@lotics/cli 0.245.0 → 0.246.0

package/docs/cli_reference.md

@@ -70,7 +70,7 @@ Per-command syntax, flags, contracts, and gotchas for the public `lotics` CLI. S
 | `lotics docs model` \| `lotics docs model/<section>[/…]` | **The model reference, from inside the binary** — the one doc this CLI carries rather than resolves, because it describes this CLI's own plan checker; listed first by `lotics docs`, at this CLI's version. Its first page is what a model composes with, the working order (jobs → entities and fields → roles → one app per job → `scaffold check` → `app preview` → `workspace build`) and the section addresses; every page of it is whole. Every top-level key of a `model.json`, every field `type` the contract admits with the config each one needs, the option / view / role / inline-template shapes, the row format (relative dates `@today` / `@month-start` with whole-day offsets; links as `"<entity-alias>:<ref>"`), the rules, the `apply` list (packages copied in after the model's own tables, each with an optional `bind` onto them), the `preset` block (a published model's branches and its at-most-two questions), the **`from` form** — `{from, variants, rename, entities, rows, field_roles, apps, apply}`, which names a preset by SLUG instead of restating it — and one complete worked example. **Offline, no account.** |
 | `lotics report '<json>'` \| `lotics report @report.json` | File a report with the Lotics team about what got in your way. **Covers the classes telemetry structurally cannot see**: a capability that does not exist (no command ran, so nothing was recorded), a command that exited 0 having done the wrong thing, an error whose message did not name the remedy, and anything that made authoring slower than it should be. **A frame, not a paragraph** — `{goal, actual, expected?, tried?, wanted?}`, `goal` and `actual` required, unknown keys dropped rather than refused. **No severity or category.** Ingest is inline JSON, `@file`, or `-` for stdin. A bare sentence is refused with the frame printed beside it, so the fix is one step; a bare invocation prints the frame BEFORE asking for a credential, since someone whose key will not resolve is exactly who has something to report. **Not spooled**: unlike telemetry it posts inline, prints whether it landed, and exits non-zero if it did not, echoing the report back so a failed send never loses it. Runs regardless of `LOTICS_TELEMETRY` — invoking it IS the consent that passive collection needs an opt-in for — but with telemetry off there are no recorded commands to attach, and it says so rather than implying context it does not have. Requires auth. Never paste records, file contents, or credentials. **Prints the id of each frame filed** — a filing nobody can cite cannot be answered about. The ids come from the server, so an instance that only logs the frames prints the count alone; the CLI never mints one of its own, which would hand back a token that resolves to nothing. |
 | `lotics app check` | Every pre-flight `deploy` runs, WITHOUT building or shipping. **First, whether this project is even based on the served version** — the one thing a deploy REFUSES outright rather than pushing (the server 409s a stale `prev_version_id`), and the one finding that invalidates every other: a stale tree and the live app are two different apps, so comparing them reports nothing trustworthy. Stale exits 1 naming both versions and stops before the rest; a project with no stamp at all — or an app with no version yet — is a first deploy, not a conflict. `deploy` runs the SAME assertion, so a stale tree fails before it pushes or builds. Then: the manifest's agent schemas against the live app row, every binding a deploy would push, aliases the source calls that nothing bound (queries, workflows AND agents), bindings this bundle stopped calling (the same transition — and the same baseline — `deploy` reports, so the two cannot disagree), capability-gated SDK calls the manifest doesn't declare, a missing icon/theme, a missing app `description` (it heads the capability catalog the chat agent reads every turn, and its absence has no other symptom), a `vite.config.ts` that never defines `global`/`__DEV__` **in a project that still ships react-native** (read off its own `package.json` — an app on the React DOM kit bundles none of it and the check would be advice to define two globals nothing reads), and a `window.open` in the app's own source (each fails ONLY in the deployed app: dev bundles with esbuild and production with rollup, so typecheck, lint, build and `app dev` are all green), an agent whose capability and its reach disagree, in EITHER direction, over any of the five declaration-bound tools (`run_app_query`/`run_app_workflow` against `query_aliases`/`workflow_aliases`; `grep_knowledge`/`read_knowledge`/`list_knowledge` against `knowledge_doc_ids`) — the tool is the capability, the list is the reach, and a tool with no reach means every call it makes is refused while the run still COMPLETES, so it surfaces as a model ignoring its prompt; read off the live row, never the manifest, which mirrors those fields but is pushed by no verb, and a notice for any alias the source computes at runtime (invisible to every check here and to `--prune`'s unbind guard). **And whether the runtime this app builds against has fallen behind what is published** — `@lotics/app-runtime` alone, since the kit an app lists sits at the runtime's range, read from `node_modules` rather than the range because a caret is minor-locked below 1.0 (`^0.13.x` can never resolve `0.14`, and `npm update` does nothing). A release LINE behind is loud and names `lotics app kit --published` and `@lotics/ui`'s `MIGRATION.md`; anything smaller is one quiet `npm update` line, because a warning that fires on every deploy is one the reader stops seeing. An app still listing `@lotics/app-sdk` is told it moved into the runtime and which verb migrates it. The lookup is bounded and every failure is silence: a version check must never become a new way for a deploy to fail. Every deploy finding is the same helper `deploy` calls, so a green check means a deploy will not complain. **And the PORTABILITY gate, the second of two rules `check` runs that a deploy does not** (the first is the undeclared call site below) — the ids an app cannot carry into another workspace, over the working tree, with the same exclusions the deploy tar applies. Two rules. **An id this workspace MINTED**, written into `src/`, a `.md` or the app's own docs — it resolves to nothing in a copy, and in prose it is an instruction the copier's agent follows; this is the one a `library publish` also refuses, on the uploaded archive. **And an id-shaped STAND-IN** too short for the generator that mints its prefix (`"opt_X"`, `"fld_a"` — quoted or in a code span, so a bare `opt_in` stays legal, and never in a test file), which only `check` runs, and which additionally reads a workflow body and the manifest: those two are exempt from the first rule because a publish INVERTS a real id there and cannot invert a fake. **And an `app_` id in `app.json` anywhere but the `app_id` a hand-off's binder writes**: a sibling app is named by its plan alias and the app it became is read off the workspace's binding at every regeneration, so a pasted one is re-pointed by nothing. Each is reported as `<file>:<line> — <id>` with the one edit that fixes it. This gate reads only the files, but the command around it still needs a resolvable credential and the live app row, so it is not an offline check. **And every bound workflow body, checked as `app workflow check` checks it** — the same isolated per-alias program against the pulled `.lotics/workflows/<alias>.globals.d.ts`, then the server's `verify_only` verdict on each body that passed, which exits 1 on any issue and warns when the server is not reached. The server verifies a body once, at the save that wrote it, so a body whose declared types have since moved stays stored, matching what is live, and is refused by the next writer — a starter copy, in somebody else's workspace. The verdict is as fresh as those types, which `pull`, `workflow pull` and `codegen` refresh. **And the app's own `npm run typecheck`**, after regenerating the `.lotics/*.d.ts` companions from the manifest — the same run a deploy makes before building, so a filter or sort key the query does not project fails here rather than at the first member's request. **Exits 1 on that, on a body the types refuse, on a failing typecheck, and on what a `deploy` would REFUSE or PUSH** — an agent schema that disagrees with the live app, any binding the project has ahead of the app (an edited workflow body or declaration, edited agent prose, a changed query), and a query or workflow `description` over the 300-character capability cap, which is a binding no `set` and no deploy will take (the rule is `@lotics/shared`'s, the same one the server refuses with, and `workflow set` / `query set` ask it before sending anything — so a long line costs one edit rather than a failed push per alias). **And every manifest query declared with NO description**, named in one line: that sentence is what a chat or MCP caller chooses between aliases by, and an alias is a JS identifier. `app create --from` deliberately writes none — a template over a shape's own English reads like a line about the business while saying nothing — so a generated app is told once, here, which lines are the author's to write. Both are things a deploy would act on, so CI gating on a green check means a deploy has nothing left to do; genuine advisories (capabilities, branding, a runtime-computed alias, orphaned bindings) stay advisory and never fail it. **`--screens` adds the rendered surface, and is its own entry below.** It runs after the typecheck and only when it passed — a type error renders nothing to measure — and its findings, and any write it refused, fold into this command's exit code. **It also refuses a call site the manifest no longer declares.** `useQuery`/`useWorkflow` keep a bare-string overload for a computed alias, so deleting or renaming an alias leaves every call site compiling and failing only when the screen renders — the one edit most likely to orphan a call site is the one the generated types cannot catch. The alias literals in `src/` are matched against `package.json#lotics.queries`/`.workflows` (the manifest, not the live row: the server still SERVES an alias whose declaration was just deleted, because a deploy never unbinds), and an undeclared one exits 1. `app codegen` prints the same finding as a warning, since it is the command an author runs right after editing the manifest. **A failing body that is byte-identical to the one the server is running is labelled as such**: its `lotics.synced.workflows.<alias>.content` baseline proves the file has not been edited since it was pushed or pulled, so the failure is a grammar migration the stored body is owed rather than a stale checkout — a link field reads as an id array, so drop `.id` or descend with `linked(…)`. Until the body is edited and `set`, the stored one keeps running as it always has. **And a query that reads past a table's ROW RULE.** A table's `private_filters` bind the CALLER, and an app query's caller is the app's OWNER — the viewer needs no table access, `app:use` is the grant — so the rule passes and every row is served. For each table a declared query reads (`GET /v1/tables/{id}`, the one surface that serves the rule), the viewer predicate — `current_member in_any_group`, or a member field's `is_current_member` / `is_not_current_member` — is attributed to the SCAN it guards: a `from_table`'s own `filter`, or an enclosing `filter` node's predicate, whose rows are the ones that scan produced. So a clause written over one table never silences the finding for the ruled table joined beside it, and every scan no clause covers is named with its table. A table this credential may not read (403, or 404 for one that is gone) is dropped — a rule that cannot be read is not evidence of one — while any other failure of that read fails the command, since "I could not ask" must never render as "there is no rule". Advisory, never part of the exit code: an app that deliberately serves the whole table to a desk of people who may all read it is legitimate, and nothing here can tell the two apart. **And it names the workflows a chat or MCP caller is offered with no description** (one `get_app_capabilities` read, the reader's own view of the app): that text is what those callers choose between aliases by, and without it they choose by the alias. It cannot be counted from the manifest — a workflow bound out of band is not declared there at all. Advisory, never part of the exit code. **It regenerates `.lotics/app_fields.ts` from the live schema before typechecking**, as a deploy does — a gitignored map from a moved schema otherwise passes. **And three claims the app makes**: a bound body's writes against `package.json#lotics.writes`, read as the step tree the server would store, exit 1 on a field no entry covers, and declaring none only warns; a description carrying `<placeholder>` syntax exits 1, since the catalogue escapes angle brackets — state the format as an example; so does one naming a desk `query_apps` does not list. |
-| `lotics app check --screens [--screen <label>] [--width <n>] [--shots <dir>] [--changed]` | **`--screens` adds the rendered surface**: the app is served the way `app dev` serves it (its real data, this key), rendered headless in Chrome (`CHROME_PATH`/`LOTICS_CHROME`, then Playwright's, then system) at 1280 and 375. **The screens are its navigation's destinations** — a `nav` landmark's `a[href]` or `role="link"` (an app's route lives in its router, so the kit's shell renders each screen as a button carrying the link role and no `href`), else the first tab strip, else the root — in that order, because a screen's own lifecycle desk draws a tablist too, and reaching for one first walks a screen's STAGES as if they were the app's. **Each screen is then WALKED THROUGH ITS OWN DOORS**, nothing configured per app: every door is named by something the document states. A record: a row stamped `data-opens="page"`, a real `a[href]`, or an EMPTY DOOR — a named control with no text and no child element, the only legal whole-row press target, which reaches a drawer register, a Schedule row and a hand-written row alike. On a record: the acts menu (`aria-haspopup`), the first child row, the facts behind the fold (`aria-expanded="false"`). On any surface: every dialog a visible primary or secondary act raises — nothing announces one, so a press is kept only where an overlay appeared. **THE RUN IS A READ, AND THE NETWORK IS WHERE THAT IS ENFORCED — never a rule about what the page may draw.** The app frame holds no credential and Chrome runs on a throwaway profile, so every call the app makes reaches the workspace through this CLI and nowhere else, and this CLI serves an ALLOWLIST of reads. Everything outside it is refused before the call leaves the machine — a workflow run, a record write, an upload, an agent run, a comment, and any op this CLI does not know, which is refused because it is not on the list rather than because anyone listed it. A refused call HEADS the report, naming the surface, the control the app had focused and the RPC by its alias (never a payload value), and fails the run on its own: a walk that provoked a write left the app in a state no reader could have put it in, so nothing measured under it means anything. What the walk does is bounded on the page as well: nothing inside an open overlay, nothing typed, no submit and nothing in a form, no act the kit marks costly (`data-tone` danger/warning) and none whose own name is the write, in either language — and nothing is FOCUSED, because focus cannot be taken from one field without leaving another, and a field that saves itself when the reader leaves it saves on exactly that. Where a reading needs the focused state, as the focus-ring rule does, Chrome is asked to PAINT `:focus-visible` and release it again: the cascade answers, focus never moves and no event is dispatched. Each overlay closes with Escape, confirmed closed; a record is descended at most twice; the walk stops at twelve surfaces per screen and NAMES each door it left. Each surface is measured once no request is in flight, and the measurable probes of `@lotics/ui` docs/reviewing.md run over the DOM, each finding printing the rule it IS — its law, its section and the one edit that answers it — so the numbers need no key. What they exempt is what the screen itself declares: a register's ordinal gutter (a column counting to the row count is the shape's numbering, which no app can treat), a strip whose list carries `data-order="sequence"` (a lifecycle rail, which composition.md permits under a screen's tabs), a hairline or `clip-path`-clipped leaf (the visually-hidden node a control plants for a screen reader), and a leaf whose own computed line clamp states a count over a sentence. **A meter counts as an encoding only where it draws a POSITION** — `aria-valuenow` inside a range with room left. A bar pinned at its own maximum — what a meter alarmed AT its maximum draws on every alarmed row — reads the same as every value above it, and one with no maximum states none; both are counted in the census's `devices` and out of its `encoded`, so "nothing drawn" and "drawn and saying nothing" never read alike. The bare values that remain are grouped into columns, each named by the heading over it, so a finding says WHICH slot draws its figures as words. **Two finding classes read what geometry cannot.** `clutter` (docs/hierarchy.md): a second primary act, a value in two places on a record, an unstated fact open beside its fold, a box reserving more lines than it holds, markers on over half a form's fields, a second accent. `right_form` (docs/templates.md's device index): a boolean as a two-option select, a day run as a repeated date column where the kit ships `Schedule`, a fold hiding a record's section, two reading columns from 1016px. Each prints a line per rule fired, with three offenders. **Four read room and counts** (reviewing.md §8l–§8o): `reserved_blank`, a register column 40px or wider that is blank (no text, no control) on more than half the rows in view, or a row-act gutter leaving 40px beside the acts on them — a row a complete register draws for nothing stored (`data-ghost`, a day nobody filed) is no row of the set; `shed_with_room`, a column the register's width gave up (the kit states them as `data-shed`) that, with the gap between two columns, fits in what its slots hold beyond what each must seat — its heading and widest ink, a fixed column's whole width, the acts most rows have; `action_off_heading`, a section's create (a control stating `data-create`) drawn in its body rather than its heading row; and `partial_count`, a count or `n of N` equal to the rows in hand of a read the server cut on the screen being walked, where the server's own count of the same alias, params and filter — asked by this CLI, a read — says otherwise. **Four read a record's anatomy** (docs/hierarchy.md), over each record body on the surface with the page or drawer heading it: `record_primary`, a primary act in a section's heading row — the record's one primary is its head's, and a section's add is secondary (two under one heading are `clutter`'s); `brief_fold`, a brief whose foot is past the first viewport of the width laid out — 800px at 1280, 667px at 375 — unscrolled; `brief_repeats`, a value of the brief's band that reads the same as a run of the head's own words (its acts aside); and `empty_tab`, the open tab with no section in its panel. A census line per screen (text runs, money strings, bare values against the devices reading, tab strips) prints first, so a clean verdict over a screen that rendered nothing cannot pass; a screen that renders no text is itself a finding, and one still changing after fifteen seconds is measured as it is. **A cold dependency optimisation is waited out, not measured**: the first paint has its own bound, far longer than settle's, since an app's modules load after its document completes and Vite holds them; only a MOUNTED, idle, textless frame is blank at once, and the finding names the wait and its blocker. It also RELOADS the page under the probe, so a width is measured again once; a second reload is a page that keeps moving and fails. **`--screen <label>` and `--width <n>` narrow a run** (repeatable, comma-separated; the label matches case-insensitively as a substring), for the author iterating on one screen who would otherwise pay a typecheck, a Vite boot and every screen at both widths on every edit; the clean verdict then names only the widths covered, and a `--screen` matching nothing is refused. **`--shots <dir>` writes what the run measured** — a `<surface>@<width>.png` and a `<surface>@<width>.json` per surface, off the SAME settled frame the probes read, so a shot and a finding can never describe different pixels. The PNG is the WHOLE surface: an app scrolls inside a box of its own, so the window is grown to the height the probe measured and put back, and an overlay a resize dismissed is shot as the viewport, the sidecar saying so (`app dev`'s header band is in it — the band the app was laid out under). The JSON is the half a picture cannot carry: the nav's first item's left edge, the title and first section heading, the primary acts by name, label/value pairs against what the folds state, values drawn twice, money that wraps, text the layout cut, and the console errors and uncaught exceptions the frame raised. Named `<nn>-<slug>` plus one `__<step>` per door taken (`__record`, `__menu`, `__child`, `__fold`, `__dialog-<n>`); `<nn>` is the walk's index, which lists the directory in reading order and keeps two labels that fold to one ASCII slug apart. The row a record surface opened from is the sidecar's `opened_from` and that screen's census line, never a file name — it is a person's data. The directory is created if missing, and refused before the dev server boots when it cannot be; `--shots` without `--screens` is refused. Findings exit 1 like the rest. **Where it renders**: the app's files are copied into the render project of its DEPENDENCY SET — `~/.lotics/render/<key>/`, keyed by the manifest's `dependencies` and `devDependencies` as written, a `file:` tarball by its bytes — the one `app preview` renders in, so the install and Vite's dependency optimisation are paid once per set rather than per app; one render holds a project at a time (`<key>.lock` — a second run waits, naming the app and pid, and takes over a dead hold). Where that install resolved another runtime or kit than the app's own, the run says which, since a deploy builds the app's own. The run ends with what each phase cost: `preflight · install|reuse · walk`. **`--changed`** walks only when something the screens READ has moved since this checkout's last walk — each part of `app.json` (every record apart), each declared query, each bundled source file, the field map, the ranges, the kit the render resolved and the CLI version, whose probes measure them — read against `.lotics/screens_check.json`, which every walk writes; with nothing moved, that walk's report is printed again under its timestamp and still fails the run. An app's screens all read its one spec, so a move walks the app whole. It walks, saying why, with no earlier walk, a different `--screen`/`--width`, or `LOTICS_UI_SRC` set (a linked working copy has no fingerprint); the workspace's rows are not an input. `--changed` without `--screens`, or beside `--shots`, is refused. |
+| `lotics app check --screens [--screen <label>] [--width <n>] [--shots <dir>] [--changed]` | **`--screens` adds the rendered surface**: the app is served the way `app dev` serves it (its real data, this key), rendered headless in Chrome (`CHROME_PATH`/`LOTICS_CHROME`, then Playwright's, then system) at 1280 and 375. **The screens are its navigation's destinations** — a `nav` landmark's `a[href]` or `role="link"` (an app's route lives in its router, so the kit's shell renders each screen as a button carrying the link role and no `href`), else the first tab strip, else the root — in that order, because a screen's own lifecycle desk draws a tablist too, and reaching for one first walks a screen's STAGES as if they were the app's. **Each screen is then WALKED THROUGH ITS OWN DOORS**, nothing configured per app: every door is named by something the document states. A record: a row stamped `data-opens="page"`, a real `a[href]`, or an EMPTY DOOR — a named control with no text and no child element, the only legal whole-row press target, which reaches a drawer register, a Schedule row and a hand-written row alike. On a record: the acts menu (`aria-haspopup`), the first child row, the facts behind the fold (`aria-expanded="false"`). On any surface: every dialog a visible primary or secondary act raises — nothing announces one, so a press is kept only where an overlay appeared. **THE RUN IS A READ, AND THE NETWORK IS WHERE THAT IS ENFORCED — never a rule about what the page may draw.** The app frame holds no credential and Chrome runs on a throwaway profile, so every call the app makes reaches the workspace through this CLI and nowhere else, and this CLI serves an ALLOWLIST of reads. Everything outside it is refused before the call leaves the machine — a workflow run, a record write, an upload, an agent run, a comment, and any op this CLI does not know, which is refused because it is not on the list rather than because anyone listed it. A refused call HEADS the report, naming the surface, the control the app had focused and the RPC by its alias (never a payload value), and fails the run on its own: a walk that provoked a write left the app in a state no reader could have put it in, so nothing measured under it means anything. What the walk does is bounded on the page as well: nothing inside an open overlay, nothing typed, no submit and nothing in a form, no act the kit marks costly (`data-tone` danger/warning) and none whose own name is the write, in either language — and nothing is FOCUSED, because focus cannot be taken from one field without leaving another, and a field that saves itself when the reader leaves it saves on exactly that. Where a reading needs the focused state, as the focus-ring rule does, Chrome is asked to PAINT `:focus-visible` and release it again: the cascade answers, focus never moves and no event is dispatched. Each overlay closes with Escape, confirmed closed; a record is descended at most twice; the walk stops at twelve surfaces per screen and NAMES each door it left. Each surface is measured once no request is in flight, and the measurable probes of `@lotics/ui` docs/reviewing.md run over the DOM, each finding printing the rule it IS — its law, its section and the one edit that answers it — so the numbers need no key. What they exempt is what the screen itself declares: a register's ordinal gutter (a column counting to the row count is the shape's numbering, which no app can treat), a strip whose list carries `data-order="sequence"` (a lifecycle rail, which composition.md permits under a screen's tabs), a hairline or `clip-path`-clipped leaf (the visually-hidden node a control plants for a screen reader), and a leaf whose own computed line clamp states a count over a sentence. **A meter counts as an encoding only where it draws a POSITION** — `aria-valuenow` inside a range with room left. A bar pinned at its own maximum — what a meter alarmed AT its maximum draws on every alarmed row — reads the same as every value above it, and one with no maximum states none; both are counted in the census's `devices` and out of its `encoded`, so "nothing drawn" and "drawn and saying nothing" never read alike. The bare values that remain are grouped into columns, each named by the heading over it, so a finding says WHICH slot draws its figures as words. **Two finding classes read what geometry cannot.** `clutter` (docs/hierarchy.md): a second primary act, a value in two places on a record, an unstated fact open beside its fold, a box reserving more lines than it holds, markers on over half a form's fields, a second accent. `right_form` (docs/templates.md's device index): a boolean as a two-option select, a day run as a repeated date column where the kit ships `Schedule`, a fold hiding a record's section, two reading columns from 1016px. Each prints a line per rule fired, with three offenders. **Four read room and counts** (reviewing.md §8l–§8o): `reserved_blank`, a register column 40px or wider that is blank (no text, no control) on more than half the rows in view, or a row-act gutter leaving 40px beside the acts on them — a row a complete register draws for nothing stored (`data-ghost`, a day nobody filed) is no row of the set; `shed_with_room`, a column the register's width gave up (the kit states them as `data-shed`) that, with the gap between two columns, fits in what its slots hold beyond what each must seat — its heading and widest ink, a fixed column's whole width, the acts most rows have; `action_off_heading`, a section's create (a control stating `data-create`) drawn in its body rather than its heading row; and `partial_count`, a count or `n of N` equal to the rows in hand of a read the server cut on the screen being walked, where the server's own count of the same alias, params and filter — asked by this CLI, a read — says otherwise. **Five read a record's anatomy** (docs/hierarchy.md), over each record body on the surface with the page or drawer heading it: `record_primary`, a primary act in a section's heading row — the record's one primary is its head's, and a section's add is secondary (two under one heading are `clutter`'s); `brief_fold`, a brief whose foot is past the first viewport of the width laid out — 800px at 1280, 667px at 375 — unscrolled; `brief_repeats`, a value of the brief's band that reads the same as a run of the head's own words (its acts aside), or a part the brief stands whose heading the tab open at rest draws too; `act_repeats`, the head's primary act heading its queue's first waiting card again in that tab; and `empty_tab`, the open tab with no section in its panel. A census line per screen (text runs, money strings, bare values against the devices reading, tab strips) prints first, so a clean verdict over a screen that rendered nothing cannot pass; a screen that renders no text is itself a finding, and one still changing after fifteen seconds is measured as it is. **A cold dependency optimisation is waited out, not measured**: the first paint has its own bound, far longer than settle's, since an app's modules load after its document completes and Vite holds them; only a MOUNTED, idle, textless frame is blank at once, and the finding names the wait and its blocker. It also RELOADS the page under the probe, so a width is measured again once; a second reload is a page that keeps moving and fails. **`--screen <label>` and `--width <n>` narrow a run** (repeatable, comma-separated; the label matches case-insensitively as a substring), for the author iterating on one screen who would otherwise pay a typecheck, a Vite boot and every screen at both widths on every edit; the clean verdict then names only the widths covered, and a `--screen` matching nothing is refused. **`--shots <dir>` writes what the run measured** — a `<surface>@<width>.png` and a `<surface>@<width>.json` per surface, off the SAME settled frame the probes read, so a shot and a finding can never describe different pixels. The PNG is the WHOLE surface: an app scrolls inside a box of its own, so the window is grown to the height the probe measured and put back, and an overlay a resize dismissed is shot as the viewport, the sidecar saying so (`app dev`'s header band is in it — the band the app was laid out under). The JSON is the half a picture cannot carry: the nav's first item's left edge, the title and first section heading, the primary acts by name, label/value pairs against what the folds state, values drawn twice, money that wraps, text the layout cut, and the console errors and uncaught exceptions the frame raised. Named `<nn>-<slug>` plus one `__<step>` per door taken (`__record`, `__menu`, `__child`, `__fold`, `__dialog-<n>`); `<nn>` is the walk's index, which lists the directory in reading order and keeps two labels that fold to one ASCII slug apart. The row a record surface opened from is the sidecar's `opened_from` and that screen's census line, never a file name — it is a person's data. The directory is created if missing, and refused before the dev server boots when it cannot be; `--shots` without `--screens` is refused. Findings exit 1 like the rest. **Where it renders**: the app's files are copied into the render project of its DEPENDENCY SET — `~/.lotics/render/<key>/`, keyed by the manifest's `dependencies` and `devDependencies` as written, a `file:` tarball by its bytes — the one `app preview` renders in, so the install and Vite's dependency optimisation are paid once per set rather than per app; one render holds a project at a time (`<key>.lock` — a second run waits, naming the app and pid, and takes over a dead hold). Where that install resolved another runtime or kit than the app's own, the run says which, since a deploy builds the app's own. The run ends with what each phase cost: `preflight · install|reuse · walk`. **`--changed`** walks only when something the screens READ has moved since this checkout's last walk — each part of `app.json` (every record apart), each declared query, each bundled source file, the field map, the ranges, the kit the render resolved and the CLI version, whose probes measure them — read against `.lotics/screens_check.json`, which every walk writes; with nothing moved, that walk's report is printed again under its timestamp and still fails the run. An app's screens all read its one spec, so a move walks the app whole. It walks, saying why, with no earlier walk, a different `--screen`/`--width`, or `LOTICS_UI_SRC` set (a linked working copy has no fingerprint); the workspace's rows are not an input. `--changed` without `--screens`, or beside `--shots`, is refused. |
 | `lotics app preview <model.json>#<app> [--shots <dir>] [--width <n>] [--screen <label>] [--kit <path>]` | **The app a plan describes, RENDERED — before a table exists, and with no credential in the process.** The model is checked offline, the named app's register is bound against the workspace the model WOULD become (synthetic `tbl_`/`fld_`/`opt_`/`grp_`/`dtl_` ids, minted positionally off the file), and every read the app makes is answered from the model's own `rows` — projected under the columns the query names, narrowed by the params it declares and sorted the way it states, so a child list under a record holds that record's rows and not every parent's. Then the SAME headless walk `app check --screens` takes: every screen, the record each row opens, its acts menu, its children, the facts behind its fold, the party a row names, and each dialog a tone-neutral act raises, at 1280 and 375, measured by the same probes. `--shots` writes a PNG and a sidecar per surface. **Exits 1 on any finding**, exactly as the check does, and prints what each phase cost — bind, install-or-reuse, walk. **A model with no `rows` is refused by name**: a register over nothing draws its empty state at both widths and measures clean, which reads as an app that is finished. **Nothing is created and nothing is reached** — no app row, no table, no workspace; the only network it needs is the npm registry, and only when the kit has moved. It renders in the cached render project of its dependency set (`~/.lotics/render/<key>/`, shared with `app check --screens` and with every model whose app lists the same set, and held by one render at a time — a second run waits, naming the app and pid holding it, and takes over a hold whose process is gone): the runtime and its kit are PUBLISHED packages, so a project is what a preview needs to install them into — the CLI has no renderer of its own. **It installs the runtime this CLI was built for**, and the kit through it, never `latest`: the spec this binary's generator writes is the one that runtime reads, and a version that published mid-session is a runtime nobody checked the plan against. A version the registry does not serve yet is refused before npm runs, naming `--kit`. **`--kit <checkout>`** (repeatable) — a checkout's ROOT, which names the app-runtime and the ui it holds, or one of those two packages — renders against that checkout instead — built, packed and proven the way `lotics app kit` installs one into an app — and the run prints that it rendered against a LOCAL CHECKOUT, first and last, so the render is never read as the published one. The project is a CACHE and never a source: `npm install` runs on the first render of a set and again only when a checkout's tarball did, and every file in it is rewritten from the plan on each run. **A document a row names by path is served** from the project's own static root, as the named file a workspace would hold, so a files section, a verdict over the pile and a required set read what the model states; a `fil_` id is carried by name. **An app read inside one subject (`scope`) is walked with the first row its switcher lists picked**, and the run names it. **`--record <ref>`** opens that row's record — a row of the entity the app's screen lists, by the ref the model gives it — where the walk otherwise opens the first; a ref naming no such row is refused with the refs there are. **Every read is answered at the BRIDGE**, not inside the page. The app SDK's design-time fixture (`registerMockFixture` + `?__mock=1`) is the wrong half of this: the flag lives in the app's own url, and the first record page is a navigation the app's router performs — so from that surface onwards the flag is gone and every read falls through anyway. The bridge is the one place every call arrives whatever the url says, so `query`, `field_options`, `members` and `context` are answered there from one reading of the rows, and the generated entry ships EXACTLY as `app create --from` writes it. **A preview writes nothing**: a workflow, an upload or an agent run meets the same read gate `app check --screens` uses, so the app draws its own refusal path rather than reporting a save nothing moved for, and the call is reported above the census the way the check reports one. **What came out empty, it names**: a `formula`, `rollup`, `lookup` or `autonumber` is computed here rather than stated in the model, and the census lists each computed column that stayed empty on EVERY row of a table that has rows — named by outcome, since a total drawn blank is either the app's own answer or a hole and only the run can say which. The census reads the CELLS and not the field types, so nothing is exempt by kind. **A computed column that came out `#ERROR:` is named too, and the run refuses it** before anything is installed: a wall of red measures clean, and a run that drew it and exited 0 would tell its author the app is fine. **It computes in UTC**, because a model states no timezone — a row's `@today`, an autonumber's `{YEAR}` and a record's own clock all land on the run's own date at UTC midnight, so two runs of one model draw the same register wherever they are made. |
 | `lotics app workflow set <alias>` | Push the edited `src/workflows/<alias>.ts` body through `set_app_workflow` (the single author of `apps.workflows`). Reads the body from disk (header + `/// <reference>` + `export {};` marker + the `__workflow` wrapper all stripped) + the typed `inputs`/`outputs` **and the `description`** from `package.json#lotics.workflows.<alias>`; the **server** re-verifies the body and echoes the bound `outputs` (declared, else DERIVED from `return({ data })`). The `description` is the one line an agent reads when choosing between the app's aliases (the workflow counterpart to a query's) — authored in the manifest so it lives beside the body in version control and rides every push; omit it and the workflow keeps whatever description it already has, so a push can never blank one set elsewhere. When the manifest declared NO `outputs`, the DERIVED echo is written back into `package.json#lotics.workflows.<alias>.outputs` (a SURGICAL write — preserves `knowledge`/`config` and every other manifest field) and that alias's types are refreshed in place, so `useWorkflow("<alias>")`'s `result.data` is typed immediately with no hand-copy and no second `lotics app codegen`; an explicitly-declared `outputs` is authoritative and never overwritten. A deploy runs this same verb for every alias whose declaration or body is ahead of the app, so this command is the one-alias spelling of what a release does, not a step a release leaves to a person. Clear error + non-zero exit on a missing file, an alias absent from the manifest, or a verify failure. A push also prints any non-blocking verify warnings, including an input the alias declares that the body never reads. A first bind MINTS the workflow row, and the id it echoes is written back into `package.json#lotics.workflows.<alias>.workflow_id` — the same surgical write the derived `outputs` gets. Without it a hand-declared alias ended up shaped unlike its siblings, so anything reading the manifest (an audit, a port to another workspace, a person comparing two blocks) had to treat a missing id as normal, which is exactly how a genuinely missing one stops being visible. **`--acknowledge-breaking-api`** carries this write out even though it breaks what the app's published API promises, snapshotting the broken contract as a new version; without it such a write is refused and every breaking change is named (see `app api`). |
 | `lotics app agent set <alias> [--acknowledge-breaking-api]` | Push `src/agents/<alias>.md` — plus `inputs`/`outputs` when `package.json#lotics.agents.<alias>` declares them — through `set_app_agent`. The agent mirror of `app workflow set`, and the deploy-free authoring path for an agent's prose and its typed edges. **It sends only those fields.** Everything else is absent, and absent means unchanged, so a declaration this CLI does not model cannot be reverted by a push from a checkout that predates it — the chat authoring agent's `knowledge_doc_ids`, another operator's `query_aliases` grant. To change one of those, call `set_app_agent` with just that field (`lotics run set_app_agent '{"app_id":…,"alias":…,"tool_names":[…]}'` — it merges), then `app pull` to bring the manifest back in step. **CREATES the alias when the app has not bound one yet**, so a new agent is authored the same way a new workflow is: write the prose, declare the typed half, push. A create needs the prose file (an agent without instructions is not an agent); it is gated on nothing else, because what keeps a binding alive is a `useAppAgentRun("<alias>")` call site in the shipped bundle — a deploy prunes an agent the bundle never names, manifest entry or not. The prose push is a conditional write against the fingerprint this project last saw, so it is refused rather than allowed to overwrite prose someone else changed. Clear error + non-zero exit when there is no prose file and nothing declared to push instead, when a create has no prose to create from, or when the file is empty once the header is stripped. **`--acknowledge-breaking-api`** carries this write out even though it breaks what the app's published API promises, snapshotting the broken contract as a new version: an agent's alias and its declared `inputs`/`outputs` are in that contract too, so renaming one or changing what it accepts breaks a caller nobody here can redeploy. Without the flag such a write is refused and every breaking change is named (see `app api`). |

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lotics/cli",
-  "version": "0.245.0",
+  "version": "0.246.0",
   "description": "Lotics SDK and CLI for AI agents",
   "type": "module",
   "bin": {