更新日志 · Ontology Atlas

2026-10-15 · Unreleased: changes since the latest release

Changed

Compact agent briefs now give a coding agent either channel as a complete handoff: the text carries the first repair call with its guards, the safety rule and the exact full-detail call, and the structure carries the typed facts once. Together they stay within 12,000 bytes, which leaves room for recorded qualifier units that did not fit before.
Reading columns in a window 1536px or wider (documents, Library pages, guides, check reports) run a 743px line instead of 629px, so a 1920px monitor shows about 102 Latin characters or 54 Hangul syllables per line and the column fills about half of its pane; windows under 1536px are unchanged.

Fixed

Compact task handoffs select capability ownership and exclusions from authored Markdown paragraphs and list items. Code examples, HTML/comments and blockquotes cannot create or cancel a capability claim, and literal-only sections cannot fall back to unrelated prose. Real legacy fields, nested boundaries, localized claims and the selected Definition statement retain their authored meaning.
Initial repository analysis excludes literal Markdown examples, comments and quotations from authored claims, and exposes omitted excerpt text through an exact hash-guarded next read with review-required trust.
The MCP bundle (.mcpb) that desktop hosts install now ships every TypeScript module as plain JavaScript, so the analysis-record tools load on a Node that does not strip types, such as Node 20 or 22.
Compact handoffs retain their exact full-body read call when source navigation is ready, so omitted conditions can be recovered before inspecting code, including when the summary budget omits every qualifier.
Updated the TLS dependency rustls to 0.23.46, which stops leaking buffered data or TLS secrets, bounds certificate decompression before authentication, and prevents a downgrade to classical key exchange (RUSTSEC-2026-0346, -0347, -0348, -0349).
Compact task handoffs preserve whole uncertainty statements with per-document provenance and omission counts, retain system gaps, and include project-level context in exact full-body recovery. CLI accepts the same 12,000-byte serialized payload as MCP and rejects inconsistent coverage metadata; capped or failed reads remain incomplete.
MCP bundle builds copy nested modules declared as package directories, preserving the verification modules needed by the packaged server.
Compact agent_brief fits its 12,000-byte budget again on the project's own vault: its meaning repair keeps the status, questions, provenance, write policy and first review call, and names the omitted later steps and stop-condition count with detail=full as recovery; full detail is unchanged. A CLI from this release reads compact and full responses from older MCP servers, while an older CLI refuses the new compact shape instead of misreading it.
Repository analysis and import inference in the MCP server no longer stop with an error when a declared workspace package has no name in its package.json; that package is now named after its folder, and two such packages in same-named folders are told apart by their full paths.
Analysis can compare Git dates when an ontology vault and its connected code live in separate repositories. It keeps document and source paths separate, offers connected-folder access and retry recovery, and leaves changed connections or incomplete evidence unknown without accepting meaning or rewriting the connection.
Compact task handoffs keep prose conditions, exclusions, and evidence limits with the statements they qualify, including across paragraph breaks. Oversized sections retain explicit omission and full-body recovery.
Compact MCP and CLI task navigation now reads authored Markdown structure before verifying exact source coordinates. Code examples, HTML/comments, quotations and nested coordinate examples cannot become work targets; example-only anchors no longer hide a later authored anchor. Scope boundaries and runner manifests exclude literal blocks while retaining genuine nested boundary conditions.
Fix code location opens the exact affected concept and recorded relative path; stale or foreign targets require remeasurement, and copied handoffs omit the private source root.
Guided map tours now highlight the requested node on their first frame, including with reduced motion, instead of briefly highlighting a previous node or the screen corner.
Compact agent handoffs include every retained uncertainty item in their text response, so source limitations and unrun tests no longer disappear after the first item. Existing excerpt and response-size bounds remain in place.
Agent setup now names the resolved vault in post-change sync guidance, including separate vault and code folders.
When a declared dependency has no witness, the MCP server now asks for the file that imports or calls the target and the symbol it uses, instead of a file and line number, so the repair it suggests no longer goes stale when lines above it move.
Project meaning checks identify a parseable document edited since finalization as requiring review, retain source and malformed-input safeguards, and guide explicit re-finalization without accepting incomplete answers.

2026-10-10 · v1.7.0: Analysis that shows its evidence

Added

Analysis can search responsibility pairs and capabilities by name without losing the graph and evidence being checked, and reveals choices six at a time with collapse and clear or Escape to recover.

Changed

Analysis opens on one responsibility pair, a bounded diagram of the capability dependencies recorded between them, and the selected declaration's evidence, replacing repeated map exploration and the global repair catalog; the selected claim stays in view through implementation inspection, its source document and an explicit correction.

Fixed

Reading a Flow request by keyboard gives it a concise name and a focus frame that the text fade no longer dims.
Copying an agent proposal for apply or the terminal includes only the selected files, reports success only after the copy finishes, and counts line changes from each file's original to final text.
Saved Flow answers recheck the evidence they cite, so an answer with unchanged evidence no longer stays marked unknown, and an older refresh can no longer replace a newer one.
The Analysis page keeps its title stable while loading and apart from the section tabs, scrolls only its reading area, keeps touch actions clear of the mobile navigation, and centres the sample empty state.
Analysis draws a capability's recorded elements as connected branches; each element opens its own document and returns to the same capability, and records without a link get no invented edge.
Running Atlas from source uses Next.js 16.3.8, which fixes six advisories, and the development server no longer exposes Next.js's own MCP endpoint to the local network.

2026-10-06 · v1.6.1: Smoother maps and reliable renderer cleanup

Changed

The map footer shows whole-map or current-scope concept counts, identifies synthetic performance data, and removes repetitive hints and the generic Flat legend.

Fixed

Returning from Library releases renderer-owned texture listeners instead of retaining detached canvases and panes.
Large maps avoid unnecessary drag physics and reuse stable 3D geometry while preserving concept positions and rendering.
Concept details reuse graph facts and count contained children in one edge scan when only document content changes.
Update TOML and source-map parsers to patched versions to avoid known denial-of-service cases.

2026-10-05 · v1.6.0: Domain structure, a clearer Flat overview and Apple Silicon only

Added

The hex board can tilt into a relief: the ▲ Relief chip or Shift and a vertical drag raises each capability by how many concepts declare they depend on it, with the number printed on its face, and the board remembers the pose.
At the overview, Galaxy draws one counted strand between each pair of domains that depend on each other.

Changed

The flat map now places each domain on a ring by how many other domains depend on it, with the project at the centre, and the lines between domains count their dependencies. Capabilities, then elements, appear as you zoom in. Hover a domain to split its lines by direction; zoom into one to see every capability named.
Galaxy now draws each domain as a galaxy of its own around the project, stays smooth at 10,000 concepts, and rests 12 seconds after your last input; Auto-arrange settles the sky again.
macOS releases now ship for Apple Silicon only: v1.5.0 is the last Intel build. It stays on its GitHub Release but receives no further updates, and on an Intel Mac a manual update check reports a connection failure. The Windows x64 beta continues.

Fixed

When a capability is the first thing selected on the map, the camera now comes to rest on its frame instead of trembling just short of it, so the map goes back to sleep.

2026-10-03 · v1.5.0: Japanese and Chinese, settings by where values live, and a steadier map

Added

Settings, Map has Drag speed and Zoom speed: five steps from 0.5x to 2x, 1x by default, kept on this computer. Drag speed sets how far the map moves for each pixel you drag in every view; Zoom speed sets how far one wheel notch, one pinch or one + or - press zooms.
The download page tells what Atlas does in one moving illustration under the first screen: a project, its domains, capabilities and elements, and an agent's write waiting for your approval. It holds still under reduced motion.
The app, the website and the download pages are available in Japanese and Simplified Chinese (/ja/, /zh/), chosen from the language switch or the browser's language. MCP, ACP, API key, Git, Markdown, frontmatter and product names stay as written in every language.
The website and the desktop app now carry the full license text of every third-party package they ship, at /third-party-licenses.txt and beside NOTICE.md in the app, and the download page footer links to it.
A schedule in Automations can be edited in place: Edit opens its own sheet on it as it stands, and saving keeps its id and run history, recomputes its next run, and asks this Mac to allow the new version when the edit lets it run more often or on more days, or read or write somewhere new.
Settings is grouped by where each value lives (This computer, This folder, App) and gains search, ⌘, to open it, text size, a Privacy · data pane listing what can leave this computer, an About pane on both surfaces, and a switch for the automatic update check.

Changed

Opening a large folder draws the map while it is still being read: the project and domains first, then capabilities and elements, with the INDEX counting the documents read; the app reads 64 Markdown files per call, so a 10,000-document vault shows its map in under a second instead of after 78 seconds in the test harness.
Settings panes start their title, group labels and cards on one left line; the left list's group labels stand apart from the rows; Privacy · data lists what can leave this computer as aligned columns (when, folder content, record, destinations); and the Korean copy uses the everyday word for version instead of a forced translation.
The hex board stays smooth at 10,000 concepts.
Technical terms such as MCP, ACP and API key stay as written instead of being translated, and an info mark beside them shows the full name and a one-line explanation in the current language.
The Harness Architecture tab, before any architecture profile exists, now shows a labelled, looping illustration of the reviewed layer profile beside its two actions, drawn with the connected source folder's real top-level folder names when the installed app has one.
The hex board draws the strokes between regions only when zoomed out to regions or with "Region names only" on; closer in, relations route only for a capability you hover or select, or for what a selected region needs.
A wheel notch or a + or - press now lands in about a tenth of a second instead of three tenths, a flick glides a third as far and stops without bouncing back, and in Strata the rings turn with the pointer instead of trailing it, still moving every frame under a mouse slower than the display.
Picking a view that leaves or enters Territories or the Hex board carries each concept from its old place to its new one instead of crossfading the two views; under reduced motion the old view fades out on the fast clock.

Fixed

Switching the map between Flat, Territories, Hex and Galaxy crossfades the two views instead of cutting, and returning to Flat from Galaxy or 3D no longer jumps in brightness as the camera lands.
Library check history's door to Automations opens the round that is selected there instead of the first schedule in the list, and the Documents lane's link back to Library check history opens with that round selected.
A round refuses an ontology change at once and names the tool, instead of waiting out its limit. A document round writes only drafts, judged as the edit tool applies each write, against the page as earlier writes left it, under any spelling of its path, and never over a node, through a symbolic link, or into a folder it did not find. After each run Atlas puts back, or removes, any page its own write left that is not a draft, keeping the text it takes out and naming the page in the run; a page changed after the run wrote it is left as it is, failing the run if it is not a draft. A service round can write the page for a document it brought in during the same run.
Galaxy's twinkle, drifting gas and shooting stars now slow to a stop about 30 seconds after the last input, as the map's comets do, so a Galaxy left open stops drawing frames instead of animating forever; any input wakes it.
Scheduled checks and reviews keep to this Mac's wall clock: on a daylight-saving night a five-minute schedule no longer goes an hour without a run, a two-hour schedule no longer drifts off its hours, and a 24-hour schedule no longer skips the day whose midnight does not exist; after the Mac moves to another time zone a daily 09:00 schedule runs at 09:00 there instead of at the old zone's hour, and that time is saved.
A large Galaxy pans more smoothly: names are placed against an index of the stars on screen instead of being checked against every star, and the Flat outlines hidden under the stars are no longer painted. At 10,000 concepts a zoomed-in pan fell from about 105 ms to under 30 ms of script per frame.
A trackpad pinch zooms the map, the hex board included, exactly as far as your fingers spread, even fast on a busy map: in the browser it used to zoom about a quarter as far, and the app now reads the pinch WebKit reports. Ctrl with a mouse wheel still zooms one notch at a time.
The hex board no longer slides a few pixels about three seconds after it opens, and panning no longer rewrites its accessible tile list on every frame.
A schedule paused while it waited behind another run no longer runs when that run ends; resuming a schedule whose time passed while it was paused waits for its next time instead of running at once as a catch-up; an overdue document schedule waits for the folder to be read instead of spending its one catch-up on a failure; and a run whose history cannot be written moves the schedule on instead of running the same window again every minute.
A failed automation run says what stopped it and what to do next instead of a raw error, a run that found no coding agent links to Agents in Automations and in Library check history, a running schedule names the step it is on, a run whose result or next time could not be written says which file refused it, and when Atlas cannot read .ontology-atlas/rounds.json the screen names that file instead of asking for a malformed schedule file to be fixed.
Entering the map from the rail fades the page to the map's canvas and reveals the map there, instead of cutting the page and flashing the loading screen, and the rail marks Map as soon as it is pressed.

2026-09-28 · v1.4.0: Security boundaries, lower memory and gray-area investigation

Added

The Mac app has an experimental Jev evidence check under Agents, then Models. With your own TypeSafe key saved in the Keychain, paste one explanation and the passage meant to support it; the exact request is shown before you send it, the transfer is recorded in the open folder first, and the answer (supported, contradicted, or cannot settle) is advice that never changes the ontology.
The launch chooser has a small companion home and a device-local journal for personal reflections, including unanswered questions. Optional keepsakes stay with those memories; the companion remains beside work status on the map.
The companion's first destination now offers a directly traversed top-down pixel sector. WASD/arrows, E, and Q move the fox, inspect nearby characters and devices, and show a walkable route to the next objective. Original code-drawn Curio and signal mote encounters, a current-concept reflection, and a visibly opening door form one complete movement-to-exit quest. A separate project-scoped save preserves sector state without changing the existing 35 automatic destinations, combat rewards, or ontology meaning. Hidden game layers leave keyboard focus, mobile controls and no-scroll layouts remain usable, and storage failures never advance the player.
The optional companion now offers camp learning expeditions through saved responsibilities, exclusions, and direct dependency declarations, with in-game source reading, preserved personal explanations, and illustrated folios distinct from the world.
The companion becomes a persistent pixel world powered by recorded ontology and wiki growth, with keyboard inventory, six permanent skills, a visible construction constellation, five-floor expeditions, seeded blessing choices, active dodge, guardian rewards, and optional auto retry. Menus preserve the world and fit without outer scrolling. Existing personal memories remain intact; fictional progress never establishes ontology accuracy or acceptance.
Saved companion learning notes now open directly from marked topics, retain their original date and uncertainty when the current source disappears, and offer explicit re-exploration without repeated experience rewards.
The companion home becomes a centered pixel adventure with a new fox, interactive daily life, real concept exploration, domain-based regions, automatic battles, equipment, a bestiary, and capped offline rewards. Existing personal memories remain available. Game progress stays separate from ontology quality and acceptance. The earlier map toolbar alignment change is restored at the owner’s request.
The companion gains one contextual camp interaction, a departure portal, 36 illustrated destinations across six regions, and 108 original creatures including 18 guardians. Routes have distinct rosters, conditions, and completion records; the field guide records actual sightings and defeats. Walking follows distance and a shared foot contact line, while pixel menu slots replace the former control placard. Existing saves and personal memories remain intact; construction and reading unlock travel independently of combat XP. Returning during guardian loot banks the clear once, and reduced-motion preferences update live.
The map has a new view, Territories, in the view picker next to Flat. It draws every capability by name around its own domain, with nothing to expand: a capability's circle grows with its elements, an amber ring marks one whose cited code changed after its document, and each domain's title counts its capabilities, elements and stale ones. Strokes between domains count how many of their capabilities depend on each other. Selecting a capability opens the usual inspector and shows its elements and its own dependencies. The desktop app reads stale from Git; the web says the state is unknown instead of guessing. A link with ?view=territories opens the view.
The optional companion now offers twelve quests tied to current project, reading, reflection, and eligible completed ACP work, with source links and explicit assistance previews. One-time relic rewards unlock guaranteed enhancement tiers for its three personal equipment slots; ordinary activities can unlock every tier. Expedition M plans rest, supply, or elite effects for the next floor. Saved claims and costs are atomic, older equipment is retained, and optional ACP requests preserve existing approval authority. Game panels retain keyboard focus after claims and pointer combat; bounded layouts avoid outer scrolling.
The map has a new view, the Hex board, in the view picker after Territories. Every capability is one hexagonal tile, each domain a region of tiles around its title tile, the project at the centre. A tile's brightness and dots count its elements, a gold rim marks one whose cited code changed after its document, and a hatched dashed tile says the evidence is unknown (the web says so rather than guess). Tiles never move once placed: a new capability fills a free cell of its region. Strokes between regions count their dependencies and run only in the gaps between tiles. Hovering a tile shows a one-line summary and its routes; selecting opens the usual inspector. "Stale only" keeps the stale tiles and names the moved file, arrow keys walk to the neighbouring tile, and the wheel scales the tiles down to one nameplate per region. A link with ?view=hex opens the view.

Changed

The Harness Architecture empty state now starts under the tabs with a title that reads above its paragraph, and the Agents list no longer points at install guides "in the list below" when there is no list below.
In a browser the Agents tab lists the 41 coding tools the Mac app looks for, read from the registry the app ships, under a filled get-the-app press; on the desktop the other-tools shelf's marks stand back until pointed at so the ready tool keeps the weight, the search dialog is titled by its errand instead of repeating the shelf heading, and group headings share the start line of the cards beneath them.
The Library's index column now has one box edge and one text start line: every door, source row, wiki card and action stands on the same edge and its words start after one shared glyph slot. With no coding agent, the wiki tab shows one notice with an Agents door instead of two disabled buttons and a paragraph. Source rows keep the format and move the byte size into the tooltip, and a checking source wears the same badge shape as the other states. An open source's name is a headline at the wiki page's title size, its outline rows are neutral with a heading glyph, the reader strip names the next file to compile, repeated counts are said once, and the reading pane fades and rises in when what it shows changes. At 1040 the graph's head stays on one row.
On the map, INDEX and the realm ledger now end under their last row instead of stretching to the floor; the meaning review names the node once in its header, says a missing connection reason once over the group, and gives each connection card its type as a bordered tag above the sentence with its actions under it; sentences are at reading size; an empty analysis history stands centred with Analyze as its action, or says why there is none on the web; and the conversation previews its first asks while the tool opens and names the picked concept in the composer.
The Agents destination speaks one control grammar: every row and heading action on both tabs, including the MCP tab's four connect buttons, is the same 32px chip with its verb's glyph, and 'Ready' is the same dot-and-tag badge on both tabs; the agents tab shows every other coding tool as a shelf of marked tiles under the ready ones, each opening the search dialog already narrowed to that tool; runtime rows lose the empty 10px strip under them, tools without a registry mark show the MCP tab's brand tile or their initials on the same plate instead of a blank square, the web tab says the Mac app finds the tools once and offers it as the one pill with MCP as a quieter link, a diagnosis result eases in instead of cutting, and the connect-check dialog states its file count once, in the status box, naming the two tools whose files it reads, with the next file to create on the line beneath.
Archive Grove's 18 enemies now use original darker pixel sprites, while compact overhead attack and target cues replace large floor rings; narrow battle framing keeps guardians visible.
The Library's Ontology, Work scope and Check history tabs read as one family. Check history opens with Work scope's headline and column and keeps one way to Automations; before anything is scheduled it shows one button to schedule a check beside the two kinds it can run, and the four results a run can leave. The away card no longer repeats the sleep gap the run list shows, and the help card beside the checks no longer pushes their names out of view. A run with no change reads "no change" in both languages. Work scope's empty state previews six concepts with their map glyphs and ends in "N more, see them all on the map". A document's meta row keeps one control height, and its properties line names the kind in words, not YAML. Every tab switch, Sources and Wiki included, fades the panel in, with reduced motion too, and run times over a minute read in minutes.
The web's door to the desktop app now carries one label everywhere, "Get the app" and its one Korean equivalent, instead of four, and the sentences beside it call it "the desktop app" (the download page offers macOS and a Windows beta, so "Mac" was not true); the line saying Windows was still coming is gone. The longer "Get the desktop app" was tried first and did not fit the phone tab bar: it wrapped onto two lines, lifting that tab's icon out of line, and at 200% text it pushed the bar past the room pages reserve for it. Every Button wears the chip corner, so a 40px primary no longer reads as a pill. The dogfood vault's ten concepts whose code changed were re-read and updated, and a leftover empty dependencies: [] was removed.
A Git step that changed three or more concepts now shows one whole concept name and the count of the rest, instead of two names cut to a few characters.
The 3D map is now lit. Strata, the default 3D view, draws each level as a glowing translucent floor in its kind's colour, with a band for every domain on the floors below it, and Neural gathers its concepts around their domains. A concept's light is its evidence: it glows when its cited code is current, dims behind an amber ring when that code changed after its document, and stays dark behind a dashed ring when the app could not date it (always, on the web). A legend names the kinds and counts each state. A click selects without moving the view; a double-click or Enter flies to the concept and Esc or Home flies back. The Cone view is gone from the picker; if it was your choice, the map opens in Strata.
The Git screen now gives the step or document you picked the largest title, named by the concepts it changed; the page title steps down to a label. The step shows a short id you can press to copy the full one, and a date in your own format instead of a timestamp. At the smallest app window the list and the document reader scroll as one page, so the reader is never squeezed to nothing. The discard button lines up with the end of the text it acts on, a document's other steps show three at a time, and neighbour names in the concept drawing no longer overlap.
Center short Wiki inquiries inside the reader pane and place saved Concept sets under Ontology, preserving existing set links, member identities, and document drafts.
Clockglass Foundry's 18 creatures now use original brass, glass, and steam pixel art. Their combat sprites meet the floor, and the three guardians have distinct full-size illustrations.
API keys and local model runners moved from Settings to a Models tab on the Agents page (Agents, Models, MCP). Ollama, LM Studio and llama.cpp each have a row at their default address that says whether the runner answered, had no models, or could not be reached after you press Check; keys can be added, checked, replaced and removed there, showing only their last four characters. The tab also counts every transfer recorded in the open folder. Settings keeps one Models row that takes you there, and the map's "add a key" button opens the tab directly.
Analysis reads as one board now. On Concepts, cards size to what they hold and the agent handoff row sits right under the tab instead of at the bottom of the window; lists in every card share one row rhythm; the boundary matrix grows its cells to fill its card; the empty Missing tab offers the map; Growth starts every card on the same left line; and Flow shows the request beside the empty explanation. On Brief, Seen up to here sits on the headline's row, a missing code repository is said once, and recent changes are grouped by kind. Captions now say what a mark means and what it tells you, without equals-sign glossaries.
The download page's first screen centres its headline, buttons and globe on one line instead of leaving an empty lower-left corner, and the headline finishes typing on time on a busy browser. The demo section starts at the same left edge as every other section, with its title beside the video. The agent conversation lines up with the three cards beside it, the app-screens list carries each screen's sentence and link at its foot, Korean captions use the text face instead of the code face, the evidence card's labels are easier to read, the facts strip wraps onto shared columns in narrower windows, and the GitHub and X buttons in the top bar are full-size buttons.
The Automations empty state now starts right under its lane tabs instead of floating mid-screen, and its preview row no longer shows placeholder dashes or repeat the preview title as an eyebrow.
Every settings pane now opens with its own heading and one sentence saying what it decides, and the side list is quieter so the settings it leads to come first. Row buttons share one size, the API Key pane has a key icon and its privacy note reads as one sentence, the Expand choices use the same joined switch as the Screen pane, and the Update row says when it could not read the version. The first-run screen centres its headline, keeps its three doors the same height, and uses larger text. On the download page, the release facts that wrap onto a second line now start at the same edge as the first.
Inkwell Marsh's 18 creatures now use original ink, reed, and parchment pixel art. Their combat sprites meet the floor, and the three guardians have distinct full-size illustrations.
In the agent conversation, work still in flight now moves: a running tool call's name and its running word, an active work group's title, and the composer's thinking status carry a light band sweeping across the letters, and each settles to still text the moment the work lands. With reduced motion the band is replaced by still, brighter text beside the existing running ring and clock.
The map's top toolbar now spans the map instead of crowding its right half. The search tools start at the map's left edge, just past INDEX or its folded tab, and the review, agent and recent-change buttons keep the right edge. The Automations button is gone from the toolbar: it only opened the Automations screen that the left rail already opens, on the same ontology schedules.
A ready expedition creature can now be clicked or tapped to cast the existing Q knowledge wave. The target ring and action slots show readiness and remaining cooldown turns, while pressing Space during an incoming strike moves the fox immediately. All actions continue through the saved game rules and earn no separate reward.
The Harness screen now reads as one page. Every view starts on the title's line and keeps the 1600px page width that other screens use, so wide windows add margin instead of stretching the content. Each data view opens with the same label, one sentence of what was found, and the rule it was counted by. Structure gets its own sentence, and the note that its numbers are not runs or scores appears once. Text in the Structure, Coverage and Guides views moves from 9.5px up to 11px. Guides rows share one height, and only a real difference between a pair wears a tag. Coverage chevrons sit beside their counts, and the answer columns now span the table instead of crowding its right edge. The agent loop is explained once, in its own card. The Architecture toolbar uses the standard control height.
The companion expedition now frames its original fox and creatures on a shared pixel combat floor. A phase-linked strike, impact, enemy windup, and dodge warning make the existing automatic turn and Space action readable. Camera bounds keep both actors visible in narrow viewports, movement stops at the foreground floor, and walk-frame changes no longer re-render the whole game world.
Analysis Wiki and Guidance read in one order. Wiki leads with the card that has findings, counts each finding with its unit, groups advisory findings under their own label, folds empty cards into one line, offers one Library link per card, and ends on a line saying the list is complete. Guidance marks the role that reaches the fewest domains, its heading outranks the domain names, empty branches are drawn dashed, every stem is drawn once at one brightness, each domain says what it is for, the folder-wide collections sit above a rule, and text mode lines its counts up in columns. The no-folder examples keep the page title as the only headline and end on a filled next-step button.
Harness Structure now shows repository instructions, agent work, and checks as one workflow, with selected file evidence beside its source on wide screens and directly below it on narrow screens.
Empty Automation, Library Collections, and check-history views now show truthful next actions and existing repository evidence without presenting examples as completed work.
Projects is a grid of equal cards, each one pressable surface with a hover step, a whole definition sentence (the flagship now shows its body definition instead of "no description yet") and its two doors in a footer, closed by a tile that says how the next project appears. The project page shares the list's frame and start line, its hero carries the whole first sentence at the title step while the overview starts after it with the Includes and Excludes bullets side by side, domain counts sit right after their bars on one line, the two AI copy blocks are one, and the flagship's web address shows its own sample instead of a not-found dead end.
The download page shows the app's other screens on one stage: the rail's names in its order beside one real capture, and pressing a name swaps the picture, its caption and its door in place. It replaces the separate architecture and library sections, and the agents copy now says Codex also runs in the in-app chat.
The Screen settings pane fits its window again: the screen guide is one row with its on/off switch and its Show again button. The API Key pane says what connecting gives you once, as its opening sentence, and lists where the key is kept, what leaves and when without repeating them. The Update pane also says when the app last checked and which build you put off. In Workspace, importing nodes from another folder is a row of its own with a Choose folder button. On the first-run screen the companion's home is a card like the others, and the download page's release facts no longer run past the column on a tablet-width window.
Automations reads as one screen: the empty, app-required and no-folder states share one stage under a single page headline, reading down one line from the next step to three facts (what runs, when, what it may change) to the schedule list, which fills the rest of the page; each sentence appears once, and outside the app the list says schedules made in the app appear there instead of counting zero. The lane count is a quiet label, an opened schedule becomes one card holding its header, latest summary, the files it touched listed once, and three matching outline actions (run now, pause, remove), removing a schedule asks in the danger tone, the sheet's default name is a real value, and the Korean copy speaks in one register.
The Analysis brief now shows one band of counts and one list of lines instead of four half-empty cards. Every line names its core and opens where it lives, a core that cannot be counted says so once, and "Get the app" appears once. Pressing "Seen up to here" folds the lines it cleared. The Concepts summary uses the same band, with counts in the same ink as the verdict and dated weekly bars.
Before a source folder is connected, Harness shows two tabs, Harness and Architecture, instead of four. Structure, Coverage and Guides all showed the same example there, so pressing between them changed nothing. All four return once a reading exists.
A Git step with two concepts whose names do not both fit now shows one whole name and "1 more" instead of two names cut to a few characters; the fit is measured on the painted row.
Toasts now stand at the bottom of the free map, between INDEX and the right panel, as one neutral box with a small tone glyph (neutral, success, warning, error); the agent's working status joins the bell in the toolbar row instead of floating under it, and leaves the toolbar while the conversation panel already shows the turn.

Fixed

In Territories, a selected capability's element list moves further along its side instead of covering its domain's title and counts.
Harness help bubbles open under the ? that opened them, can be hovered into and close on Escape; the architecture toolbar keeps one 40px height and one close shape, its copy confirmation no longer stretches over the evidence rail, and below 1280px a pressed role scrolls to its answer. An open coverage cell pushes rows down instead of covering them. Removing a schedule asks in place of its buttons and closes on Escape, the ontology review sheet uses the documents sheet's cadence picker, and the empty list drops its dead add box. In Analysis the guidance popup stays under the title, the row menu keeps one line per item, the Korean handoff footer is no longer monospace, and the preview and Flow actions share one button and copy control.
Reading a DOCX or XLSX source in the browser stops at the size each entry of its archive declares, and at 256 MiB when an entry declares none, so a crafted file can no longer exhaust the page's memory.
With the agent panel open, Strata and Neural name only the relations the flat map would name (the selected or pointed-at one, the focused concept's, and the top-level ones) instead of up to 24 containment labels over the layers. Strata names each layer beside its own rim; a layer with no clear place beside it is left to the colour key along the bottom, and the list in the bottom-right corner is gone.
Pressing Escape the moment the map's view picker opens now gives focus back to the view button, as a later Escape always did, instead of leaving the keyboard nowhere once the picker closed.
On the map, the right-click menu takes the keyboard, so the arrow keys move through its actions, not the map. Territories keeps every name it shows inside the free map: on a small window the territories draw closer, and a name that still does not fit appears on hover instead of under INDEX. Its legend stays between the panels and its evidence note is never cut off. The line hover card no longer lands on INDEX or the tool buttons, and the lit 3D legend is a strip the nodes stay above. The node and relation panels share one width and one button style. The tour drops its dead Back button and keeps its cards off the nodes they explain. Closing dialogs and folding INDEX keep your keyboard place. Adding documents to an existing map is called "Add to the map", and every map view stays in the address.
The agent conversation dock no longer draws its tool and mode pickers over the status word on a narrow dock, and its buttons share one size. Escape keeps a typed sentence on the map and in the Library alike, and a sentence put away on the map is still there when the dock opens again. After Stop, New conversation or a permission answer, focus is in the composer, and the Library dock returns focus to its door. You can type while the tool connects. A write you declined says so instead of "failed", and tool rows no longer show a function name. On short windows the permission card leaves the call it asks about in view, the close button stays in the corner, and a failed save shows one sentence with the detail folded.
The hex board opened with a node already selected, by a link to that node or by switching to Hex with something chosen, draws every tile beside the inspector instead of an empty canvas, and opened by its address it sits right of INDEX at full height rather than in a strip under the panel. In the app, the INDEX recent-changes window, the map spotlight and a node's changed-ago date read each document's date from Git: its last commit, or its file's date when Git shows it edited or new since, so a cloned, checked-out or restored folder no longer reads as all changed today, and nothing is counted until Git answers. INDEX opens the rows above whatever is selected, from a link, the map or a search that was then cleared, and scrolls that row into view.
In Territories, a capability name shown only on hover or under a selection now turns to stay inside the free map, so selecting a capability beside the inspector no longer cuts off the name of a capability it depends on at the window's edge.
Changing a document's kind no longer leaves it in other documents' lists for its old kind: a domain that listed a capability lists it among its elements once that capability becomes an element, in the app whether or not the file moves, and through MCP reclassify_concept and merge_concepts. The quick edit names each document whose list changes before Save, and a notice names them after. When a document's own kind keeps no list for the new kind, the entry stays where it was and that document's page says so, as it now does for any entry sitting in a list for another kind.
Pausing or removing an automation now ends its run at once, even while the coding agent is still starting: the "running now" line clears, the run history says the run was stopped, the pause stays a pause, and every other schedule can run again without reopening the app. A run that never gets an answer now gives up after twenty minutes instead of blocking the others. An ontology review that had no coding agent is recorded as failed with the reason, in your language, instead of an English sentence claiming it completed, and reviews already recorded that way no longer show it. Automations' run history now explains a run that had no agent or was stopped, in the same words as the Library's check history.
A document's page has Rename and Delete buttons beside its address, opening the app's own dialogs: rename asks for a name, shows the new address and keeps the file's own slug in step; delete names the documents that still point at it, and their pages then flag the missing target. The quick edit also changes each language's display name, and a kind change moves the file into its new kind's folder in the same save. A save refused because the file changed elsewhere says so in your language, on the page and in the map's explanation editor, which now keeps your text, and no green "edited" notice covers that refusal. A concept added under a domain from the map records the domain as domains/<name>, like every other writer.
On a small window, Territories folds more capability names to hover so every circle stays inside the free map, instead of spreading a larger map out with its names under INDEX and off the window.
Push on the Git screen no longer commits waiting changes on its own; it opens the commit step first, with sending ticked. Commit, discard and restore confirms share one button size, take focus, close with Escape and hand focus back, and a step's time reads in your language. The remote buttons read in Korean on /ko, and their explanation, with the git command it runs, opens on keyboard focus as well as hover. The project page's agent dock stays one window tall, so its header and message box are both on screen. Agents and MCP buttons share one size, the connection check names what it counts, and the add-connector dialog returns focus and answers an empty search first. New projects count only the fields asked for, copy buttons keep their width, and the map's code-folder card keeps its border and honest wording.
The guide and changelog keep a full reading measure at laptop widths (the article was 64px wide at 1040), the changelog lists each fact as its own bullet under its category instead of one 10,000-character paragraph, Korean pages say the body is the English original and mark it lang=en, sidebar entry titles no longer start with a stray separator, the 404 and the render-error screen share one centred stage in the visitor's language with one primary action (home on the web, project search in the app) and the 404 wears the gateway nav, and the nav's GitHub and X marks are 32px targets. The 404 waits for the visitor's language and surface instead of flashing English or the gateway nav first, and the error screen asks for an error ID only when it shows one.
Removing the last relation of a key no longer leaves the key behind: the map's relation editor, MCP remove_relation and replace_relation, and the CLI's remove-relation now delete a list emptied by the removal, and relation_notes with its last reason, where they used to write relates: [] and relation_notes: { }; only a kind's own scaffold list, such as a capability's elements, returns to the [] creating the node writes. The change review says "None" where it printed [] or null. The launch chooser lists the folders that still exist first and gathers the ones that no longer exist into one quiet line at the end of the list, whose review names each path and forgets one or all of them only when pressed; the list is drawn once, from the reachability check, instead of redrawing a frame after it appears. The rail switcher's picker reads "Open another folder".
The app chrome keeps its geometry and focus while it works. Switching folders keeps the folder chip in the rail and names the folder being opened instead of a blank workbench. The folder switcher animates out, returns focus to the chip, closes when Tab leaves it, and opens in the free map beside INDEX and under the map toolbar, so it covers neither the INDEX panel, the folded INDEX tab nor a toolbar control. The web Get the app tile matches the gear. The shortcut sheet gives its height to the shortcuts and lists the map once. In Settings, checking for updates keeps focus and shows failures in the warning tone on the row line, key forms use one control size, both Finder actions read Show in Finder, the web hides the app-only Updates pane, and a language switch reopens the same pane.
In the installed app, Escape closes the palette, sheets and dialogs under every macOS input source: the app notices each Escape press itself and hands the page the one keypress a Korean input source can keep from it, never a second one when the page already has it. The settings sheet takes focus when it opens by a click, so Escape and Tab work at once, and a click on the dim beside it closes it. The Git screen tells three states apart instead of saying there is no remote: with no remote it still offers to connect one; a branch never sent to an existing origin gets a one-press first send that records where it goes; a detached HEAD names the commit and says sending needs a branch. Nothing on that screen can replace an existing origin any more, and Push with nothing new to commit now sends the commits already made.
On the web, opening a Library source opens a copy only of a PDF, an image or plain text; an HTML, SVG or any other file is saved to downloads instead, because a copy opened in the page would run in the site's origin. A remote image in a page is no longer loaded: its place names the host and opens it in a new tab only when pressed. The production site now carries a Content-Security-Policy that refuses remote images, connections, plugins and form targets. Each program in the connector catalogue runs the version checked on its date, and its row says which.
The map's toolbar keeps its tools on the first line: the path and trail chips follow the tools and move under them when there is no room, and changing the view no longer makes the tools jump lines. The trail popover opens toward the map instead of under INDEX. The view picker focuses the current view and gives focus back after a choice; the agent panel closes on Escape and returns focus to its button. The path chip keeps "no path" or the hop count visible, its copy button matches its clear button, and a copied path links to the map in your language. On the sample the recent-changes tooltip says what pressing does. The tour's try-it card clears the toolbar and fades out, and the companion says the sample is not counted instead of showing zeros.
On the map's hex board, relation lines stay in the free map: they no longer run along the top of the canvas under the toolbar, under the INDEX tab, the inspector or the legend, and a line whose target lies under a panel stops short with an arrow pointing at it; the board also keeps a margin under the toolbar when it fits.
A compiled wiki page's source citations now read as places instead of raw addresses: a fact that cites line 5 of its one original says "line 5" where the page used to print src:sources/budget.md#l5, and a page built from several originals names the file too ("release-plan.md · line 9"), matching the words the Source pane and search results already use. A label the author wrote is kept as written. With no coding agent on this computer, the Check results page now shows the same notice and Connect an agent link as the Wiki list beside it, instead of a disabled Check button with no way forward. The English Library tab and its page are now called Work scopes, the same name the Korean interface uses.
The map's constellation popover now hangs from its own button and stays in the free map, so it no longer covers the INDEX panel or its folded tab at any width. A vault that has never saved a constellation, or holds an empty constellations file, opens to the empty state instead of "could not read your saved constellations", and a failed save no longer relabels the list as a read failure. When a read does fail, the notice is a readable card with one Reload action, and Create is either pressable or says why it is not. In the node inspector the kind chip, the domain chip and the close button share one 32px chip height, radius and type size and end on one right edge.
The shortcut sheet (?) and search (⌘K) now open on every screen with the navigation rail, not only the map, a project page and Insights. Off the map the search is named for what it searches, concepts, not "this map", and ? answers right after Esc closes another dialog. The sheet lists one search key instead of ⇧⌘K as a second search, and lists D under the map, the only screen that answers it. Saving or removing a model key, choosing or dropping a local runner, and copying a project's link are confirmed by the pressed row or button and read out to screen readers, not by a toast over the page's own text. A conversation still connecting has Stop beside Send: it ends the attempt, keeps the draft, says nothing was sent and offers Connect again, and the Library's conversation stays stopped when closed and reopened.
Screens say an empty state once and draw a page's first press in one shape. On Automations an empty lane's schedule list is now only its frame (title, count and column heads), so "nothing scheduled yet" is said by the stage title alone. Harness → Architecture, before any source inspection, shows one empty state across the delta and observation columns that names Inspect source, in place of seven dashed boxes, seven hollow marks and two column notes. The bell inbox leaves out a duration under one second and shows longer ones in one rule, the next unit only when it is not zero (2m 5s, 1h). No description in the map's view picker ends on a lone word. The web-only doors on Harness, Agents and Models are the standard primary button instead of a pill, and Automations' first press is the same 32px button.
A connector or Library round that a folder arrives with switched on, whether cloned, synced or pulled, no longer attaches or runs until this Mac allows it. The MCP tab and Automations list each waiting item with everything it would run and one press to allow it, and ask again when the folder changes it. After this update, connectors and rounds switched on before wait for that one press each. A connector can use only its own keychain entries.
The map's top toolbar no longer draws its buttons on top of one another. With the meaning-review panel or the agent dock narrowing the map, the search tools, a path and the trail used to slide under the buttons on the right; now the row keeps every button whole, moves the search tools to a second line when both groups do not fit, and shortens the path's name instead (the full name is on hover). The mascot button now has the same square as the rest of the row, and every button in the row is spaced the same.
Expedition panels now pause automatic combat and sprite motion until their exit finishes, then resume the unfinished turn; game closure still allows bounded offline catch-up.
The Library graph names each concept in the screen's language: a Korean screen shows a concept's Korean name where the document records one, instead of its English title.
"Just start" and "Create a new folder" on the first-run screen now leave the starter you chose in the folder: the map's starter nodes, the wiki template and the agent guides are written before the folder is first shown, and "Just start" says where it made the folder. Both used to open an empty folder with no word at all, because the screen that pressed them was gone before it could write. If the starter cannot be written, a message says so and points to Settings › Workspace. The door also names the right place (Ontology Atlas in your home folder, not Documents), and the map's frame meter reads entirely in the screen's language.
On the Git screen, pressing a step under "Other steps that changed this document" now opens that step on the same document, in the same view, brought into sight. It used to open on the step's first document, so "Restore this version" there could put back README.md instead of the document you were following. The restore button now names the file it puts back, and a step whose file list does not hold that document, such as a merge, says so and selects nothing in its place.
In the Library, what opens stays beside what opened it. Automated checks explains itself in a line under its link instead of a clipped help box over that link. The explanations on the list's buttons open beside the list, wrapped, not over the side bar and search box, and those buttons are named by the words they show. How to use no longer covers the list on a narrow window. Below 1280 pixels the conversation opens beside the list, takes the keyboard, closes on Escape and gives the keyboard back. Buttons in Library popups and graph cards match the ones that open them. Find documents with nothing to suggest offers one way forward and a close button. The page header names the waiting file once, graph counts drop whole clauses, and New page no longer jumps on the first key.
On the Git screen, a step's "Concepts changed" no longer counts the vault README as a concept, and a document-kind mark no longer borrows the element's glyph. The default concepts lens now shows what the step wrote in the chosen concept's document, marked in place, above that document's other steps. After connecting a remote, the connect button stops saying "Connecting…" as soon as the address is saved and the header offers the branch's first send from that moment; a remote action's result is announced with the screen it describes, a sent save is reported in one sentence, and a failed send says git's reason.
The project name at the centre of the map now sits on a ground plate, so the relation running straight down from the project no longer shows between the words of its name.
The map's overview sits in the middle of the free space between INDEX and the right-hand controls again, with the same room on both sides, whether INDEX is open or folded and with the agent panel open. It had framed a concept it does not draw at the overview (the most-referenced one, folded into a crowded domain), which pushed the drawing up to about 110 px to one side, and it left more room on the right than on the left. The top toolbar and the right-hand controls now stand on the same line and margin as INDEX, and the map is drawn at the canvas's exact size, so its lines and names are no longer slightly stretched after it opens.
Territories folds a capability name to hover rather than draw a dependency count or a domain title on top of another label or outside the free map.
The map's project panel is the project's from its first frame: while its code evidence is read it names the section and says it is reading, and when the answer lands nothing above the buttons jumps; opening a project also no longer walks its whole code folder when there is no receipt to compare against. View on map on the Projects list and View topology on a project's page open the project itself on the map, where its code folder is connected, instead of the project drawer, which cannot connect one. An editable project name stays the page's level-1 heading for screen readers, with the press that edits it inside.
In the Library, a graph card shows a long page name whole on up to two lines, with its kind underneath, instead of cutting it off. The "since you left" card on Automated checks gives every page the same button, and a refused request now has a button that brings its entry into view in the ledger.
The analysis brief no longer prints a partial count while the guidance scan or the code-evidence walk is still reading. Its headline says it is counting, behind the same two marks the list uses and with a light moving over the words, and the numbers and the list of what changed since arrive together once every part has reported. A later recount, such as coming back from another question, keeps the last numbers on screen and marks them as being recounted instead of changing them without a sign.
Domain names on the map now sit on the same ground plate as the project name, so the fan of relations converging on a domain no longer runs through its name.
Choosing where a path starts (right-click, Path) now keeps every concept you could pick as its end on screen, framed between INDEX and the right-hand controls, instead of zooming in on the start and pushing its neighbours off the canvas. Territories and Neural rest in the middle of the free map like the other views, the flat overview stays centred at tablet widths with INDEX open, and the 3D views keep clear of the right-hand controls. Escape pressed while the settings gear has focus reaches the map again, as it does from the map's other controls.

Removed

The installed app no longer watches for Escape itself or hands the page a stand-in key press: the Korean input source was never keeping Escape from the app. The presses that went missing were swallowed by a screen-control tool while it ran; without it the app received every Escape on its own, and one press still closes the palette, a sheet or a dialog.

2026-09-22 · v1.3.0: Library rounds, scheduled refinement and clearer analysis

Added

The Git screen and the rail's Git dot follow the open folder while they are on screen: an edit by a person, an agent or a terminal is re-read at once, read-only and coalesced, instead of waiting for the next visit or the next window focus.
A step's detail lists the other steps that changed the focused document and jumps to any of them, reading deeper into the list when the step has not been loaded yet.
Library gains a Rounds tab in the installed app: a consistency round hashes every cited source and checks every page on a cadence with no agent turn, redrafting only stale sources when asked; a service round re-reads one attached connector once per pass and redrafts what changed. A round is approved once, in a sheet that states the exact scope and the daily agent-turn bill; a pass refuses anything outside sources/ and template-fitting wiki pages and names the refusal. The stage opens on a since-you-left card, then a time-axis ledger where held passes are quiet lines, changes are cards, and sleep is a hatched gap; a missed window catches up once. Rounds live in .ontology-atlas/ on this Mac.
The project detail's overview tab draws the body as a brief: each ## section becomes a numbered block, an ordered list a strip of steps, a bullet list rows, and wikilinks doors to the cited document; a body without sections stays prose. Below it, the page hands an agent the exact instructions to lay the overview out in four sections and change only the body under the write guard, for review in the Library. The sample project carries the four sections.
The Git screen restores one document: discard its uncommitted changes, or bring back the content it had at a past step. Each door states what is lost before it runs, nothing runs before the confirm, and a restore lands as an uncommitted change the person can read line by line and discard. A restore is refused when the document's identity at that point differs from the one on disk.
A Library round's cadence is now a rail you drag (1, 5, 10, 15 or 30 minutes, or 1 to 24 hours) instead of four fixed chips, so a Slack incident room can be re-read every ten minutes while a release is out. A round also names the places it watches: folders inside this folder and one or more connected services, each narrowed to a channel, space or database, and the sheet reads the whole rule back as one sentence before you allow it.
In the installed app with a guarded agent runtime ready, the project page's "hand the overview to an agent" ask opens an agent dock beside the page and seats the brief instructions as the first turn, with every file write still stopping at the permission card; the web and an app without a runtime keep the copy path.
The Library graph becomes a map of islands once a folder passes four hundred marks: an island per concept, its pages at the centre and the files they were written from around them, an Unsorted island for pages naming no concept and an Unread island for files no page has read. Every island carries its name and count; a stale page is an amber dot; no line is drawn at rest. Pressing an island opens it as the columns picture with a chip and Escape back to the map. On any folder whose page column folds, the folded pages stand in stacks, each with the files its pages read directly to its left, so no citation crosses another stack's names. Measured at 11,240 marks: 42 islands, every one named, under 5 ms a frame at rest.

Changed

Measured Guidance now offers diagram and text views of the same domain-scoped declarations, with keyboard-accessible evidence details that preserve source identities and distinguish declared scope from execution results.
A project is drawn by its locale display name (display_<locale>) on every screen that names it: the projects list, the project page's heading, breadcrumb and document title, the map drawer, the map's path chips and both search palettes. The map, the INDEX and the Library already did; the list and the page said the canonical title beside them. Editing the heading in place edits the word on screen: the display key when one exists, the canonical name otherwise. Search matches the canonical name and every display name.
The Library's islands map lays the files no page has read as a shore: a staggered band of dots under the islands, as wide as they stand, instead of the largest disc in the middle. The source list says its most common state once above the rows, in that state's tone, and keeps a badge only on rows that differ; the opened reader no longer draws a focus ring around the whole pane.
On the Git screen the action that restores one document to a commit now reads as a control at rest: it resolved to the same text token as the sentences beside it (#82828a against a heading at #8a8f98), so only a hover told the screen's one disk-writing action apart from a caption. It now carries the discard door's chip grammar and one ramp step more ink, with its label still on the column's start line.
The Git screen's list of steps reads ten at a time and ends with a fact instead of stopping: while older steps exist the last row fetches the next ten in place, and once the first step is on screen a line says so. The depth a person opens is remembered per folder.
When an agent session ends, the app now writes the first few lines the agent process printed on its error stream into the app log, and says so when a session that ended within ten seconds printed nothing at all. Until now those lines existed only inside the conversation that died with them, so a press that did nothing left no trace to read.
The map's relation panel now links to its source document by the document's name alone. It used to put an arrow between the file name and the word for opening it, which is the decoration the design rules refuse for a link that stays inside the app; a screen reader still hears that pressing it opens the document.
Discarding the walk you have taken on the map now takes two presses. The trail is not saved anywhere and cannot be rebuilt, yet a single press on a small button erased it with no warning and no way back; the button now arms first and says so, and the arming releases itself after a few seconds.
The project page opens with a composition board instead of the project document: one cell each for Ontology, Library and Harness, with the figures that surface can honestly give, a line saying how connected the map is and where it is thin, and a door into that surface. The domain rows follow in a card beside the rail rather than behind a tab, the two tabs are gone, the document is summarised to its opening paragraph and a contents line of its section titles, and the page's agent hand-offs are one card instead of two.
The Library's collections tab now says where a collection comes from: its empty state explains that a set of concepts picked on the map's Galaxy view, with a name and a purpose, is what lands here, and its doors read "Pick on the map" and "View on the map" instead of naming the Galaxy alone.
Four Korean sentences in the Library drop transliterated or technical words for the product's own: the local-model address is no longer a "runner", Compile is called by the Korean word the Compile stage already uses, and the automatic-check sheet names the check kinds as its own two options do instead of a word for "consistency".
Agents and MCP are one destination: /agents carries two header tabs (agents · MCP); MCP keeps its two sections on ?mcp=. The rail drops the MCP tile, and /mcp with any of its parameters redirects into the tab, so the ontology-atlas://mcp?install= deep link and older links still land. The Library strip no longer repeats its name beside the tabs, and its index column seats the search field where the repeated list name stood.
What the map's ego focus dims now recedes: nodes outside the selection and its neighbours sink to one rest alpha, with their lines, chips and (for domains) names, instead of a colour ramp too close to the normal ink to see. The double-click line in the features list now describes what double-click does.
The destination that keeps the folder's history is named Git in both languages, because durable history, restore and discard are git's and the screen no longer hides the word.
The Git screen reads an uncommitted change as the document it is: the changed documents as chips named by their concept, then the chosen document whole in the reading face with the changed lines marked in place. Front matter stays apart and collapses when none of it changed, and the same reader draws a past step's file.
A local automatic check in the Library runs once as it is saved, so its first result stands on the screen right away instead of waiting for the next clock boundary or a press on "run now"; the saved toast says so. A service check, which spends an agent turn per pass, keeps its first pass on the clock the person chose.
The Library's Korean tab names now say what the tabs hold. The Collections tab's Korean name reads "work scope" (a set of concepts saved from the galaxy for one piece of work) instead of a bare "collection", and the Rounds tab's reads "automatic check" (the checks the app runs on a schedule while it is open) instead of a word for "patrol"; one pass of a check is called a "run" rather than a "round". English names are unchanged.
The project detail page opens its own Markdown file: the top-bar documents door leads to this project's document instead of the vault root, the footer names the file's path beside the slug, "View topology" stands first and filled as the page's primary action, and a zero document count is no longer drawn.
The Library graph is now a flow picture: sources stand in a column on the left, the pages written from them in the middle level with the files they cite, and the concepts those pages name on the right, with the lines running left to right as S-curves. The same folder draws the same picture on every visit and every machine. A long folder lays itself out as a taller list the camera fits; past an 18px row a column of pages folds into sub-columns with room for every name, and a folder of hundreds of files folds its file band into a grid that names a file on hover or zoom. File names stand left of their squares, page and concept names right of their discs. Dragging a mark now moves the whole picture; the legend under the canvas says what stands where.
The Library's new-automatic-check sheet now reads back, in one sentence above its controls, what the check will do and when, and the sentence follows every press. The sheet is also capped at the window height and scrolls inside it, so its save button can no longer fall below a short window.
The empty Collections and Rounds tabs draw one dashed ghost row with the names of their columns instead of three grey loading bars: the bars read as a load that never ends, and an empty destination is finished, not loading.
With exactly one project in the folder, the Projects door opens that project instead of a list with one row: the desktop rail, its keyboard shortcut and the mobile tab bar all read the same answer, and the list is still one press away as the breadcrumb at the top of every project page. A folder with no project, or with several, opens the list as before.
Double-clicking a map node opens or folds its children (the same act as its +N chip) and keeps the node selected. Before, the second click of a double-click deselected the node, so the gesture did nothing. A second quick click on a leaf keeps the selection too.
On a wide monitor the Library graph legend keeps a reading measure: from 1536px the line is bounded to the document column width instead of running the full pane, so the teaching sentence wraps into readable lines. At 1512 and below nothing moves.
The Library's source list and wiki shelf keep only the rows in view in the DOM. On a 3,000-file folder the index cost 33,000 nodes and a 450 ms task on the rail press, and scrolling hit a 400 ms stall; it is now about forty rows, the scroller keeps the whole list's height, and scrolling stays under a frame.
The Agents destination no longer has a header tab strip: MCP is the last section of the Agents page, with its own share/connectors switch. The ?tab=mcp address and the /mcp/ redirect scroll to that section, so older links and the installed app's deep link keep working.
The map's two expand controls no longer share a name. The button in the toolbar keeps "Expand all", which opens every folded group; the bar above a selected node now reads the bare verb "Expand", the mirror of its own "Collapse", so the one that stays inside the selected node is the one without "all".
The Library's source list and wiki shelf are one tab stop each: the arrow keys, PageDown and PageUp, Home and End walk the rows, End reaches the last of four thousand, and Tab leaves the list. Before, Tab moved through every row.
A step whose title Atlas wrote automatically now reads in the reader's language everywhere it is drawn, in the list's reason column and as the detail's headline, with the original title kept one line down as the audit trail.
On the web, the folder reread that keeps the Library and the map current now waits at least twenty times its own cost between polls (capped at a minute): a 12,000-file folder reread costs about 230 ms, so it still polls every 5 s, and no folder can make the poll take more than a twentieth of the time.

Fixed

The status line above an open Library page counts what the home counts, in the home's words: how many originals are not compiled and how many have changed, instead of one summed "waiting" figure that disagreed with the numbers a press away.
The line under the Library's wiki list that counts pages missing the template now counts the rows on screen: under a search it counts the matches, instead of the whole folder's figure standing under a list of none.
Zooming the wheel out of an opened island returns the Library map and leaves it standing at its fit: the wheel steps that follow the return belong to the return, not to the map, so the archipelago no longer ends up small in the middle of a dark canvas.
An opened island in the Library graph counts what its picture shows: its files, its pages and its cites. The folder caption used to run there and read 0 concepts and 0 mentions, a count of the concept the island itself is.
Counts in the Library past a thousand are written one way everywhere: the status line, the home's clauses, the graph caption, the tab counts and the index chips all group thousands, in Korean as in English, so a folder of 3,000 originals no longer reads "1,400" in one place and "1400" in the next.
Selecting a map node now draws its neighbours that live in other folded domains, with their names and lines, so the ego graph matches the relations the panel lists; each folded parent's chip claims only what still folds. The focus camera also comes to rest: its target is clamped to the leash the map keeps around the selected node, where before a far-off target kept the map redrawing at full frame rate for the whole selection.
On a narrow window the map no longer shrinks to a speck. The camera used to reserve a fixed 470 pixels for the side panel and the tool rail whatever the window size, which is more than a phone screen is wide; it now reserves at most half of the frame, and never less than what an open panel actually covers, so the graph uses the same share of the canvas on a phone or tablet as it does on a desktop.
Viewing a single area of the map now calls that area by the same name everywhere on screen. The chip, the node on the canvas and the panel's list all used the name written for your language, while the panel's heading used the underlying English one.
The map's meaning-review chip now says whether the panel is open and closes it when pressed again. It lit up while the panel was open but announced nothing to a screen reader, and pressing it again did nothing, so the map stayed narrowed with no way back from that control.
The INDEX tree now tells a screen reader how deep each row sits and where it falls among its siblings. Its rows are drawn as one flat list with indentation, so without that a reader announced a project and its domains as peers, and an expanded domain's capabilities as more of the same.
Asking the map how one concept reaches another now names both ends of the answer. The path's nodes kept full ink but competed for label space as ordinary concepts, so an end sitting inside a crowded domain was drawn without its name.
The map's focus leash, which keeps a selected node from leaving the frame, is now sized to the screen: half the free extent beside the open panels less an edge pad, per axis, with the old token as its floor. The selection frame respects it when it picks a zoom, so a node whose neighbours sit far across the map is framed with all of them beside the detail panel instead of under it.
Restoring one document now refuses when the file on disk cannot be read at all, so the uid and slug guard can never be skipped, and the app-scoped keychain mirror is only treated as done when the write command really succeeded. On the map, the path and constellation lens sink a capability or element name exactly once, as it already did for a domain name, and the Record screen's discard door stands in the document header at a readable size instead of under the whole document.
In the Library flow picture a page or concept name shortens with an ellipsis before it leaves its side of the column, instead of dropping under or left of its disc onto the rows below when the canvas is narrow (the conversation dock open beside a 1512 window).
Pressing the map's agent chip again now closes the agent panel and gives the map its width back. It opened the panel and then ignored every later press while still showing itself as active.
The Library source list and wiki shelf follow their selection: a page opened from the graph card or a citation, or a file opened from a card, is scrolled into view in its list and becomes the keyboard stop, even when its row was not rendered yet. Before, on a large folder the list stayed at its top with nothing marked.
The saved-questions popover in the Library no longer says its name twice: the section heading inside it is kept for screen readers and hidden from sight, since the popover title already carries it.
A Library search that matches neither an original nor a page now says so in one sentence naming the word, under whichever list is open, instead of leaving the list blank under a caption of zeros.
On the Library's islands map, an island's name no longer stands on a neighbouring island's marks: it tries the seats under, over, right and left of its own island and takes the first clear one, may at most brush a neighbour's shore, and otherwise waits for the camera to close in. A name that fits across its island's body now sits inside it.
Closing the keyboard-shortcuts sheet now puts the keyboard back on the button that opened it. That button disappears while the sheet is up, so the sheet had nothing to return to and dropped focus at the start of the page content instead.
The project detail page reads in one column and keeps its name: the overview body stands in a centred reading column at the reading size instead of a 14px run left-aligned in a card twice its width, the four hero controls take their own row below the name until the xl width (the name column was 80px at 768), and the heading is read by its content (screen readers heard "Project name" where the project's name belonged).
A question typed into the Library's conversation can now be saved as a page: the session carries the Library's citation rule (every fact cited in the wiki's own form) for whatever the person types, not only for the question door's brief. When saving is refused, the reason now stays under the save button until the next answer instead of leaving with a toast.
A map node drawn near the bottom of the canvas keeps its name: when the label's usual slot below the node falls under the floor band reserved for the corner readout, the name sits above the node instead of being dropped.
The number on a domain is now the number opening it reveals: a concept several domains reference is counted once, under the domain the containment spine holds it in, and a cluster chip's +N states the whole subtree it hides instead of only its first rank. Opening a fan near an edge reframes the camera once, so the parent and every child it opened stand inside the visible region rather than crushed against the top of the window or behind the INDEX panel. In the Library's column picture, a page name runs on into the concept column's width when that column holds no rows, instead of being cut while the space beside it stands empty. The Insights guide's second page points at the headline the screen opens on, so the guide is two pages again.
In the browser, the Library graph card for a file says Open a copy on its reveal door, the same word the source pane uses, instead of Show in Finder: the browser has no Finder and hands over the file it was granted.
A concept whose name cannot fit below it now tries the space above before the map gives up and draws it unnamed. Opening every group in a folder of twenty concepts left one of them anonymous even though half the canvas was empty.
The keyboard walks the Library flow picture in its reading order, column by column and top to bottom; before, the arrows followed the model order and three presses from the top landed on the seventh square.
A conversation turn stopped before the agent said anything now leaves one line under the request, Stopped before the agent answered, instead of a bare request bubble that read as an agent that never replied. A stop after the agent had spoken, or while a tool was running, adds nothing new.
On the web, the "Get the app" door at the end of the Library's "writing needs the app" note, and its twin under Find documents, now wear the link colour the Rounds tab's door already wears, instead of the grey of the sentence they end.
The map's saved-constellations chip now points at a named region that screen readers can find and announce. It used to say it would open a dialog while opening a panel with no name and no link back to the chip, so a reader was told something expanded and had nothing to move to.
Inside the Library's reading column, the open original's note, the wiki page's problem cards and the check report's sentences now run to the column's edge like the body does. Set in smaller type, each carried its own line cap that resolved shorter than the column, so one card showed three right edges.
On a saved answer in the Library, the evidence block above the answer now runs its lines to the column's edge like the answer's body does; set in small type, each line carried its own cap that resolved shorter than the column and drew a second right edge over the page.
The room a flow page name is cut to now keeps the same breathing gap the collision pass tests with, so a shortened name passes on its own side instead of still falling to the left.
A wikilink in a page Summary or a concept description now reads as the words a reader would see in the Library card, topology labels and node drawers: About capabilities/payments instead of About [[capabilities/payments]].

2026-09-15 · v1.2.5: Galaxy skies, saved constellations and reliable map restoration

Added

Save named constellations of existing ontology concepts with a purpose, revisit their real stars in Galaxy, open the same sets in Library Collections, and prepare an explicit agent handoff. Read-only MCP and source CLI recover each saved set by ID while preserving missing identities and separating user scope from current meaning and impact evidence.

Changed

Galaxy stars now flare independently on intermittent 4–8 second cadences. Each brief corona-and-glint flare leaves the contrast-safe hot core, graph position, and hit target unchanged, and reduced motion holds every star at a steady readable state.
Galaxy now presents every real concept in a stable three-arm ontology sky, reveals actual relationships on attention, approaches selected stars without losing context, and keeps its compact map utilities identifiable by tooltips.
Galaxy now keeps its ontology stars anchored while cached diffuse gas flows gently inside the arms, luminance breathes, and shooting stars use varied entry-seeded background paths; reduced motion freezes the atmosphere.
The Projects destination now presents a compact index built for choosing a project: each row shows the linked project name, its explicitly authored one-line purpose, a recent update, and clearly labelled actions for details and the map. Graph counts, domain composition, recent document activity, and CLI or agent creation instructions no longer compete with project selection; project composition, editing, and construction review remain on the detail page.

Fixed

Switching into Galaxy now paints stars immediately, keeps each mode's camera view, and restores the prior Flat zoom after layout settles, including reduced motion and a late viewport resize. Map loading uses a restrained light cluster with a static reduced-motion equivalent, replacing the rotating ring and document-box illustration.

2026-09-15 · v1.2.4: Neural maps, clearer workbench interactions and traceable meaning review

Changed

Opt-in compact task handoffs preserve complete qualifiers and bounded source evidence, refusing conflicting or unsupported capability claims. MCP schema compatibility and validation diagnostics retain typed evidence without granting new write authority.
Task meaning review retains complete qualifiers, omitted coverage, correction and deferral. Supported single-patch decisions and correlated writer/readback evidence are preserved in immutable History records; semantic acceptance, write permission, code checks and deployment remain separate.
The folder chooser uses one integrated list with opening and creation controls together. Map entry gives visible loading feedback, and the 3D picker uses the shared menu and control design.
Neural replaces Cloud with relation-driven groups and restrained cell light. All three 3D arrangements share bounded projected curves for rendering, picking and inspection, with readable labels, reduced-motion transitions and a visible return to the overview.
Library now holds Sources, Wiki, and Ontology in one destination while keeping each file on disk as the source of truth. Markdown source headings open their cited lines, the Ontology tab starts with graph-backed documents and finds localized names or canonical paths, and legacy /docs/?slug=… links and unsaved drafts survive tab changes. Wiki creation previews the actual Markdown path and section structure in a centered dialog, the current work receipt stays separate from its history, and a successful app write produces one actionable notification instead of a duplicate folder-watch toast.

Fixed

The full-detail surface a node opens into behaves like a place you can read and leave: the body is rendered as the Markdown it is instead of printing ##, - and backticks at the reader, the 1/2/3 reach steps are a visible control with a name and arrow keys rather than three grey numerals in a sentence, opening the document leaves a crumb back to the node it was opened from, and pressing edit-relations closes the card so the editor it opens on the map is actually on screen.
Search and related concepts retain keyboard focus, including immediate focus indication in the installed macOS app. Map controls respect INDEX and detail overlays and existing touch targets; selected agents survive asynchronous detection. Accent previews, notification wrapping, analysis and Harness wording, and project validation and save feedback now reflect the actual state.
Updated the TLS dependency rustls to 0.23.45 to reject TLS 1.3 handshake messages sent at the wrong encryption level (RUSTSEC-2026-0285).

2026-09-14 · v1.2.3: The Architecture tab becomes Harness, and launching asks which folder to open

Added

The Architecture tab is now Harness and answers three questions about what this repository tells its agents, the guides each tool actually reads with a source cited for every tool and file claim, the structure those guides describe, and the sensors watching the work, with each row bounded by what a file can prove so a wired hook says its script exists rather than that it ran
When Atlas knows two or more folders on this computer, launching now asks which one to open instead of silently reopening the last. Each row says how many documents and concepts it held when it was last read, when that was, and whether it can be opened right now; a folder that has been renamed, moved, unmounted or refused says so before it is pressed rather than failing after. One known folder still opens directly, and dropping a folder from the list is what returns a launch to opening directly, so there is no setting to find. Whichever way the launch went, the open folder's name now sits at the top of the rail on every destination, and pressing it opens the same list plus a picker for a folder Atlas has not seen.
The download page and the third-party notices now state that Claude Code, Codex, Cursor and the other product names Atlas mentions are trademarks of their owners and that Atlas is not affiliated with or endorsed by any of them, so a reader can tell a statement of what Atlas connects to from a claim of partnership
Repository analysis accepts bounded sourceReads and returns exact implementation ranges with file hashes, continuation and refusal state. Proposal replay rechecks selected bytes and binds them into the opt-in lifecycle source digest. Source citations resolve observed provenance without becoming semantic authority or authorizing writes; legacy calls and vault-only read_source remain unchanged.

Changed

The Harness tab now opens on a coverage matrix whose rows are the areas the ontology records for this repository and whose columns say what tells, gates and watches each one, where a file appears in an area only because a path it declares reaches a path the ontology records, an empty cell is written as a sentence beside the ontology's own record of what that area is for, anything declaring no path stands once in a row above the areas instead of being repeated down every column, and beside the matrix every authored Markdown file is split by whether a guide is pointing at it, so a document nothing names is visible rather than silently unread
The meaning-corpus report exposes candidate discovery, golden-oracle consistency, and unmeasured independent construction in JSON and readable output, deriving effective thresholds without changing existing pass/fail behavior. The V1–V4 execution plan registers 19 concrete items, including six separate human-review UX commitments; BACKLOG owns their live status.
Aligned README, download and guide copy with Atlas's human value: understanding, evidence-calibrated judgment, and control through shared codebase meaning. Product and PO guidance now share an Atlas-specific thesis and distinguish current task context/write review from unproven construction, Meaning Diff, and repeat-use goals; existing solo/two-reviewer routing is unchanged.
Bootstrap guidance now investigates bounded source predicates, state changes, attempted effects, failure paths and verification before treating a structural index as insufficient meaning evidence. It preserves conditions, counterexamples and unknowns, keeps unsupported domain assignments outside proposals, and retains independent qualification and exact human acceptance.
the Harness coverage matrix says what its marks mean: a square is filled when something names a domain and drawn empty when nothing does, each column carries a card with the number of domains it has no answer for, a mirrored guard appears once with the tools that read it, and a read under a second shows no wait screen at all.
Bootstrap guidance distinguishes supported tentative responsibility hypotheses from accepted meaning. It carries source-derived conditions, counter-boundaries and uncertainty into candidate bodies under existing authority/qualification gates, checks reader limits before requests, and keeps unsupported assignments outside proposals.

Fixed

On the Harness tab, the panel that explains how checks are counted is readable again: it was drawn under the view below it, so the guides heading, the file table and the matrix printed straight through the explanation and the two texts read as one smear.
Construction qualification now advertises the existing nonempty purpose-authority arrays in its public schema, so empty decisions, non-goals and source references are rejected consistently. The handoff contract explains that accepted quantifier classifications retain the sealed rationale and source-reference order; existing runtime approval and equality gates are unchanged.

2026-09-13 · v1.2.2: switching destinations no longer reloads the app, and every failure a person meets is a sentence

Changed

The notification panel now answers three questions in one place with the counts engraved in its tabs, what waits on you, what got done and what happened when, instead of one undifferentiated event log, and opening it no longer marks every unread result read
Check page format answers with one sentence and one action per problem, naming pages and originals by their titles and the place in words, and keeping the codes, line anchors and tool notes whole one press away behind a fold
A press on the Library picture opens a card anchored beside the mark rather than leaving the picture for the page, carrying the mark's identity, one sentence, a facts line that goes amber with the stale citation count, the marks at the other end of its relations with each one's state, and doors to open the page or ask for a redraft
verification is faster and narrower: pre-push runs what a push can break, CI shards its slowest job, and tests wait for the condition rather than the clock.

Fixed

the folder check report reads as sentences, states each rule once instead of repeating it for every finding, and its table of contents lands on the section it names
a failure a person can see is now a sentence in their own language rather than a bare code or an English message, and an action the page already knows cannot succeed is offered disabled, with the reason and the next step readable before the press
Fixed a second LLM call being refused with audit-log-busy on a healthy vault when Atlas happened to be starting a child process at that moment.
the Library picture fills more of its canvas, the fit control moves when pressed and rests once the picture is framed, a mark never loses its name, and an unverified citation's amber mark is visible on a small folder and calm on a large one
in the installed app a rail press is a route change again rather than a reload of the whole app, so the window no longer goes blank on the way (0.1-0.3 s on a 104-document folder), and no longer shows another folder's screen, complete with a download button, on the way in.

2026-09-12 · v1.2.1: the Library opens on its graph, and the graph stays readable at three hundred pages

Added: the Library computes the structural check itself and keeps it, behind a report door naming whether a check runs, what it covered and what it has not; the README points at the MCP directory listings.

Changed: the Library's home is its graph, with the guide and its strip's clauses as doors; the graph keeps the map's scale, names every page and sets files as satellites of their hubs, so three hundred pages stay readable; one text edge per column over a 16px reading column; the index grid, header, tooltip and toast take plain labels and their places; the type ramp is in rem, so text zoom reaches it.

Fixed: touch floors reach hybrid pointers; an arriving route takes input at once; the source pane clears its bottom and keeps sha256 on one line; the source inventory reads the 8000 entries the dogfood repository outgrew; the installed app never offers its own download.