ts) is the classification a pull starts with, run in this order:

1. git status --porcelain=v1 -z — any tracked change (an untracked file, code ??, never blocks a merge) means { kind: 'dirty', paths }.
2. No @{u} (no upstream configured) means { kind: 'no-upstream', branch }.
3. git fetch. A failure means { kind: 'error', message }.
4. HEAD already an ancestor of @{u} means { kind: 'up-to-date' }.
5. @{u} already an ancestor of HEAD means a fast-forward: git merge --ff-only @{u}, then { kind: 'fast-forwarded' } or { kind: 'error', message }.
6. Otherwise the histories have diverged. The merge base and the project file's text at all three revisions (git merge-base, then git show <rev>:<path>) are read before anything touches the work tree, so nothing that follows can change what actually gets merged.
7. git merge --no-commit --no-ff <ref>. If it never even started (MERGE_HEAD absent — unrelated histories, a hook refusing) that is { kind: 'error', message }, without ever calling merge --abort (which would itself fail with "There is no merge to abort" — checking MERGE_HEAD first is what keeps the next point true).
8. If anything other than the project file is left unmerged, SaiLoR does not know how to help — it knows how to merge an annotation JSON, not a PDF or a .gitignore. The git merge is aborted (git merge --abort) and { kind: 'conflict-elsewhere', paths } is returned; nothing is half-done.
9. Otherwise: { kind: 'merge', ref, base, ours, theirs } — the three texts, handed to mergeProjects (below).

Contract: git:pullBegin always returns with the repository in exactly one of two states — not mid-merge, for every outcome except 'merge', or mid-merge with nothing unmerged except the project file, for 'merge'. It never returns leaving a half-merge the renderer did not ask for.

git:pullFinish(root, relPath, text) always writes text — the merged project's own serializeProject output — over whatever git's own line-based attempt produced, then git add + git commit --no-edit. git commit after a merge with MERGE_HEAD set is what records both parents and tolerates an empty tree change, which is why the merge commit is finished this way rather than with commit-tree or merge -m; --no-edit takes git's own prepared MERGE_MSG, and GIT_EDITOR=true (above) is the backstop if it ever tried to open one anyway.

Two gates before a pull touches anything: on-disk clean, and in-memory clean

git:pullBegin's 'dirty' check answers one question: is the file on disk clean by git's own status. It says nothing about the reviewer's unsaved, in-memory annotations — those exist only in the React state and are invisible to git entirely. gitStore.ts's runPull() therefore refuses outright, before ever calling beginPull, when useStore.getState().dirty is true: a fast-forward or a finished merge reloads the project file from disk (reloadOpenProject, which is exactly openRecent(path) — the file changed underneath the open project, so the in-memory copy is stale either way), and without this guard that reload would silently discard whatever the reviewer had not yet saved. This is the single most important line in the whole feature: get it wrong and a pull can lose a reviewer's unsaved work with no warning at all. GitDialog.tsx shows a dirty-banner with a Save project button and disables Pull (and Commit — committing the file on disk while the in-memory copy disagrees with it is its own kind of confusing) while it's up.

The merge (src/git/merge.ts)

The full reasoning lives in data-model.md's "Merging two copies of a project"; this is the shape of it from the code side.

• The one rule (merge3): a side that did not change a value away from the base does not get a vote on it. Applied identically from the project's own title down to a single annotation field — not a special case, the actual algorithm. Returns a conflict only when both sides changed a value, to different things.
• Exact equality, not comparable(). unanimous.ts's comparable() folds case and whitespace because it is answering "did the reviewers say the same thing"; a merge is answering "did I change this value since the base", and folding away a capitalization fix would silently revert the reviewer's own edit in favor of the remote's. merge.ts uses plain === on FieldValue, and merge.test.ts pins the distinction directly (a base→ours case-only fix survives unopposed; the same fix made differently on both sides is a real conflict).
• Absent reads as empty (valueAt). pruneTree drops only the trailing empty instances on save, so an instance that is empty-but-present and one that is simply missing are the same thing on disk — and must merge the same way, or a field one side filled in from nothing would wrongly conflict against an "absent" base, and an entry the remote deleted would come back. This is also what makes instance removal fall out of the ordinary field-level rule with no instance-level logic at all: a removed entry reads as all-empty on that side, the rule takes the empties, and pruneTree drops the now-trailing instance on the next save.
• Repeatable arrays are unioned by index and never compacted. count is the union of all three sides' lengths (clamped to def.max); position is never closed up, because position carries meaning — consolidation lines up each reviewer's entries by index (src/consolidate/apply.ts), and closing a gap here would silently re-point that alignment.
• A conflicted field holds our value in merged until it is resolved. If the resolution dialog is ever bypassed, the file still holds the local reviewer's own work — the safe side. It is not a decision on the merge's part: GitMergeDialog marks every conflicted row undecided regardless, and will not finish the merge until each one has actually been answered.
• What refuses, and why. A change to version, config.schema, config.ai, config.reviewers, config.screening, or a root/paper extra key, made differently on both sides, refuses the whole merge rather than guessing a field-level answer — each of these re-shapes the file (most obviously the schema, which decides the shape of every annotation tree; config.screening for the same reason, since it decides config.schema via screeningSchemaDefs in the first place), and a left/middle/right conflict row cannot ask "which taxonomy is right". Project.title is deliberately not on this list: it is one string, a conflict row expresses it perfectly, and refusing an entire merge over two people renaming the review would be absurd. Two more refuse for two different reasons: provenance is a nested record no FieldConflict shape can express, not something that reshapes the file — the common case (only one side ever sets it) still resolves cleanly with no refusal at all. protocol is here for the same reason — a nested record no FieldConflict row can express, and half-dropping a reviewer's authored protocol is worse than asking them to reconcile it.
• The paper-deletion asymmetry. A paper one side deleted and the other side changed is kept, with a note, never silently deleted — a field-level UI cannot ask "keep or delete this paper", and the two failure directions are not symmetric (a paper nobody wanted is one click from gone; deleted annotated work is just gone). Contrast with reviews: a reviewer's tree is never deleted by a merge, only by both sides having already dropped it — the same rule normalizeReviews already applies on load (see "Lowering the reviewer count" in the schema guide), because a reviewer's tree is someone's labour, not a project-author decision the way a paper is.
• aiUsage is a union, not a three-way merge, and deliberately does not consult base — it is an append-only disclosure log with no delete operation, so a record either side still holds must survive regardless of what the base looked like.
• A boolean can never conflict. Paper.equal, a set spelled as an array, merges per-path via merge3<boolean> — a boolean has only two values, so "both sides changed it, differently" cannot happen; merge3's first branch (eq(ours, theirs)) always takes it.
• Paper.abstract and abstractFromPdf are ordinary paper-level fields, merged independently. Each gets its own merge3 call and its own possible conflict row, the same as title/doi/pdf. A known, accepted gap: resolving an abstract conflict does not retroactively touch abstractFromPdf — if only one side changed the flag, merge3's "one side changed" branch takes it before the reviewer ever sees the separate abstract conflict, so the resolved text and the disclosure flag can end up describing two different edits. merge.test.ts demonstrates and pins this exact case rather than hiding it. This was a genuine bug until fixed: both fields were entirely absent from mergePaper's field list, canonicalPaper, and applyOne — every pull silently dropped every paper's abstract, and canonicalPaper's omission meant an abstract-only edit could make paperUnchanged wrongly true, risking the paper being dropped as part of the paper-deletion asymmetry above.

applyResolutions writes the reviewer's per-conflict choices into the merged project with immer's produce — not structuredClone (not something to bet on under every runtime this ships to) and not JSON.parse(JSON.stringify(...)) (which drops undefined-valued keys that both deepEqualJson and the round-trip test care about). deepEqualJson itself is exported from src/model/project.ts rather than redefined in merge.ts — the same "one shared implementation" rule as comparable().

Field-level commit review

Before this existed, git → Commit only offered a whole-file checkbox: tick the project JSON or don't. src/git/changes.ts breaks the open project's own file down into the individual fields that actually changed, so a reviewer can decide field by field whether to commit it now (Use), leave it as a still-uncommitted local change to revisit later (Ignore), or revert it (Discard) — without touching everything else in the file.

The comparison is structural, not textual. detectFieldChanges(head, working) compares two parsed Projects (HEAD's own copy of the file, read via git:headContent — git show HEAD:relPath — against the working tree's, read via the side-effect-free git:workingContent), walking annotation trees the same recursive way merge.ts's three-way walk does and reusing its conflictId/MergeTree shapes for row identity. This is deliberately immune to JSON formatting/key-order noise — the same "compare parsed values, not text" choice the codebase already made for needsShapeMigration. Unlike merge.ts, there is no three-way question here: only the working tree has changed, so every difference is something the reviewer decides about, not something that might resolve itself.

A structural difference refuses field-level review entirely, falling back to the plain file checkbox — the same refusal list mergeProjects uses (config.schema, config.reviewers, config.ai, config.screening, version, provenance, protocol, root extra): once the schema itself might differ, "which fields changed" stops being a question with a field-level answer. provenance and protocol are here for a different reason than the rest — they don't reshape anything, they're just nested records no FieldConflict row can express.

Coupled fields are bundled into one row. PAPER_META_BUNDLES maps abstract to its hidden dependent, abstractFromPdf — the disclosure flag is not an independent fact a reviewer chooses among, it just describes whichever abstract value ends up committed, so it never gets a row of its own; it silently follows the primary field's disposition (applyFieldWithBundle, which reads the raw boolean directly from the source Project rather than trying to derive it from the bundled field's own display string).

Discarding is deferred, and it is a real write. Picking Discard on a row does not touch the file on disk — it only records the decision, in gitStore's panel.fieldReview.decisions. Nothing is reverted until the reviewer actually presses Commit: composeContents(head, working, changes, decisions) then builds two divergent Projects from the same three inputs — committed (what gets staged: HEAD's own value for every discard/ignore row, the working tree's value for every use row) and workingOut (what the file on disk ends up holding afterward: HEAD's value written back in for every discard row, unchanged everywhere else). A paper added locally follows the same shape — discard deletes it from workingOut; a paper removed locally follows it too — discard restores it from head into workingOut.

Standalone discard without committing. When every changed row is marked Discard (nothing is marked Use), the primary button in GitDialog relabels from "Commit" to "Discard all" and turns danger-red. Pressing it calls runDiscard (gitStore.ts), which composes only workingOut (the reverted file content) and writes it directly via git:writeWorking — a separate IPC handler that writes text to the working file without staging or committing anything. This lets a reviewer revert all local edits in one action without going through the commit ceremony. The per-row Discard button was also moved to the row's right edge (opposite Use/Ignore) so the three dispositions read left-to-right as a single visual axis.

Partial-file staging has no native git primitive, so it is a write → commit → write-back sequence. git:commitPartial (electron/main.ts) writes committedText to the file, git add + git commits it (along with whatever else is in otherPaths, the ordinary whole-file selections), then — in a finally block, unconditionally, whether or not the commit itself succeeded — writes workingText back over it. The finally is load-bearing: without it, a failed commit would leave the working file stuck holding content that was never actually staged as anything.

Decisions survive an incidental refresh. Clicking the panel's own ↻ (or runPull/runCommit implicitly calling refreshStatus) recomputes detectFieldChanges from scratch; refreshFieldReview (gitStore.ts) carries forward any decision whose row id is still present in the new result and drops the rest — an accidental re-scan must never silently reset a reviewer's careful per-row choices, but a row that stopped being a change (its id vanished) has nothing left to carry the decision about.

Testing

src/git/merge.test.ts builds every fixture through the real loadProject, never a hand-assembled Project, so base/ours/theirs are exactly as schema-normalized and empty-skeleton-shaped as mergeProjects' actual caller hands it. It covers the field-level guarantee in both directions, repeatable-node growth (colliding and non-colliding), the interior-gap and instance-removal invariants, the multi-reviewer headline case (disjoint edits by two reviewers, zero conflicts), the paper add/remove asymmetry, every refusal, the aiUsage union, the Paper.equal boolean-set merge, the abstract/abstractFromPdf merge (including the documented resolve-order gap above), applyResolutions, and a full round-trip through serializeProject/loadProject. src/git/changes.test.ts covers detectFieldChanges (structural refusal, no changes, paper-metadata diffing, the abstract/abstractFromPdf bundle including its no-primary-row fallback, annotation tree fields including nested repeatable groups and per-reviewer trees, paper add/remove) and composeContents (all three dispositions on a field, on an added paper, and on a removed paper — including "discard restores a removed paper" — the bundle applying as one unit, and round-trip stability). src/state/gitStore.test.ts covers runPull's orchestration against a fake GitPlatform: each pull classification, a refused merge (aborts and reports why), an unparseable revision (aborts, writes nothing), and the dirty guard refusing to even call beginPull — plus refreshFieldReview's branches (untracked, unreadable, unparseable, structural, a genuine detected change, decisions surviving and being dropped across a refresh), setFieldDisposition/ setAllFieldDispositions, and runCommit's commitPartial branch (composed content, success with its reload, and a failure surfacing the error without one).

Serializer round-trip guard. src/state/editorStore.test.ts includes a test that verifies buildProjectJson (the editor's serializer) and serializeProject (the core's) agree on every root field — if one writes a field the other silently drops, the test fails. This catches the class of bug where a field is lost depending on which path last saved, the same shape of bug that once dropped abstract in the merge layer.

Rejected: streamed clone progress and a cancel button

execFile buffers the whole child process output; a live progress bar (git's own --progress percentage lines) would need a second IPC channel, a subscription lifecycle, and parsing a carriage-return-animated stderr stream — real work for a cosmetic improvement. Cancelling a clone mid-flight would also leave a partial .git directory that would need cleaning up. A spinner plus an elapsed-seconds counter already say "this has not frozen", which is the actual requirement, and the network timeout (GIT_NETWORK_TIMEOUT_MS, 15 minutes) is the backstop against a clone that really has hung.

Why the merge dialog has no Escape

GitMergeDialog is the one modal in the app with no Escape, no backdrop-click, and no × in the header — a deliberate deviation from the app's own modal pattern. The repository is genuinely mid-merge for as long as it is open; dismissing it the ordinary way would leave the reviewer's checkout in a state they cannot get out of without the command line, which is the one outcome a merge UI must never produce. Only Cancel merge (git merge --abort) and Finish merge (disabled until every row is decided) leave.

Electron Main Process

electron/main.ts is a thin main process:

• Window: BrowserWindow defaulting to 1920×1080, context isolation enabled, node integration disabled, preload script loaded. The taskbar/dock icon is set from build/icon.png via nativeImage (and app.dock.setIcon on macOS so it shows in dev, not just the packaged bundle).
• Window-state persistence: size, position, and maximized state are saved to window-state.json in app.getPath('userData') and restored on the next launch. Writes are debounced on resize/move/maximize/unmaximize (400 ms) and flushed synchronously on close; maximized/fullscreen windows store their getNormalBounds() so restore returns to the user's chosen size. A saved position is only reused if it still overlaps a connected display (screen.getAllDisplays()), so a disconnected monitor can't strand the window off-screen; otherwise only the size is applied and the window is centered. Absent/corrupt state falls back to the 1920×1080 default.
• Dev vs prod: loads VITE_DEV_SERVER_URL in dev, dist/index.html in production.
• External links: setWindowOpenHandler sends target="_blank" links (external links in PDFs) to the system browser via shell.openExternal and denies the popup; will-navigate prevents any off-app navigation of the window itself. Only http:/https:/mailto: URLs are passed to the OS.
• slr-file:// protocol: registered as privileged (secure, stream, fetch API, CORS-enabled). Handler resolves paths relative to projectDir with traversal guard (path.resolve + prefix check). Returns 403 for traversal attempts, 404 for missing files. corsEnabled is required, not cosmetic: the renderer's origin (dev server, or file:// when packaged) differs from slr-file://, so loading a PDF is a cross-origin request. Without it Chromium rejects the request before protocol.handle runs, and pdf.js surfaces the opaque failure as Unexpected server response (0).
• IPC handlers:
  • project:open — dialog.showOpenDialog → readFile
  • project:openPath — reads a file by absolute path (for re-opening recent projects); returns null if the file is missing or unreadable
  • project:save — writeFile to given path
  • project:saveAs — dialog.showSaveDialog → writeFile
  • project:setDir — sets projectDir from the project file's directory
  • project:pickSavePath — dialog.showSaveDialog to choose where a project JSON should live; writes nothing (the project editor picks a location before there is a file)
  • pdf:pick — dialog.showOpenDialog with multiSelections to add PDFs; returns absolute paths
  • pdf:pickFolder — dialog.showOpenDialog with properties: ['openDirectory'], then a recursive readdir walk collecting every .pdf (a directory that can't be read is skipped, not fatal); returns absolute paths
  • reference:pick — dialog.showOpenDialog filtered to .bib/.ris/.json; returns { text, name } for src/model/references.ts to parse, or null if cancelled
  • pdf:read — raw bytes of a PDF by absolute path, so the editor can read its title/authors. Deliberately not confined to the project directory (unlike slr-file://): the user may add PDFs from anywhere and picked them through a native dialog.
  • paths:relative — path.relative(dirname(fromFile), to) for each target, POSIX-separated. This is what makes a paper's pdf relative to the JSON, and what re-derives the paths when the JSON moves.
  • paths:absolute — the inverse of paths:relative: path.resolve(dirname(fromFile), rel) for each target. Backs absolutePdfPaths, used when importing papers from a screening project.
  • paths:sibling — path.join(dirname(sourceFile), fileName). Backs siblingProjectLocation, the default save location for a new project started from screening results.
  • app:setDirty — the renderer reports its unsaved-changes state (drives the quit dialog)
  • app:saveComplete — the renderer reports the result of a save it was asked to run before quitting
  • llm:configs / llm:saveConfig / llm:deleteConfig — the LLM targets in userData/llm-config.json. The renderer is handed publicConfigs(): everything except the key, plus hasKey.
  • llm:call / llm:abort — send a renderer-built request with net.fetch after substituting the real key and checking the target origin against the config's baseUrl (see AI-assisted annotation above). llm:abort aborts an in-flight call by its requestId, since an AbortSignal cannot cross IPC.
  • git:probe — git --version; { available, version, error }
  • git:pickCloneDir — dialog.showOpenDialog for the clone destination folder
  • git:clone — validates the URL/destination (src/git/url.ts), then git clone -- <url> <dest>
  • git:pickProjectIn — dialog.showOpenDialog with defaultPath: dir, the mechanism that opens the picker already inside a freshly cloned repository
  • git:info — rev-parse --is-inside-work-tree / --show-toplevel / --show-prefix, branch, upstream, whether HEAD exists
  • git:status — raw status --porcelain=v1 -z + diff --no-color HEAD --, parsed on the renderer side (src/git/output.ts)
  • git:headContent / git:workingContent — the project file's text at HEAD (git show HEAD:relPath) and on disk (a plain, side-effect-free read); the two inputs detectFieldChanges compares for the commit panel's field-level review
  • git:commitPartial — the write → add + commit → write-back sequence field-level commit review needs, since git has no native partial-file staging primitive; see "Field-level commit review" above
  • git:commit — pathspec-limited add then commit -m
  • git:push — plain git push; a missing upstream surfaces git's own message rather than inventing --set-upstream
  • git:pullBegin / git:pullFinish / git:pullAbort — the pull classification and its two ways to conclude; see "Git" above for the full command sequence
  • git:writeWorking — writes composeContents's workingOut directly to the working file, letting a reviewer discard local edits without committing (see "Field-level commit review" above)
• Menu: custom template with File, Edit, View, Window menus.
  • The Edit menu is hand-built: Undo/Redo send app:undo / app:redo to the renderer (routing to the store's history) rather than the native text-undo role, so undo works app-wide; cut/copy/paste/selectAll keep their native roles.
  • The View menu is hand-built (not the default { role: 'viewMenu' }) and deliberately omits zoom roles so that Ctrl +/-/0 reach the renderer for PDF zoom (and Ctrl+Shift +/-/0 for app font scaling) instead of triggering native browser/Electron zoom.
• Unsaved-changes quit flow: a window close handler (promptUnsavedChanges) intercepts the close/quit when isDirty is set, and shows a native Save / Don't Save / Cancel dialog. "Save" asks the renderer to save (app:requestSave) and closes once it reports back; "Don't Save" closes discarding changes. A before-quit flag lets the guard resume app.quit() after confirmation (so Cmd+Q fully quits on macOS, where destroying the window alone would not).

electron/preload.ts uses contextBridge.exposeInMainWorld('slr', ...) to expose IPC-backed methods: openProject, openPath (read file by absolute path), saveProject, saveProjectAs, setProjectDir, plus the quit/menu coordination — setDirty, onRequestSave, saveComplete, onUndo, onRedo — and the fifteen git* methods (gitProbe, gitPickCloneDir, gitClone, gitPickProjectIn, gitInfo, gitStatus, gitHeadContent, gitWorkingContent, gitCommitPartial, gitCommit, gitPush, gitPullBegin, gitPullFinish, gitPullAbort, gitWriteWorking) mirroring the git:* IPC handlers one for one. This window.slr object is the detection signal for isElectron().

Hooks

• useAutosave (src/hooks/useAutosave.ts): Periodically saves unsaved annotation changes every 5 minutes when autosaveEnabled is on (opt-in, default off, toggled from the Save menu). Skipped while the project editor is open, since save() here writes the annotation project state, not the editor's own draft. Only fires when dirty && !busy.

• useKeybindings (src/hooks/useKeybindings.ts): Global keyboard shortcuts registered on window.keydown. Handles open (Ctrl/Cmd+O), save (Ctrl/Cmd+S), save-as (Ctrl/Cmd+Shift+S), paper navigation (Alt+↓/↑, [/]), zoom/font (Ctrl/Cmd + +/=/-/0 → PDF zoom; add Shift → app font size), and help (F1). Paper navigation skips when typing in a field (unless Alt is held). Zoom/font detection matches e.key and e.code to handle numpad and international layouts; reset is detected by the digit-0 e.code (Shift-independent) to avoid the German Shift+0 → = clash. Copy/cut/paste/undo are left to the browser/Electron Edit menu. Paper stepping ([/] and Alt+Arrow) walks the filtered/searched paper list by querying the DOM for .paper-list [role="option"][data-paper-id] rows, falling back to the raw project order only when the sidebar is collapsed — so keyboard stepping and the visible list never disagree about what "next" means.

  Bare-key bindings are suppressed while anything blocking is on screen. The screening decision keys (I/E/U, 1–9) and paper navigation act on the paper behind a dialog, invisibly, so aModalIsOpen() gates them on a DOM query. The selector lists the blocking surfaces — .modal-overlay, .error-overlay, .menu — rather than asserting that every dialog renders a .modal-overlay, which was the original rule and was false twice over: a failed save renders .error-overlay, and an open Dropdown renders .menu, so pressing e behind a save-failure overlay, or typing the first letter of a project you were hunting for in the Open menu, silently excluded the current paper and auto-advanced. Modifier combos stay live behind a modal on purpose.

  Ctrl/Cmd+S commits the focused edit first, in the project editor. The schema field name and the screening reason label both hang their confirm-before-you-lose-answers guards on blur, and a keyboard save moves no focus — so without this, Ctrl+S was a way to land a destructive rename without ever being asked. Blur handlers and the zustand writes they make are synchronous, so the save sees the result, including a rename the reviewer just declined. Scoped to the editor: annotation fields have no such guard, and taking someone's cursor mid-sentence to save would be pure cost.

• useDirtyGuard (src/hooks/useDirtyGuard.ts): Browser only. Registers a beforeunload listener that calls e.preventDefault() when either the project or the editor's draft is dirty, triggering the browser's "unsaved changes" confirmation. Both, for the same reason useElectronCloseGuard's isDirty() checks both: an unsaved schema draft is no less lost on a tab close than an unsaved annotation, and while the editor is open it is the only thing on screen. It is skipped under Electron (isElectron()), because a beforeunload that returns a value there silently cancels the quit with no dialog — Electron handles unsaved changes via a native dialog in the main process instead (see useElectronCloseGuard and the quit flow below).

• useElectronCloseGuard (src/hooks/useElectronCloseGuard.ts): Electron only. Wires the renderer to the main process for a clean quit and for the Edit menu: it pushes the current dirty state to main (slr.setDirty), runs a save when main asks after the user picks "Save" in the native close dialog (slr.onRequestSave → save() → slr.saveComplete(ok)), and routes the Edit-menu Undo/Redo (slr.onUndo / slr.onRedo) to the store's undo() / redo().

Settings & Theming (src/state/settings.ts)

App appearance is controlled by a settings module that persists to localStorage:

• Theme: 'light' | 'dark' — defaults to OS preference (prefers-color-scheme) on first load. applyTheme() sets document.documentElement.dataset.theme. CSS uses :root[data-theme='dark'] selectors so the theme is user-controlled, not OS-only.
• Font scale: float from 0.7 to 2.0 (step 0.1). applyFontScale() sets the --app-font-scale CSS variable on <html>. The base font size is calc(14px * var(--app-font-scale)) and most text sizes use rem units so they scale proportionally. The PDF paper itself is always rendered on a white background regardless of theme.
• Pane widths: the left (paper list) and right (annotations) pane widths are persisted via loadPaneWidths() / savePaneWidths() (localStorage, clamped). App.tsx holds them in state, applies them as the workspace grid template, and updates them as the Splitter drag handles are dragged.
• Autosave: loadAutosaveEnabled() reads from localStorage (slr.autosave key), returns true only when the stored value is '1' — off by default. saveAutosaveEnabled(enabled) persists changes. The store reads the initial value on app start and exposes setAutosaveEnabled() to toggle it from the Save menu. useAutosave (see Hooks above) is the consumer.

src/main.tsx calls applyTheme(loadTheme()) and applyFontScale(loadFontScale()) before rendering React to avoid a flash of unstyled or wrong-theme content on startup.

App Initialization (src/App.tsx)

1. useKeybindings() and useDirtyGuard() are called at the top level.
2. On mount, checks ?project=<url> query parameter — if present, calls loadFromUrl(url) for server-deployment auto-loading.
3. Renders Toolbar always. If a project is loaded, renders the three-pane workspace; otherwise shows a welcome screen with an "Open project…" button and, if there are recent projects, a clickable list of them wired to openRecent(id).
4. ErrorPanel is always rendered (renders null when no error).

Build Configuration

vite.config.ts uses base: './' so the built SPA works from both a server subpath and file:// (for Electron). The ELECTRON=1 environment variable conditionally adds the vite-plugin-electron plugin which builds electron/main.ts and electron/preload.ts (preload emitted as .cjs since the package is type: module).

Change Guidance

• Adding a new platform operation: Add it to PlatformAdapter, implement in both ElectronAdapter and BrowserAdapter, then call from the store or components.

• Adding a new annotation field type: Update FieldType in schema.ts, emptyValue() in annotations.ts, and Field.tsx rendering. Consider validation in the zod schema. Also teach the AI layer about it: isUnanswered() in src/llm/fields.ts, coerce() in src/llm/parse.ts, and the type rules in src/llm/prompt.ts.

• Adding a new LLM provider: Add it to Provider in src/llm/types.ts and to PROVIDERS in src/llm/providers.ts (base URL, whether it is editable, whether it can take a PDF, tokenParam), then handle its request shape in buildRequest and its response shape in extractText / extractError. Also add a branch to buildModelsRequest / parseModelsResponse in src/llm/models.ts — its list-models endpoint, auth, and pagination scheme are almost never identical to another provider's chat endpoint even when the chat shape is "OpenAI-compatible" (Mistral's /v1/models returns a bare array, not {data:[...]}; OpenRouter paginates, Groq doesn't). Decide reasoning-effort support last, and only from what you can confirm — a model-ID pattern if the provider's own docs name specific models, or nothing at all (null) if you can't confirm the field/values, per the DeepSeek/openai-compatible precedent above. Nothing else needs to change — PROVIDER_LIST drives the settings dropdown and the "which providers can take a PDF" hint automatically, so the platform and UI layers are provider-agnostic.

  Verify against the provider's own current docs before writing buildRequest, not against another provider's shape. The bug that made tokenParam a field in the first place: this app originally sent max_tokens to every OpenAI-shaped provider, and OpenAI's own API now rejects that on current models — "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead." xAI and Groq have followed the same rename (both have reasoning-model variants, which is what drove OpenAI's original change); OpenRouter, Mistral, DeepSeek and generic self-hosted OpenAI-compatible servers still document max_tokens and do not confirm the newer name. Four vendors that all claim "OpenAI-compatible" disagreed on one required parameter — assume the same is possible for endpoint paths, PDF support, auth header names, list-models shapes, and reasoning-effort contracts, and check each freshly. supportsPdf in particular is not "does the vendor support PDFs at all" but "can a single request carry the bytes inline the way this app's PaperPart does" — Mistral's PDF support needs a fetchable URL (no inline-base64 form in chat completions), and xAI's needs an upload-then-reference flow across two requests, so both are false here despite the vendor supporting PDFs some other way.

• Changing the schema format: it is described to the model in SCHEMA_FORMAT_DOC (src/llm/prompt.ts), which mirrors docs/annotation-schema.md §3 by hand. Both must move together, or the model reads an unfamiliar schema against a stale description.

• Adding a new keyboard shortcut: Add to useKeybindings.ts. Check isEditable() if it should be ignored inside input fields.

• Changing the Electron IPC surface: Update preload.ts (the SlrBridge interface), electron/main.ts (IPC handler), and src/platform/electron.ts (adapter method). All three must stay in sync.

• Adding a git operation: add the new operation as an enumerated handler in electron/main.ts (never widen the renderer's power to name an argv — see "The renderer never names an argv" above), the matching bridge method in preload.ts, the method on GitPlatform in src/git/types.ts, and the pass-through in ElectronAdapter's git object (src/platform/electron.ts) — a private readonly field, not a fresh object literal per getGit() call, since getPlatform() is a singleton and a new object every call would make every useGitStore selector see a "different" platform and churn. Keep the decision (what the operation does, what it validates) in src/git/ where it is unit-testable; keep the argv in electron/main.ts where it is enforced.

• Adding a new settings/appearance option: Add to src/state/settings.ts (load/apply functions), add state + actions to the store, add UI controls to Toolbar.tsx, and add keybindings if needed.

• Touching anything that reads or writes the current paper's annotation data: route it through currentTree() (src/state/store.ts) rather than paper.annotations directly, or it will silently ignore reviewer selection on a multi-reviewer project. Grep for .annotations across src/ and check each hit against "should this see the active reviewer's tree or always the consolidated one". The direct reads that remain outside src/model/ are deliberate: currentTree()'s own body; ConsolidationDialog, which reads the consolidated tree on purpose (it is showing you what the final answer currently is); editorStore, which edits the file's papers and has no reviewer concept; and null-project fallbacks.

• Reading a screening project's schema: always go through loadProject/Project.schema, never the raw file. config.screening.reasons is the single source of truth; config.schema in the on-disk JSON is a projection of it that loadProject ignores and serializeProject rewrites on every save (see Screening above). A tool that reads the raw JSON's config.schema directly (rather than through this app) will see the derived snapshot from whenever the file was last saved, not something it can trust to reflect a hand-edited reasons list until the file is re-saved.