Cairn Documentation
Version History
Release notes and feature updates. Cairn follows semantic versioning: MAJOR.MINOR.PATCH.
Onboarding, Part 4: the model forms as you watch, and the cairn comes with you
The interview shows your model taking shape and what each answer would do to it, and the cairn the first run ends on follows you into the workspace, where every change says what it added and what comes next.
- While Cairn reads your description or your answers, the page shows what it is writing as it writes it: what it understood, the questions, the sharpened description, and the actors and requirements, beside your own words.
- Pointing at an option, or choosing it, draws what it would add in Your model so far: a new actor, the requirements it brings.
- Each round's summary leads with what your answers did, and marks the same facts in your model. Facts Cairn only rephrased are named in one line instead of listed as changes.
- The question screens use the width of the window, with the model panel given more room.
- The status bar keeps the system's cairn; click it to open Completeness on the system.
- Change Review names the completeness checks the operations you ticked would meet, before you apply them.
- After a change that meets a check that was open, a card shows the stone that grew and your next move, with Ask Cairn. A complete cairn is celebrated, and points to the part with the most left to do.
- The Completeness lens draws the node's cairn and starts with its next move; the Overview's From your interview card draws the system's, with why each move matters.
- Your systems shows each system's cairn and its next move.
- A fact renamed between rounds read as removed and lost your choice to keep it. Facts keep their identity from round to round now.
- Every assumption said You weren't sure. Only one from a question you answered Not sure says so now; the rest say what Cairn assumed them from.
- The Behavior specialist can add up to 20 transitions at once, up from 12, so a machine of eight states keeps its fault exits and recoveries.
No false “Changed on another device” after a first run
A system built on one computer could say it had been changed on another device.
- Adding the first proposal at the end of a first run recorded the change twice, so the system's backup read as changed in two places, and the status bar said Changed on another device. It is recorded once now.
- A first run's backups, once the system is built, are checked against your account like any other.
Hardening, Part 2: one system in many places
A system open on two computers, in two tabs or under two accounts no longer loses either copy's work, and the workspace reads and works better from the keyboard.
- Opening a backed-up system checks your account first. A newer copy there opens instead, with a word. If both copies changed since they last matched, Cairn asks which to carry on with and keeps the other as a separate system, so nothing is lost.
- A backup that would replace changes made on another device waits instead, and the status bar says so: Changed on another device · Choose.
- A system open in two tabs is saved from one at a time. The other says so, and Use it here moves it, with the first tab's last changes and its notebook's.
- An interview answered on a phone and a computer at once keeps both: the one further along carries on.
- Signing out saves your last edits and sends waiting backups first, and so does a page that is hidden or closed. A dropped connection no longer sends you to sign in.
- Your own API keys are forgotten when you sign out, so the next person on this browser can't use them.
- A colleague's exported system imports as your own, with backup off until you turn it on.
- When Cairn can't apply a change, it says what and why in your words: “The data type ‘Telemetry Record’ needs a category.”
- Change Review leaves out a change it can't apply, with its reason (“Apply 3 of 4”), instead of asking you to force it.
- Undoing a delete brings back the part's simulations.
- A request that reaches your monthly allowance midway stops and says so.
- Escape closes the window on top, and only that one; focus returns to where you were. Lists in the tools work from the keyboard.
- Traceability is ⌘⇧X (browsers keep ⌘⇧T). Its Link adds the part that satisfies a requirement.
- Quality says when filters hide findings. Images being drawn show only on their own part. A file not on this computer says so.
Hardening, Part 1: nothing fails quietly
What review shows is what Apply accepts, your work is kept even when saving or backing up fails, and when something does fail, Cairn says so where you are.
- An AI proposal is checked by the same guard that decides Apply, so anything Apply would refuse is marked in review before you click it.
- A proposal you parked, or the first proposal you put off, is brought up to date when you open it: ids the model has taken since are renamed. A review is never applied to a different project from the one it was proposed for.
- Splitting a node, Apply Valid Only, and deselecting part of a cascade now apply what is left instead of being refused; what can't go without what you left out is dropped, with the reason.
- The trace above a review names each step of a plan and its time. The AI's reasoning speaks about your system, not about itself.
- If this device can't save, the status bar shows Not saved · Retry until it can.
- Leaving a project no longer drops its last minute of backup. A failed backup is tried again on its own, and shows until one goes up. A restore becomes your account's latest copy at once.
- Importing a JSON export of a project already on this device brings it in as a copy.
- Older projects missing a description load with an empty one, so their entities can be edited again.
- Messages show above every window, and on Your systems too: a refused change, a failed export, upload or download, a system no longer on this computer, an account that couldn't be reached.
- The first build stops, rather than leave out your last answers when they couldn't be added, and says when the project hasn't reached your account yet.
- A failed parameter suggestion says why, instead of looking like no suggestions.
Onboarding, Part 3: a first run that teaches, and ends on your system
The interview explains every choice, the build shows its work, and the first run ends on your system's day one instead of a list of numbers.
- Each option says what choosing it means, and the one you choose opens on its consequence for the design and what it adds to the model: an actor, a requirement, or the part it shapes.
- Why this matters tells you what a question decides, and Not sure says what Cairn will assume before you choose it.
- While Cairn writes the next round, each question appears as it is written.
- The build tells you what it is doing with your own names: the actors placed, the requirements written.
- The first concept image of your system, drawn during the build, is shown the moment it is ready, beside the sketch.
- The Architect is asked directly, in one call, so the first parts come sooner, and nothing in the proposal can depend on a step that never arrived.
- The first run ends on a page about your system, on a computer and on a phone: what you built and from how many answers, its completeness drawn as a cairn of stones, and your next three moves, each with why it matters.
- On a computer, a move opens the workspace with its request ready in ⌘K, and four lens previews of your own model open their lens. On a phone, the link to your computer comes first.
- The Overview's From your interview card offers the same next moves.
- A first proposal you put off opens up to date with your model: ids the model has taken since are renamed, so Apply no longer refuses it.
- When the AI numbers new entities from 001 and your model already has some, the renamed ones no longer collide with the proposal's own, and the signals that use them follow.
- A change Cairn can't apply now says why above Change Review, instead of behind it.
Fixes for the first run
The interview, the first proposal and adding it, found by running the first run on a phone.
- When the AI's reply to your description or your answers can't be read, Cairn asks once more before showing an error, so one bad reply no longer stops the interview. If the second can't be read either, the error says so in plain words, and Try again picks up where you were.
- Switching to another app on a phone while the first parts are proposed no longer loses them: a request the phone dropped is asked again when you come back. If a step of the proposal is lost anyway, the first build says so and offers Ask again, instead of showing what was left (connections with no parts to join) as the proposal. The same holds for any AI request whose steps build on each other.
- A proposal that adds new data types can be added again. The AI didn't know a data type's fields in a new project and left out its category, which stopped the whole proposal; it is now told the fields, and Cairn fills in any it leaves out. When a proposal can't be added, the interview now says why instead of doing nothing.
Public site, Part 2: the real product, page by page
The demo is now the app itself, and every page about what Cairn does shows the real workspace and claims only what it does today.
- The demo opens the real workspace on the rover template: the tree, every lens, the inspector, Explorer, Quality, History and Trace. Nothing is saved, and Reset loads the rover again.
- Three AI requests are scripted (Hub Motors' parts, the Battery Pack's requirements and a state machine for the cargo lock), so the real router, specialists, checks and Change Review run with no account or key. Anything else says what the demo can answer.
- A tour points at the tree, the lens bar, the command bar and the tools.
- On a phone or a narrow window, the demo shows pictures of the app and the way to start a model.
- Eleven feature pages, six of them new: the interview, decomposition, architecture with N² and interface lists, behavior, design decisions, requirements, traceability, the verification matrix, completeness, Change Review and History, and exports. Each shows the app on the rover.
- Use cases walk through a workflow in each team's own domain, and say what Cairn doesn't do for them yet, starting with a shared workspace.
- Comparisons describe Cairn only, say when they were checked, and carry trademark notes. Cameo Systems Modeler is named in full.
- The glossary adds N² chart, VCRM, ChangeSet, TRL and flowdown, and says where Cairn does nothing for a term.
- The blog says what shipped where posts once said it was coming, the pipeline post describes the pipeline as it is, three false lines are gone, other tools' prices are out, posts are listed newest first, articles read in dark mode, and four more posts have their own social cards.
- The retired concept pages now lead to the feature pages that replaced them.
- The quality score is now the share of the model with nothing to fix. Every node, requirement, interface, state, transition, verification and data type counts once; an error on one zeroes it, each warning halves it and each note takes a tenth off. Before, each finding took a fixed number of points off, so any model of real size scored 0, the rover included.
Public site, Part 1: one door, told true
The public site catches up with the product. Every page leads to the same door, “What are you building?”, says only what Cairn does today, and shows it in pictures of the real app, in light and dark.
- A new home page opens on the door: describe a system, sign in with an email code, and the interview starts, on a phone or a computer. Start a model in the header leads there from every page, phones included.
- The home page walks through the interview, the first build, the lenses, Change Review and History, traceability, where your data lives and which models run, each with a shot of the app.
- About now says what Cairn is and where it stands. Pricing gives the early-access terms only: the allowance, your own key, the models, and plans at launch, with no prices until there are any.
- The reel stays at /reel for talks, rebuilt on shots of the app, with the lens bar as it is and the door on its last slide.
- Story and the Concepts pages are retired and redirect: Story to About, the concepts to the lens reference and Why Dead Paths Matter.
- Dark mode works across the shared header, footer, home, About, Pricing and the reel, and the footer has a theme switch.
- Public pages load in a fraction of the time: they no longer download the workbench, so the first page carries about a sixth of the script it did.
- A skip link, one main landmark per page, labelled navigation with the current page marked, slides out of view kept out of the tab order, and motion that stops when your system asks for less.
- Change Review names requirements, verifications and use cases by their titles, not their ids.
- History names the model that wrote an AI change, not the one that routed it, and shows each step's cost.
- The rover template's card says four wheels, as its model does.
- Your profile's plan pane counts against the early-access allowance, as the server does, and shows no prices until plans exist.
- Session recordings mask every input and all text, and Do Not Track is honoured, as the privacy policy says.
- The privacy policy and Local-First guide describe backup as it works since 1.17, and list every service Cairn uses.
A demo that stays up, and whole titles in the feeds
Two fixes found by the public site audit that opens the 1.19 refresh.
- In the demo's Behavior lens, picking a state and then switching machine no longer replaces the whole page with an error; the new machine opens on its first state.
- Article titles and subtitles with an apostrophe reach the RSS feed, llms.txt and social cards whole, where they were cut off at the apostrophe (“Why AI Can\”). An article whose entry is written differently now fails the build instead of quietly dropping out of the sitemap and feeds.
AI models: the current Sonnet, Haiku and Opus
Every AI call moves to Claude Sonnet 5.5, Haiku 5.5 and Opus 5.5, and asks for them the way they expect: how much to think instead of a temperature, room for that thinking, and a plain answer when Claude declines.
- The router runs on Claude Haiku 5.5, specialists on Claude Sonnet 5.5, and Claude Opus 5.5 is the choice for your own key and high-quality 3D models.
- Each call sets an effort: low for routing, checks and the interview, medium for proposals, simulations and exports, high for high-quality 3D models. Room for thinking goes on top of what a reply needs, so long answers aren't cut short.
- Replies are read by block type, so a reply that opens with the model's thinking still reads as its answer, in a stream or whole.
- Your own key and Cairn's hosted AI send the same request, built in one place, so the two can't drift.
- Sonnet and Opus calls turn on Anthropic's server-side fallback: a request their safeguards decline is retried on the model Anthropic recommends for it, and the reply continues where it stopped.
- A request still declined proposes nothing, and the palette says which safeguard declined it and how to go on. A declined router hands the request to keyword routing, and the specialist gets its own turn.
- Change Review's pipeline steps name the model that answered, Keywords when the router fell back, rather than reading Sonnet for everything that wasn't Haiku.
- Pricing for the current models: Sonnet 5.5 at $2 / $10 per million tokens and Opus 5.5 at $4 / $20, both with cache reads at 5%, and Haiku 5.5 at $0.10 / $0.50, or $0.50 / $2.50 once a prompt passes 100K tokens.
- Usage counts every attempt a fallback made, each at its own model's rates, and is recorded under the model that answered. A proposal's estimate in Change Review adds up each call's own cost, cached tokens included, rather than pricing its token count at Sonnet 4 rates.
- The AI Pipeline reference gains Models and When Claude Declines, and the FAQ says which models Cairn uses.
- The docs' pipeline steps, tables, notes, code blocks and hover states were light on the dark theme; they follow the theme now.
The newer copy, wherever a system is opened
Three fixes to 1.17, from its last review, for systems that move between a phone and a computer.
- The link emailed from a phone opens the newer of this computer's copy and your account's, so parts added on the phone since the computer last opened the system are there.
- Settling the first proposal, or Done with these on the Overview, counts as a change, so a decision isn't undone by an older copy from your account when an upload failed.
- On a phone, Email me a link to this page from the workspace names the open system, so the computer opens it rather than your list of systems.
Onboarding, Part 2: a first run you can start on a phone
The interview and the first build, rebuilt around one question at a time. Start on a phone or a computer, stop at any point and carry on from the other, see what each round changed, and choose Cairn's first parts one by one before any is added.
- One way in, What are you building?: on the empty welcome screen, behind Begin a new project, and at the interview itself. Examples come from the templates, ⌘Enter starts, and Start from a blank model skips the interview.
- Back up to my account as I go is ticked by default, so the interview is saved to your account as you answer.
- One question at a time, two to four a round, most important first. Options answer by click or by number and Enter; Not sure. Assume for me says what Cairn will do; answers already given sit above the question with Change.
- Your model so far shows the name (rename it there), how clear the idea is, a sketch of the system and its actors, and each actor and requirement with a switch to keep it.
- After each round, Here's what your answers added: each actor and requirement that is new, now stated or reworded, with the answer it came from; the description with its new sentences marked; and what is no longer in the model. Build now is there at every step.
- Reviewed in the interview: every proposed part and interface has a tick, and the sketch shows what you left out. An interface whose other end you left out can't be added without it. A partial choice is recorded as a partial accept.
- Open without them leaves the proposal waiting on the project. The Overview's new From your interview card offers it again, with the next steps the interview points to, and Change Review settles it the same way.
- The interview is a project from its first reply, saved on this device at once and to your account a few seconds after each answer. Save and finish later stops it.
- The welcome screen's Pick up where you left off lists paused interviews, here or in your account, and systems started on a phone. Whichever copy is newer opens.
- The interview and the first build run on a phone, with the model in a sheet and the buttons in reach. Your systems lists what is in your account, interviews to carry on first.
- Once built, Email me a link for my computer sends a sign-in link to your own address that opens the system. Other app pages on a phone say they need a computer, with a link that opens them there. Signing in carries on to the page you asked for.
- History's Origin entry says which device the interview started on and how the first proposal was settled.
- The tour no longer opens for a new account; Take the tour is on the welcome screen, and it ends at the interview.
- The tour's quick preview of your own idea, and the AI endpoint only it used, are gone: the interview does that job, and keeps what it builds.
- The Begin a new project button reads in dark theme.
Onboarding: an interview that becomes the model
The inception interview and the first build of a model, rebuilt. What the interview learns now reaches the model as actors and requirements you choose to keep, and the subsystems AI drafts wait for your review instead of being applied for you.
- Each round settles the system's actors and requirements alongside the description, each marked Stated or Assumed and tied to the answer it came from. Assumptions are no longer written into the description.
- Every question can be answered Not sure, and Cairn records what it assumed instead.
- Options are radio groups, past rounds open by keyboard, and the screens follow the dark theme. The header shows the real app version.
- Before building, edit the description and keep or drop each actor and requirement. What you said is kept; assumptions start unticked.
- The kept actors and requirements go into the model in one ChangeSet. The subsystems the AI proposes open in Change Review in the workspace, so nothing it drafted is added until you apply it. The brief is no longer written during the build.
- A failed proposal says why and can be asked again on the same project; a retry never creates a second one.
- History's Origin entry lists each fact the interview settled, what it became (or that you dropped it), and what the build added and proposed.
- The onboarding tour reads its numbers, parts, requirements and gaps from the rover template instead of figures written into the page, and its last step shows Completeness instead of an impact analysis Cairn doesn't have.
Images, connections and use case titles held to what they claim
Three fixes to 1.15, from its last review.
- An image reads Out of date when its node changes type, since the type sets the shot: a wide shot of a system, a close-up of a part. Images made in 1.15.0 read out of date once, for this reason.
- In Completeness, a part linked only to an alternative a trade study set aside no longer counts as connected. Inside the alternative, its own parts still connect to each other.
- A use case can't be renamed to the title of another on the same node.
Lenses, Part 7: Operational, Completeness and Visuals
The last three lenses, rebuilt on what the model records. Use cases tie to the model's actors and requirements, completeness is a set of checks rather than a score against made-up targets, and an image knows when what it was drawn from has changed.
- Each actor a use case names links to the external actor it matches, on the node or above it. A name nothing matches is flagged, and can be added to the model from the lens.
- Requirements are traced to the use case they come from, from the inspector.
- Every list (actors, before, steps, after) is edited in the inspector, and a use case can be added by hand. The lens is a keyboard list.
- AI names actors the way the model does, and Change Review warns on any it invents or on a repeated title, new or renamed.
- Checks along four axes, each read from what another lens shows: described, parts, the state machine, use cases, requirements traced and testable, parts connected, actors modeled, verification.
- Each gap opens the lens that fixes it, or asks AI where it can help. Checks that don't apply to a node are left out instead of scored.
- The tree shows each node's gaps, Overview counts them and suggests the next one, and the lens no longer opens the inspector by itself.
- An image records what it was drawn from, and reads Out of date when that changes. Before, every image drawn from a description read as stale on a node with a brief.
- A new 3D model replaces the old one only once it is built, shows only on the node it belongs to, and choosing no reference image is respected.
- The tabs are keyboard tabs, buttons say what they do, deletes ask first, and colours follow the theme.
AI alternatives and assessments checked before review
Two fixes to the AI paths of 1.14, from its last review.
- Change Review warns when a suggested alternative is weighed against a part that is not a live option beside it, or depends on a part that is not in the model. Before, it showed up only in the Dendritic lens after you accepted it.
- An AI maturity assessment replaces evidence that older projects kept with the stage, even when it gives none, so the old text can't come back under a new stage.
Decisions and readiness that claim only what is recorded
The sixth lens deep dive, on the two lenses that read a design's history and its foundations. Both made things up: every node in the Dendritic lens read as an "engineering decision", and the causality pyramid counted pruned alternatives as prerequisites while dropping anything three levels down. The causality analysis was AI prose written into the model with no review. Both lenses now show what is recorded, say where nothing is, and can be edited by hand.
- The lens lists every decision in the design: for each node with parts, what was chosen, the alternatives set aside and what is still open. Before, it was a second copy of the System Tree with a detail panel.
- An alternative says what it was weighed against ("instead of Battery Pack"), as well as its reason, first principle, decision type and phase. The type and phase show only where they are recorded, instead of defaulting to "engineering decision".
- Decisions that record no alternative are counted and filtered, with ◆ Suggest alternatives on each. A reason missing, or a link to a node that is gone, is flagged.
- The inspector edits the evaluation by hand: chosen, open or set aside, what it was weighed against, why it lost, its first principle, type and phase. Each change can be undone.
- Decisions are a keyboard list (↑ ↓ Enter), in the theme's colours. Opening the lens no longer opens the inspector.
- The pyramid leaves out alternatives set aside and outside actors, and puts everything three or more levels down in parts and materials, which before were dropped.
- Maturity rolls up: a node can be no further along than the least mature thing it rests on. The lens names that limit and the stones that set it, and says when the node claims more. The rover reads as in development, but rests on Thermal Management at concept.
- Stones not assessed say so, and how far along they can be at most. A subsystem with no parts is flagged.
- The inspector sets a stone's stage, confidence and evidence by hand. ◆ Assess maturity asks a new Causality specialist for the rest. Its proposal arrives in Change Review and never changes a stage a person set. The old "Analyze Causality" prose, saved without review, is gone.
- Stones are buttons, in the theme's colours, and opening the lens no longer opens the inspector.
- Change Review shows each parameter an update changes, with long prose such as a prune reason in full.
- The Overview's Causality card shows the limit and what sets it. The Dendritic card counts alternatives, and decisions with none.
Briefs that keep what the AI leaves out
Three fixes to the brief from the last review of 1.13.
- When an AI rewrite of the brief leaves sections out of its answer, those sections keep their text. Before, each one it skipped was proposed as empty.
- Operating Modes is written from the use cases as well as the state machine, so changing a use case marks it out of date.
- Change Review shows an AI update that only changes which entities a section cites.
A brief and a narrative that know what they were written from
The fifth lens deep dive, on the two lenses that tell a node in prose. Prose drifts from the model the moment the model moves, and nothing said so: the rover's brief read as out of date on a fresh project, an AI rewrite could wipe a node's parameters, and the narrative drew every subsystem as an island and saved the AI's output straight into the model. Both lenses now say what they were written from, when that has changed, and send every AI rewrite through Change Review.
- Each of the eight sections says what it covers and is written from named parts of the model: Purpose & Scope from the node and its requirements, Interface Summary from its interfaces, and so on. When one of those changes, the section is marked out of date with the reason ("interfaces changed"), and only the sections it touches.
- The header's one AI action does what the brief needs most: update the out-of-date sections, write the missing ones, or rewrite it. ◆ Rewrite on a section asks for just that one, and the AI copies every other section as it is.
- Text you write locks its section, and the AI keeps a locked section as written even if it tries to change it. A locked section the model has moved on from says so quietly.
- The entities a section cites are links; one that is no longer in the model is struck through. Text being typed is never replaced by a change arriving from elsewhere, and Esc puts a section back.
- An AI rewrite of the brief no longer drops the node's parameters and other properties: an update that sends only some of a node's properties is merged with the rest.
- The narrative tells how a node's parts work together: its children, less the alternatives a trade study set aside, joined by the interfaces between them at any depth. Before, it read the whole subtree, so the system node, every subsystem and each pruned alternative appeared as islands.
- The parts and links are drawn before any AI runs. ◆ Write the narrative is a request to a new Narrative specialist, which picks the mainstay and a verb for each link and writes the sentence and a few paragraphs. Every id and direction it returns is checked against the links, and the result arrives in Change Review instead of being saved straight into the model.
- The story is stored by link with a fingerprint of what it was told from, so the lens says when the node's description, its parts or the links between them have changed, and ◆ Update the narrative retells what that touches.
- The systemigram is drawn in the theme's colours, with labels placed clear of the parts. Parts are buttons: click or Enter opens one in the inspector, double-click opens its own narrative. A Links table lists who does what for whom, and the parts the story leaves out are listed with a way to connect them. Export draws on the theme's background.
- The inspector shows where a part sits in its parent's story: its role, what it does for its neighbours, and the mainstay with it marked. Opening the lens no longer opens the inspector by itself, and the scroll wheel scrolls the page again.
- The rover template ships with its narrative told.
- Change Review shows each changed brief section, and the narrative's sentence, story and links, old text above new. Before, an AI update showed no fields at all, because it arrives without what it replaces; the model as it is now stands in for it.
- The Overview's Brief and Narrative cards say how many sections are out of date, and whether the narrative is told and current.
Behavior checks that say exactly what competes, and why
Three fixes to the state machine checks that 1.12 introduced.
- Competing transitions are flagged only when nothing tells them apart: either has no guard, or both have the same one. Before, with three transitions on one event guarded
x,yandy, the one guardedxwas flagged too. - Opening a transition from the Transitions list's Issues shows why it was flagged, including a final state it leaves or an unreachable state it leads to. The list and the inspector now share one definition.
- The review of an AI proposal checks every machine the proposal changes, even when the AI leaves out what an item was before the change. A deleted or moved initial state is no longer missed.
State machines you can read, check and build by hand
The fourth lens deep dive. A state machine has to answer three questions: where does it start, what can happen next, and can it get stuck? The Behavior lens now answers them. The rover's diagram used to draw seven states 40 px apart with their trigger labels piled on top of each other. It now lays the machine out top to bottom in the order things happen, and every label can be read.
- States are laid out top to bottom by how many steps they are from the initial state. Connectors bend around states in their way, and each trigger label slides along its connector until it clears the states and the other labels. A transition's guard shows when you hover it.
- Opening the lens no longer writes to the model: states without a position are laid out in memory, and Auto layout applies the same layout. The rover template and AI-generated machines leave placement to it.
- Dragging a state for the first time keeps every other state where it is drawn. Before, they jumped into a row underneath it. The Architecture diagram had the same problem, and it is fixed there too.
- Drag from a state's dot to another state to add a transition. Right-click the canvas to add a state where you clicked, or a state to add a transition from or to it.
- The simulator shows each trigger's guard, counts the states entered and transitions fired, and keeps a log of the run.
- Transitions is the state transition table: trigger, from → to, [guard], / action, and any problem with it. Filter to the ones with issues, search with /, move with ↑ ↓ and open one with Enter.
- Events is a state × event matrix. Each cell says where an event takes a state. An empty cell means the state ignores that event; press Enter on it to start that transition with its state and event filled in. Cells where two transitions compete are flagged, and final states are hatched.
- The lens checks the machine for: no initial state, or several; states the initial state can't reach; dead ends; final states with a way out; transitions with no trigger; and transitions out of one state on the same event with no guard to choose between them.
- Reachability follows the transitions from the initial state. Before, Quality called a state reachable if anything pointed at it, so two states that only reach each other passed.
- Quality, the Overview's Behavior card and the review of AI proposals use the same checks, so a problem has the same name everywhere it appears.
- + State and + Transition add by hand from any view. The transition form suggests the machine's events and the signals coming into the node. A node with no machine yet can start one from its first state.
- The inspector lists a state's transitions in and out, each one click away, and changes a transition's From and To within its machine. Both show the problems the lens flags.
- The "Add Sub-States" AI button is gone: the model has no nested states to put them in. Typical duration no longer shows its unit twice.
Architecture at every level, an N² chart and an interface list
The third lens deep dive. Interfaces are usually drawn between assemblies, but a node is read one level at a time. The Architecture lens now shows an interface wherever it matters: as a connection between the two blocks that hold its ends, or as a link to a node outside the one you're viewing. In the rover template every interface runs between assemblies, so its top level used to show nine unconnected blocks. It now shows eight connections.
- Connections made through the blocks' parts are drawn dashed, with the number of interfaces they carry. Nodes outside the viewed node sit on the left and right edges, at the end of the interfaces that cross its boundary.
- Arrows show which way signals flow, read from each signal's direction. Connectors run border to border, so they stay readable wherever you drag the blocks.
- Click a connector for the interfaces it carries; one click on any of them opens it. Double-click a block to open its parts, or an outside node to go to it.
- A node with no parts is shown in its context: itself in the middle and everything it connects to around it.
- Blocks say how many interfaces sit inside them, and flag the ones nothing reaches as "Not connected". Interfaces drawn on a decomposed node itself, rather than on the part that carries them, are listed as "on the node itself" so you can route them to that part.
- Opening the lens no longer writes to the model. Blocks without a position are laid out in memory. Before, the first look at a node added an "Initial layout" entry to History. Auto layout now arranges blocks left to right in the direction things flow.
- N² puts the blocks on the diagonal; the cell in row A, column B holds what flows from A to B, with an Outside row and column for the boundary. Click an empty cell to start an interface between two blocks. Arrow keys move between cells.
- Interfaces lists every interface at the level, with where it sits (between blocks, through parts, across the boundary, inside a block), its protocol and any issues. Rows expand (→ / ←) to show each signal, which way it flows, its data type and rate. It replaces the Interface Control Document table under the diagram, which only listed direct connections.
- Quality flagged every subsystem whose interfaces belong to its assemblies as "has no interfaces". It now flags a node only when no interface reaches it or its parts. It also reports interfaces with no signals, no protocol, untyped signals or a signal named twice.
- The Overview's Architecture card and interface list count what the lens shows.
- New interfaces and nodes made on the canvas continue the project's id sequence, and a new interface opens in the inspector so you can add its signals.
Requirements that flow down, and a real verification matrix
The second lens deep dive: Requirements and Verification, rebuilt around how requirements actually work. They derive from requirements above them, are satisfied by parts of the design, and are closed by verification activities.
- Each row shows the statement, type, priority, what satisfies it, and its verification methods as I, A, D or T letters coloured by status. It also says where it comes from (↑ parent), how many requirements derive from it, and how many acceptance criteria it has.
- A new Flowdown view nests derived requirements under their parents across levels, tagged with the node each one lives on. A search keeps the matches' parents in view.
- Gap chips find what needs work: untraced, no verification, failing, and wording. Type, priority and a search over ids, titles and statements narrow further.
- Wording checks follow the INCOSE guide: a statement without "shall", several "shall"s in one, unverifiable words ("user-friendly", "as needed", "TBD"…), and performance with no number. Quality reports the same checks.
- Adding a requirement takes a statement and the parent it derives from, and continues the project's id sequence. ◆ Flow down asks AI to derive the children's requirements.
- The list updates as soon as AI adds requirements; before, it could stay stale. Long lists no longer animate in for seconds. Rows are keyboard-navigable: ↑ ↓, Enter, and / to search.
- The Matrix is a verification cross-reference matrix: one row per requirement, a column for Inspection, Analysis, Demonstration and Test, and the requirement's state. Click an empty cell to add that method.
- A closure bar splits requirements into passed, in progress, failing and no method. It replaces the old "0% verified" figure, which ignored everything already planned.
- Activities replaces Test Cases. It covers every method, not just tests, puts failing work first, and lets you set status in place.
- A requirement shows and edits its links: what it is derived from, what was flowed down from it, the nodes that satisfy it (link or unlink), and the activities that verify it (add one by method). The rationale is editable.
- A verification shows the acceptance criteria it must demonstrate. Its method and results are now editable.
- System Tree names stay on one line and show in full on hover. Rows no longer jump when the hover actions appear.
- Opening a different project starts on its Overview. Before, the lens, the open tool and the Explorer's state carried over from the last project.
A lens bar you can drive from the keyboard, and Overview as the hub
The first lens deep dive. It covers what every lens shares (how you switch, what each tab promises, what happens when the model changes under you) and rebuilds Overview as the way into all the others.
- A card for every lens says what is behind it: requirements traced, verification coverage, states and transitions, use cases, brief sections out of date, completeness, pruned alternatives. Click a card to open that lens.
- Rename the node and rewrite its description where you read them. Each edit is one History entry.
- A suggested next step comes from the node's weakest completeness axis, and an empty node offers three ways to start. Both open ⌘K with a prompt you can edit before it runs.
- Children show their health on their top edge. A child with no requirements is no longer shown as healthy. Property values show their units.
- Budget rollups are computed once per change in linear time, and a damaged parent loop can no longer hang them. The Simulate button, which appeared on nodes that couldn't be simulated, is gone; Simulation lives in Tools.
- Lenses are grouped as Model, Story and Analysis. Model lenses show a count (requirements, children, states, verifications, use cases) so you can tell what is there before you click.
- Every lens has a letter (hover a tab to see it), and 1–9 follow the new bar order. Arrow keys, Home and End move between tabs. The full list is under ?.
- Narrative and Dendritic are offered only for nodes with children. A lens that doesn't apply falls back to Overview, and an unknown
?lens=link opens Overview. - Each node opens its lens fresh, scrolled to the top, instead of carrying over filters and scroll from the last node.
- "Traced" means one thing everywhere: a trace link that hasn't been rejected touches the requirement. Overview, Requirements and Quality all use it. Rejected links no longer count toward coverage anywhere, including the Traceability matrix and the Markdown export, so some numbers may drop.
- Deleting the selected node (or undoing its creation) moves the selection to its nearest surviving parent instead of leaving a blank workspace. Deleting what the inspector shows closes it.
- Every AI button in the lenses, inspectors and tools opens ⌘K with its prompt, even if ⌘K is already open. Before, some closed it and dropped the prompt.
A model you can trust, a tree you can work in, and the Model Explorer
The first of a series of deep passes over the core of Cairn. This one starts with the model itself: what it promises about its own integrity, the outline you navigate it by, and a new deterministic view of the whole model as editable data. The Simulation notebook also now saves with the project, replacing the notebook described in 1.7.0-beta.
- A new tool showing every entity of every kind as tables you can search, sort and edit in place. Search matches the names behind references too, and Under <node> limits a table to one branch of the system.
- Every reference is a link, and each record lists what points at it: children, requirements, interfaces in and out, verifications, trace links, and where a unit or data type is used.
- + New creates a valid entity with the next id in your project's sequence. Edits and deletes are ordinary ChangeSets, with the same guard and cascade preview as everywhere else.
- An Integrity view lists every broken reference, parent loop, duplicate id and mis-kinded trace link. Open the Explorer from Tools, with ⌘⇧E, or from a tree node's menu.
- One type size, with indent guides and clearer type glyphs. The guide under the selected node is highlighted. Collapsed parents show how many nodes they hold, and a red dot marks nodes with Quality errors.
- Filter the tree by name or id (/); matches keep their parents and are highlighted.
- Reorder and re-parent by drag-and-drop, with ⌥+arrows, or from the row menu. Each move is one undoable History entry and can never put a node inside its own subtree. Rename with F2; hover a row for add and menu buttons.
- The tree can be resized, and the tools list folds away so the outline gets the height. The tree is much faster on large models.
- Undo and Restore now put back exactly what the model held. Previously, undoing some AI edits could blank the entity, and undoing a drag on the Architecture or Behavior canvas left the node where it was.
- A ChangeSet that creates something and then refines it, or adds a subsystem and moves nodes under it, is no longer rejected. The guard also catches parent loops, a second root, and trace links that point at nothing.
- When the guard refuses a change, Cairn says why instead of reporting success, ChangeSet review stays open, and nothing is added to History.
- An AI edit that sends only the fields it changed no longer wipes the others. Before this fix, such an edit could detach a node from the tree.
- AI deletes now include the verifications and trace links that depend on what is deleted, as separate operations you can see in review. Previously these were left pointing at nothing.
- Quality reports every broken reference in the model, not only broken interface endpoints. Projects from older versions and imports are normalized when they open.
- The notebook, its threads, and its runs save to the project as you work. Open one from the notebook bar or from Assets and the cells, figures, and conversation come back. Nothing reruns just to show you the last result.
- Each saved notebook records the model snapshot it was computed against.
- Switching projects saves pending notebook edits first. Finish or stop a run before switching; if saving fails, the project stays open. Restoring an archive also keeps its notebook files pending cloud backup.
- If the project changes after a save, the notebook is marked stale. A run stays on the snapshot it started with. The notebook does not swap in new project data while a run is still going.
- Update to latest project, once the notebook is idle, keeps the earlier result and reruns against the current model. The earlier result stays in the bar and in Assets. Opening it restores that snapshot's outputs and does not rerun them.
- The notebook agent uses a key you provide. An Anthropic key already in Settings is reused. OpenAI and xAI keys are entered in the notebook and stay in the browser. Cairn does not run this agent through the managed pipeline.
Simulation is now an analysis partner — a live agentic notebook
Superseded by the 1.8.0-beta notebook. At the time, the one-shot script tool was replaced by a notebook with an analyst that reads the model, writes and runs Python cell by cell, looks at the figures, and narrates the result.
- Open Simulation and the notebook is already seeded from the selected system: a first cell materializes your real parameters — values, units, and where each one came from — as a Python
modelnamespace. The analysis builds on real numbers, never guesses. - For large, detailed projects the agent drills in on demand: ask it to model a subsystem and it reads that subsystem's full parameters, requirements, and interfaces — without drowning in the whole tree.
- Describe an analysis in plain language. The agent stages a markdown framing cell and the code that backs it, runs it, and — crucially — sees the output: figures come back to it as images, tables as text. It reads tracebacks and fixes the cell instead of barreling past errors.
- Real Python in your browser — numpy, pandas, scipy, scikit-learn, matplotlib, and more, plus interactive Plotly. A live variable inspector shows what's in the session.
- The notebook is an artifact, not a transcript: save it as a portable
.ipynbasset that exports with your project, or download it. Result files the agent publishes — a cleaned dataset, a summary table — land in the Asset browser as first-class outputs.
- This notebook could use your Anthropic key. The October 2026 notebook is bring-your-own-key only and does not go through the managed pipeline.
Responsive Pipeline — AI runs that stream live and stop on demand
AI generations now stream in as the model writes them, so you can tell an actively-working run from one still waiting — and a Cancel button stops any run mid-flight, halting the request instead of letting it finish and be thrown away.
- Specialist responses stream in as they generate — the run indicator shows a live “receiving response” state with a growing character count, not a static spinner.
- Works on both paths — your own API key (direct) and the managed service — with usage metered accurately even if you navigate away mid-stream.
- A Cancel button ends a run in flight and actually aborts the underlying request — no wasted tokens, and the app returns cleanly to idle rather than surfacing an error.
Native Change Intelligence — your model's history, legible in place
The History tool is now a two-pane workspace: a filterable timeline of every change drives a five-tab detail view. This is model history — not a chat log, not a code diff. Changes read as engineering edits, and you can roll the model back to any earlier point without losing the record of how you got there.
- A newest-first timeline filters by All / AI / User / Partial / Undone, with a specialist badge, summary, and an applied-versus-proposed count on every entry.
- Selecting an entry opens five tabs — Summary, Diff, Files, Impact, Provenance — each a view of data Cairn already stores. Nothing new is persisted.
- Diffs are semantic — requirement priority must → should, interface gained signal CAN_RX — field-level, not red/green line noise.
- Provenance is thin-but-true: who or what proposed a change, what was applied versus dropped, and why. Manual edits show no fabricated model, tokens, or cost.
- Restore to this point reverses every later change in one pass and records the restore itself — nothing is erased, so the restore is undoable too.
- A preview shows exactly what will be removed, reverted, or restored — and what is never touched (asset blobs, usage ledger, project name, prior cloud snapshots) — before you confirm.
Artifact Review — see exactly what leaves or enters your workspace
Every full-project export and every .cairn.zip import from the sidebar now opens a review modal first. You see what is about to leave — or land in — your workspace, and confirm it explicitly. No silent reads, no silent writes.
- Before a download fires, review a scorecard, the full file tree with inline previews, the asset list with per-file inclusion status, the applied change history, and any warnings.
- Toggle between the structured backup and a flat single-file backup without losing your place.
- Imports preview the incoming project before anything is written — tabs for the file tree, assets, change history, and warnings.
- Choose to create a new project or replace an existing one. Replacing requires typing the target project's name to confirm.
- Multi-project backups show a per-project list up front and a succeeded / skipped / failed summary after import.
- Importing a project that shares an ID with an existing one no longer overwrites it — IDs are detected and rewritten so both projects survive.
- Replacing a project happens in a single transaction: if any part fails, nothing changes.
- Unrecognized or corrupt archives explain what was checked and why the import is blocked, instead of failing silently.
The Depth Release — cross-lens navigation, validator warnings, and lens deepening
The Depth Release came out of a systematic audit of all 12 lenses, and acted on what that audit surfaced: the lenses learned to link to one another, several gained genuinely new capability, the AI pipeline got a validator pass that flags structural mistakes in Change Review, and a documentation sweep brought every described capability back in line with what the product actually does.
- The Brief and Narrative lenses now cross-link from the Inspector — when a node has both, each offers a one-click jump to the other.
- Architecture and Dendritic link bidirectionally: jump from a node's children to its pruned design alternatives and back.
- The Completeness lens gained a fourth axis — Verification — alongside Entities, Processes, and Relationships.
- When Completeness flags a gap, the weakest axis now links straight to the lens that owns it — "this node is thin" becomes one click to the fix.
- Dendritic lens: filter alternatives by decision type (physics, engineering, mission, and so on). The filter bar appears whenever a subtree carries more than one type.
- Narrative lens: export the systemigram diagram as a PNG image.
- Verification lens: toggle coverage between this node alone and the full subtree, so tree-level gaps are visible at a glance.
- Operational lens: a dedicated AI specialist now generates use cases — focused operational scenarios with actors, preconditions, postconditions, and ordered steps — in place of the general-purpose fallback.
- AI-generated ChangeSets now surface warnings in Change Review when they violate a specialist's declared invariants — state-count caps, naming conventions, reachability, create limits, and similar structural rules. The warnings are advisory: they never block an apply, they just make it visible when the AI strays from its own rules.
- Invariant checks run for seven specialists — Architect, Requirements, Behavior, Verification, Brief, Dendritic, and the Operational specialist introduced in this release.
- Three lens empty-state buttons that looked clickable but weren't — in the Architecture, Behavior, and Verification lenses — now actually open the command palette.
- Causality analysis no longer evaporates when you switch away from the lens and back.
- The Verification and Operational lenses refresh immediately after an AI ChangeSet is applied, instead of needing a manual nudge.
- Fixed a routing bug where the word "narrative" sent requests to the Brief specialist instead of the Narrative one.
- Fixed 3D mesh property wiring in the Visuals lens.
- Removed roughly 20 dead fields the type definitions still carried but nothing read.
- Swept the documentation across all 12 lenses so every capability the docs describe matches what the product actually does.
- All 12 lenses are now covered in the in-app help system.
Polish bundle — confirm dialogs, export labels, Inspector edits, pipeline tests
A bundle of accumulated polish and hardening. The remaining twowindow.confirm dialogs are gone, asset delete now warns about referrers instead of asking twice, and the two confusingly-similar Export buttons finally tell you which one keeps your history. State and Transition Inspectors finish their click-to-edit story (type, durations, guard, action, timing). Under the hood, the AI pipeline gets its first 12 tests, the asset URL cache moves to its own module so storage auto-invalidates on overwrite, and the MVP-fallback router stage no longer looks like missing data in the trace.
- The "Leave the interview?" prompt during inception is now a Cairn-styled modal instead of a native browser dialog. Same goes for the asset delete — one click opens a danger-styled confirm that surfaces how many visuals, mesh scripts, or simulation scripts reference the asset, so you know what you're about to leave dangling.
- Removed dead onboarding code (an old WelcomeScreen surface that was no longer wired up).
- Settings → Export is now labeled "Model Export (no history or assets)" with a pointer to the sidebar's Export Full Projectfor lossless backups. The sidebar button is renamed to make the
.cairn.zipdistinction visible (with a hover tooltip spelling out what's included). No payload changes — this is purely about making the choice obvious.
- First tests for the AI orchestrator dispatch (every specialist route + the default-fallback contract), the user-message envelope shape that prompt caching depends on, and the context-serialization tag order. 384 / 384 vitest passes (was 372).
- State and Transition Inspectors are now fully click-to-edit. State
type,typicalDuration, anddurationUnit; Transitionguard,action, and the full timing block (typicalDuration/durationUnit/minDuration/maxDuration/dutyCycle) all open inline on click. Closes the promise the Behavior right-click Edit item had been making.
- Asset URL cache extracted into its own module so the storage primitive auto-invalidates on overwrite. Removes a class of stale-image bugs from future code paths.
- When the AI router falls back to the in-code MVP plan, the pipeline trace now labels that stage
mvp-fallbackinstead of leaving the model field blank with 0 / 0 tokens — small clarity win during dogfood metering checks.
AI provenance, History polish & palette cleanup
A focused follow-up to 1.2.0-beta's cherry-pick. Three findings from the post-release dogfood landed together: AI rationale that used to vanish on Apply now persists on every changeset; History values that looked truncated-to-data-loss now hover for the full text (or click to expand for very long ones); and the command palette no longer hangs in a stale state after Apply. Plus the same tooltip pattern lands in cascade-confirm dialogs and the Dendritic lens.
- Every applied AI changeset now carries the reasoning and the typed issue list (errors / review / auto-fixed / info) you saw in the Change Review modal at apply time. Three weeks later you can still ask the History panel "why did the AI add REQ-029?" and get an answer.
- History detail panel renders AI reasoning(collapsed-by-default italic block) and AI review notes (collapsed-by-default four-bucket list) whenever the data is present. User-authored changesets render neither — there's nothing to surface.
- Cloud sync threads the same data through to a new
ai_reviewcolumn on the changesets table, so audit trail stays consistent across local and cloud surfaces.
- Long create-field and update-diff values in the History detail panel render with the usual
…AND now surface their full content via a hover tooltip. Non-truncated values get no tooltip — visible text isn't echoed. - Click-to-expand for very long content. When the full text exceeds the readable range of a native browser tooltip (~280 characters), the value swaps from hover to a click-to-expand button. Click expands inline with paragraph structure preserved; click again to collapse. Keyboard accessible via the focusable button.
- The same hover-for-full-text affordance lands on long entity names in the cascade-confirm and cherry-pick-confirm dialogs, and on Dendritic lens node descriptions.
- Apply (in any of its three forms — full accept, Apply Valid Only, cherry-pick partial) now closes the command palette and resets pipeline state. Reopening the palette no longer shows stale progress from the just-applied run.
- The What's next? follow-up flow is unchanged — it still closes and reopens the palette with your queued prompt after a brief delay.
- Filter History by AI flags — "show me only changesets where the AI flagged something". The data is now persisted; the filter UI is on the backlog.
- Settings export / sidebar export alignment. Two exports with the same surface name produce different shapes today — clarifying that lands naturally in the
.cairn.zipv2 directory-tree refactor.
Cherry-pick in the Change Review modal
Until now, the Change Review modal was all-or-nothing — accept the full proposal, force-apply with errors, or reject. Beta users kept asking for one in-between path: "most of this is right, but drop these two ops." This release ships that path. Each operation gets its own checkbox; the modal handles the integrity mechanics so deselecting one op doesn't leave another with a dangling reference.
- Each operation in the Change Review modal carries a checkbox. The footer button reads "Apply N of M" when you've selected a subset; reverts to "Apply Changes" when everything is checked or nothing is. Disabled when nothing is selected.
- Errors are unchecked by default. The footer button stays green as long as no errors remain in your selection.
- Per-section bulk-select chip in each section header: Select all / Deselect all, with a mixed state when only some ops in the section are selected. Right-pane toolbar carries a global cycle: All → None → Errors off → All. Keyboard s advances the cycle.
- Park-and-reopen preserves your selection. Click the × on the modal to set a review aside; reopen via the indicator chip and your checkboxes are exactly where you left them.
- When deselecting an operation would invalidate other ops in the same ChangeSet (e.g. you uncheck a new node that another op references), the modal computes the cascade and opens a confirmation dialog listing each auto-dropped op with its specific reason: "Depends on NOD-X, which you unchecked" or "Updates REQ-007, whose create you unchecked."
- Cancel preserves your selection; confirm applies the trimmed ChangeSet with full audit metadata. Edge case: if the cascade would auto-drop everything you selected, the modal toasts and cancels rather than silently writing an empty change to history.
- Starting a new pipeline while a parked review carries unreviewed selections fires a discard prompt: "Your in-progress cherry-pick selections will be lost." Cancel keeps the review intact; confirm proceeds with the new run. The prompt only fires when you've actually modified the selection — accepting the default "all selected" layout doesn't prompt.
- Timeline cards for partial-accept ChangeSets show a secondary line: "Applied 7 of 10 operations (2 auto-dropped)." Hover the auto-dropped suffix to see the per-op rationale that fired during cascade.
- Cloud sync and analytics record the same metadata so the audit trail stays consistent across both paths.
- The Force Apply overlay's Apply Valid Only shortcut now records the same partial-accept metadata that cherry-pick does. History rendering, cloud sync, and analytics all see a uniform shape across the two paths.
- The shortcut still applies all non-error ops in one click — useful when you trust the validator's error verdict and don't want to manually pick. Cherry-pick is the surgical option from the main modal; Apply Valid Only stays as the fast lane.
- Live cascade revalidation — surface auto-drop status inline as you toggle checkboxes, instead of waiting until apply time. Phase B of the cherry-pick rollout; deferred to a future release for dogfood feedback.
- Pipeline cancellation from the palette — still pending the design pass on partial-work disposition + Anthropic billing.
Change Review Severity Refactor
A focused fix to a UX bug dogfooding flagged: warnings-only changesets used to turn the Apply button amber and read "Apply with Warnings", even when every warning was a benign auto-fixed note. Beta users were hesitating on changesets that were entirely safe. This release introduces a typed severity model so the modal can tell real errors apart from FYI notes — and the Apply button stops crying wolf.
- Warnings-only changesets keep the green "Apply Changes" button. The amber "Apply with Warnings" middle state is gone — only real errors flip the button to red Force Apply, which keeps its existing confirmation overlay.
- Force Apply behavior is unchanged. The same overlay lists each error verbatim and offers Apply Despite Errors / Apply Valid Only / Cancel.
- The single "Warnings" section is replaced by up to four severity-keyed buckets: Errors, Warnings, Auto-fixed, Info. Empty buckets render nothing.
- Errors and Warnings open by default — they need attention. Auto-fixed and Info default collapsed — they're trust signals and FYI noise, not attention grabs. Click the section head to toggle, or use the "Expand all" toolbar button.
- Auto-fixed bucket shows what the validator quietly fixed for you (auto-renamed duplicate IDs, stamped missing timestamps, set default sort order). Trust signal — visually unobtrusive.
- Info bucket surfaces low-priority notes that previously got dropped on the floor. Now visible if you want to see them, hidden by default.
- When a compound (multi-specialist) action has a subtask fail mid-run, the failure now surfaces as an Error rather than a Warning. The Apply button reads Force Apply — you have to consciously acknowledge the partial run rather than slide past it on the amber path. Follow-up retry from a clearer prompt is usually the right move.
- Pick & choose individual operation acceptance in the Change Review modal — the typed severity model laid in this release is what cherry-pick uses to derive its "default = errors unchecked" rule cleanly.
- Pipeline cancellation from the palette — still pending the design pass on partial-work disposition + Anthropic billing.
Lens & UI Polish
A focused polish pass on three surfaces dogfooding flagged as friction points after the Editability Release: the Behavior lens, the Change Review modal, and the command palette. Nothing here changes the underlying data model or the AI pipeline contract — only how you interact with them.
- State nodes, the Behavior canvas, and transition table rows now expose the same right-click menus the Architecture lens has had since v1.1.0-beta. State menu: Edit · Add transition from/to here · Delete (with cascade preview) · Delete with AI review. Final states omit "Add transition from here". Canvas: Add state. Transition row: Edit · Delete · Delete with AI review.
- Keyboard Delete / Backspace on a selected state triggers the same cascade-aware delete. Parity with Architecture canvas.
- AI-driven menu items disable while a behavior simulation is active — same disable pattern Architecture uses while a pipeline is in flight. Manual items (Edit, Delete) stay enabled.
- Two-pane layout: warnings + AI reasoning on the left, operations + What's Next on the right. Both panes scroll independently. Below an 880px viewport the panes collapse to a single column.
- Per-section "Expand all descriptions" chip in section headers, plus per-row click-to-expand on long descriptions (default two-line clamp). Items with attached warnings render expanded automatically.
- Item-ID references inside warnings are now clickable. Click jumps the right pane to that operation, opens its section if collapsed, expands its description, and briefly flashes the row.
- Bottom scroll-progress fill tracks position in the right pane. Keyboard e cycles all right-pane sections expanded ↔ collapsed.
- Auto-growing textarea — long prefilled prompts from "What's Next?" pills are fully readable now (capped at ~6 lines, scrolls beyond). The single-line input was hiding everything past the first ~80 characters.
- "Prefilled from review modal" caption beneath the input when the prompt was injected from a follow-up pill, so users know where the text came from and can edit before submitting.
- Pipeline progress shows a specialist-specific duration hint ("typically 30–60s for decomposition work") so the user has a sense of "still working, not stuck."
- Clarification questions render as horizontal chip buttons instead of native radio/checkbox controls, with a banner explaining what's happening, the original prompt visible above the questions, and a "{n} of {m} answered" progress footer.
- Preserved: closing the palette while a pipeline is running keeps the run alive. Re-open via the status indicator chip and you're back where you left off.
- State and Transition Inspector field editing — today the state name and transition trigger are the only click-to-edit fields; description, type, timing annotations, guard, and action become editable in the next release.
- Pipeline cancellation — there's no path today to abort an in-flight AI call from the palette. The disposition of partial work and Anthropic billing implications need a small design pass before this lands.
- Manual creation forms for transitions and states — the Behavior lens right-click items currently route through ⌘K rather than opening inline forms (Architecture got the inline form for Connect; Behavior's equivalents are next).
- "Pick & choose" individual operation acceptance in the Change Review modal — the two-pane layout accommodates it; the selection mechanic ships separately.
Editability Release
1.0.0-beta required ⌘K and the AI to change the model. 1.1.0-beta makes every artifact in the model directly editable, and gives every action a deterministic right-click or Inspector path alongside the existing ⌘K command palette.
- Click-to-edit on every Inspector field — node names, descriptions, requirement priorities, interface protocols, state timing, signal metadata, and so on. Edits commit immediately.
- Verification records gained a short Title field, populated by AI on creation and editable in the Inspector.
- Tree right-click menu: Add child, Rename, Decompose with AI, Delete (with cascade preview), Delete with AI review.
- Architecture canvas right-click menu: Add child, Decompose with AI, Connect to… (manual), Connect with AI, Delete (with cascade preview), Delete with AI review.
- Every delete that has downstream dependents — node, requirement, interface, state, transition, and verification — now shows what else will go before you confirm. Trace links, child nodes, dependent verifications, and orphaned signals all surface in the preview.
- Signals use a lighter two-click arm-then-confirm pattern instead, since they have no cross-entity dependents.
- Manual interface creation via canvas right-click → Connect to… (manual), or via the parent node's Inspector → + New Interface.
- Signals are inline-editable inside an interface: a collapsed row per signal, click the chevron to expand the full editing grid (data type, direction, rate, simulation metadata).
- Two-click Remove on signals — first click arms, second click within three seconds removes; protects against stray clicks.
- Pipeline status chip in the toolbar shows running / error / clarification-needed / ready states.
- When the Router's confidence is low, it now emits structured clarification questions (text, single-choice, or multi-choice) directly in the palette. Answering re-runs the same prompt with your answers appended.
- Refactor the .cairn.zip export so it round-trips cleanly through GitHub (file structure, JSON formatting, asset directory layout).
- Manual creation forms for states and transitions — Requirements and Verification lenses already have inline + Add buttons; the Behavior lens does not yet.
- Right-click parity on the Behavior lens — states and transitions don't yet expose a context menu on the canvas.
- Curated unit dropdowns on signal metadata (currently free-text), cross-referenced against the project's unit catalog.
Initial Release
Cairn's first public release. The items below represent the primary feature set at launch — see the Roadmap section for what's coming next.
- 12 analytical lenses: Overview, Brief, Visuals, Requirements, Architecture, Causality, Completeness, Narrative, Dendritic, Behavior, Verification, Operational
- Hierarchical system modeling with five node types: system, subsystem, assembly, part, external
- Requirements management with six types: functional, performance, interface, safety, environmental, constraint
- Interface and signal definitions with protocol and rate metadata
- State machine editor with states, transitions, guards, actions, and timing annotations
- Engineering properties catalog with budget rollups (mass, power, cost) and TRL tracking
- AI property suggestions with per-item accept/dismiss workflow
- Verification records with test, analysis, demonstration, and inspection methods
- Traceability tool with satisfies, verifies, derives, depends_on, and implements link types
- AI decomposition and generation via ⌘K command palette
- 17 AI specialists routed through a 5-stage pipeline (Router → Context → Specialist → Validator → Review)
- ChangeSet governance: AI proposes, user reviews operation-by-operation before applying
- 2D concept renders via Gemini image generation with 6 style kits
- 3D mesh generation via AI-authored Three.js code with MeshBuilder
- Monte Carlo simulation via in-browser Python with AI-generated scripts (launch behavior; the 1.8.0-beta notebook entry is current)
- Export: JSON, Markdown, CSV, full ZIP with assets
- AI-powered export: PPTX presentations and DOCX reports via Claude Skills API
- Local-first architecture: all model data stored in browser via IndexedDB/Dexie
- Quality linter with 16+ automated rules across nodes, requirements, interfaces, and traces
- Supabase authentication with magic link sign-in
- Four Questions framework for systematic model interrogation (Causality, Completeness, Narrative, Dendritic)
- Harney's Pyramid of Causality for technology dependency and maturity analysis
- Dendritic decision tracking with pruned alternatives as first-class model entities
- Systemigram narrative visualization with mainstay identification (Boardman/McDermott)
- Three-axis completeness scoring based on Pace/Loper simulation fidelity dimensions
- Cloud sync for multi-device workflows
- Real-time collaboration
- GitHub export with ChangeSet-to-commit mapping
- AI-assisted trade study tool with Pugh/AHP methodology