Repository navigation
CACCHT Explorer (cfabric-browser): reading, searching and comparing corpora - #18
Merged
Merged
Conversation
S.search() and S.study() accept a CancelToken (or a timeout shorthand). The engine checks the token between study phases, while spinning atoms and edges, inside quantifier sub-searches, and once per candidate of the outermost stitching loop, raising SearchCancelled / SearchTimeout. This is groundwork for the browser, which must be able to stop runaway queries. Also corrects the docstrings that said a generator is returned when a limit is given; it is the other way round. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CorpusRegistry holds named corpora, loads each at most once (also under concurrent access), supports lazy loading via register(), and exposes a per-corpus search lock, since S keeps per-query state on the Api. The MCP CorpusManager now subclasses it, keeping its existing behaviour that load() replaces a corpus of the same name. The browser will share the same registry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The display layer that Context-Fabric was missing compared to TF's tf.advanced, without any HTML: - DisplayConfig reads a corpus' TF app config.yaml (found next to the data) and computes TF's defaults for whatever it leaves open: levels, labels, feature lists, hidden/base/condense types, writing direction, section separators, none values, provenance. - unravel/render_node/render_tuple build render trees: the nested fragments of a node, cut at the boundaries of other nodes, dressed with labels, texts, section headings, features and highlights, and serialisable to JSON for the browser. - section_ref/node_from_section_ref handle references like Genesis 1:1 with the corpus' own separators. The unravel algorithm is a port of tf.advanced.unravel. Checked against TF on 3944 random nodes (all types, pretty/plain, folded/unfolded) of BHSA, DSS, LXX, CUC, Descartes and Peshitta: all trees identical. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A new package, libs/browser, with the server half of the Context-Fabric browser (phase 1 of the browser plan): - cfabric-browser command: corpora by name (bhsa, ETCBC/dss), TF-layout repository or data directory; binds to 127.0.0.1. - FastAPI JSON API: corpus cards, features, section tree, passages and nodes as render trees, node inspector, template check, search with server-sent status events, paging, results in context, TSV export. Endpoints mirror the MCP tools where they overlap; no server paths are exposed. - SearchPool: searches in spawn-based worker processes that share the memory-mapped corpora. Parallel, cancellable (cooperative token, with a watchdog that restarts unresponsive workers), time limits, result cap, streaming batches, canonical sort when done. - Structured template problems (line, message, feature/type). Problems inside quantifiers are only logged by the engine (as in TF), which then returns results as if the quantifier excluded nothing; the browser reports them as errors instead. The interface (a Vite/React app) follows; until it is built, / shows a placeholder and the API docs are at /docs. CI gets a test-browser job. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Vite + React + TypeScript interface for cfabric-browser, built into cfabric_browser/static (not in git; CI and the release workflow build it, and the publish job checks it is in the wheel). - Read: book and chapter navigator, passages as running text or as structure (pretty display), previous/next across chapters, opens at the start of the corpus; right-to-left text set in its own script. - Search: CodeMirror editor for TF templates with highlighting, autocomplete of types and features, and live error marking; live status with Stop; virtualised result list that loads pages as you scroll; results unfold in context with hits gilded (deeper gold for a hit inside a hit); recent searches; TSV download. - Inspector: all features of a node (empty values and internal features summarised), relations, what it lies in and contains. - Pretty displays draw render-tree edges with meaning: solid where a node starts or ends, dashed at a gap, open where it continues. - State in the URL (corpus, view, passage, template, search, result); display preferences per browser; light and dark themes; works at phone width. - Fonts bundled for offline use (Source Serif 4, JetBrains Mono, Noto Serif Hebrew, Gentium Plus, Noto Sans Syriac, Amiri); Ezra SIL or SBL Hebrew are used when installed. Server changes: passages report their section path and previous/next sections; the node inspector leaves out omap@ version-mapping edges. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The results TSV gets, per node of a result, a column for each feature mentioned on that node's line of the template; `sp*` adds a feature without narrowing the search, as in Text-Fabric. `?excel=true` writes UTF-16 with a byte-order mark, which Excel opens with Hebrew and other scripts intact. The interface offers both downloads and explains the `feature*` trick. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a corpus is loaded from .tf files (its first load, before the .cfm cache exists), feature metadata uses TF's `valueType` key, but Text read only the .cfm key `value_type`. Every section level was then typed `str`, so references such as "Genesis 1:1" could not be resolved to nodes in that session. Read both keys; the regression test loads a fresh copy twice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Book concordance (books.py): book names across corpora mapped to OSIS
ids (1_Samuel, 1Sam, Samuel_1, ...). Only safe mappings: e.g. the LXX's
2Esdr (Ezra + Nehemiah) and Theodotion's Daniel are not mapped.
- GET /api/corpora/{c}/align?from=&ref=: the counterpart of a passage in
another corpus, falling back to the enclosing chapter (exact: false)
when a verse is missing there, as with 1 Sam 17:12 in the LXX.
Verses are matched by number; versification differences (e.g. the
Psalms in the LXX) are not corrected yet.
- Library: GET /api/library lists corpora found under ~/text-fabric-data
and ~/github; POST /api/corpora opens one by library id (never by
path) and tells the search workers about it. --no-library switches
this off for servers reachable by others.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Workspace of up to four panes side by side, each with its own corpus, view, passage, template and search; all in the URL (c0, v0, ref0, ...), and older single-view links still open as the first pane. - Sync passages: opening a passage in one pane opens the corresponding passage in the other reading panes via the alignment API, with a note when only the chapter exists there (e.g. 1 Sam 17:12 in the LXX). - Open a corpus: a library dialog lists corpora found on this computer and opens one in a new pane. - Search all corpora: runs the template in every open corpus, shows the live count per corpus (or why the template does not fit), and opens any of them in its own pane. - Narrow panes fold the passage list away and put the search editor above the results; result rows no longer repeat a verse heading as text when the reference column already shows it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`DT-UCPH/sp@7.2.4` loads that version, `sp@latest` the newest on this computer; without `@` the version named in the corpus' app config is used, as before. Asking for a version that is not on disk lists the versions that are. The library's open endpoint accepts the same syntax, and the corpus card now reports the data version actually loaded rather than the app config's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings in ~/.cfabric/display/<org>/<repo>.yaml, or for one version <repo>@<version>.yaml, are merged over the corpus' Text-Fabric app config. This covers corpora whose app config lags behind their data (SP 7.2.4 has clause and phrase layers that its app config does not describe yet) without editing downloaded files, which Text-Fabric replaces on updates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Results / Statistics switch in the search view. Statistics count, over the results (or the results so far while a search runs): - the values of a feature on one part of the result tuples, by distinct node; by default a feature the template leaves open (`lex*`) rather than one it fixes (`sp=verb`); "no value" and the long tail are shown as muted bars; - the results per book or chapter, in text order, by the section of each result's first slot (so nodes that cross sections still count); bars link to the passage. Charts are single-series horizontal bars in a chart blue derived from woad and checked for lightness, chroma and contrast in both themes, with counts at the tips and a table view. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
export.tsv takes `extra=<part>:<feature>,...` (repeatable) to add feature columns for one part of the results, after the features the template mentions; unknown parts or features are refused. In the interface, the two download links become a "Download..." dialog listing the features of each part (template features ticked and fixed), with a filter, and a choice between TSV and TSV for Excel. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Workspaces menu: save the current workspace (its URL: panes, corpora, passages, templates, sync) under a name in this browser; open or delete saved ones; export to and import from .cfabric.json files. - Import also accepts Text-Fabric browser job files: the query opens in a search pane and runs, the first section in a reading pane, provided the job's corpus is open (otherwise it says which one to open). - Panes rerun their template when the server no longer knows their search id (searches live in server memory), so saved workspaces and shared links keep working after a restart. - The top bar wraps its buttons to a second row on narrow screens instead of breaking their labels. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After `sp=` (or `sp=verb|`) the editor offers the values that occur, most frequent first, with their counts. Hovering a feature name shows its description, the object types it is on, how many values it has and the most frequent ones; hovering an object type shows its count. Feature descriptions come from the existing feature endpoint and are cached per corpus. Tooltips are built with textContent only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A "Cite" button in the reading header and in the search status line opens a citation to copy: the passage, or the search template with its number of results; the corpus with title, version and DOI; and the browser and context-fabric versions with the date. /api/health now reports the installed context-fabric version for this. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Copies of one corpus (same app title, e.g. DT-UCPH/sp and its fork saulocantanhede/sp, or DT-UCPH/cuc and cuc_) are listed once, keeping the one with the newest data; the others are named as copies and can still be opened by id. - Each entry carries the version that opening it loads, and the dialog shows that one instead of the newest on disk. - Corpora can be hidden from the list with `library: hide:` in ~/.cfabric/browser.yaml or with --hide ORG/REPO. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Where slots are below word level (SP, DSS), a click in the plain text opened the sign under the cursor. Now the base-type fragment (word) is the click target, and base words in the pretty view are clickable too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With "Show differences" on, reading panes are compared with the first one, verse by verse, on their consonants (g_cons_utf8, or the letters of the original-script text). Vowels, accents and word division don't count. Words that differ only in matres lectionis are marked as spelling, other differences as variants; a plus is marked in the text that has it. New endpoint: GET /api/compare. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The DSS has no book, chapter or verse sections; its biblical words carry book/chapter/verse features. A book index turns these into a second way to read the corpus: per verse, the words of each scroll that contains it. New endpoints /books and /witnesses; align can target the book view. Reconstructed signs are flagged so the client can dim them, and text the DSS also lists as non-biblical (biblical=2) is labelled. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A DSS reading pane gets a "Scrolls | Bible books" switch. In book view the navigator lists the biblical books and chapters, and each verse shows the text of every scroll that contains it, with reconstructed text dimmed and a label for text also listed as non-biblical. A scroll name opens its fragment in the scroll view. Synchronised reading aligns book view panes with MT and SP. The pane's choice is kept in the URL (by1=books). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With "Show differences", a DSS pane in book view is compared scroll by scroll with the MT or SP pane. Reconstructed letters keep the texts aligned but never count, and letters a fragment lacks count only between two preserved letters, not where the text is lost at an edge or next to a reconstruction. The MT/SP pane marks a word where any scroll differs, with a tooltip naming the scrolls. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The DSS transcription has placeholder signs: ε (type missing) for lost text that is not reconstructed, and # and ? (type unc) for letters too damaged to read. They are now shown as […] and ◦, also in the scroll view, and the book view explains them. In comparisons with MT and SP an illegible letter counts as unknown, so MT is not marked against it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The DSS transcription marks morpheme breaks with a sign of its own (type sep, ׳), as in עמ׳י. They are not in the scrolls, so the browser no longer shows them, in the book view and the scroll view alike. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With MT, SP and a DSS pane in book view open, each place where MT and SP differ is read in every scroll that has the verse, between two preserved anchor letters: the scroll agrees with MT, with SP, or with neither, or is left out where its text is lost or reconstructed. Differences in the same words are one place, so a scroll reading מארות against MT מארת and SP מאורות agrees with neither. The DSS pane lists the places per verse with a chip per scroll; the marked words get tooltips. New endpoint /api/agreement. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With Sync passages on, a reading pane that has no passage yet (a short link that names the passage only once, or a newly added pane) opens the counterpart of the first pane that has one, so links like ?c0=bhsa&ref0=Deuteronomy+5&c1=sp&c2=dss&by2=books&sync=1 work. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…th screenshots made by a script Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… moves from the release notes to the guide Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ned; Reader's panel no longer covers the menus Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…elling in the lexicon Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… their vowels Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…he guide, a release checklist, and a test that the guide's labels exist Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…itself, read-only tools, the app's name on proposals - cfabric-browser mcp answers initialize, ping and the lists of tools, prompts and resources from the same server definition, and finds or starts the Explorer in the background; only a tool call waits for it (Codex gives a server 10 seconds to start). - All tools but propose_search carry readOnlyHint, so apps ask only for the proposal. - A proposal without model is named after the app (clientInfo.name via X-Cfabric-Client from the helper and the Claude Desktop extension, or the User-Agent). - A trial that runs out of time says how to narrow the template. Phase 2 of chatgpt-plan.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rrectly - scripts/ask_questions.py asks Claude Code (claude -p) or Codex (codex exec) each request of the question set through a running Explorer, runs each proposal in full and scores it. Claude: 30 of 30. - An atom under /have/ is a new one, so /where/ word /have/ word sp=nmpr only asks that one proper name exists. The BHSA guide, the user guide (Searching) and cfabric-mcp's syntax guide now name the atom (w:word ... w sp=nmpr) and give the /without/ form; the subjects of proper names only are 5,893, not 10,484. - Question set: 29 had the same mistake (22,447), 27 counted pairs of words instead of verses (60); `also` holds other right readings (15, 26, 28), checked against the BHSA like the rest. - README: connecting the ChatGPT desktop app and Codex; the runner. Phase 3 of chatgpt-plan.md (Claude side). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The chapter on asking Claude says how to connect the ChatGPT desktop app (Settings → MCP servers, STDIO, the program with mcp) and Codex (codex mcp add), with a longer tool_timeout_sec for the first question, and that ChatGPT on the web and phones can't reach the program. Release notes: asking ChatGPT, the corrected "only proper names" example, the helper that no longer gives up while the Explorer starts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The desktop app's Add server form may have no field for arguments, and the Mac program's path has a space, so the guide now gives the address (Streamable HTTP, http://127.0.0.1:8107/mcp) and codex mcp add --url; the config.toml block with command, args and tool_timeout_sec for letting ChatGPT start the Explorer, with the Mac path and how to open ~/.codex; that the desktop app may offer it only when working with Codex; and the sentence about waiting longer rewritten. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Server: refuse requests from other web sites and foreign host names (guard.py); the search watchdog no longer hands a queued search to the worker it ends or overwrites a finished search, and a corpus that fails to load no longer restarts workers without end; /api/corpora no longer waits for a corpus being prepared; translation builds keep the old one when they fail and replace only their own folder, and the library skips half-built folders; partial downloads are removed; the stdio helper waits before starting the Explorer again; the lexicon's caches are locked. Interface: the Reader shows the right chapter after > then <; Reader, Interlinear and Concordance no longer hang on a corpus that isn't open; the Interlinear scrolls to the URL's verse once; the first page after a change of order or search is fetched directly; search history and pane state no longer carry over; a failed interface switch says so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The secret in ~/.cfabric/running.json protected only POST /api/proposals, not /mcp as its documentation said. Other web sites are now kept out of all routes by guard.py, and a secret that changes at every start can't be given to apps that connect by address. running.json keeps the address; the screenshot script no longer needs CFABRIC_SECRET. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Interface: Panes and the template editor load when first shown (278 kB at the start instead of 792 kB); the hovered word lives in a small store that only the card follows, so the text isn't drawn again on every hover; corpus data stays fresh once fetched (staleTime), while searches, the library and jobs are still asked again when shown. Server: the Concordance is extended with new results instead of rebuilt per batch; innermost nodes are found from node spans instead of slot sets (book/word: 110 s -> 0.2 s for 3,000 results), shared with Parallel's search results (Spans); the three-way comparison aligns each scroll once per verse; the DSS book index asks for scroll names once per scroll (7.7 s -> 1.3 s); per-corpus indexes are built once for requests that come together. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
api.py was one function of about 1,700 lines. Its routes now live in cfabric_browser/routes/ (library, reading, parallel, search, proposals, lexicon), each a build(ctx) on the shared Context in routes/common.py, with their bodies unchanged. The OpenAPI schema is identical (54 operations, same parameters and docs). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…times net.py: certificates, requests under the browser's name, checksums and the checked download with retry, for corpus downloads, translations and dictionaries (three copies before; translations' private _ssl_context was imported across packages). words.py: the word type, the words of a node, the text format in the original script, and BookSpans (the book of a word by bisection), for the comparison, the Concordance, the Interlinear and the lexicon. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Frontend: useUrlState replaces five copies of the URL-state code (and no longer writes the history inside a state updater); useHealth replaces four raw fetches of /api/health; useCloseOutside replaces six copies of the close-on-click-outside/Escape effect; corpus cards are fetched once per corpus (corpusCard on one shared queryClient), also when switching interfaces; openReference and Tree, unused, are gone. Server: CorpusRegistry keeps register, get, get_api, list_corpora and current, what the browser and cfabric-mcp's tools use. Tests of the search pool for cancelling queued and running searches and for which finished searches are kept. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
caccht-explorer is the same command as cfabric-browser under the app's name. The extra desktop is split into window (pywebview, certifi) and build (also PyInstaller), so that context-fabric[explorer] can depend on cfabric-browser[window] without bringing PyInstaller. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
codykingham
approved these changes
Nov 2, 2026
codykingham
left a comment
codykingham
left a comment
Collaborator
There was a problem hiding this comment.
This PR offers a valuable addition with the explorer application, along with several CF bugfixes. LGTM!
MartijnNaaijer
marked this pull request as ready for review
November 2, 2026 15:26
This was referenced Nov 2, 2026
tonyjurg
pushed a commit
to tonyjurg/FORK_context-fabric
that referenced
this pull request
Nov 2, 2026
The ruleset on master (since 2026-10-28) only accepts changes through pull requests, so the bot's direct push after merging Context-Fabric#18 was refused. The regenerated docs now go to the branch auto/api-docs with a pull request, which later runs update. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds CACCHT Explorer (package
cfabric-browser, inlibs/explorer). It is the successor of the Text-Fabric browser: a FastAPI server with a React interface for reading, searching and comparing corpora side by side, offline. It also runs as a standalone program for Windows, macOS and Linux. The program has been tried as a series of pre-releases, fromdesktop-v0.1.0-test.1todesktop-v0.1.0-test.16.This is a draft: a few things need agreement first (see To agree on below). It is 112 commits, built on a chain of branches (
browser-phase1→browser-parallel→browser-interfaces→browser-jps→browser-downloads→browser-bdb→browser-rust). If you'd rather review it in parts, I can split it along that chain.What it does
Context-Fabric/bdbv1.0, CC BY-NC 4.0.cfabric-browser add jps1917).cfabric-browser download./mcp, andcfabric-browser mcpover stdio, also as a Claude Desktop extension). An AI app such as Claude or ChatGPT writes a search template and proposes it; the user sees it and runs it. Results always come from the CF search engine. Tools are shared withcfabric-mcp. A BHSA guide with 42 worked examples comes with it./guide/and on caccht.github.io/explorer/.Works on either core
The Explorer uses only the public API, so it runs on the Python core (0.5) and on the Rust core (
context-fabric==0.6.0rc1). Sincebrowser-rust, a few pieces that used to be added to the core live in the Explorer instead: display (render trees, a port of TF'stf.advancedwithout HTML), the corpus registry and cancellation. Searches go throughS.search, and the worker process checks for cancellation between results. The standalone program is built on 0.6.0rc1.Changes outside
libs/explorerlibs/core:T.sectionFeatureTypeswasstrfor every level on the first load from.tffiles, so references likeGenesis 1:1failed in that session.S.search/S.fetchdocstrings hadlimitthe wrong way round.libs/mcp:tools.using(registry)lets the tools run on another registry of corpora, here the Explorer's.search_syntax_guide("quantifiers")explained/where/ … /have/wrongly..github/workflows/ci.yml: atest-browserjob with a matrixcore: [python, rust]..github/workflows/publish.yml:browser-v*tags publishcfabric-browserto PyPI. The job checks that the built interface is in the wheel..github/workflows/browser-desktop.ymlbuilds the standalone program on Windows, macOS (arm64 and x64) and Linux, and runs five packaging tests on each: smoke, translation, download, dictionary and MCP.desktop-v*tags publish a GitHub pre-release..gitignore:.cfr/(the Rust core's cache) and the built interface (libs/explorer/cfabric_browser/static/).site/: a docs page Explorer (/docs/explorer).To agree on
context-fabric0.6.0, and keep it working on 0.5 for now?explorerextra in the Rust core'spyproject.toml:explorer = ["cfabric-browser[window]"]. Thenpip install "context-fabric[explorer,mcp]"gives everything. This belongs oncf-rust, not in this PR.cfabric-mcp0.2 withtools.using. The Explorer needs it, but its requirement still sayscfabric-mcp>=0.1.7, which lacks it. That requirement should become>=0.2once 0.2 is out.cfabric-mcp0.2, thencfabric-browser0.1.0 (never published yet), thencontext-fabric0.6.0 with the extras. After that, trypip installin clean environments on all platforms in CI.Tests
cd libs/explorer && pytest tests) and the frontend's Vitest tests (npm test).npm run buildtype-checks.tests/fixtures/) and corpora written in temporary directories; no real corpora or network are needed.CFABRIC_REAL_CORPORA=1adds a question set of 30 BHSA requests for the AI search guide.ci.ymlruns on this work. The branches only ranbrowser-desktop.yml, which is green on all four platforms for this commit.Trying it
Or download a standalone program from the latest pre-release,
desktop-v0.1.0-test.16.🤖 Generated with Claude Code